Hugging Face's logo

bartowski
/
LensVLM-9B-GGUF

Image-Text-to-Text
GGUF
vision-language-model
long-context
visual-text-compression
conversational
Model card Files
xet
 Community
bartowski commited on
Commit
5e67b17
·
verified ·
1 Parent(s): 887a237

Upload folder using huggingface_hub

Browse files
Files changed (36) hide show
  1. layouts/LensVLM-9B-IQ2_M.layout.json +200 -0
  2. layouts/LensVLM-9B-IQ2_M.tensor-types.txt +250 -0
  3. layouts/LensVLM-9B-IQ3_M.layout.json +199 -0
  4. layouts/LensVLM-9B-IQ3_M.tensor-types.txt +250 -0
  5. layouts/LensVLM-9B-IQ3_XS.layout.json +195 -0
  6. layouts/LensVLM-9B-IQ3_XS.tensor-types.txt +250 -0
  7. layouts/LensVLM-9B-IQ3_XXS.layout.json +200 -0
  8. layouts/LensVLM-9B-IQ3_XXS.tensor-types.txt +250 -0
  9. layouts/LensVLM-9B-IQ4_NL.layout.json +197 -0
  10. layouts/LensVLM-9B-IQ4_NL.tensor-types.txt +250 -0
  11. layouts/LensVLM-9B-IQ4_XS.layout.json +193 -0
  12. layouts/LensVLM-9B-IQ4_XS.tensor-types.txt +250 -0
  13. layouts/LensVLM-9B-Q2_K.layout.json +197 -0
  14. layouts/LensVLM-9B-Q2_K.tensor-types.txt +250 -0
  15. layouts/LensVLM-9B-Q3_K_L.layout.json +200 -0
  16. layouts/LensVLM-9B-Q3_K_L.tensor-types.txt +250 -0
  17. layouts/LensVLM-9B-Q3_K_M.layout.json +201 -0
  18. layouts/LensVLM-9B-Q3_K_M.tensor-types.txt +250 -0
  19. layouts/LensVLM-9B-Q3_K_S.layout.json +193 -0
  20. layouts/LensVLM-9B-Q3_K_S.tensor-types.txt +250 -0
  21. layouts/LensVLM-9B-Q4_K_L.layout.json +198 -0
  22. layouts/LensVLM-9B-Q4_K_L.tensor-types.txt +250 -0
  23. layouts/LensVLM-9B-Q4_K_M.layout.json +196 -0
  24. layouts/LensVLM-9B-Q4_K_M.tensor-types.txt +250 -0
  25. layouts/LensVLM-9B-Q4_K_S.layout.json +194 -0
  26. layouts/LensVLM-9B-Q4_K_S.tensor-types.txt +250 -0
  27. layouts/LensVLM-9B-Q5_K_M.layout.json +193 -0
  28. layouts/LensVLM-9B-Q5_K_M.tensor-types.txt +250 -0
  29. layouts/LensVLM-9B-Q5_K_S.layout.json +191 -0
  30. layouts/LensVLM-9B-Q5_K_S.tensor-types.txt +250 -0
  31. layouts/LensVLM-9B-Q6_K.layout.json +192 -0
  32. layouts/LensVLM-9B-Q6_K.tensor-types.txt +250 -0
  33. layouts/LensVLM-9B-Q6_K_L.layout.json +192 -0
  34. layouts/LensVLM-9B-Q6_K_L.tensor-types.txt +250 -0
  35. layouts/LensVLM-9B-Q6_K_S.layout.json +191 -0
  36. layouts/LensVLM-9B-Q6_K_S.tensor-types.txt +250 -0
layouts/LensVLM-9B-IQ2_M.layout.json ADDED
@@ -0,0 +1,200 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "IQ2_M",
5
+ "base_type": "iq2_s",
6
+ "family": "iq",
7
+ "recipe": "ident",
8
+ "variant": "shr0.70",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "M",
12
+ "rung_share": 0.7,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.1861,
17
+ "type": "q4_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.1163,
21
+ "type": "q3_k"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q4_k",
131
+ "token_embd.weight": "q3_k"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 3729047552,
135
+ "map_body+embd": 3496525824,
136
+ "ratio": 0.9376,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.7,
140
+ "body_bpw": 2.879,
141
+ "body_bpw_heuristic": 3.001,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 3525937152,
144
+ "file_bpw_pred": 3.15,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "iq2_s": 17,
148
+ "iq4_xs": 5,
149
+ "q5_k": 1,
150
+ "iq3_xxs": 1
151
+ },
152
+ "attn_k": {
153
+ "iq4_xs": 7,
154
+ "q5_k": 1
155
+ },
156
+ "attn_output": {
157
+ "iq4_xs": 7,
158
+ "q5_k": 1
159
+ },
160
+ "attn_q": {
161
+ "iq2_s": 7,
162
+ "iq4_xs": 1
163
+ },
164
+ "attn_qkv": {
165
+ "iq2_s": 22,
166
+ "iq4_xs": 1,
167
+ "iq3_s": 1
168
+ },
169
+ "attn_v": {
170
+ "q5_k": 8
171
+ },
172
+ "ffn_down": {
173
+ "iq2_s": 29,
174
+ "iq4_xs": 2,
175
+ "iq3_s": 1
176
+ },
177
+ "ffn_gate": {
178
+ "iq2_s": 31,
179
+ "iq4_xs": 1
180
+ },
181
+ "ffn_up": {
182
+ "iq2_s": 23,
183
+ "iq3_s": 8,
184
+ "iq4_xs": 1
185
+ },
186
+ "ssm_out": {
187
+ "iq4_xs": 23,
188
+ "q5_k": 1
189
+ }
190
+ },
191
+ "prior": "layout/prior.json",
192
+ "pin_rule": "floor",
193
+ "embd_cap": "heur",
194
+ "generator": "auto_quant_v2.layout",
195
+ "generator_key": "93ecba14028d5e4b",
196
+ "quant_type": "IQ2_M",
197
+ "quantize_ftype": "IQ2_M",
198
+ "llama_cpp_version": "b11081",
199
+ "commit": "3724aa1c0eb238a683e222c6f4df4b0e71ed3d74"
200
+ }
layouts/LensVLM-9B-IQ2_M.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q4_k
50
+ ^token_embd\.weight$=q3_k
51
+ ^blk\.0\.attn_gate\.weight$=q5_k
52
+ ^blk\.0\.attn_qkv\.weight$=iq4_xs
53
+ ^blk\.0\.ffn_down\.weight$=iq4_xs
54
+ ^blk\.0\.ffn_gate\.weight$=iq2_s
55
+ ^blk\.0\.ffn_up\.weight$=iq3_s
56
+ ^blk\.0\.ssm_out\.weight$=q5_k
57
+ ^blk\.1\.attn_gate\.weight$=iq4_xs
58
+ ^blk\.1\.attn_qkv\.weight$=iq3_s
59
+ ^blk\.1\.ffn_down\.weight$=iq2_s
60
+ ^blk\.1\.ffn_gate\.weight$=iq2_s
61
+ ^blk\.1\.ffn_up\.weight$=iq3_s
62
+ ^blk\.1\.ssm_out\.weight$=iq4_xs
63
+ ^blk\.2\.attn_gate\.weight$=iq4_xs
64
+ ^blk\.2\.attn_qkv\.weight$=iq2_s
65
+ ^blk\.2\.ffn_down\.weight$=iq2_s
66
+ ^blk\.2\.ffn_gate\.weight$=iq2_s
67
+ ^blk\.2\.ffn_up\.weight$=iq3_s
68
+ ^blk\.2\.ssm_out\.weight$=iq4_xs
69
+ ^blk\.3\.attn_k\.weight$=iq4_xs
70
+ ^blk\.3\.attn_output\.weight$=iq4_xs
71
+ ^blk\.3\.attn_q\.weight$=iq2_s
72
+ ^blk\.3\.attn_v\.weight$=q5_k
73
+ ^blk\.3\.ffn_down\.weight$=iq2_s
74
+ ^blk\.3\.ffn_gate\.weight$=iq2_s
75
+ ^blk\.3\.ffn_up\.weight$=iq3_s
76
+ ^blk\.4\.attn_gate\.weight$=iq4_xs
77
+ ^blk\.4\.attn_qkv\.weight$=iq2_s
78
+ ^blk\.4\.ffn_down\.weight$=iq2_s
79
+ ^blk\.4\.ffn_gate\.weight$=iq2_s
80
+ ^blk\.4\.ffn_up\.weight$=iq3_s
81
+ ^blk\.4\.ssm_out\.weight$=iq4_xs
82
+ ^blk\.5\.attn_gate\.weight$=iq4_xs
83
+ ^blk\.5\.attn_qkv\.weight$=iq2_s
84
+ ^blk\.5\.ffn_down\.weight$=iq2_s
85
+ ^blk\.5\.ffn_gate\.weight$=iq2_s
86
+ ^blk\.5\.ffn_up\.weight$=iq3_s
87
+ ^blk\.5\.ssm_out\.weight$=iq4_xs
88
+ ^blk\.6\.attn_gate\.weight$=iq4_xs
89
+ ^blk\.6\.attn_qkv\.weight$=iq2_s
90
+ ^blk\.6\.ffn_down\.weight$=iq2_s
91
+ ^blk\.6\.ffn_gate\.weight$=iq2_s
92
+ ^blk\.6\.ffn_up\.weight$=iq3_s
93
+ ^blk\.6\.ssm_out\.weight$=iq4_xs
94
+ ^blk\.7\.attn_k\.weight$=iq4_xs
95
+ ^blk\.7\.attn_output\.weight$=iq4_xs
96
+ ^blk\.7\.attn_q\.weight$=iq2_s
97
+ ^blk\.7\.attn_v\.weight$=q5_k
98
+ ^blk\.7\.ffn_down\.weight$=iq2_s
99
+ ^blk\.7\.ffn_gate\.weight$=iq2_s
100
+ ^blk\.7\.ffn_up\.weight$=iq3_s
101
+ ^blk\.8\.attn_gate\.weight$=iq2_s
102
+ ^blk\.8\.attn_qkv\.weight$=iq2_s
103
+ ^blk\.8\.ffn_down\.weight$=iq2_s
104
+ ^blk\.8\.ffn_gate\.weight$=iq2_s
105
+ ^blk\.8\.ffn_up\.weight$=iq2_s
106
+ ^blk\.8\.ssm_out\.weight$=iq4_xs
107
+ ^blk\.9\.attn_gate\.weight$=iq2_s
108
+ ^blk\.9\.attn_qkv\.weight$=iq2_s
109
+ ^blk\.9\.ffn_down\.weight$=iq2_s
110
+ ^blk\.9\.ffn_gate\.weight$=iq2_s
111
+ ^blk\.9\.ffn_up\.weight$=iq2_s
112
+ ^blk\.9\.ssm_out\.weight$=iq4_xs
113
+ ^blk\.10\.attn_gate\.weight$=iq2_s
114
+ ^blk\.10\.attn_qkv\.weight$=iq2_s
115
+ ^blk\.10\.ffn_down\.weight$=iq2_s
116
+ ^blk\.10\.ffn_gate\.weight$=iq2_s
117
+ ^blk\.10\.ffn_up\.weight$=iq2_s
118
+ ^blk\.10\.ssm_out\.weight$=iq4_xs
119
+ ^blk\.11\.attn_k\.weight$=iq4_xs
120
+ ^blk\.11\.attn_output\.weight$=iq4_xs
121
+ ^blk\.11\.attn_q\.weight$=iq2_s
122
+ ^blk\.11\.attn_v\.weight$=q5_k
123
+ ^blk\.11\.ffn_down\.weight$=iq2_s
124
+ ^blk\.11\.ffn_gate\.weight$=iq2_s
125
+ ^blk\.11\.ffn_up\.weight$=iq2_s
126
+ ^blk\.12\.attn_gate\.weight$=iq2_s
127
+ ^blk\.12\.attn_qkv\.weight$=iq2_s
128
+ ^blk\.12\.ffn_down\.weight$=iq2_s
129
+ ^blk\.12\.ffn_gate\.weight$=iq2_s
130
+ ^blk\.12\.ffn_up\.weight$=iq2_s
131
+ ^blk\.12\.ssm_out\.weight$=iq4_xs
132
+ ^blk\.13\.attn_gate\.weight$=iq2_s
133
+ ^blk\.13\.attn_qkv\.weight$=iq2_s
134
+ ^blk\.13\.ffn_down\.weight$=iq2_s
135
+ ^blk\.13\.ffn_gate\.weight$=iq2_s
136
+ ^blk\.13\.ffn_up\.weight$=iq2_s
137
+ ^blk\.13\.ssm_out\.weight$=iq4_xs
138
+ ^blk\.14\.attn_gate\.weight$=iq2_s
139
+ ^blk\.14\.attn_qkv\.weight$=iq2_s
140
+ ^blk\.14\.ffn_down\.weight$=iq2_s
141
+ ^blk\.14\.ffn_gate\.weight$=iq2_s
142
+ ^blk\.14\.ffn_up\.weight$=iq2_s
143
+ ^blk\.14\.ssm_out\.weight$=iq4_xs
144
+ ^blk\.15\.attn_k\.weight$=iq4_xs
145
+ ^blk\.15\.attn_output\.weight$=iq4_xs
146
+ ^blk\.15\.attn_q\.weight$=iq2_s
147
+ ^blk\.15\.attn_v\.weight$=q5_k
148
+ ^blk\.15\.ffn_down\.weight$=iq2_s
149
+ ^blk\.15\.ffn_gate\.weight$=iq2_s
150
+ ^blk\.15\.ffn_up\.weight$=iq2_s
151
+ ^blk\.16\.attn_gate\.weight$=iq2_s
152
+ ^blk\.16\.attn_qkv\.weight$=iq2_s
153
+ ^blk\.16\.ffn_down\.weight$=iq2_s
154
+ ^blk\.16\.ffn_gate\.weight$=iq2_s
155
+ ^blk\.16\.ffn_up\.weight$=iq2_s
156
+ ^blk\.16\.ssm_out\.weight$=iq4_xs
157
+ ^blk\.17\.attn_gate\.weight$=iq2_s
158
+ ^blk\.17\.attn_qkv\.weight$=iq2_s
159
+ ^blk\.17\.ffn_down\.weight$=iq2_s
160
+ ^blk\.17\.ffn_gate\.weight$=iq2_s
161
+ ^blk\.17\.ffn_up\.weight$=iq2_s
162
+ ^blk\.17\.ssm_out\.weight$=iq4_xs
163
+ ^blk\.18\.attn_gate\.weight$=iq2_s
164
+ ^blk\.18\.attn_qkv\.weight$=iq2_s
165
+ ^blk\.18\.ffn_down\.weight$=iq2_s
166
+ ^blk\.18\.ffn_gate\.weight$=iq2_s
167
+ ^blk\.18\.ffn_up\.weight$=iq2_s
168
+ ^blk\.18\.ssm_out\.weight$=iq4_xs
169
+ ^blk\.19\.attn_k\.weight$=iq4_xs
170
+ ^blk\.19\.attn_output\.weight$=iq4_xs
171
+ ^blk\.19\.attn_q\.weight$=iq2_s
172
+ ^blk\.19\.attn_v\.weight$=q5_k
173
+ ^blk\.19\.ffn_down\.weight$=iq2_s
174
+ ^blk\.19\.ffn_gate\.weight$=iq2_s
175
+ ^blk\.19\.ffn_up\.weight$=iq2_s
176
+ ^blk\.20\.attn_gate\.weight$=iq2_s
177
+ ^blk\.20\.attn_qkv\.weight$=iq2_s
178
+ ^blk\.20\.ffn_down\.weight$=iq2_s
179
+ ^blk\.20\.ffn_gate\.weight$=iq2_s
180
+ ^blk\.20\.ffn_up\.weight$=iq2_s
181
+ ^blk\.20\.ssm_out\.weight$=iq4_xs
182
+ ^blk\.21\.attn_gate\.weight$=iq2_s
183
+ ^blk\.21\.attn_qkv\.weight$=iq2_s
184
+ ^blk\.21\.ffn_down\.weight$=iq2_s
185
+ ^blk\.21\.ffn_gate\.weight$=iq2_s
186
+ ^blk\.21\.ffn_up\.weight$=iq2_s
187
+ ^blk\.21\.ssm_out\.weight$=iq4_xs
188
+ ^blk\.22\.attn_gate\.weight$=iq2_s
189
+ ^blk\.22\.attn_qkv\.weight$=iq2_s
190
+ ^blk\.22\.ffn_down\.weight$=iq2_s
191
+ ^blk\.22\.ffn_gate\.weight$=iq2_s
192
+ ^blk\.22\.ffn_up\.weight$=iq2_s
193
+ ^blk\.22\.ssm_out\.weight$=iq4_xs
194
+ ^blk\.23\.attn_k\.weight$=iq4_xs
195
+ ^blk\.23\.attn_output\.weight$=iq4_xs
196
+ ^blk\.23\.attn_q\.weight$=iq2_s
197
+ ^blk\.23\.attn_v\.weight$=q5_k
198
+ ^blk\.23\.ffn_down\.weight$=iq2_s
199
+ ^blk\.23\.ffn_gate\.weight$=iq2_s
200
+ ^blk\.23\.ffn_up\.weight$=iq2_s
201
+ ^blk\.24\.attn_gate\.weight$=iq3_xxs
202
+ ^blk\.24\.attn_qkv\.weight$=iq2_s
203
+ ^blk\.24\.ffn_down\.weight$=iq2_s
204
+ ^blk\.24\.ffn_gate\.weight$=iq2_s
205
+ ^blk\.24\.ffn_up\.weight$=iq2_s
206
+ ^blk\.24\.ssm_out\.weight$=iq4_xs
207
+ ^blk\.25\.attn_gate\.weight$=iq2_s
208
+ ^blk\.25\.attn_qkv\.weight$=iq2_s
209
+ ^blk\.25\.ffn_down\.weight$=iq2_s
210
+ ^blk\.25\.ffn_gate\.weight$=iq2_s
211
+ ^blk\.25\.ffn_up\.weight$=iq2_s
212
+ ^blk\.25\.ssm_out\.weight$=iq4_xs
213
+ ^blk\.26\.attn_gate\.weight$=iq2_s
214
+ ^blk\.26\.attn_qkv\.weight$=iq2_s
215
+ ^blk\.26\.ffn_down\.weight$=iq2_s
216
+ ^blk\.26\.ffn_gate\.weight$=iq2_s
217
+ ^blk\.26\.ffn_up\.weight$=iq2_s
218
+ ^blk\.26\.ssm_out\.weight$=iq4_xs
219
+ ^blk\.27\.attn_k\.weight$=q5_k
220
+ ^blk\.27\.attn_output\.weight$=iq4_xs
221
+ ^blk\.27\.attn_q\.weight$=iq2_s
222
+ ^blk\.27\.attn_v\.weight$=q5_k
223
+ ^blk\.27\.ffn_down\.weight$=iq3_s
224
+ ^blk\.27\.ffn_gate\.weight$=iq2_s
225
+ ^blk\.27\.ffn_up\.weight$=iq2_s
226
+ ^blk\.28\.attn_gate\.weight$=iq2_s
227
+ ^blk\.28\.attn_qkv\.weight$=iq2_s
228
+ ^blk\.28\.ffn_down\.weight$=iq2_s
229
+ ^blk\.28\.ffn_gate\.weight$=iq2_s
230
+ ^blk\.28\.ffn_up\.weight$=iq2_s
231
+ ^blk\.28\.ssm_out\.weight$=iq4_xs
232
+ ^blk\.29\.attn_gate\.weight$=iq2_s
233
+ ^blk\.29\.attn_qkv\.weight$=iq2_s
234
+ ^blk\.29\.ffn_down\.weight$=iq2_s
235
+ ^blk\.29\.ffn_gate\.weight$=iq2_s
236
+ ^blk\.29\.ffn_up\.weight$=iq2_s
237
+ ^blk\.29\.ssm_out\.weight$=iq4_xs
238
+ ^blk\.30\.attn_gate\.weight$=iq2_s
239
+ ^blk\.30\.attn_qkv\.weight$=iq2_s
240
+ ^blk\.30\.ffn_down\.weight$=iq2_s
241
+ ^blk\.30\.ffn_gate\.weight$=iq2_s
242
+ ^blk\.30\.ffn_up\.weight$=iq2_s
243
+ ^blk\.30\.ssm_out\.weight$=iq4_xs
244
+ ^blk\.31\.attn_k\.weight$=iq4_xs
245
+ ^blk\.31\.attn_output\.weight$=q5_k
246
+ ^blk\.31\.attn_q\.weight$=iq4_xs
247
+ ^blk\.31\.attn_v\.weight$=q5_k
248
+ ^blk\.31\.ffn_down\.weight$=iq4_xs
249
+ ^blk\.31\.ffn_gate\.weight$=iq4_xs
250
+ ^blk\.31\.ffn_up\.weight$=iq4_xs
layouts/LensVLM-9B-IQ3_M.layout.json ADDED
@@ -0,0 +1,199 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "IQ3_M",
5
+ "base_type": "iq3_s",
6
+ "family": "iq",
7
+ "recipe": "ident",
8
+ "variant": "shr0.50",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "L",
12
+ "rung_share": 0.5,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.1771,
17
+ "type": "q4_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.0928,
21
+ "type": "q4_k"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q4_k",
131
+ "token_embd.weight": "q4_k"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 4682153984,
135
+ "map_body+embd": 4806082560,
136
+ "ratio": 1.0265,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.5,
140
+ "body_bpw": 4.238,
141
+ "body_bpw_heuristic": 3.948,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 4835493888,
144
+ "file_bpw_pred": 4.32,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "q5_k": 11,
148
+ "iq3_s": 11,
149
+ "q8_0": 1,
150
+ "q4_k": 1
151
+ },
152
+ "attn_k": {
153
+ "q5_k": 6,
154
+ "q6_k": 2
155
+ },
156
+ "attn_output": {
157
+ "q6_k": 6,
158
+ "q5_k": 2
159
+ },
160
+ "attn_q": {
161
+ "iq3_s": 6,
162
+ "q5_k": 2
163
+ },
164
+ "attn_qkv": {
165
+ "iq3_s": 12,
166
+ "q5_k": 6,
167
+ "q4_k": 6
168
+ },
169
+ "attn_v": {
170
+ "q6_k": 4,
171
+ "q8_0": 4
172
+ },
173
+ "ffn_down": {
174
+ "q5_k": 16,
175
+ "iq3_s": 16
176
+ },
177
+ "ffn_gate": {
178
+ "iq3_s": 31,
179
+ "q5_k": 1
180
+ },
181
+ "ffn_up": {
182
+ "iq3_s": 22,
183
+ "q5_k": 10
184
+ },
185
+ "ssm_out": {
186
+ "q5_k": 23,
187
+ "q8_0": 1
188
+ }
189
+ },
190
+ "prior": "layout/prior.json",
191
+ "pin_rule": "floor",
192
+ "embd_cap": "heur",
193
+ "generator": "auto_quant_v2.layout",
194
+ "generator_key": "93ecba14028d5e4b",
195
+ "quant_type": "IQ3_M",
196
+ "quantize_ftype": "IQ3_M",
197
+ "llama_cpp_version": "b11081",
198
+ "commit": "d7cfc9bb610f62821944853c508648097409f4a7"
199
+ }
layouts/LensVLM-9B-IQ3_M.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q4_k
50
+ ^token_embd\.weight$=q4_k
51
+ ^blk\.0\.attn_gate\.weight$=q8_0
52
+ ^blk\.0\.attn_qkv\.weight$=q5_k
53
+ ^blk\.0\.ffn_down\.weight$=q5_k
54
+ ^blk\.0\.ffn_gate\.weight$=iq3_s
55
+ ^blk\.0\.ffn_up\.weight$=q5_k
56
+ ^blk\.0\.ssm_out\.weight$=q8_0
57
+ ^blk\.1\.attn_gate\.weight$=q5_k
58
+ ^blk\.1\.attn_qkv\.weight$=q5_k
59
+ ^blk\.1\.ffn_down\.weight$=q5_k
60
+ ^blk\.1\.ffn_gate\.weight$=iq3_s
61
+ ^blk\.1\.ffn_up\.weight$=q5_k
62
+ ^blk\.1\.ssm_out\.weight$=q5_k
63
+ ^blk\.2\.attn_gate\.weight$=q5_k
64
+ ^blk\.2\.attn_qkv\.weight$=q5_k
65
+ ^blk\.2\.ffn_down\.weight$=q5_k
66
+ ^blk\.2\.ffn_gate\.weight$=iq3_s
67
+ ^blk\.2\.ffn_up\.weight$=q5_k
68
+ ^blk\.2\.ssm_out\.weight$=q5_k
69
+ ^blk\.3\.attn_k\.weight$=q5_k
70
+ ^blk\.3\.attn_output\.weight$=q5_k
71
+ ^blk\.3\.attn_q\.weight$=iq3_s
72
+ ^blk\.3\.attn_v\.weight$=q6_k
73
+ ^blk\.3\.ffn_down\.weight$=q5_k
74
+ ^blk\.3\.ffn_gate\.weight$=iq3_s
75
+ ^blk\.3\.ffn_up\.weight$=q5_k
76
+ ^blk\.4\.attn_gate\.weight$=q5_k
77
+ ^blk\.4\.attn_qkv\.weight$=q5_k
78
+ ^blk\.4\.ffn_down\.weight$=q5_k
79
+ ^blk\.4\.ffn_gate\.weight$=iq3_s
80
+ ^blk\.4\.ffn_up\.weight$=q5_k
81
+ ^blk\.4\.ssm_out\.weight$=q5_k
82
+ ^blk\.5\.attn_gate\.weight$=q5_k
83
+ ^blk\.5\.attn_qkv\.weight$=q5_k
84
+ ^blk\.5\.ffn_down\.weight$=q5_k
85
+ ^blk\.5\.ffn_gate\.weight$=iq3_s
86
+ ^blk\.5\.ffn_up\.weight$=q5_k
87
+ ^blk\.5\.ssm_out\.weight$=q5_k
88
+ ^blk\.6\.attn_gate\.weight$=q5_k
89
+ ^blk\.6\.attn_qkv\.weight$=q5_k
90
+ ^blk\.6\.ffn_down\.weight$=q5_k
91
+ ^blk\.6\.ffn_gate\.weight$=iq3_s
92
+ ^blk\.6\.ffn_up\.weight$=q5_k
93
+ ^blk\.6\.ssm_out\.weight$=q5_k
94
+ ^blk\.7\.attn_k\.weight$=q5_k
95
+ ^blk\.7\.attn_output\.weight$=q5_k
96
+ ^blk\.7\.attn_q\.weight$=iq3_s
97
+ ^blk\.7\.attn_v\.weight$=q6_k
98
+ ^blk\.7\.ffn_down\.weight$=q5_k
99
+ ^blk\.7\.ffn_gate\.weight$=iq3_s
100
+ ^blk\.7\.ffn_up\.weight$=q5_k
101
+ ^blk\.8\.attn_gate\.weight$=q4_k
102
+ ^blk\.8\.attn_qkv\.weight$=q4_k
103
+ ^blk\.8\.ffn_down\.weight$=iq3_s
104
+ ^blk\.8\.ffn_gate\.weight$=iq3_s
105
+ ^blk\.8\.ffn_up\.weight$=iq3_s
106
+ ^blk\.8\.ssm_out\.weight$=q5_k
107
+ ^blk\.9\.attn_gate\.weight$=iq3_s
108
+ ^blk\.9\.attn_qkv\.weight$=q4_k
109
+ ^blk\.9\.ffn_down\.weight$=iq3_s
110
+ ^blk\.9\.ffn_gate\.weight$=iq3_s
111
+ ^blk\.9\.ffn_up\.weight$=iq3_s
112
+ ^blk\.9\.ssm_out\.weight$=q5_k
113
+ ^blk\.10\.attn_gate\.weight$=iq3_s
114
+ ^blk\.10\.attn_qkv\.weight$=q4_k
115
+ ^blk\.10\.ffn_down\.weight$=iq3_s
116
+ ^blk\.10\.ffn_gate\.weight$=iq3_s
117
+ ^blk\.10\.ffn_up\.weight$=iq3_s
118
+ ^blk\.10\.ssm_out\.weight$=q5_k
119
+ ^blk\.11\.attn_k\.weight$=q6_k
120
+ ^blk\.11\.attn_output\.weight$=q6_k
121
+ ^blk\.11\.attn_q\.weight$=iq3_s
122
+ ^blk\.11\.attn_v\.weight$=q8_0
123
+ ^blk\.11\.ffn_down\.weight$=iq3_s
124
+ ^blk\.11\.ffn_gate\.weight$=iq3_s
125
+ ^blk\.11\.ffn_up\.weight$=iq3_s
126
+ ^blk\.12\.attn_gate\.weight$=iq3_s
127
+ ^blk\.12\.attn_qkv\.weight$=q4_k
128
+ ^blk\.12\.ffn_down\.weight$=iq3_s
129
+ ^blk\.12\.ffn_gate\.weight$=iq3_s
130
+ ^blk\.12\.ffn_up\.weight$=iq3_s
131
+ ^blk\.12\.ssm_out\.weight$=q5_k
132
+ ^blk\.13\.attn_gate\.weight$=iq3_s
133
+ ^blk\.13\.attn_qkv\.weight$=q4_k
134
+ ^blk\.13\.ffn_down\.weight$=iq3_s
135
+ ^blk\.13\.ffn_gate\.weight$=iq3_s
136
+ ^blk\.13\.ffn_up\.weight$=iq3_s
137
+ ^blk\.13\.ssm_out\.weight$=q5_k
138
+ ^blk\.14\.attn_gate\.weight$=iq3_s
139
+ ^blk\.14\.attn_qkv\.weight$=q4_k
140
+ ^blk\.14\.ffn_down\.weight$=iq3_s
141
+ ^blk\.14\.ffn_gate\.weight$=iq3_s
142
+ ^blk\.14\.ffn_up\.weight$=iq3_s
143
+ ^blk\.14\.ssm_out\.weight$=q5_k
144
+ ^blk\.15\.attn_k\.weight$=q5_k
145
+ ^blk\.15\.attn_output\.weight$=q6_k
146
+ ^blk\.15\.attn_q\.weight$=iq3_s
147
+ ^blk\.15\.attn_v\.weight$=q8_0
148
+ ^blk\.15\.ffn_down\.weight$=iq3_s
149
+ ^blk\.15\.ffn_gate\.weight$=iq3_s
150
+ ^blk\.15\.ffn_up\.weight$=iq3_s
151
+ ^blk\.16\.attn_gate\.weight$=iq3_s
152
+ ^blk\.16\.attn_qkv\.weight$=iq3_s
153
+ ^blk\.16\.ffn_down\.weight$=iq3_s
154
+ ^blk\.16\.ffn_gate\.weight$=iq3_s
155
+ ^blk\.16\.ffn_up\.weight$=iq3_s
156
+ ^blk\.16\.ssm_out\.weight$=q5_k
157
+ ^blk\.17\.attn_gate\.weight$=iq3_s
158
+ ^blk\.17\.attn_qkv\.weight$=iq3_s
159
+ ^blk\.17\.ffn_down\.weight$=iq3_s
160
+ ^blk\.17\.ffn_gate\.weight$=iq3_s
161
+ ^blk\.17\.ffn_up\.weight$=iq3_s
162
+ ^blk\.17\.ssm_out\.weight$=q5_k
163
+ ^blk\.18\.attn_gate\.weight$=iq3_s
164
+ ^blk\.18\.attn_qkv\.weight$=iq3_s
165
+ ^blk\.18\.ffn_down\.weight$=iq3_s
166
+ ^blk\.18\.ffn_gate\.weight$=iq3_s
167
+ ^blk\.18\.ffn_up\.weight$=iq3_s
168
+ ^blk\.18\.ssm_out\.weight$=q5_k
169
+ ^blk\.19\.attn_k\.weight$=q5_k
170
+ ^blk\.19\.attn_output\.weight$=q6_k
171
+ ^blk\.19\.attn_q\.weight$=iq3_s
172
+ ^blk\.19\.attn_v\.weight$=q6_k
173
+ ^blk\.19\.ffn_down\.weight$=iq3_s
174
+ ^blk\.19\.ffn_gate\.weight$=iq3_s
175
+ ^blk\.19\.ffn_up\.weight$=iq3_s
176
+ ^blk\.20\.attn_gate\.weight$=iq3_s
177
+ ^blk\.20\.attn_qkv\.weight$=iq3_s
178
+ ^blk\.20\.ffn_down\.weight$=iq3_s
179
+ ^blk\.20\.ffn_gate\.weight$=iq3_s
180
+ ^blk\.20\.ffn_up\.weight$=iq3_s
181
+ ^blk\.20\.ssm_out\.weight$=q5_k
182
+ ^blk\.21\.attn_gate\.weight$=iq3_s
183
+ ^blk\.21\.attn_qkv\.weight$=iq3_s
184
+ ^blk\.21\.ffn_down\.weight$=iq3_s
185
+ ^blk\.21\.ffn_gate\.weight$=iq3_s
186
+ ^blk\.21\.ffn_up\.weight$=iq3_s
187
+ ^blk\.21\.ssm_out\.weight$=q5_k
188
+ ^blk\.22\.attn_gate\.weight$=iq3_s
189
+ ^blk\.22\.attn_qkv\.weight$=iq3_s
190
+ ^blk\.22\.ffn_down\.weight$=iq3_s
191
+ ^blk\.22\.ffn_gate\.weight$=iq3_s
192
+ ^blk\.22\.ffn_up\.weight$=iq3_s
193
+ ^blk\.22\.ssm_out\.weight$=q5_k
194
+ ^blk\.23\.attn_k\.weight$=q5_k
195
+ ^blk\.23\.attn_output\.weight$=q6_k
196
+ ^blk\.23\.attn_q\.weight$=iq3_s
197
+ ^blk\.23\.attn_v\.weight$=q6_k
198
+ ^blk\.23\.ffn_down\.weight$=iq3_s
199
+ ^blk\.23\.ffn_gate\.weight$=iq3_s
200
+ ^blk\.23\.ffn_up\.weight$=iq3_s
201
+ ^blk\.24\.attn_gate\.weight$=q5_k
202
+ ^blk\.24\.attn_qkv\.weight$=iq3_s
203
+ ^blk\.24\.ffn_down\.weight$=q5_k
204
+ ^blk\.24\.ffn_gate\.weight$=iq3_s
205
+ ^blk\.24\.ffn_up\.weight$=iq3_s
206
+ ^blk\.24\.ssm_out\.weight$=q5_k
207
+ ^blk\.25\.attn_gate\.weight$=q5_k
208
+ ^blk\.25\.attn_qkv\.weight$=iq3_s
209
+ ^blk\.25\.ffn_down\.weight$=q5_k
210
+ ^blk\.25\.ffn_gate\.weight$=iq3_s
211
+ ^blk\.25\.ffn_up\.weight$=iq3_s
212
+ ^blk\.25\.ssm_out\.weight$=q5_k
213
+ ^blk\.26\.attn_gate\.weight$=q5_k
214
+ ^blk\.26\.attn_qkv\.weight$=iq3_s
215
+ ^blk\.26\.ffn_down\.weight$=q5_k
216
+ ^blk\.26\.ffn_gate\.weight$=iq3_s
217
+ ^blk\.26\.ffn_up\.weight$=iq3_s
218
+ ^blk\.26\.ssm_out\.weight$=q5_k
219
+ ^blk\.27\.attn_k\.weight$=q6_k
220
+ ^blk\.27\.attn_output\.weight$=q6_k
221
+ ^blk\.27\.attn_q\.weight$=q5_k
222
+ ^blk\.27\.attn_v\.weight$=q8_0
223
+ ^blk\.27\.ffn_down\.weight$=q5_k
224
+ ^blk\.27\.ffn_gate\.weight$=iq3_s
225
+ ^blk\.27\.ffn_up\.weight$=q5_k
226
+ ^blk\.28\.attn_gate\.weight$=q5_k
227
+ ^blk\.28\.attn_qkv\.weight$=iq3_s
228
+ ^blk\.28\.ffn_down\.weight$=q5_k
229
+ ^blk\.28\.ffn_gate\.weight$=iq3_s
230
+ ^blk\.28\.ffn_up\.weight$=iq3_s
231
+ ^blk\.28\.ssm_out\.weight$=q5_k
232
+ ^blk\.29\.attn_gate\.weight$=q5_k
233
+ ^blk\.29\.attn_qkv\.weight$=iq3_s
234
+ ^blk\.29\.ffn_down\.weight$=q5_k
235
+ ^blk\.29\.ffn_gate\.weight$=iq3_s
236
+ ^blk\.29\.ffn_up\.weight$=iq3_s
237
+ ^blk\.29\.ssm_out\.weight$=q5_k
238
+ ^blk\.30\.attn_gate\.weight$=q5_k
239
+ ^blk\.30\.attn_qkv\.weight$=iq3_s
240
+ ^blk\.30\.ffn_down\.weight$=q5_k
241
+ ^blk\.30\.ffn_gate\.weight$=iq3_s
242
+ ^blk\.30\.ffn_up\.weight$=iq3_s
243
+ ^blk\.30\.ssm_out\.weight$=q5_k
244
+ ^blk\.31\.attn_k\.weight$=q5_k
245
+ ^blk\.31\.attn_output\.weight$=q6_k
246
+ ^blk\.31\.attn_q\.weight$=q5_k
247
+ ^blk\.31\.attn_v\.weight$=q8_0
248
+ ^blk\.31\.ffn_down\.weight$=q5_k
249
+ ^blk\.31\.ffn_gate\.weight$=q5_k
250
+ ^blk\.31\.ffn_up\.weight$=q5_k
layouts/LensVLM-9B-IQ3_XS.layout.json ADDED
@@ -0,0 +1,195 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "IQ3_XS",
5
+ "base_type": "iq3_s",
6
+ "family": "iq",
7
+ "recipe": "ident",
8
+ "variant": "shr0.90",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "S",
12
+ "rung_share": 0.9,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.1834,
17
+ "type": "q4_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.0961,
21
+ "type": "q4_k"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q4_k",
131
+ "token_embd.weight": "q4_k"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 4520148992,
135
+ "map_body+embd": 4227661824,
136
+ "ratio": 0.9353,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.9,
140
+ "body_bpw": 3.569,
141
+ "body_bpw_heuristic": 3.76,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 4257073152,
144
+ "file_bpw_pred": 3.804,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "iq3_s": 23,
148
+ "q5_k": 1
149
+ },
150
+ "attn_k": {
151
+ "q5_k": 7,
152
+ "q4_k": 1
153
+ },
154
+ "attn_output": {
155
+ "q5_k": 8
156
+ },
157
+ "attn_q": {
158
+ "iq3_s": 8
159
+ },
160
+ "attn_qkv": {
161
+ "iq3_s": 23,
162
+ "q5_k": 1
163
+ },
164
+ "attn_v": {
165
+ "q5_k": 6,
166
+ "q6_k": 2
167
+ },
168
+ "ffn_down": {
169
+ "iq3_s": 31,
170
+ "q5_k": 1
171
+ },
172
+ "ffn_gate": {
173
+ "iq3_s": 32
174
+ },
175
+ "ffn_up": {
176
+ "iq3_s": 31,
177
+ "q5_k": 1
178
+ },
179
+ "ssm_out": {
180
+ "iq3_s": 18,
181
+ "q4_k": 3,
182
+ "q5_k": 2,
183
+ "q6_k": 1
184
+ }
185
+ },
186
+ "prior": "layout/prior.json",
187
+ "pin_rule": "floor",
188
+ "embd_cap": "heur",
189
+ "generator": "auto_quant_v2.layout",
190
+ "generator_key": "93ecba14028d5e4b",
191
+ "quant_type": "IQ3_XS",
192
+ "quantize_ftype": "IQ3_XS",
193
+ "llama_cpp_version": "b11081",
194
+ "commit": "d7cfc9bb610f62821944853c508648097409f4a7"
195
+ }
layouts/LensVLM-9B-IQ3_XS.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q4_k
50
+ ^token_embd\.weight$=q4_k
51
+ ^blk\.0\.attn_gate\.weight$=q5_k
52
+ ^blk\.0\.attn_qkv\.weight$=q5_k
53
+ ^blk\.0\.ffn_down\.weight$=iq3_s
54
+ ^blk\.0\.ffn_gate\.weight$=iq3_s
55
+ ^blk\.0\.ffn_up\.weight$=iq3_s
56
+ ^blk\.0\.ssm_out\.weight$=q6_k
57
+ ^blk\.1\.attn_gate\.weight$=iq3_s
58
+ ^blk\.1\.attn_qkv\.weight$=iq3_s
59
+ ^blk\.1\.ffn_down\.weight$=iq3_s
60
+ ^blk\.1\.ffn_gate\.weight$=iq3_s
61
+ ^blk\.1\.ffn_up\.weight$=iq3_s
62
+ ^blk\.1\.ssm_out\.weight$=q5_k
63
+ ^blk\.2\.attn_gate\.weight$=iq3_s
64
+ ^blk\.2\.attn_qkv\.weight$=iq3_s
65
+ ^blk\.2\.ffn_down\.weight$=iq3_s
66
+ ^blk\.2\.ffn_gate\.weight$=iq3_s
67
+ ^blk\.2\.ffn_up\.weight$=iq3_s
68
+ ^blk\.2\.ssm_out\.weight$=q5_k
69
+ ^blk\.3\.attn_k\.weight$=q5_k
70
+ ^blk\.3\.attn_output\.weight$=q5_k
71
+ ^blk\.3\.attn_q\.weight$=iq3_s
72
+ ^blk\.3\.attn_v\.weight$=q5_k
73
+ ^blk\.3\.ffn_down\.weight$=iq3_s
74
+ ^blk\.3\.ffn_gate\.weight$=iq3_s
75
+ ^blk\.3\.ffn_up\.weight$=iq3_s
76
+ ^blk\.4\.attn_gate\.weight$=iq3_s
77
+ ^blk\.4\.attn_qkv\.weight$=iq3_s
78
+ ^blk\.4\.ffn_down\.weight$=iq3_s
79
+ ^blk\.4\.ffn_gate\.weight$=iq3_s
80
+ ^blk\.4\.ffn_up\.weight$=iq3_s
81
+ ^blk\.4\.ssm_out\.weight$=q4_k
82
+ ^blk\.5\.attn_gate\.weight$=iq3_s
83
+ ^blk\.5\.attn_qkv\.weight$=iq3_s
84
+ ^blk\.5\.ffn_down\.weight$=iq3_s
85
+ ^blk\.5\.ffn_gate\.weight$=iq3_s
86
+ ^blk\.5\.ffn_up\.weight$=iq3_s
87
+ ^blk\.5\.ssm_out\.weight$=q4_k
88
+ ^blk\.6\.attn_gate\.weight$=iq3_s
89
+ ^blk\.6\.attn_qkv\.weight$=iq3_s
90
+ ^blk\.6\.ffn_down\.weight$=iq3_s
91
+ ^blk\.6\.ffn_gate\.weight$=iq3_s
92
+ ^blk\.6\.ffn_up\.weight$=iq3_s
93
+ ^blk\.6\.ssm_out\.weight$=q4_k
94
+ ^blk\.7\.attn_k\.weight$=q5_k
95
+ ^blk\.7\.attn_output\.weight$=q5_k
96
+ ^blk\.7\.attn_q\.weight$=iq3_s
97
+ ^blk\.7\.attn_v\.weight$=q5_k
98
+ ^blk\.7\.ffn_down\.weight$=iq3_s
99
+ ^blk\.7\.ffn_gate\.weight$=iq3_s
100
+ ^blk\.7\.ffn_up\.weight$=iq3_s
101
+ ^blk\.8\.attn_gate\.weight$=iq3_s
102
+ ^blk\.8\.attn_qkv\.weight$=iq3_s
103
+ ^blk\.8\.ffn_down\.weight$=iq3_s
104
+ ^blk\.8\.ffn_gate\.weight$=iq3_s
105
+ ^blk\.8\.ffn_up\.weight$=iq3_s
106
+ ^blk\.8\.ssm_out\.weight$=iq3_s
107
+ ^blk\.9\.attn_gate\.weight$=iq3_s
108
+ ^blk\.9\.attn_qkv\.weight$=iq3_s
109
+ ^blk\.9\.ffn_down\.weight$=iq3_s
110
+ ^blk\.9\.ffn_gate\.weight$=iq3_s
111
+ ^blk\.9\.ffn_up\.weight$=iq3_s
112
+ ^blk\.9\.ssm_out\.weight$=iq3_s
113
+ ^blk\.10\.attn_gate\.weight$=iq3_s
114
+ ^blk\.10\.attn_qkv\.weight$=iq3_s
115
+ ^blk\.10\.ffn_down\.weight$=iq3_s
116
+ ^blk\.10\.ffn_gate\.weight$=iq3_s
117
+ ^blk\.10\.ffn_up\.weight$=iq3_s
118
+ ^blk\.10\.ssm_out\.weight$=iq3_s
119
+ ^blk\.11\.attn_k\.weight$=q5_k
120
+ ^blk\.11\.attn_output\.weight$=q5_k
121
+ ^blk\.11\.attn_q\.weight$=iq3_s
122
+ ^blk\.11\.attn_v\.weight$=q5_k
123
+ ^blk\.11\.ffn_down\.weight$=iq3_s
124
+ ^blk\.11\.ffn_gate\.weight$=iq3_s
125
+ ^blk\.11\.ffn_up\.weight$=iq3_s
126
+ ^blk\.12\.attn_gate\.weight$=iq3_s
127
+ ^blk\.12\.attn_qkv\.weight$=iq3_s
128
+ ^blk\.12\.ffn_down\.weight$=iq3_s
129
+ ^blk\.12\.ffn_gate\.weight$=iq3_s
130
+ ^blk\.12\.ffn_up\.weight$=iq3_s
131
+ ^blk\.12\.ssm_out\.weight$=iq3_s
132
+ ^blk\.13\.attn_gate\.weight$=iq3_s
133
+ ^blk\.13\.attn_qkv\.weight$=iq3_s
134
+ ^blk\.13\.ffn_down\.weight$=iq3_s
135
+ ^blk\.13\.ffn_gate\.weight$=iq3_s
136
+ ^blk\.13\.ffn_up\.weight$=iq3_s
137
+ ^blk\.13\.ssm_out\.weight$=iq3_s
138
+ ^blk\.14\.attn_gate\.weight$=iq3_s
139
+ ^blk\.14\.attn_qkv\.weight$=iq3_s
140
+ ^blk\.14\.ffn_down\.weight$=iq3_s
141
+ ^blk\.14\.ffn_gate\.weight$=iq3_s
142
+ ^blk\.14\.ffn_up\.weight$=iq3_s
143
+ ^blk\.14\.ssm_out\.weight$=iq3_s
144
+ ^blk\.15\.attn_k\.weight$=q5_k
145
+ ^blk\.15\.attn_output\.weight$=q5_k
146
+ ^blk\.15\.attn_q\.weight$=iq3_s
147
+ ^blk\.15\.attn_v\.weight$=q5_k
148
+ ^blk\.15\.ffn_down\.weight$=iq3_s
149
+ ^blk\.15\.ffn_gate\.weight$=iq3_s
150
+ ^blk\.15\.ffn_up\.weight$=iq3_s
151
+ ^blk\.16\.attn_gate\.weight$=iq3_s
152
+ ^blk\.16\.attn_qkv\.weight$=iq3_s
153
+ ^blk\.16\.ffn_down\.weight$=iq3_s
154
+ ^blk\.16\.ffn_gate\.weight$=iq3_s
155
+ ^blk\.16\.ffn_up\.weight$=iq3_s
156
+ ^blk\.16\.ssm_out\.weight$=iq3_s
157
+ ^blk\.17\.attn_gate\.weight$=iq3_s
158
+ ^blk\.17\.attn_qkv\.weight$=iq3_s
159
+ ^blk\.17\.ffn_down\.weight$=iq3_s
160
+ ^blk\.17\.ffn_gate\.weight$=iq3_s
161
+ ^blk\.17\.ffn_up\.weight$=iq3_s
162
+ ^blk\.17\.ssm_out\.weight$=iq3_s
163
+ ^blk\.18\.attn_gate\.weight$=iq3_s
164
+ ^blk\.18\.attn_qkv\.weight$=iq3_s
165
+ ^blk\.18\.ffn_down\.weight$=iq3_s
166
+ ^blk\.18\.ffn_gate\.weight$=iq3_s
167
+ ^blk\.18\.ffn_up\.weight$=iq3_s
168
+ ^blk\.18\.ssm_out\.weight$=iq3_s
169
+ ^blk\.19\.attn_k\.weight$=q5_k
170
+ ^blk\.19\.attn_output\.weight$=q5_k
171
+ ^blk\.19\.attn_q\.weight$=iq3_s
172
+ ^blk\.19\.attn_v\.weight$=q5_k
173
+ ^blk\.19\.ffn_down\.weight$=iq3_s
174
+ ^blk\.19\.ffn_gate\.weight$=iq3_s
175
+ ^blk\.19\.ffn_up\.weight$=iq3_s
176
+ ^blk\.20\.attn_gate\.weight$=iq3_s
177
+ ^blk\.20\.attn_qkv\.weight$=iq3_s
178
+ ^blk\.20\.ffn_down\.weight$=iq3_s
179
+ ^blk\.20\.ffn_gate\.weight$=iq3_s
180
+ ^blk\.20\.ffn_up\.weight$=iq3_s
181
+ ^blk\.20\.ssm_out\.weight$=iq3_s
182
+ ^blk\.21\.attn_gate\.weight$=iq3_s
183
+ ^blk\.21\.attn_qkv\.weight$=iq3_s
184
+ ^blk\.21\.ffn_down\.weight$=iq3_s
185
+ ^blk\.21\.ffn_gate\.weight$=iq3_s
186
+ ^blk\.21\.ffn_up\.weight$=iq3_s
187
+ ^blk\.21\.ssm_out\.weight$=iq3_s
188
+ ^blk\.22\.attn_gate\.weight$=iq3_s
189
+ ^blk\.22\.attn_qkv\.weight$=iq3_s
190
+ ^blk\.22\.ffn_down\.weight$=iq3_s
191
+ ^blk\.22\.ffn_gate\.weight$=iq3_s
192
+ ^blk\.22\.ffn_up\.weight$=iq3_s
193
+ ^blk\.22\.ssm_out\.weight$=iq3_s
194
+ ^blk\.23\.attn_k\.weight$=q5_k
195
+ ^blk\.23\.attn_output\.weight$=q5_k
196
+ ^blk\.23\.attn_q\.weight$=iq3_s
197
+ ^blk\.23\.attn_v\.weight$=q5_k
198
+ ^blk\.23\.ffn_down\.weight$=iq3_s
199
+ ^blk\.23\.ffn_gate\.weight$=iq3_s
200
+ ^blk\.23\.ffn_up\.weight$=iq3_s
201
+ ^blk\.24\.attn_gate\.weight$=iq3_s
202
+ ^blk\.24\.attn_qkv\.weight$=iq3_s
203
+ ^blk\.24\.ffn_down\.weight$=iq3_s
204
+ ^blk\.24\.ffn_gate\.weight$=iq3_s
205
+ ^blk\.24\.ffn_up\.weight$=iq3_s
206
+ ^blk\.24\.ssm_out\.weight$=iq3_s
207
+ ^blk\.25\.attn_gate\.weight$=iq3_s
208
+ ^blk\.25\.attn_qkv\.weight$=iq3_s
209
+ ^blk\.25\.ffn_down\.weight$=iq3_s
210
+ ^blk\.25\.ffn_gate\.weight$=iq3_s
211
+ ^blk\.25\.ffn_up\.weight$=iq3_s
212
+ ^blk\.25\.ssm_out\.weight$=iq3_s
213
+ ^blk\.26\.attn_gate\.weight$=iq3_s
214
+ ^blk\.26\.attn_qkv\.weight$=iq3_s
215
+ ^blk\.26\.ffn_down\.weight$=iq3_s
216
+ ^blk\.26\.ffn_gate\.weight$=iq3_s
217
+ ^blk\.26\.ffn_up\.weight$=iq3_s
218
+ ^blk\.26\.ssm_out\.weight$=iq3_s
219
+ ^blk\.27\.attn_k\.weight$=q5_k
220
+ ^blk\.27\.attn_output\.weight$=q5_k
221
+ ^blk\.27\.attn_q\.weight$=iq3_s
222
+ ^blk\.27\.attn_v\.weight$=q6_k
223
+ ^blk\.27\.ffn_down\.weight$=iq3_s
224
+ ^blk\.27\.ffn_gate\.weight$=iq3_s
225
+ ^blk\.27\.ffn_up\.weight$=iq3_s
226
+ ^blk\.28\.attn_gate\.weight$=iq3_s
227
+ ^blk\.28\.attn_qkv\.weight$=iq3_s
228
+ ^blk\.28\.ffn_down\.weight$=iq3_s
229
+ ^blk\.28\.ffn_gate\.weight$=iq3_s
230
+ ^blk\.28\.ffn_up\.weight$=iq3_s
231
+ ^blk\.28\.ssm_out\.weight$=iq3_s
232
+ ^blk\.29\.attn_gate\.weight$=iq3_s
233
+ ^blk\.29\.attn_qkv\.weight$=iq3_s
234
+ ^blk\.29\.ffn_down\.weight$=iq3_s
235
+ ^blk\.29\.ffn_gate\.weight$=iq3_s
236
+ ^blk\.29\.ffn_up\.weight$=iq3_s
237
+ ^blk\.29\.ssm_out\.weight$=iq3_s
238
+ ^blk\.30\.attn_gate\.weight$=iq3_s
239
+ ^blk\.30\.attn_qkv\.weight$=iq3_s
240
+ ^blk\.30\.ffn_down\.weight$=iq3_s
241
+ ^blk\.30\.ffn_gate\.weight$=iq3_s
242
+ ^blk\.30\.ffn_up\.weight$=iq3_s
243
+ ^blk\.30\.ssm_out\.weight$=iq3_s
244
+ ^blk\.31\.attn_k\.weight$=q4_k
245
+ ^blk\.31\.attn_output\.weight$=q5_k
246
+ ^blk\.31\.attn_q\.weight$=iq3_s
247
+ ^blk\.31\.attn_v\.weight$=q6_k
248
+ ^blk\.31\.ffn_down\.weight$=q5_k
249
+ ^blk\.31\.ffn_gate\.weight$=iq3_s
250
+ ^blk\.31\.ffn_up\.weight$=q5_k
layouts/LensVLM-9B-IQ3_XXS.layout.json ADDED
@@ -0,0 +1,200 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "IQ3_XXS",
5
+ "base_type": "iq3_xxs",
6
+ "family": "iq",
7
+ "recipe": "ident",
8
+ "variant": "shr0.70",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "M",
12
+ "rung_share": 0.7,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.164,
17
+ "type": "q4_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.1025,
21
+ "type": "q4_k"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q4_k",
131
+ "token_embd.weight": "q4_k"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 4235640832,
135
+ "map_body+embd": 4098457600,
136
+ "ratio": 0.9676,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.7,
140
+ "body_bpw": 3.419,
141
+ "body_bpw_heuristic": 3.587,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 4127868928,
144
+ "file_bpw_pred": 3.688,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "iq3_xxs": 17,
148
+ "iq4_xs": 5,
149
+ "q5_k": 1,
150
+ "iq3_s": 1
151
+ },
152
+ "attn_k": {
153
+ "q5_k": 8
154
+ },
155
+ "attn_output": {
156
+ "q5_k": 8
157
+ },
158
+ "attn_q": {
159
+ "iq3_xxs": 7,
160
+ "iq4_xs": 1
161
+ },
162
+ "attn_qkv": {
163
+ "iq3_xxs": 21,
164
+ "iq4_xs": 2,
165
+ "q5_k": 1
166
+ },
167
+ "attn_v": {
168
+ "q5_k": 5,
169
+ "q6_k": 3
170
+ },
171
+ "ffn_down": {
172
+ "iq3_xxs": 29,
173
+ "iq4_xs": 2,
174
+ "q5_k": 1
175
+ },
176
+ "ffn_gate": {
177
+ "iq3_xxs": 31,
178
+ "q4_k": 1
179
+ },
180
+ "ffn_up": {
181
+ "iq3_xxs": 23,
182
+ "iq4_xs": 8,
183
+ "q5_k": 1
184
+ },
185
+ "ssm_out": {
186
+ "q4_k": 18,
187
+ "q5_k": 5,
188
+ "q6_k": 1
189
+ }
190
+ },
191
+ "prior": "layout/prior.json",
192
+ "pin_rule": "floor",
193
+ "embd_cap": "heur",
194
+ "generator": "auto_quant_v2.layout",
195
+ "generator_key": "93ecba14028d5e4b",
196
+ "quant_type": "IQ3_XXS",
197
+ "quantize_ftype": "IQ3_XXS",
198
+ "llama_cpp_version": "b11081",
199
+ "commit": "d7cfc9bb610f62821944853c508648097409f4a7"
200
+ }
layouts/LensVLM-9B-IQ3_XXS.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q4_k
50
+ ^token_embd\.weight$=q4_k
51
+ ^blk\.0\.attn_gate\.weight$=q5_k
52
+ ^blk\.0\.attn_qkv\.weight$=q5_k
53
+ ^blk\.0\.ffn_down\.weight$=iq4_xs
54
+ ^blk\.0\.ffn_gate\.weight$=iq3_xxs
55
+ ^blk\.0\.ffn_up\.weight$=iq4_xs
56
+ ^blk\.0\.ssm_out\.weight$=q6_k
57
+ ^blk\.1\.attn_gate\.weight$=iq4_xs
58
+ ^blk\.1\.attn_qkv\.weight$=iq4_xs
59
+ ^blk\.1\.ffn_down\.weight$=iq3_xxs
60
+ ^blk\.1\.ffn_gate\.weight$=iq3_xxs
61
+ ^blk\.1\.ffn_up\.weight$=iq4_xs
62
+ ^blk\.1\.ssm_out\.weight$=q5_k
63
+ ^blk\.2\.attn_gate\.weight$=iq4_xs
64
+ ^blk\.2\.attn_qkv\.weight$=iq4_xs
65
+ ^blk\.2\.ffn_down\.weight$=iq3_xxs
66
+ ^blk\.2\.ffn_gate\.weight$=iq3_xxs
67
+ ^blk\.2\.ffn_up\.weight$=iq4_xs
68
+ ^blk\.2\.ssm_out\.weight$=q5_k
69
+ ^blk\.3\.attn_k\.weight$=q5_k
70
+ ^blk\.3\.attn_output\.weight$=q5_k
71
+ ^blk\.3\.attn_q\.weight$=iq3_xxs
72
+ ^blk\.3\.attn_v\.weight$=q5_k
73
+ ^blk\.3\.ffn_down\.weight$=iq3_xxs
74
+ ^blk\.3\.ffn_gate\.weight$=iq3_xxs
75
+ ^blk\.3\.ffn_up\.weight$=iq4_xs
76
+ ^blk\.4\.attn_gate\.weight$=iq4_xs
77
+ ^blk\.4\.attn_qkv\.weight$=iq3_xxs
78
+ ^blk\.4\.ffn_down\.weight$=iq3_xxs
79
+ ^blk\.4\.ffn_gate\.weight$=iq3_xxs
80
+ ^blk\.4\.ffn_up\.weight$=iq4_xs
81
+ ^blk\.4\.ssm_out\.weight$=q5_k
82
+ ^blk\.5\.attn_gate\.weight$=iq4_xs
83
+ ^blk\.5\.attn_qkv\.weight$=iq3_xxs
84
+ ^blk\.5\.ffn_down\.weight$=iq3_xxs
85
+ ^blk\.5\.ffn_gate\.weight$=iq3_xxs
86
+ ^blk\.5\.ffn_up\.weight$=iq4_xs
87
+ ^blk\.5\.ssm_out\.weight$=q5_k
88
+ ^blk\.6\.attn_gate\.weight$=iq4_xs
89
+ ^blk\.6\.attn_qkv\.weight$=iq3_xxs
90
+ ^blk\.6\.ffn_down\.weight$=iq3_xxs
91
+ ^blk\.6\.ffn_gate\.weight$=iq3_xxs
92
+ ^blk\.6\.ffn_up\.weight$=iq4_xs
93
+ ^blk\.6\.ssm_out\.weight$=q5_k
94
+ ^blk\.7\.attn_k\.weight$=q5_k
95
+ ^blk\.7\.attn_output\.weight$=q5_k
96
+ ^blk\.7\.attn_q\.weight$=iq3_xxs
97
+ ^blk\.7\.attn_v\.weight$=q5_k
98
+ ^blk\.7\.ffn_down\.weight$=iq3_xxs
99
+ ^blk\.7\.ffn_gate\.weight$=iq3_xxs
100
+ ^blk\.7\.ffn_up\.weight$=iq4_xs
101
+ ^blk\.8\.attn_gate\.weight$=iq3_xxs
102
+ ^blk\.8\.attn_qkv\.weight$=iq3_xxs
103
+ ^blk\.8\.ffn_down\.weight$=iq3_xxs
104
+ ^blk\.8\.ffn_gate\.weight$=iq3_xxs
105
+ ^blk\.8\.ffn_up\.weight$=iq3_xxs
106
+ ^blk\.8\.ssm_out\.weight$=q4_k
107
+ ^blk\.9\.attn_gate\.weight$=iq3_xxs
108
+ ^blk\.9\.attn_qkv\.weight$=iq3_xxs
109
+ ^blk\.9\.ffn_down\.weight$=iq3_xxs
110
+ ^blk\.9\.ffn_gate\.weight$=iq3_xxs
111
+ ^blk\.9\.ffn_up\.weight$=iq3_xxs
112
+ ^blk\.9\.ssm_out\.weight$=q4_k
113
+ ^blk\.10\.attn_gate\.weight$=iq3_xxs
114
+ ^blk\.10\.attn_qkv\.weight$=iq3_xxs
115
+ ^blk\.10\.ffn_down\.weight$=iq3_xxs
116
+ ^blk\.10\.ffn_gate\.weight$=iq3_xxs
117
+ ^blk\.10\.ffn_up\.weight$=iq3_xxs
118
+ ^blk\.10\.ssm_out\.weight$=q4_k
119
+ ^blk\.11\.attn_k\.weight$=q5_k
120
+ ^blk\.11\.attn_output\.weight$=q5_k
121
+ ^blk\.11\.attn_q\.weight$=iq3_xxs
122
+ ^blk\.11\.attn_v\.weight$=q6_k
123
+ ^blk\.11\.ffn_down\.weight$=iq3_xxs
124
+ ^blk\.11\.ffn_gate\.weight$=iq3_xxs
125
+ ^blk\.11\.ffn_up\.weight$=iq3_xxs
126
+ ^blk\.12\.attn_gate\.weight$=iq3_xxs
127
+ ^blk\.12\.attn_qkv\.weight$=iq3_xxs
128
+ ^blk\.12\.ffn_down\.weight$=iq3_xxs
129
+ ^blk\.12\.ffn_gate\.weight$=iq3_xxs
130
+ ^blk\.12\.ffn_up\.weight$=iq3_xxs
131
+ ^blk\.12\.ssm_out\.weight$=q4_k
132
+ ^blk\.13\.attn_gate\.weight$=iq3_xxs
133
+ ^blk\.13\.attn_qkv\.weight$=iq3_xxs
134
+ ^blk\.13\.ffn_down\.weight$=iq3_xxs
135
+ ^blk\.13\.ffn_gate\.weight$=iq3_xxs
136
+ ^blk\.13\.ffn_up\.weight$=iq3_xxs
137
+ ^blk\.13\.ssm_out\.weight$=q4_k
138
+ ^blk\.14\.attn_gate\.weight$=iq3_xxs
139
+ ^blk\.14\.attn_qkv\.weight$=iq3_xxs
140
+ ^blk\.14\.ffn_down\.weight$=iq3_xxs
141
+ ^blk\.14\.ffn_gate\.weight$=iq3_xxs
142
+ ^blk\.14\.ffn_up\.weight$=iq3_xxs
143
+ ^blk\.14\.ssm_out\.weight$=q4_k
144
+ ^blk\.15\.attn_k\.weight$=q5_k
145
+ ^blk\.15\.attn_output\.weight$=q5_k
146
+ ^blk\.15\.attn_q\.weight$=iq3_xxs
147
+ ^blk\.15\.attn_v\.weight$=q5_k
148
+ ^blk\.15\.ffn_down\.weight$=iq3_xxs
149
+ ^blk\.15\.ffn_gate\.weight$=iq3_xxs
150
+ ^blk\.15\.ffn_up\.weight$=iq3_xxs
151
+ ^blk\.16\.attn_gate\.weight$=iq3_xxs
152
+ ^blk\.16\.attn_qkv\.weight$=iq3_xxs
153
+ ^blk\.16\.ffn_down\.weight$=iq3_xxs
154
+ ^blk\.16\.ffn_gate\.weight$=iq3_xxs
155
+ ^blk\.16\.ffn_up\.weight$=iq3_xxs
156
+ ^blk\.16\.ssm_out\.weight$=q4_k
157
+ ^blk\.17\.attn_gate\.weight$=iq3_xxs
158
+ ^blk\.17\.attn_qkv\.weight$=iq3_xxs
159
+ ^blk\.17\.ffn_down\.weight$=iq3_xxs
160
+ ^blk\.17\.ffn_gate\.weight$=iq3_xxs
161
+ ^blk\.17\.ffn_up\.weight$=iq3_xxs
162
+ ^blk\.17\.ssm_out\.weight$=q4_k
163
+ ^blk\.18\.attn_gate\.weight$=iq3_xxs
164
+ ^blk\.18\.attn_qkv\.weight$=iq3_xxs
165
+ ^blk\.18\.ffn_down\.weight$=iq3_xxs
166
+ ^blk\.18\.ffn_gate\.weight$=iq3_xxs
167
+ ^blk\.18\.ffn_up\.weight$=iq3_xxs
168
+ ^blk\.18\.ssm_out\.weight$=q4_k
169
+ ^blk\.19\.attn_k\.weight$=q5_k
170
+ ^blk\.19\.attn_output\.weight$=q5_k
171
+ ^blk\.19\.attn_q\.weight$=iq3_xxs
172
+ ^blk\.19\.attn_v\.weight$=q5_k
173
+ ^blk\.19\.ffn_down\.weight$=iq3_xxs
174
+ ^blk\.19\.ffn_gate\.weight$=iq3_xxs
175
+ ^blk\.19\.ffn_up\.weight$=iq3_xxs
176
+ ^blk\.20\.attn_gate\.weight$=iq3_xxs
177
+ ^blk\.20\.attn_qkv\.weight$=iq3_xxs
178
+ ^blk\.20\.ffn_down\.weight$=iq3_xxs
179
+ ^blk\.20\.ffn_gate\.weight$=iq3_xxs
180
+ ^blk\.20\.ffn_up\.weight$=iq3_xxs
181
+ ^blk\.20\.ssm_out\.weight$=q4_k
182
+ ^blk\.21\.attn_gate\.weight$=iq3_xxs
183
+ ^blk\.21\.attn_qkv\.weight$=iq3_xxs
184
+ ^blk\.21\.ffn_down\.weight$=iq3_xxs
185
+ ^blk\.21\.ffn_gate\.weight$=iq3_xxs
186
+ ^blk\.21\.ffn_up\.weight$=iq3_xxs
187
+ ^blk\.21\.ssm_out\.weight$=q4_k
188
+ ^blk\.22\.attn_gate\.weight$=iq3_xxs
189
+ ^blk\.22\.attn_qkv\.weight$=iq3_xxs
190
+ ^blk\.22\.ffn_down\.weight$=iq3_xxs
191
+ ^blk\.22\.ffn_gate\.weight$=iq3_xxs
192
+ ^blk\.22\.ffn_up\.weight$=iq3_xxs
193
+ ^blk\.22\.ssm_out\.weight$=q4_k
194
+ ^blk\.23\.attn_k\.weight$=q5_k
195
+ ^blk\.23\.attn_output\.weight$=q5_k
196
+ ^blk\.23\.attn_q\.weight$=iq3_xxs
197
+ ^blk\.23\.attn_v\.weight$=q5_k
198
+ ^blk\.23\.ffn_down\.weight$=iq3_xxs
199
+ ^blk\.23\.ffn_gate\.weight$=iq3_xxs
200
+ ^blk\.23\.ffn_up\.weight$=iq3_xxs
201
+ ^blk\.24\.attn_gate\.weight$=iq3_s
202
+ ^blk\.24\.attn_qkv\.weight$=iq3_xxs
203
+ ^blk\.24\.ffn_down\.weight$=iq3_xxs
204
+ ^blk\.24\.ffn_gate\.weight$=iq3_xxs
205
+ ^blk\.24\.ffn_up\.weight$=iq3_xxs
206
+ ^blk\.24\.ssm_out\.weight$=q4_k
207
+ ^blk\.25\.attn_gate\.weight$=iq3_xxs
208
+ ^blk\.25\.attn_qkv\.weight$=iq3_xxs
209
+ ^blk\.25\.ffn_down\.weight$=iq3_xxs
210
+ ^blk\.25\.ffn_gate\.weight$=iq3_xxs
211
+ ^blk\.25\.ffn_up\.weight$=iq3_xxs
212
+ ^blk\.25\.ssm_out\.weight$=q4_k
213
+ ^blk\.26\.attn_gate\.weight$=iq3_xxs
214
+ ^blk\.26\.attn_qkv\.weight$=iq3_xxs
215
+ ^blk\.26\.ffn_down\.weight$=iq3_xxs
216
+ ^blk\.26\.ffn_gate\.weight$=iq3_xxs
217
+ ^blk\.26\.ffn_up\.weight$=iq3_xxs
218
+ ^blk\.26\.ssm_out\.weight$=q4_k
219
+ ^blk\.27\.attn_k\.weight$=q5_k
220
+ ^blk\.27\.attn_output\.weight$=q5_k
221
+ ^blk\.27\.attn_q\.weight$=iq3_xxs
222
+ ^blk\.27\.attn_v\.weight$=q6_k
223
+ ^blk\.27\.ffn_down\.weight$=iq4_xs
224
+ ^blk\.27\.ffn_gate\.weight$=iq3_xxs
225
+ ^blk\.27\.ffn_up\.weight$=iq3_xxs
226
+ ^blk\.28\.attn_gate\.weight$=iq3_xxs
227
+ ^blk\.28\.attn_qkv\.weight$=iq3_xxs
228
+ ^blk\.28\.ffn_down\.weight$=iq3_xxs
229
+ ^blk\.28\.ffn_gate\.weight$=iq3_xxs
230
+ ^blk\.28\.ffn_up\.weight$=iq3_xxs
231
+ ^blk\.28\.ssm_out\.weight$=q4_k
232
+ ^blk\.29\.attn_gate\.weight$=iq3_xxs
233
+ ^blk\.29\.attn_qkv\.weight$=iq3_xxs
234
+ ^blk\.29\.ffn_down\.weight$=iq3_xxs
235
+ ^blk\.29\.ffn_gate\.weight$=iq3_xxs
236
+ ^blk\.29\.ffn_up\.weight$=iq3_xxs
237
+ ^blk\.29\.ssm_out\.weight$=q4_k
238
+ ^blk\.30\.attn_gate\.weight$=iq3_xxs
239
+ ^blk\.30\.attn_qkv\.weight$=iq3_xxs
240
+ ^blk\.30\.ffn_down\.weight$=iq3_xxs
241
+ ^blk\.30\.ffn_gate\.weight$=iq3_xxs
242
+ ^blk\.30\.ffn_up\.weight$=iq3_xxs
243
+ ^blk\.30\.ssm_out\.weight$=q4_k
244
+ ^blk\.31\.attn_k\.weight$=q5_k
245
+ ^blk\.31\.attn_output\.weight$=q5_k
246
+ ^blk\.31\.attn_q\.weight$=iq4_xs
247
+ ^blk\.31\.attn_v\.weight$=q6_k
248
+ ^blk\.31\.ffn_down\.weight$=q5_k
249
+ ^blk\.31\.ffn_gate\.weight$=q4_k
250
+ ^blk\.31\.ffn_up\.weight$=q5_k
layouts/LensVLM-9B-IQ4_NL.layout.json ADDED
@@ -0,0 +1,197 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "IQ4_NL",
5
+ "base_type": "iq4_nl",
6
+ "family": "iq",
7
+ "recipe": "ident",
8
+ "variant": "shr0.70",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "M",
12
+ "rung_share": 0.7,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.1526,
17
+ "type": "q6_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.1046,
21
+ "type": "iq4_nl"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q6_k",
131
+ "token_embd.weight": "iq4_nl"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 5438521344,
135
+ "map_body+embd": 5784682496,
136
+ "ratio": 1.0636,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.701,
140
+ "body_bpw": 5.067,
141
+ "body_bpw_heuristic": 4.667,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 5814093824,
144
+ "file_bpw_pred": 5.195,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "iq4_nl": 18,
148
+ "q6_k": 5,
149
+ "q8_0": 1
150
+ },
151
+ "attn_k": {
152
+ "q8_0": 8
153
+ },
154
+ "attn_output": {
155
+ "q8_0": 8
156
+ },
157
+ "attn_q": {
158
+ "iq4_nl": 7,
159
+ "q6_k": 1
160
+ },
161
+ "attn_qkv": {
162
+ "iq4_nl": 22,
163
+ "q8_0": 1,
164
+ "q6_k": 1
165
+ },
166
+ "attn_v": {
167
+ "q8_0": 8
168
+ },
169
+ "ffn_down": {
170
+ "iq4_nl": 29,
171
+ "q6_k": 2,
172
+ "q8_0": 1
173
+ },
174
+ "ffn_gate": {
175
+ "iq4_nl": 31,
176
+ "q6_k": 1
177
+ },
178
+ "ffn_up": {
179
+ "iq4_nl": 23,
180
+ "q6_k": 8,
181
+ "q8_0": 1
182
+ },
183
+ "ssm_out": {
184
+ "q6_k": 17,
185
+ "q8_0": 7
186
+ }
187
+ },
188
+ "prior": "layout/prior.json",
189
+ "pin_rule": "floor",
190
+ "embd_cap": "heur",
191
+ "generator": "auto_quant_v2.layout",
192
+ "generator_key": "93ecba14028d5e4b",
193
+ "quant_type": "IQ4_NL",
194
+ "quantize_ftype": "IQ4_NL",
195
+ "llama_cpp_version": "b11081",
196
+ "commit": "d7cfc9bb610f62821944853c508648097409f4a7"
197
+ }
layouts/LensVLM-9B-IQ4_NL.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q6_k
50
+ ^token_embd\.weight$=iq4_nl
51
+ ^blk\.0\.attn_gate\.weight$=q8_0
52
+ ^blk\.0\.attn_qkv\.weight$=q8_0
53
+ ^blk\.0\.ffn_down\.weight$=q6_k
54
+ ^blk\.0\.ffn_gate\.weight$=iq4_nl
55
+ ^blk\.0\.ffn_up\.weight$=q6_k
56
+ ^blk\.0\.ssm_out\.weight$=q8_0
57
+ ^blk\.1\.attn_gate\.weight$=q6_k
58
+ ^blk\.1\.attn_qkv\.weight$=q6_k
59
+ ^blk\.1\.ffn_down\.weight$=iq4_nl
60
+ ^blk\.1\.ffn_gate\.weight$=iq4_nl
61
+ ^blk\.1\.ffn_up\.weight$=q6_k
62
+ ^blk\.1\.ssm_out\.weight$=q8_0
63
+ ^blk\.2\.attn_gate\.weight$=q6_k
64
+ ^blk\.2\.attn_qkv\.weight$=iq4_nl
65
+ ^blk\.2\.ffn_down\.weight$=iq4_nl
66
+ ^blk\.2\.ffn_gate\.weight$=iq4_nl
67
+ ^blk\.2\.ffn_up\.weight$=q6_k
68
+ ^blk\.2\.ssm_out\.weight$=q8_0
69
+ ^blk\.3\.attn_k\.weight$=q8_0
70
+ ^blk\.3\.attn_output\.weight$=q8_0
71
+ ^blk\.3\.attn_q\.weight$=iq4_nl
72
+ ^blk\.3\.attn_v\.weight$=q8_0
73
+ ^blk\.3\.ffn_down\.weight$=iq4_nl
74
+ ^blk\.3\.ffn_gate\.weight$=iq4_nl
75
+ ^blk\.3\.ffn_up\.weight$=q6_k
76
+ ^blk\.4\.attn_gate\.weight$=q6_k
77
+ ^blk\.4\.attn_qkv\.weight$=iq4_nl
78
+ ^blk\.4\.ffn_down\.weight$=iq4_nl
79
+ ^blk\.4\.ffn_gate\.weight$=iq4_nl
80
+ ^blk\.4\.ffn_up\.weight$=q6_k
81
+ ^blk\.4\.ssm_out\.weight$=q8_0
82
+ ^blk\.5\.attn_gate\.weight$=q6_k
83
+ ^blk\.5\.attn_qkv\.weight$=iq4_nl
84
+ ^blk\.5\.ffn_down\.weight$=iq4_nl
85
+ ^blk\.5\.ffn_gate\.weight$=iq4_nl
86
+ ^blk\.5\.ffn_up\.weight$=q6_k
87
+ ^blk\.5\.ssm_out\.weight$=q8_0
88
+ ^blk\.6\.attn_gate\.weight$=q6_k
89
+ ^blk\.6\.attn_qkv\.weight$=iq4_nl
90
+ ^blk\.6\.ffn_down\.weight$=iq4_nl
91
+ ^blk\.6\.ffn_gate\.weight$=iq4_nl
92
+ ^blk\.6\.ffn_up\.weight$=q6_k
93
+ ^blk\.6\.ssm_out\.weight$=q8_0
94
+ ^blk\.7\.attn_k\.weight$=q8_0
95
+ ^blk\.7\.attn_output\.weight$=q8_0
96
+ ^blk\.7\.attn_q\.weight$=iq4_nl
97
+ ^blk\.7\.attn_v\.weight$=q8_0
98
+ ^blk\.7\.ffn_down\.weight$=iq4_nl
99
+ ^blk\.7\.ffn_gate\.weight$=iq4_nl
100
+ ^blk\.7\.ffn_up\.weight$=q6_k
101
+ ^blk\.8\.attn_gate\.weight$=iq4_nl
102
+ ^blk\.8\.attn_qkv\.weight$=iq4_nl
103
+ ^blk\.8\.ffn_down\.weight$=iq4_nl
104
+ ^blk\.8\.ffn_gate\.weight$=iq4_nl
105
+ ^blk\.8\.ffn_up\.weight$=iq4_nl
106
+ ^blk\.8\.ssm_out\.weight$=q6_k
107
+ ^blk\.9\.attn_gate\.weight$=iq4_nl
108
+ ^blk\.9\.attn_qkv\.weight$=iq4_nl
109
+ ^blk\.9\.ffn_down\.weight$=iq4_nl
110
+ ^blk\.9\.ffn_gate\.weight$=iq4_nl
111
+ ^blk\.9\.ffn_up\.weight$=iq4_nl
112
+ ^blk\.9\.ssm_out\.weight$=q6_k
113
+ ^blk\.10\.attn_gate\.weight$=iq4_nl
114
+ ^blk\.10\.attn_qkv\.weight$=iq4_nl
115
+ ^blk\.10\.ffn_down\.weight$=iq4_nl
116
+ ^blk\.10\.ffn_gate\.weight$=iq4_nl
117
+ ^blk\.10\.ffn_up\.weight$=iq4_nl
118
+ ^blk\.10\.ssm_out\.weight$=q6_k
119
+ ^blk\.11\.attn_k\.weight$=q8_0
120
+ ^blk\.11\.attn_output\.weight$=q8_0
121
+ ^blk\.11\.attn_q\.weight$=iq4_nl
122
+ ^blk\.11\.attn_v\.weight$=q8_0
123
+ ^blk\.11\.ffn_down\.weight$=iq4_nl
124
+ ^blk\.11\.ffn_gate\.weight$=iq4_nl
125
+ ^blk\.11\.ffn_up\.weight$=iq4_nl
126
+ ^blk\.12\.attn_gate\.weight$=iq4_nl
127
+ ^blk\.12\.attn_qkv\.weight$=iq4_nl
128
+ ^blk\.12\.ffn_down\.weight$=iq4_nl
129
+ ^blk\.12\.ffn_gate\.weight$=iq4_nl
130
+ ^blk\.12\.ffn_up\.weight$=iq4_nl
131
+ ^blk\.12\.ssm_out\.weight$=q6_k
132
+ ^blk\.13\.attn_gate\.weight$=iq4_nl
133
+ ^blk\.13\.attn_qkv\.weight$=iq4_nl
134
+ ^blk\.13\.ffn_down\.weight$=iq4_nl
135
+ ^blk\.13\.ffn_gate\.weight$=iq4_nl
136
+ ^blk\.13\.ffn_up\.weight$=iq4_nl
137
+ ^blk\.13\.ssm_out\.weight$=q6_k
138
+ ^blk\.14\.attn_gate\.weight$=iq4_nl
139
+ ^blk\.14\.attn_qkv\.weight$=iq4_nl
140
+ ^blk\.14\.ffn_down\.weight$=iq4_nl
141
+ ^blk\.14\.ffn_gate\.weight$=iq4_nl
142
+ ^blk\.14\.ffn_up\.weight$=iq4_nl
143
+ ^blk\.14\.ssm_out\.weight$=q6_k
144
+ ^blk\.15\.attn_k\.weight$=q8_0
145
+ ^blk\.15\.attn_output\.weight$=q8_0
146
+ ^blk\.15\.attn_q\.weight$=iq4_nl
147
+ ^blk\.15\.attn_v\.weight$=q8_0
148
+ ^blk\.15\.ffn_down\.weight$=iq4_nl
149
+ ^blk\.15\.ffn_gate\.weight$=iq4_nl
150
+ ^blk\.15\.ffn_up\.weight$=iq4_nl
151
+ ^blk\.16\.attn_gate\.weight$=iq4_nl
152
+ ^blk\.16\.attn_qkv\.weight$=iq4_nl
153
+ ^blk\.16\.ffn_down\.weight$=iq4_nl
154
+ ^blk\.16\.ffn_gate\.weight$=iq4_nl
155
+ ^blk\.16\.ffn_up\.weight$=iq4_nl
156
+ ^blk\.16\.ssm_out\.weight$=q6_k
157
+ ^blk\.17\.attn_gate\.weight$=iq4_nl
158
+ ^blk\.17\.attn_qkv\.weight$=iq4_nl
159
+ ^blk\.17\.ffn_down\.weight$=iq4_nl
160
+ ^blk\.17\.ffn_gate\.weight$=iq4_nl
161
+ ^blk\.17\.ffn_up\.weight$=iq4_nl
162
+ ^blk\.17\.ssm_out\.weight$=q6_k
163
+ ^blk\.18\.attn_gate\.weight$=iq4_nl
164
+ ^blk\.18\.attn_qkv\.weight$=iq4_nl
165
+ ^blk\.18\.ffn_down\.weight$=iq4_nl
166
+ ^blk\.18\.ffn_gate\.weight$=iq4_nl
167
+ ^blk\.18\.ffn_up\.weight$=iq4_nl
168
+ ^blk\.18\.ssm_out\.weight$=q6_k
169
+ ^blk\.19\.attn_k\.weight$=q8_0
170
+ ^blk\.19\.attn_output\.weight$=q8_0
171
+ ^blk\.19\.attn_q\.weight$=iq4_nl
172
+ ^blk\.19\.attn_v\.weight$=q8_0
173
+ ^blk\.19\.ffn_down\.weight$=iq4_nl
174
+ ^blk\.19\.ffn_gate\.weight$=iq4_nl
175
+ ^blk\.19\.ffn_up\.weight$=iq4_nl
176
+ ^blk\.20\.attn_gate\.weight$=iq4_nl
177
+ ^blk\.20\.attn_qkv\.weight$=iq4_nl
178
+ ^blk\.20\.ffn_down\.weight$=iq4_nl
179
+ ^blk\.20\.ffn_gate\.weight$=iq4_nl
180
+ ^blk\.20\.ffn_up\.weight$=iq4_nl
181
+ ^blk\.20\.ssm_out\.weight$=q6_k
182
+ ^blk\.21\.attn_gate\.weight$=iq4_nl
183
+ ^blk\.21\.attn_qkv\.weight$=iq4_nl
184
+ ^blk\.21\.ffn_down\.weight$=iq4_nl
185
+ ^blk\.21\.ffn_gate\.weight$=iq4_nl
186
+ ^blk\.21\.ffn_up\.weight$=iq4_nl
187
+ ^blk\.21\.ssm_out\.weight$=q6_k
188
+ ^blk\.22\.attn_gate\.weight$=iq4_nl
189
+ ^blk\.22\.attn_qkv\.weight$=iq4_nl
190
+ ^blk\.22\.ffn_down\.weight$=iq4_nl
191
+ ^blk\.22\.ffn_gate\.weight$=iq4_nl
192
+ ^blk\.22\.ffn_up\.weight$=iq4_nl
193
+ ^blk\.22\.ssm_out\.weight$=q6_k
194
+ ^blk\.23\.attn_k\.weight$=q8_0
195
+ ^blk\.23\.attn_output\.weight$=q8_0
196
+ ^blk\.23\.attn_q\.weight$=iq4_nl
197
+ ^blk\.23\.attn_v\.weight$=q8_0
198
+ ^blk\.23\.ffn_down\.weight$=iq4_nl
199
+ ^blk\.23\.ffn_gate\.weight$=iq4_nl
200
+ ^blk\.23\.ffn_up\.weight$=iq4_nl
201
+ ^blk\.24\.attn_gate\.weight$=iq4_nl
202
+ ^blk\.24\.attn_qkv\.weight$=iq4_nl
203
+ ^blk\.24\.ffn_down\.weight$=iq4_nl
204
+ ^blk\.24\.ffn_gate\.weight$=iq4_nl
205
+ ^blk\.24\.ffn_up\.weight$=iq4_nl
206
+ ^blk\.24\.ssm_out\.weight$=q8_0
207
+ ^blk\.25\.attn_gate\.weight$=iq4_nl
208
+ ^blk\.25\.attn_qkv\.weight$=iq4_nl
209
+ ^blk\.25\.ffn_down\.weight$=iq4_nl
210
+ ^blk\.25\.ffn_gate\.weight$=iq4_nl
211
+ ^blk\.25\.ffn_up\.weight$=iq4_nl
212
+ ^blk\.25\.ssm_out\.weight$=q6_k
213
+ ^blk\.26\.attn_gate\.weight$=iq4_nl
214
+ ^blk\.26\.attn_qkv\.weight$=iq4_nl
215
+ ^blk\.26\.ffn_down\.weight$=iq4_nl
216
+ ^blk\.26\.ffn_gate\.weight$=iq4_nl
217
+ ^blk\.26\.ffn_up\.weight$=iq4_nl
218
+ ^blk\.26\.ssm_out\.weight$=q6_k
219
+ ^blk\.27\.attn_k\.weight$=q8_0
220
+ ^blk\.27\.attn_output\.weight$=q8_0
221
+ ^blk\.27\.attn_q\.weight$=iq4_nl
222
+ ^blk\.27\.attn_v\.weight$=q8_0
223
+ ^blk\.27\.ffn_down\.weight$=q6_k
224
+ ^blk\.27\.ffn_gate\.weight$=iq4_nl
225
+ ^blk\.27\.ffn_up\.weight$=iq4_nl
226
+ ^blk\.28\.attn_gate\.weight$=iq4_nl
227
+ ^blk\.28\.attn_qkv\.weight$=iq4_nl
228
+ ^blk\.28\.ffn_down\.weight$=iq4_nl
229
+ ^blk\.28\.ffn_gate\.weight$=iq4_nl
230
+ ^blk\.28\.ffn_up\.weight$=iq4_nl
231
+ ^blk\.28\.ssm_out\.weight$=q6_k
232
+ ^blk\.29\.attn_gate\.weight$=iq4_nl
233
+ ^blk\.29\.attn_qkv\.weight$=iq4_nl
234
+ ^blk\.29\.ffn_down\.weight$=iq4_nl
235
+ ^blk\.29\.ffn_gate\.weight$=iq4_nl
236
+ ^blk\.29\.ffn_up\.weight$=iq4_nl
237
+ ^blk\.29\.ssm_out\.weight$=q6_k
238
+ ^blk\.30\.attn_gate\.weight$=iq4_nl
239
+ ^blk\.30\.attn_qkv\.weight$=iq4_nl
240
+ ^blk\.30\.ffn_down\.weight$=iq4_nl
241
+ ^blk\.30\.ffn_gate\.weight$=iq4_nl
242
+ ^blk\.30\.ffn_up\.weight$=iq4_nl
243
+ ^blk\.30\.ssm_out\.weight$=q6_k
244
+ ^blk\.31\.attn_k\.weight$=q8_0
245
+ ^blk\.31\.attn_output\.weight$=q8_0
246
+ ^blk\.31\.attn_q\.weight$=q6_k
247
+ ^blk\.31\.attn_v\.weight$=q8_0
248
+ ^blk\.31\.ffn_down\.weight$=q8_0
249
+ ^blk\.31\.ffn_gate\.weight$=q6_k
250
+ ^blk\.31\.ffn_up\.weight$=q8_0
layouts/LensVLM-9B-IQ4_XS.layout.json ADDED
@@ -0,0 +1,193 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "IQ4_XS",
5
+ "base_type": "iq4_xs",
6
+ "family": "iq",
7
+ "recipe": "ident",
8
+ "variant": "shr0.90",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "S",
12
+ "rung_share": 0.9,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.1595,
17
+ "type": "q6_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.1033,
21
+ "type": "iq4_xs"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q6_k",
131
+ "token_embd.weight": "iq4_xs"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 5202264064,
135
+ "map_body+embd": 5186928640,
136
+ "ratio": 0.9971,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.9,
140
+ "body_bpw": 4.412,
141
+ "body_bpw_heuristic": 4.43,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 5216339968,
144
+ "file_bpw_pred": 4.661,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "iq4_xs": 23,
148
+ "q8_0": 1
149
+ },
150
+ "attn_k": {
151
+ "q6_k": 8
152
+ },
153
+ "attn_output": {
154
+ "q6_k": 7,
155
+ "q8_0": 1
156
+ },
157
+ "attn_q": {
158
+ "iq4_xs": 8
159
+ },
160
+ "attn_qkv": {
161
+ "iq4_xs": 23,
162
+ "q6_k": 1
163
+ },
164
+ "attn_v": {
165
+ "q8_0": 8
166
+ },
167
+ "ffn_down": {
168
+ "iq4_xs": 31,
169
+ "q6_k": 1
170
+ },
171
+ "ffn_gate": {
172
+ "iq4_xs": 32
173
+ },
174
+ "ffn_up": {
175
+ "iq4_xs": 31,
176
+ "q6_k": 1
177
+ },
178
+ "ssm_out": {
179
+ "iq4_xs": 18,
180
+ "q5_k": 5,
181
+ "q8_0": 1
182
+ }
183
+ },
184
+ "prior": "layout/prior.json",
185
+ "pin_rule": "floor",
186
+ "embd_cap": "heur",
187
+ "generator": "auto_quant_v2.layout",
188
+ "generator_key": "93ecba14028d5e4b",
189
+ "quant_type": "IQ4_XS",
190
+ "quantize_ftype": "IQ4_XS",
191
+ "llama_cpp_version": "b11081",
192
+ "commit": "d7cfc9bb610f62821944853c508648097409f4a7"
193
+ }
layouts/LensVLM-9B-IQ4_XS.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q6_k
50
+ ^token_embd\.weight$=iq4_xs
51
+ ^blk\.0\.attn_gate\.weight$=q8_0
52
+ ^blk\.0\.attn_qkv\.weight$=q6_k
53
+ ^blk\.0\.ffn_down\.weight$=iq4_xs
54
+ ^blk\.0\.ffn_gate\.weight$=iq4_xs
55
+ ^blk\.0\.ffn_up\.weight$=iq4_xs
56
+ ^blk\.0\.ssm_out\.weight$=q8_0
57
+ ^blk\.1\.attn_gate\.weight$=iq4_xs
58
+ ^blk\.1\.attn_qkv\.weight$=iq4_xs
59
+ ^blk\.1\.ffn_down\.weight$=iq4_xs
60
+ ^blk\.1\.ffn_gate\.weight$=iq4_xs
61
+ ^blk\.1\.ffn_up\.weight$=iq4_xs
62
+ ^blk\.1\.ssm_out\.weight$=q5_k
63
+ ^blk\.2\.attn_gate\.weight$=iq4_xs
64
+ ^blk\.2\.attn_qkv\.weight$=iq4_xs
65
+ ^blk\.2\.ffn_down\.weight$=iq4_xs
66
+ ^blk\.2\.ffn_gate\.weight$=iq4_xs
67
+ ^blk\.2\.ffn_up\.weight$=iq4_xs
68
+ ^blk\.2\.ssm_out\.weight$=q5_k
69
+ ^blk\.3\.attn_k\.weight$=q6_k
70
+ ^blk\.3\.attn_output\.weight$=q6_k
71
+ ^blk\.3\.attn_q\.weight$=iq4_xs
72
+ ^blk\.3\.attn_v\.weight$=q8_0
73
+ ^blk\.3\.ffn_down\.weight$=iq4_xs
74
+ ^blk\.3\.ffn_gate\.weight$=iq4_xs
75
+ ^blk\.3\.ffn_up\.weight$=iq4_xs
76
+ ^blk\.4\.attn_gate\.weight$=iq4_xs
77
+ ^blk\.4\.attn_qkv\.weight$=iq4_xs
78
+ ^blk\.4\.ffn_down\.weight$=iq4_xs
79
+ ^blk\.4\.ffn_gate\.weight$=iq4_xs
80
+ ^blk\.4\.ffn_up\.weight$=iq4_xs
81
+ ^blk\.4\.ssm_out\.weight$=q5_k
82
+ ^blk\.5\.attn_gate\.weight$=iq4_xs
83
+ ^blk\.5\.attn_qkv\.weight$=iq4_xs
84
+ ^blk\.5\.ffn_down\.weight$=iq4_xs
85
+ ^blk\.5\.ffn_gate\.weight$=iq4_xs
86
+ ^blk\.5\.ffn_up\.weight$=iq4_xs
87
+ ^blk\.5\.ssm_out\.weight$=q5_k
88
+ ^blk\.6\.attn_gate\.weight$=iq4_xs
89
+ ^blk\.6\.attn_qkv\.weight$=iq4_xs
90
+ ^blk\.6\.ffn_down\.weight$=iq4_xs
91
+ ^blk\.6\.ffn_gate\.weight$=iq4_xs
92
+ ^blk\.6\.ffn_up\.weight$=iq4_xs
93
+ ^blk\.6\.ssm_out\.weight$=q5_k
94
+ ^blk\.7\.attn_k\.weight$=q6_k
95
+ ^blk\.7\.attn_output\.weight$=q6_k
96
+ ^blk\.7\.attn_q\.weight$=iq4_xs
97
+ ^blk\.7\.attn_v\.weight$=q8_0
98
+ ^blk\.7\.ffn_down\.weight$=iq4_xs
99
+ ^blk\.7\.ffn_gate\.weight$=iq4_xs
100
+ ^blk\.7\.ffn_up\.weight$=iq4_xs
101
+ ^blk\.8\.attn_gate\.weight$=iq4_xs
102
+ ^blk\.8\.attn_qkv\.weight$=iq4_xs
103
+ ^blk\.8\.ffn_down\.weight$=iq4_xs
104
+ ^blk\.8\.ffn_gate\.weight$=iq4_xs
105
+ ^blk\.8\.ffn_up\.weight$=iq4_xs
106
+ ^blk\.8\.ssm_out\.weight$=iq4_xs
107
+ ^blk\.9\.attn_gate\.weight$=iq4_xs
108
+ ^blk\.9\.attn_qkv\.weight$=iq4_xs
109
+ ^blk\.9\.ffn_down\.weight$=iq4_xs
110
+ ^blk\.9\.ffn_gate\.weight$=iq4_xs
111
+ ^blk\.9\.ffn_up\.weight$=iq4_xs
112
+ ^blk\.9\.ssm_out\.weight$=iq4_xs
113
+ ^blk\.10\.attn_gate\.weight$=iq4_xs
114
+ ^blk\.10\.attn_qkv\.weight$=iq4_xs
115
+ ^blk\.10\.ffn_down\.weight$=iq4_xs
116
+ ^blk\.10\.ffn_gate\.weight$=iq4_xs
117
+ ^blk\.10\.ffn_up\.weight$=iq4_xs
118
+ ^blk\.10\.ssm_out\.weight$=iq4_xs
119
+ ^blk\.11\.attn_k\.weight$=q6_k
120
+ ^blk\.11\.attn_output\.weight$=q6_k
121
+ ^blk\.11\.attn_q\.weight$=iq4_xs
122
+ ^blk\.11\.attn_v\.weight$=q8_0
123
+ ^blk\.11\.ffn_down\.weight$=iq4_xs
124
+ ^blk\.11\.ffn_gate\.weight$=iq4_xs
125
+ ^blk\.11\.ffn_up\.weight$=iq4_xs
126
+ ^blk\.12\.attn_gate\.weight$=iq4_xs
127
+ ^blk\.12\.attn_qkv\.weight$=iq4_xs
128
+ ^blk\.12\.ffn_down\.weight$=iq4_xs
129
+ ^blk\.12\.ffn_gate\.weight$=iq4_xs
130
+ ^blk\.12\.ffn_up\.weight$=iq4_xs
131
+ ^blk\.12\.ssm_out\.weight$=iq4_xs
132
+ ^blk\.13\.attn_gate\.weight$=iq4_xs
133
+ ^blk\.13\.attn_qkv\.weight$=iq4_xs
134
+ ^blk\.13\.ffn_down\.weight$=iq4_xs
135
+ ^blk\.13\.ffn_gate\.weight$=iq4_xs
136
+ ^blk\.13\.ffn_up\.weight$=iq4_xs
137
+ ^blk\.13\.ssm_out\.weight$=iq4_xs
138
+ ^blk\.14\.attn_gate\.weight$=iq4_xs
139
+ ^blk\.14\.attn_qkv\.weight$=iq4_xs
140
+ ^blk\.14\.ffn_down\.weight$=iq4_xs
141
+ ^blk\.14\.ffn_gate\.weight$=iq4_xs
142
+ ^blk\.14\.ffn_up\.weight$=iq4_xs
143
+ ^blk\.14\.ssm_out\.weight$=iq4_xs
144
+ ^blk\.15\.attn_k\.weight$=q6_k
145
+ ^blk\.15\.attn_output\.weight$=q6_k
146
+ ^blk\.15\.attn_q\.weight$=iq4_xs
147
+ ^blk\.15\.attn_v\.weight$=q8_0
148
+ ^blk\.15\.ffn_down\.weight$=iq4_xs
149
+ ^blk\.15\.ffn_gate\.weight$=iq4_xs
150
+ ^blk\.15\.ffn_up\.weight$=iq4_xs
151
+ ^blk\.16\.attn_gate\.weight$=iq4_xs
152
+ ^blk\.16\.attn_qkv\.weight$=iq4_xs
153
+ ^blk\.16\.ffn_down\.weight$=iq4_xs
154
+ ^blk\.16\.ffn_gate\.weight$=iq4_xs
155
+ ^blk\.16\.ffn_up\.weight$=iq4_xs
156
+ ^blk\.16\.ssm_out\.weight$=iq4_xs
157
+ ^blk\.17\.attn_gate\.weight$=iq4_xs
158
+ ^blk\.17\.attn_qkv\.weight$=iq4_xs
159
+ ^blk\.17\.ffn_down\.weight$=iq4_xs
160
+ ^blk\.17\.ffn_gate\.weight$=iq4_xs
161
+ ^blk\.17\.ffn_up\.weight$=iq4_xs
162
+ ^blk\.17\.ssm_out\.weight$=iq4_xs
163
+ ^blk\.18\.attn_gate\.weight$=iq4_xs
164
+ ^blk\.18\.attn_qkv\.weight$=iq4_xs
165
+ ^blk\.18\.ffn_down\.weight$=iq4_xs
166
+ ^blk\.18\.ffn_gate\.weight$=iq4_xs
167
+ ^blk\.18\.ffn_up\.weight$=iq4_xs
168
+ ^blk\.18\.ssm_out\.weight$=iq4_xs
169
+ ^blk\.19\.attn_k\.weight$=q6_k
170
+ ^blk\.19\.attn_output\.weight$=q6_k
171
+ ^blk\.19\.attn_q\.weight$=iq4_xs
172
+ ^blk\.19\.attn_v\.weight$=q8_0
173
+ ^blk\.19\.ffn_down\.weight$=iq4_xs
174
+ ^blk\.19\.ffn_gate\.weight$=iq4_xs
175
+ ^blk\.19\.ffn_up\.weight$=iq4_xs
176
+ ^blk\.20\.attn_gate\.weight$=iq4_xs
177
+ ^blk\.20\.attn_qkv\.weight$=iq4_xs
178
+ ^blk\.20\.ffn_down\.weight$=iq4_xs
179
+ ^blk\.20\.ffn_gate\.weight$=iq4_xs
180
+ ^blk\.20\.ffn_up\.weight$=iq4_xs
181
+ ^blk\.20\.ssm_out\.weight$=iq4_xs
182
+ ^blk\.21\.attn_gate\.weight$=iq4_xs
183
+ ^blk\.21\.attn_qkv\.weight$=iq4_xs
184
+ ^blk\.21\.ffn_down\.weight$=iq4_xs
185
+ ^blk\.21\.ffn_gate\.weight$=iq4_xs
186
+ ^blk\.21\.ffn_up\.weight$=iq4_xs
187
+ ^blk\.21\.ssm_out\.weight$=iq4_xs
188
+ ^blk\.22\.attn_gate\.weight$=iq4_xs
189
+ ^blk\.22\.attn_qkv\.weight$=iq4_xs
190
+ ^blk\.22\.ffn_down\.weight$=iq4_xs
191
+ ^blk\.22\.ffn_gate\.weight$=iq4_xs
192
+ ^blk\.22\.ffn_up\.weight$=iq4_xs
193
+ ^blk\.22\.ssm_out\.weight$=iq4_xs
194
+ ^blk\.23\.attn_k\.weight$=q6_k
195
+ ^blk\.23\.attn_output\.weight$=q6_k
196
+ ^blk\.23\.attn_q\.weight$=iq4_xs
197
+ ^blk\.23\.attn_v\.weight$=q8_0
198
+ ^blk\.23\.ffn_down\.weight$=iq4_xs
199
+ ^blk\.23\.ffn_gate\.weight$=iq4_xs
200
+ ^blk\.23\.ffn_up\.weight$=iq4_xs
201
+ ^blk\.24\.attn_gate\.weight$=iq4_xs
202
+ ^blk\.24\.attn_qkv\.weight$=iq4_xs
203
+ ^blk\.24\.ffn_down\.weight$=iq4_xs
204
+ ^blk\.24\.ffn_gate\.weight$=iq4_xs
205
+ ^blk\.24\.ffn_up\.weight$=iq4_xs
206
+ ^blk\.24\.ssm_out\.weight$=iq4_xs
207
+ ^blk\.25\.attn_gate\.weight$=iq4_xs
208
+ ^blk\.25\.attn_qkv\.weight$=iq4_xs
209
+ ^blk\.25\.ffn_down\.weight$=iq4_xs
210
+ ^blk\.25\.ffn_gate\.weight$=iq4_xs
211
+ ^blk\.25\.ffn_up\.weight$=iq4_xs
212
+ ^blk\.25\.ssm_out\.weight$=iq4_xs
213
+ ^blk\.26\.attn_gate\.weight$=iq4_xs
214
+ ^blk\.26\.attn_qkv\.weight$=iq4_xs
215
+ ^blk\.26\.ffn_down\.weight$=iq4_xs
216
+ ^blk\.26\.ffn_gate\.weight$=iq4_xs
217
+ ^blk\.26\.ffn_up\.weight$=iq4_xs
218
+ ^blk\.26\.ssm_out\.weight$=iq4_xs
219
+ ^blk\.27\.attn_k\.weight$=q6_k
220
+ ^blk\.27\.attn_output\.weight$=q6_k
221
+ ^blk\.27\.attn_q\.weight$=iq4_xs
222
+ ^blk\.27\.attn_v\.weight$=q8_0
223
+ ^blk\.27\.ffn_down\.weight$=iq4_xs
224
+ ^blk\.27\.ffn_gate\.weight$=iq4_xs
225
+ ^blk\.27\.ffn_up\.weight$=iq4_xs
226
+ ^blk\.28\.attn_gate\.weight$=iq4_xs
227
+ ^blk\.28\.attn_qkv\.weight$=iq4_xs
228
+ ^blk\.28\.ffn_down\.weight$=iq4_xs
229
+ ^blk\.28\.ffn_gate\.weight$=iq4_xs
230
+ ^blk\.28\.ffn_up\.weight$=iq4_xs
231
+ ^blk\.28\.ssm_out\.weight$=iq4_xs
232
+ ^blk\.29\.attn_gate\.weight$=iq4_xs
233
+ ^blk\.29\.attn_qkv\.weight$=iq4_xs
234
+ ^blk\.29\.ffn_down\.weight$=iq4_xs
235
+ ^blk\.29\.ffn_gate\.weight$=iq4_xs
236
+ ^blk\.29\.ffn_up\.weight$=iq4_xs
237
+ ^blk\.29\.ssm_out\.weight$=iq4_xs
238
+ ^blk\.30\.attn_gate\.weight$=iq4_xs
239
+ ^blk\.30\.attn_qkv\.weight$=iq4_xs
240
+ ^blk\.30\.ffn_down\.weight$=iq4_xs
241
+ ^blk\.30\.ffn_gate\.weight$=iq4_xs
242
+ ^blk\.30\.ffn_up\.weight$=iq4_xs
243
+ ^blk\.30\.ssm_out\.weight$=iq4_xs
244
+ ^blk\.31\.attn_k\.weight$=q6_k
245
+ ^blk\.31\.attn_output\.weight$=q8_0
246
+ ^blk\.31\.attn_q\.weight$=iq4_xs
247
+ ^blk\.31\.attn_v\.weight$=q8_0
248
+ ^blk\.31\.ffn_down\.weight$=q6_k
249
+ ^blk\.31\.ffn_gate\.weight$=iq4_xs
250
+ ^blk\.31\.ffn_up\.weight$=q6_k
layouts/LensVLM-9B-Q2_K.layout.json ADDED
@@ -0,0 +1,197 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "Q2_K",
5
+ "base_type": "q2_k",
6
+ "family": "k",
7
+ "recipe": "ident",
8
+ "variant": "shr0.70",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "M",
12
+ "rung_share": 0.7,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.2058,
17
+ "type": "q4_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.0823,
21
+ "type": "q3_k"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q4_k",
131
+ "token_embd.weight": "q3_k"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 4023894016,
135
+ "map_body+embd": 3604004864,
136
+ "ratio": 0.8957,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.7,
140
+ "body_bpw": 3.003,
141
+ "body_bpw_heuristic": 3.305,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 3633416192,
144
+ "file_bpw_pred": 3.246,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "q2_k": 18,
148
+ "q4_k": 5,
149
+ "q5_k": 1
150
+ },
151
+ "attn_k": {
152
+ "q4_k": 7,
153
+ "q5_k": 1
154
+ },
155
+ "attn_output": {
156
+ "q4_k": 7,
157
+ "q5_k": 1
158
+ },
159
+ "attn_q": {
160
+ "q2_k": 7,
161
+ "q4_k": 1
162
+ },
163
+ "attn_qkv": {
164
+ "q2_k": 23,
165
+ "q4_k": 1
166
+ },
167
+ "attn_v": {
168
+ "q5_k": 8
169
+ },
170
+ "ffn_down": {
171
+ "q2_k": 30,
172
+ "q4_k": 2
173
+ },
174
+ "ffn_gate": {
175
+ "q2_k": 31,
176
+ "q4_k": 1
177
+ },
178
+ "ffn_up": {
179
+ "q2_k": 23,
180
+ "q4_k": 8,
181
+ "q3_k": 1
182
+ },
183
+ "ssm_out": {
184
+ "q4_k": 23,
185
+ "q5_k": 1
186
+ }
187
+ },
188
+ "prior": "layout/prior.json",
189
+ "pin_rule": "floor",
190
+ "embd_cap": "heur",
191
+ "generator": "auto_quant_v2.layout",
192
+ "generator_key": "93ecba14028d5e4b",
193
+ "quant_type": "Q2_K",
194
+ "quantize_ftype": "Q2_K",
195
+ "llama_cpp_version": "b11081",
196
+ "commit": "3724aa1c0eb238a683e222c6f4df4b0e71ed3d74"
197
+ }
layouts/LensVLM-9B-Q2_K.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q4_k
50
+ ^token_embd\.weight$=q3_k
51
+ ^blk\.0\.attn_gate\.weight$=q5_k
52
+ ^blk\.0\.attn_qkv\.weight$=q4_k
53
+ ^blk\.0\.ffn_down\.weight$=q4_k
54
+ ^blk\.0\.ffn_gate\.weight$=q2_k
55
+ ^blk\.0\.ffn_up\.weight$=q3_k
56
+ ^blk\.0\.ssm_out\.weight$=q5_k
57
+ ^blk\.1\.attn_gate\.weight$=q4_k
58
+ ^blk\.1\.attn_qkv\.weight$=q2_k
59
+ ^blk\.1\.ffn_down\.weight$=q2_k
60
+ ^blk\.1\.ffn_gate\.weight$=q2_k
61
+ ^blk\.1\.ffn_up\.weight$=q4_k
62
+ ^blk\.1\.ssm_out\.weight$=q4_k
63
+ ^blk\.2\.attn_gate\.weight$=q4_k
64
+ ^blk\.2\.attn_qkv\.weight$=q2_k
65
+ ^blk\.2\.ffn_down\.weight$=q2_k
66
+ ^blk\.2\.ffn_gate\.weight$=q2_k
67
+ ^blk\.2\.ffn_up\.weight$=q4_k
68
+ ^blk\.2\.ssm_out\.weight$=q4_k
69
+ ^blk\.3\.attn_k\.weight$=q4_k
70
+ ^blk\.3\.attn_output\.weight$=q4_k
71
+ ^blk\.3\.attn_q\.weight$=q2_k
72
+ ^blk\.3\.attn_v\.weight$=q5_k
73
+ ^blk\.3\.ffn_down\.weight$=q2_k
74
+ ^blk\.3\.ffn_gate\.weight$=q2_k
75
+ ^blk\.3\.ffn_up\.weight$=q4_k
76
+ ^blk\.4\.attn_gate\.weight$=q4_k
77
+ ^blk\.4\.attn_qkv\.weight$=q2_k
78
+ ^blk\.4\.ffn_down\.weight$=q2_k
79
+ ^blk\.4\.ffn_gate\.weight$=q2_k
80
+ ^blk\.4\.ffn_up\.weight$=q4_k
81
+ ^blk\.4\.ssm_out\.weight$=q4_k
82
+ ^blk\.5\.attn_gate\.weight$=q4_k
83
+ ^blk\.5\.attn_qkv\.weight$=q2_k
84
+ ^blk\.5\.ffn_down\.weight$=q2_k
85
+ ^blk\.5\.ffn_gate\.weight$=q2_k
86
+ ^blk\.5\.ffn_up\.weight$=q4_k
87
+ ^blk\.5\.ssm_out\.weight$=q4_k
88
+ ^blk\.6\.attn_gate\.weight$=q4_k
89
+ ^blk\.6\.attn_qkv\.weight$=q2_k
90
+ ^blk\.6\.ffn_down\.weight$=q2_k
91
+ ^blk\.6\.ffn_gate\.weight$=q2_k
92
+ ^blk\.6\.ffn_up\.weight$=q4_k
93
+ ^blk\.6\.ssm_out\.weight$=q4_k
94
+ ^blk\.7\.attn_k\.weight$=q4_k
95
+ ^blk\.7\.attn_output\.weight$=q4_k
96
+ ^blk\.7\.attn_q\.weight$=q2_k
97
+ ^blk\.7\.attn_v\.weight$=q5_k
98
+ ^blk\.7\.ffn_down\.weight$=q2_k
99
+ ^blk\.7\.ffn_gate\.weight$=q2_k
100
+ ^blk\.7\.ffn_up\.weight$=q4_k
101
+ ^blk\.8\.attn_gate\.weight$=q2_k
102
+ ^blk\.8\.attn_qkv\.weight$=q2_k
103
+ ^blk\.8\.ffn_down\.weight$=q2_k
104
+ ^blk\.8\.ffn_gate\.weight$=q2_k
105
+ ^blk\.8\.ffn_up\.weight$=q2_k
106
+ ^blk\.8\.ssm_out\.weight$=q4_k
107
+ ^blk\.9\.attn_gate\.weight$=q2_k
108
+ ^blk\.9\.attn_qkv\.weight$=q2_k
109
+ ^blk\.9\.ffn_down\.weight$=q2_k
110
+ ^blk\.9\.ffn_gate\.weight$=q2_k
111
+ ^blk\.9\.ffn_up\.weight$=q2_k
112
+ ^blk\.9\.ssm_out\.weight$=q4_k
113
+ ^blk\.10\.attn_gate\.weight$=q2_k
114
+ ^blk\.10\.attn_qkv\.weight$=q2_k
115
+ ^blk\.10\.ffn_down\.weight$=q2_k
116
+ ^blk\.10\.ffn_gate\.weight$=q2_k
117
+ ^blk\.10\.ffn_up\.weight$=q2_k
118
+ ^blk\.10\.ssm_out\.weight$=q4_k
119
+ ^blk\.11\.attn_k\.weight$=q4_k
120
+ ^blk\.11\.attn_output\.weight$=q4_k
121
+ ^blk\.11\.attn_q\.weight$=q2_k
122
+ ^blk\.11\.attn_v\.weight$=q5_k
123
+ ^blk\.11\.ffn_down\.weight$=q2_k
124
+ ^blk\.11\.ffn_gate\.weight$=q2_k
125
+ ^blk\.11\.ffn_up\.weight$=q2_k
126
+ ^blk\.12\.attn_gate\.weight$=q2_k
127
+ ^blk\.12\.attn_qkv\.weight$=q2_k
128
+ ^blk\.12\.ffn_down\.weight$=q2_k
129
+ ^blk\.12\.ffn_gate\.weight$=q2_k
130
+ ^blk\.12\.ffn_up\.weight$=q2_k
131
+ ^blk\.12\.ssm_out\.weight$=q4_k
132
+ ^blk\.13\.attn_gate\.weight$=q2_k
133
+ ^blk\.13\.attn_qkv\.weight$=q2_k
134
+ ^blk\.13\.ffn_down\.weight$=q2_k
135
+ ^blk\.13\.ffn_gate\.weight$=q2_k
136
+ ^blk\.13\.ffn_up\.weight$=q2_k
137
+ ^blk\.13\.ssm_out\.weight$=q4_k
138
+ ^blk\.14\.attn_gate\.weight$=q2_k
139
+ ^blk\.14\.attn_qkv\.weight$=q2_k
140
+ ^blk\.14\.ffn_down\.weight$=q2_k
141
+ ^blk\.14\.ffn_gate\.weight$=q2_k
142
+ ^blk\.14\.ffn_up\.weight$=q2_k
143
+ ^blk\.14\.ssm_out\.weight$=q4_k
144
+ ^blk\.15\.attn_k\.weight$=q4_k
145
+ ^blk\.15\.attn_output\.weight$=q4_k
146
+ ^blk\.15\.attn_q\.weight$=q2_k
147
+ ^blk\.15\.attn_v\.weight$=q5_k
148
+ ^blk\.15\.ffn_down\.weight$=q2_k
149
+ ^blk\.15\.ffn_gate\.weight$=q2_k
150
+ ^blk\.15\.ffn_up\.weight$=q2_k
151
+ ^blk\.16\.attn_gate\.weight$=q2_k
152
+ ^blk\.16\.attn_qkv\.weight$=q2_k
153
+ ^blk\.16\.ffn_down\.weight$=q2_k
154
+ ^blk\.16\.ffn_gate\.weight$=q2_k
155
+ ^blk\.16\.ffn_up\.weight$=q2_k
156
+ ^blk\.16\.ssm_out\.weight$=q4_k
157
+ ^blk\.17\.attn_gate\.weight$=q2_k
158
+ ^blk\.17\.attn_qkv\.weight$=q2_k
159
+ ^blk\.17\.ffn_down\.weight$=q2_k
160
+ ^blk\.17\.ffn_gate\.weight$=q2_k
161
+ ^blk\.17\.ffn_up\.weight$=q2_k
162
+ ^blk\.17\.ssm_out\.weight$=q4_k
163
+ ^blk\.18\.attn_gate\.weight$=q2_k
164
+ ^blk\.18\.attn_qkv\.weight$=q2_k
165
+ ^blk\.18\.ffn_down\.weight$=q2_k
166
+ ^blk\.18\.ffn_gate\.weight$=q2_k
167
+ ^blk\.18\.ffn_up\.weight$=q2_k
168
+ ^blk\.18\.ssm_out\.weight$=q4_k
169
+ ^blk\.19\.attn_k\.weight$=q4_k
170
+ ^blk\.19\.attn_output\.weight$=q4_k
171
+ ^blk\.19\.attn_q\.weight$=q2_k
172
+ ^blk\.19\.attn_v\.weight$=q5_k
173
+ ^blk\.19\.ffn_down\.weight$=q2_k
174
+ ^blk\.19\.ffn_gate\.weight$=q2_k
175
+ ^blk\.19\.ffn_up\.weight$=q2_k
176
+ ^blk\.20\.attn_gate\.weight$=q2_k
177
+ ^blk\.20\.attn_qkv\.weight$=q2_k
178
+ ^blk\.20\.ffn_down\.weight$=q2_k
179
+ ^blk\.20\.ffn_gate\.weight$=q2_k
180
+ ^blk\.20\.ffn_up\.weight$=q2_k
181
+ ^blk\.20\.ssm_out\.weight$=q4_k
182
+ ^blk\.21\.attn_gate\.weight$=q2_k
183
+ ^blk\.21\.attn_qkv\.weight$=q2_k
184
+ ^blk\.21\.ffn_down\.weight$=q2_k
185
+ ^blk\.21\.ffn_gate\.weight$=q2_k
186
+ ^blk\.21\.ffn_up\.weight$=q2_k
187
+ ^blk\.21\.ssm_out\.weight$=q4_k
188
+ ^blk\.22\.attn_gate\.weight$=q2_k
189
+ ^blk\.22\.attn_qkv\.weight$=q2_k
190
+ ^blk\.22\.ffn_down\.weight$=q2_k
191
+ ^blk\.22\.ffn_gate\.weight$=q2_k
192
+ ^blk\.22\.ffn_up\.weight$=q2_k
193
+ ^blk\.22\.ssm_out\.weight$=q4_k
194
+ ^blk\.23\.attn_k\.weight$=q4_k
195
+ ^blk\.23\.attn_output\.weight$=q4_k
196
+ ^blk\.23\.attn_q\.weight$=q2_k
197
+ ^blk\.23\.attn_v\.weight$=q5_k
198
+ ^blk\.23\.ffn_down\.weight$=q2_k
199
+ ^blk\.23\.ffn_gate\.weight$=q2_k
200
+ ^blk\.23\.ffn_up\.weight$=q2_k
201
+ ^blk\.24\.attn_gate\.weight$=q2_k
202
+ ^blk\.24\.attn_qkv\.weight$=q2_k
203
+ ^blk\.24\.ffn_down\.weight$=q2_k
204
+ ^blk\.24\.ffn_gate\.weight$=q2_k
205
+ ^blk\.24\.ffn_up\.weight$=q2_k
206
+ ^blk\.24\.ssm_out\.weight$=q4_k
207
+ ^blk\.25\.attn_gate\.weight$=q2_k
208
+ ^blk\.25\.attn_qkv\.weight$=q2_k
209
+ ^blk\.25\.ffn_down\.weight$=q2_k
210
+ ^blk\.25\.ffn_gate\.weight$=q2_k
211
+ ^blk\.25\.ffn_up\.weight$=q2_k
212
+ ^blk\.25\.ssm_out\.weight$=q4_k
213
+ ^blk\.26\.attn_gate\.weight$=q2_k
214
+ ^blk\.26\.attn_qkv\.weight$=q2_k
215
+ ^blk\.26\.ffn_down\.weight$=q2_k
216
+ ^blk\.26\.ffn_gate\.weight$=q2_k
217
+ ^blk\.26\.ffn_up\.weight$=q2_k
218
+ ^blk\.26\.ssm_out\.weight$=q4_k
219
+ ^blk\.27\.attn_k\.weight$=q5_k
220
+ ^blk\.27\.attn_output\.weight$=q4_k
221
+ ^blk\.27\.attn_q\.weight$=q2_k
222
+ ^blk\.27\.attn_v\.weight$=q5_k
223
+ ^blk\.27\.ffn_down\.weight$=q2_k
224
+ ^blk\.27\.ffn_gate\.weight$=q2_k
225
+ ^blk\.27\.ffn_up\.weight$=q2_k
226
+ ^blk\.28\.attn_gate\.weight$=q2_k
227
+ ^blk\.28\.attn_qkv\.weight$=q2_k
228
+ ^blk\.28\.ffn_down\.weight$=q2_k
229
+ ^blk\.28\.ffn_gate\.weight$=q2_k
230
+ ^blk\.28\.ffn_up\.weight$=q2_k
231
+ ^blk\.28\.ssm_out\.weight$=q4_k
232
+ ^blk\.29\.attn_gate\.weight$=q2_k
233
+ ^blk\.29\.attn_qkv\.weight$=q2_k
234
+ ^blk\.29\.ffn_down\.weight$=q2_k
235
+ ^blk\.29\.ffn_gate\.weight$=q2_k
236
+ ^blk\.29\.ffn_up\.weight$=q2_k
237
+ ^blk\.29\.ssm_out\.weight$=q4_k
238
+ ^blk\.30\.attn_gate\.weight$=q2_k
239
+ ^blk\.30\.attn_qkv\.weight$=q2_k
240
+ ^blk\.30\.ffn_down\.weight$=q2_k
241
+ ^blk\.30\.ffn_gate\.weight$=q2_k
242
+ ^blk\.30\.ffn_up\.weight$=q2_k
243
+ ^blk\.30\.ssm_out\.weight$=q4_k
244
+ ^blk\.31\.attn_k\.weight$=q4_k
245
+ ^blk\.31\.attn_output\.weight$=q5_k
246
+ ^blk\.31\.attn_q\.weight$=q4_k
247
+ ^blk\.31\.attn_v\.weight$=q5_k
248
+ ^blk\.31\.ffn_down\.weight$=q4_k
249
+ ^blk\.31\.ffn_gate\.weight$=q4_k
250
+ ^blk\.31\.ffn_up\.weight$=q4_k
layouts/LensVLM-9B-Q3_K_L.layout.json ADDED
@@ -0,0 +1,200 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "Q3_K_L",
5
+ "base_type": "q3_k",
6
+ "family": "k",
7
+ "recipe": "ident",
8
+ "variant": "shr0.50",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "L",
12
+ "rung_share": 0.5,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.1636,
17
+ "type": "q4_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.0857,
21
+ "type": "q4_k"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q4_k",
131
+ "token_embd.weight": "q4_k"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 5070651392,
135
+ "map_body+embd": 4618780672,
136
+ "ratio": 0.9109,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.5,
140
+ "body_bpw": 4.021,
141
+ "body_bpw_heuristic": 4.397,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 4648192000,
144
+ "file_bpw_pred": 4.153,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "q3_k": 12,
148
+ "q4_k": 6,
149
+ "q5_k": 5,
150
+ "q6_k": 1
151
+ },
152
+ "attn_k": {
153
+ "q5_k": 6,
154
+ "q6_k": 2
155
+ },
156
+ "attn_output": {
157
+ "q5_k": 7,
158
+ "q6_k": 1
159
+ },
160
+ "attn_q": {
161
+ "q3_k": 5,
162
+ "q4_k": 3
163
+ },
164
+ "attn_qkv": {
165
+ "q4_k": 17,
166
+ "q3_k": 6,
167
+ "q5_k": 1
168
+ },
169
+ "attn_v": {
170
+ "q6_k": 8
171
+ },
172
+ "ffn_down": {
173
+ "q3_k": 16,
174
+ "q4_k": 15,
175
+ "q5_k": 1
176
+ },
177
+ "ffn_gate": {
178
+ "q3_k": 31,
179
+ "q5_k": 1
180
+ },
181
+ "ffn_up": {
182
+ "q3_k": 22,
183
+ "q4_k": 9,
184
+ "q5_k": 1
185
+ },
186
+ "ssm_out": {
187
+ "q5_k": 23,
188
+ "q6_k": 1
189
+ }
190
+ },
191
+ "prior": "layout/prior.json",
192
+ "pin_rule": "floor",
193
+ "embd_cap": "heur",
194
+ "generator": "auto_quant_v2.layout",
195
+ "generator_key": "93ecba14028d5e4b",
196
+ "quant_type": "Q3_K_L",
197
+ "quantize_ftype": "Q3_K_L",
198
+ "llama_cpp_version": "b11081",
199
+ "commit": "d7cfc9bb610f62821944853c508648097409f4a7"
200
+ }
layouts/LensVLM-9B-Q3_K_L.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q4_k
50
+ ^token_embd\.weight$=q4_k
51
+ ^blk\.0\.attn_gate\.weight$=q6_k
52
+ ^blk\.0\.attn_qkv\.weight$=q5_k
53
+ ^blk\.0\.ffn_down\.weight$=q4_k
54
+ ^blk\.0\.ffn_gate\.weight$=q3_k
55
+ ^blk\.0\.ffn_up\.weight$=q4_k
56
+ ^blk\.0\.ssm_out\.weight$=q6_k
57
+ ^blk\.1\.attn_gate\.weight$=q5_k
58
+ ^blk\.1\.attn_qkv\.weight$=q4_k
59
+ ^blk\.1\.ffn_down\.weight$=q4_k
60
+ ^blk\.1\.ffn_gate\.weight$=q3_k
61
+ ^blk\.1\.ffn_up\.weight$=q4_k
62
+ ^blk\.1\.ssm_out\.weight$=q5_k
63
+ ^blk\.2\.attn_gate\.weight$=q5_k
64
+ ^blk\.2\.attn_qkv\.weight$=q4_k
65
+ ^blk\.2\.ffn_down\.weight$=q4_k
66
+ ^blk\.2\.ffn_gate\.weight$=q3_k
67
+ ^blk\.2\.ffn_up\.weight$=q4_k
68
+ ^blk\.2\.ssm_out\.weight$=q5_k
69
+ ^blk\.3\.attn_k\.weight$=q5_k
70
+ ^blk\.3\.attn_output\.weight$=q5_k
71
+ ^blk\.3\.attn_q\.weight$=q4_k
72
+ ^blk\.3\.attn_v\.weight$=q6_k
73
+ ^blk\.3\.ffn_down\.weight$=q4_k
74
+ ^blk\.3\.ffn_gate\.weight$=q3_k
75
+ ^blk\.3\.ffn_up\.weight$=q4_k
76
+ ^blk\.4\.attn_gate\.weight$=q5_k
77
+ ^blk\.4\.attn_qkv\.weight$=q4_k
78
+ ^blk\.4\.ffn_down\.weight$=q4_k
79
+ ^blk\.4\.ffn_gate\.weight$=q3_k
80
+ ^blk\.4\.ffn_up\.weight$=q4_k
81
+ ^blk\.4\.ssm_out\.weight$=q5_k
82
+ ^blk\.5\.attn_gate\.weight$=q5_k
83
+ ^blk\.5\.attn_qkv\.weight$=q4_k
84
+ ^blk\.5\.ffn_down\.weight$=q4_k
85
+ ^blk\.5\.ffn_gate\.weight$=q3_k
86
+ ^blk\.5\.ffn_up\.weight$=q4_k
87
+ ^blk\.5\.ssm_out\.weight$=q5_k
88
+ ^blk\.6\.attn_gate\.weight$=q5_k
89
+ ^blk\.6\.attn_qkv\.weight$=q4_k
90
+ ^blk\.6\.ffn_down\.weight$=q4_k
91
+ ^blk\.6\.ffn_gate\.weight$=q3_k
92
+ ^blk\.6\.ffn_up\.weight$=q4_k
93
+ ^blk\.6\.ssm_out\.weight$=q5_k
94
+ ^blk\.7\.attn_k\.weight$=q5_k
95
+ ^blk\.7\.attn_output\.weight$=q5_k
96
+ ^blk\.7\.attn_q\.weight$=q3_k
97
+ ^blk\.7\.attn_v\.weight$=q6_k
98
+ ^blk\.7\.ffn_down\.weight$=q4_k
99
+ ^blk\.7\.ffn_gate\.weight$=q3_k
100
+ ^blk\.7\.ffn_up\.weight$=q4_k
101
+ ^blk\.8\.attn_gate\.weight$=q3_k
102
+ ^blk\.8\.attn_qkv\.weight$=q4_k
103
+ ^blk\.8\.ffn_down\.weight$=q3_k
104
+ ^blk\.8\.ffn_gate\.weight$=q3_k
105
+ ^blk\.8\.ffn_up\.weight$=q3_k
106
+ ^blk\.8\.ssm_out\.weight$=q5_k
107
+ ^blk\.9\.attn_gate\.weight$=q3_k
108
+ ^blk\.9\.attn_qkv\.weight$=q4_k
109
+ ^blk\.9\.ffn_down\.weight$=q3_k
110
+ ^blk\.9\.ffn_gate\.weight$=q3_k
111
+ ^blk\.9\.ffn_up\.weight$=q3_k
112
+ ^blk\.9\.ssm_out\.weight$=q5_k
113
+ ^blk\.10\.attn_gate\.weight$=q3_k
114
+ ^blk\.10\.attn_qkv\.weight$=q4_k
115
+ ^blk\.10\.ffn_down\.weight$=q3_k
116
+ ^blk\.10\.ffn_gate\.weight$=q3_k
117
+ ^blk\.10\.ffn_up\.weight$=q3_k
118
+ ^blk\.10\.ssm_out\.weight$=q5_k
119
+ ^blk\.11\.attn_k\.weight$=q6_k
120
+ ^blk\.11\.attn_output\.weight$=q5_k
121
+ ^blk\.11\.attn_q\.weight$=q3_k
122
+ ^blk\.11\.attn_v\.weight$=q6_k
123
+ ^blk\.11\.ffn_down\.weight$=q3_k
124
+ ^blk\.11\.ffn_gate\.weight$=q3_k
125
+ ^blk\.11\.ffn_up\.weight$=q3_k
126
+ ^blk\.12\.attn_gate\.weight$=q3_k
127
+ ^blk\.12\.attn_qkv\.weight$=q4_k
128
+ ^blk\.12\.ffn_down\.weight$=q3_k
129
+ ^blk\.12\.ffn_gate\.weight$=q3_k
130
+ ^blk\.12\.ffn_up\.weight$=q3_k
131
+ ^blk\.12\.ssm_out\.weight$=q5_k
132
+ ^blk\.13\.attn_gate\.weight$=q3_k
133
+ ^blk\.13\.attn_qkv\.weight$=q4_k
134
+ ^blk\.13\.ffn_down\.weight$=q3_k
135
+ ^blk\.13\.ffn_gate\.weight$=q3_k
136
+ ^blk\.13\.ffn_up\.weight$=q3_k
137
+ ^blk\.13\.ssm_out\.weight$=q5_k
138
+ ^blk\.14\.attn_gate\.weight$=q3_k
139
+ ^blk\.14\.attn_qkv\.weight$=q4_k
140
+ ^blk\.14\.ffn_down\.weight$=q3_k
141
+ ^blk\.14\.ffn_gate\.weight$=q3_k
142
+ ^blk\.14\.ffn_up\.weight$=q3_k
143
+ ^blk\.14\.ssm_out\.weight$=q5_k
144
+ ^blk\.15\.attn_k\.weight$=q5_k
145
+ ^blk\.15\.attn_output\.weight$=q5_k
146
+ ^blk\.15\.attn_q\.weight$=q3_k
147
+ ^blk\.15\.attn_v\.weight$=q6_k
148
+ ^blk\.15\.ffn_down\.weight$=q3_k
149
+ ^blk\.15\.ffn_gate\.weight$=q3_k
150
+ ^blk\.15\.ffn_up\.weight$=q3_k
151
+ ^blk\.16\.attn_gate\.weight$=q3_k
152
+ ^blk\.16\.attn_qkv\.weight$=q4_k
153
+ ^blk\.16\.ffn_down\.weight$=q3_k
154
+ ^blk\.16\.ffn_gate\.weight$=q3_k
155
+ ^blk\.16\.ffn_up\.weight$=q3_k
156
+ ^blk\.16\.ssm_out\.weight$=q5_k
157
+ ^blk\.17\.attn_gate\.weight$=q3_k
158
+ ^blk\.17\.attn_qkv\.weight$=q4_k
159
+ ^blk\.17\.ffn_down\.weight$=q3_k
160
+ ^blk\.17\.ffn_gate\.weight$=q3_k
161
+ ^blk\.17\.ffn_up\.weight$=q3_k
162
+ ^blk\.17\.ssm_out\.weight$=q5_k
163
+ ^blk\.18\.attn_gate\.weight$=q3_k
164
+ ^blk\.18\.attn_qkv\.weight$=q4_k
165
+ ^blk\.18\.ffn_down\.weight$=q3_k
166
+ ^blk\.18\.ffn_gate\.weight$=q3_k
167
+ ^blk\.18\.ffn_up\.weight$=q3_k
168
+ ^blk\.18\.ssm_out\.weight$=q5_k
169
+ ^blk\.19\.attn_k\.weight$=q5_k
170
+ ^blk\.19\.attn_output\.weight$=q5_k
171
+ ^blk\.19\.attn_q\.weight$=q3_k
172
+ ^blk\.19\.attn_v\.weight$=q6_k
173
+ ^blk\.19\.ffn_down\.weight$=q3_k
174
+ ^blk\.19\.ffn_gate\.weight$=q3_k
175
+ ^blk\.19\.ffn_up\.weight$=q3_k
176
+ ^blk\.20\.attn_gate\.weight$=q3_k
177
+ ^blk\.20\.attn_qkv\.weight$=q4_k
178
+ ^blk\.20\.ffn_down\.weight$=q3_k
179
+ ^blk\.20\.ffn_gate\.weight$=q3_k
180
+ ^blk\.20\.ffn_up\.weight$=q3_k
181
+ ^blk\.20\.ssm_out\.weight$=q5_k
182
+ ^blk\.21\.attn_gate\.weight$=q3_k
183
+ ^blk\.21\.attn_qkv\.weight$=q4_k
184
+ ^blk\.21\.ffn_down\.weight$=q3_k
185
+ ^blk\.21\.ffn_gate\.weight$=q3_k
186
+ ^blk\.21\.ffn_up\.weight$=q3_k
187
+ ^blk\.21\.ssm_out\.weight$=q5_k
188
+ ^blk\.22\.attn_gate\.weight$=q3_k
189
+ ^blk\.22\.attn_qkv\.weight$=q4_k
190
+ ^blk\.22\.ffn_down\.weight$=q3_k
191
+ ^blk\.22\.ffn_gate\.weight$=q3_k
192
+ ^blk\.22\.ffn_up\.weight$=q3_k
193
+ ^blk\.22\.ssm_out\.weight$=q5_k
194
+ ^blk\.23\.attn_k\.weight$=q5_k
195
+ ^blk\.23\.attn_output\.weight$=q5_k
196
+ ^blk\.23\.attn_q\.weight$=q3_k
197
+ ^blk\.23\.attn_v\.weight$=q6_k
198
+ ^blk\.23\.ffn_down\.weight$=q3_k
199
+ ^blk\.23\.ffn_gate\.weight$=q3_k
200
+ ^blk\.23\.ffn_up\.weight$=q3_k
201
+ ^blk\.24\.attn_gate\.weight$=q4_k
202
+ ^blk\.24\.attn_qkv\.weight$=q3_k
203
+ ^blk\.24\.ffn_down\.weight$=q4_k
204
+ ^blk\.24\.ffn_gate\.weight$=q3_k
205
+ ^blk\.24\.ffn_up\.weight$=q3_k
206
+ ^blk\.24\.ssm_out\.weight$=q5_k
207
+ ^blk\.25\.attn_gate\.weight$=q4_k
208
+ ^blk\.25\.attn_qkv\.weight$=q3_k
209
+ ^blk\.25\.ffn_down\.weight$=q4_k
210
+ ^blk\.25\.ffn_gate\.weight$=q3_k
211
+ ^blk\.25\.ffn_up\.weight$=q3_k
212
+ ^blk\.25\.ssm_out\.weight$=q5_k
213
+ ^blk\.26\.attn_gate\.weight$=q4_k
214
+ ^blk\.26\.attn_qkv\.weight$=q3_k
215
+ ^blk\.26\.ffn_down\.weight$=q4_k
216
+ ^blk\.26\.ffn_gate\.weight$=q3_k
217
+ ^blk\.26\.ffn_up\.weight$=q3_k
218
+ ^blk\.26\.ssm_out\.weight$=q5_k
219
+ ^blk\.27\.attn_k\.weight$=q6_k
220
+ ^blk\.27\.attn_output\.weight$=q5_k
221
+ ^blk\.27\.attn_q\.weight$=q4_k
222
+ ^blk\.27\.attn_v\.weight$=q6_k
223
+ ^blk\.27\.ffn_down\.weight$=q4_k
224
+ ^blk\.27\.ffn_gate\.weight$=q3_k
225
+ ^blk\.27\.ffn_up\.weight$=q4_k
226
+ ^blk\.28\.attn_gate\.weight$=q4_k
227
+ ^blk\.28\.attn_qkv\.weight$=q3_k
228
+ ^blk\.28\.ffn_down\.weight$=q4_k
229
+ ^blk\.28\.ffn_gate\.weight$=q3_k
230
+ ^blk\.28\.ffn_up\.weight$=q3_k
231
+ ^blk\.28\.ssm_out\.weight$=q5_k
232
+ ^blk\.29\.attn_gate\.weight$=q4_k
233
+ ^blk\.29\.attn_qkv\.weight$=q3_k
234
+ ^blk\.29\.ffn_down\.weight$=q4_k
235
+ ^blk\.29\.ffn_gate\.weight$=q3_k
236
+ ^blk\.29\.ffn_up\.weight$=q3_k
237
+ ^blk\.29\.ssm_out\.weight$=q5_k
238
+ ^blk\.30\.attn_gate\.weight$=q4_k
239
+ ^blk\.30\.attn_qkv\.weight$=q3_k
240
+ ^blk\.30\.ffn_down\.weight$=q4_k
241
+ ^blk\.30\.ffn_gate\.weight$=q3_k
242
+ ^blk\.30\.ffn_up\.weight$=q3_k
243
+ ^blk\.30\.ssm_out\.weight$=q5_k
244
+ ^blk\.31\.attn_k\.weight$=q5_k
245
+ ^blk\.31\.attn_output\.weight$=q6_k
246
+ ^blk\.31\.attn_q\.weight$=q4_k
247
+ ^blk\.31\.attn_v\.weight$=q6_k
248
+ ^blk\.31\.ffn_down\.weight$=q5_k
249
+ ^blk\.31\.ffn_gate\.weight$=q5_k
250
+ ^blk\.31\.ffn_up\.weight$=q5_k
layouts/LensVLM-9B-Q3_K_M.layout.json ADDED
@@ -0,0 +1,201 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "Q3_K_M",
5
+ "base_type": "q3_k",
6
+ "family": "k",
7
+ "recipe": "ident",
8
+ "variant": "shr0.70",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "M",
12
+ "rung_share": 0.7,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.17,
17
+ "type": "q4_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.089,
21
+ "type": "q4_k"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q4_k",
131
+ "token_embd.weight": "q4_k"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 4879810560,
135
+ "map_body+embd": 4439572480,
136
+ "ratio": 0.9098,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.7,
140
+ "body_bpw": 3.814,
141
+ "body_bpw_heuristic": 4.176,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 4468983808,
144
+ "file_bpw_pred": 3.993,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "q3_k": 17,
148
+ "q5_k": 3,
149
+ "q4_k": 3,
150
+ "q6_k": 1
151
+ },
152
+ "attn_k": {
153
+ "q5_k": 6,
154
+ "q6_k": 2
155
+ },
156
+ "attn_output": {
157
+ "q5_k": 7,
158
+ "q6_k": 1
159
+ },
160
+ "attn_q": {
161
+ "q3_k": 7,
162
+ "q4_k": 1
163
+ },
164
+ "attn_qkv": {
165
+ "q3_k": 20,
166
+ "q4_k": 3,
167
+ "q5_k": 1
168
+ },
169
+ "attn_v": {
170
+ "q6_k": 7,
171
+ "q8_0": 1
172
+ },
173
+ "ffn_down": {
174
+ "q3_k": 29,
175
+ "q4_k": 2,
176
+ "q5_k": 1
177
+ },
178
+ "ffn_gate": {
179
+ "q3_k": 31,
180
+ "q5_k": 1
181
+ },
182
+ "ffn_up": {
183
+ "q3_k": 23,
184
+ "q4_k": 8,
185
+ "q5_k": 1
186
+ },
187
+ "ssm_out": {
188
+ "q5_k": 23,
189
+ "q6_k": 1
190
+ }
191
+ },
192
+ "prior": "layout/prior.json",
193
+ "pin_rule": "floor",
194
+ "embd_cap": "heur",
195
+ "generator": "auto_quant_v2.layout",
196
+ "generator_key": "93ecba14028d5e4b",
197
+ "quant_type": "Q3_K_M",
198
+ "quantize_ftype": "Q3_K_M",
199
+ "llama_cpp_version": "b11081",
200
+ "commit": "d7cfc9bb610f62821944853c508648097409f4a7"
201
+ }
layouts/LensVLM-9B-Q3_K_M.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q4_k
50
+ ^token_embd\.weight$=q4_k
51
+ ^blk\.0\.attn_gate\.weight$=q6_k
52
+ ^blk\.0\.attn_qkv\.weight$=q5_k
53
+ ^blk\.0\.ffn_down\.weight$=q4_k
54
+ ^blk\.0\.ffn_gate\.weight$=q3_k
55
+ ^blk\.0\.ffn_up\.weight$=q4_k
56
+ ^blk\.0\.ssm_out\.weight$=q6_k
57
+ ^blk\.1\.attn_gate\.weight$=q5_k
58
+ ^blk\.1\.attn_qkv\.weight$=q4_k
59
+ ^blk\.1\.ffn_down\.weight$=q3_k
60
+ ^blk\.1\.ffn_gate\.weight$=q3_k
61
+ ^blk\.1\.ffn_up\.weight$=q4_k
62
+ ^blk\.1\.ssm_out\.weight$=q5_k
63
+ ^blk\.2\.attn_gate\.weight$=q5_k
64
+ ^blk\.2\.attn_qkv\.weight$=q4_k
65
+ ^blk\.2\.ffn_down\.weight$=q3_k
66
+ ^blk\.2\.ffn_gate\.weight$=q3_k
67
+ ^blk\.2\.ffn_up\.weight$=q4_k
68
+ ^blk\.2\.ssm_out\.weight$=q5_k
69
+ ^blk\.3\.attn_k\.weight$=q5_k
70
+ ^blk\.3\.attn_output\.weight$=q5_k
71
+ ^blk\.3\.attn_q\.weight$=q3_k
72
+ ^blk\.3\.attn_v\.weight$=q6_k
73
+ ^blk\.3\.ffn_down\.weight$=q3_k
74
+ ^blk\.3\.ffn_gate\.weight$=q3_k
75
+ ^blk\.3\.ffn_up\.weight$=q4_k
76
+ ^blk\.4\.attn_gate\.weight$=q5_k
77
+ ^blk\.4\.attn_qkv\.weight$=q4_k
78
+ ^blk\.4\.ffn_down\.weight$=q3_k
79
+ ^blk\.4\.ffn_gate\.weight$=q3_k
80
+ ^blk\.4\.ffn_up\.weight$=q4_k
81
+ ^blk\.4\.ssm_out\.weight$=q5_k
82
+ ^blk\.5\.attn_gate\.weight$=q4_k
83
+ ^blk\.5\.attn_qkv\.weight$=q3_k
84
+ ^blk\.5\.ffn_down\.weight$=q3_k
85
+ ^blk\.5\.ffn_gate\.weight$=q3_k
86
+ ^blk\.5\.ffn_up\.weight$=q4_k
87
+ ^blk\.5\.ssm_out\.weight$=q5_k
88
+ ^blk\.6\.attn_gate\.weight$=q4_k
89
+ ^blk\.6\.attn_qkv\.weight$=q3_k
90
+ ^blk\.6\.ffn_down\.weight$=q3_k
91
+ ^blk\.6\.ffn_gate\.weight$=q3_k
92
+ ^blk\.6\.ffn_up\.weight$=q4_k
93
+ ^blk\.6\.ssm_out\.weight$=q5_k
94
+ ^blk\.7\.attn_k\.weight$=q5_k
95
+ ^blk\.7\.attn_output\.weight$=q5_k
96
+ ^blk\.7\.attn_q\.weight$=q3_k
97
+ ^blk\.7\.attn_v\.weight$=q6_k
98
+ ^blk\.7\.ffn_down\.weight$=q3_k
99
+ ^blk\.7\.ffn_gate\.weight$=q3_k
100
+ ^blk\.7\.ffn_up\.weight$=q4_k
101
+ ^blk\.8\.attn_gate\.weight$=q3_k
102
+ ^blk\.8\.attn_qkv\.weight$=q3_k
103
+ ^blk\.8\.ffn_down\.weight$=q3_k
104
+ ^blk\.8\.ffn_gate\.weight$=q3_k
105
+ ^blk\.8\.ffn_up\.weight$=q3_k
106
+ ^blk\.8\.ssm_out\.weight$=q5_k
107
+ ^blk\.9\.attn_gate\.weight$=q3_k
108
+ ^blk\.9\.attn_qkv\.weight$=q3_k
109
+ ^blk\.9\.ffn_down\.weight$=q3_k
110
+ ^blk\.9\.ffn_gate\.weight$=q3_k
111
+ ^blk\.9\.ffn_up\.weight$=q3_k
112
+ ^blk\.9\.ssm_out\.weight$=q5_k
113
+ ^blk\.10\.attn_gate\.weight$=q3_k
114
+ ^blk\.10\.attn_qkv\.weight$=q3_k
115
+ ^blk\.10\.ffn_down\.weight$=q3_k
116
+ ^blk\.10\.ffn_gate\.weight$=q3_k
117
+ ^blk\.10\.ffn_up\.weight$=q3_k
118
+ ^blk\.10\.ssm_out\.weight$=q5_k
119
+ ^blk\.11\.attn_k\.weight$=q6_k
120
+ ^blk\.11\.attn_output\.weight$=q5_k
121
+ ^blk\.11\.attn_q\.weight$=q3_k
122
+ ^blk\.11\.attn_v\.weight$=q6_k
123
+ ^blk\.11\.ffn_down\.weight$=q3_k
124
+ ^blk\.11\.ffn_gate\.weight$=q3_k
125
+ ^blk\.11\.ffn_up\.weight$=q3_k
126
+ ^blk\.12\.attn_gate\.weight$=q3_k
127
+ ^blk\.12\.attn_qkv\.weight$=q3_k
128
+ ^blk\.12\.ffn_down\.weight$=q3_k
129
+ ^blk\.12\.ffn_gate\.weight$=q3_k
130
+ ^blk\.12\.ffn_up\.weight$=q3_k
131
+ ^blk\.12\.ssm_out\.weight$=q5_k
132
+ ^blk\.13\.attn_gate\.weight$=q3_k
133
+ ^blk\.13\.attn_qkv\.weight$=q3_k
134
+ ^blk\.13\.ffn_down\.weight$=q3_k
135
+ ^blk\.13\.ffn_gate\.weight$=q3_k
136
+ ^blk\.13\.ffn_up\.weight$=q3_k
137
+ ^blk\.13\.ssm_out\.weight$=q5_k
138
+ ^blk\.14\.attn_gate\.weight$=q3_k
139
+ ^blk\.14\.attn_qkv\.weight$=q3_k
140
+ ^blk\.14\.ffn_down\.weight$=q3_k
141
+ ^blk\.14\.ffn_gate\.weight$=q3_k
142
+ ^blk\.14\.ffn_up\.weight$=q3_k
143
+ ^blk\.14\.ssm_out\.weight$=q5_k
144
+ ^blk\.15\.attn_k\.weight$=q5_k
145
+ ^blk\.15\.attn_output\.weight$=q5_k
146
+ ^blk\.15\.attn_q\.weight$=q3_k
147
+ ^blk\.15\.attn_v\.weight$=q6_k
148
+ ^blk\.15\.ffn_down\.weight$=q3_k
149
+ ^blk\.15\.ffn_gate\.weight$=q3_k
150
+ ^blk\.15\.ffn_up\.weight$=q3_k
151
+ ^blk\.16\.attn_gate\.weight$=q3_k
152
+ ^blk\.16\.attn_qkv\.weight$=q3_k
153
+ ^blk\.16\.ffn_down\.weight$=q3_k
154
+ ^blk\.16\.ffn_gate\.weight$=q3_k
155
+ ^blk\.16\.ffn_up\.weight$=q3_k
156
+ ^blk\.16\.ssm_out\.weight$=q5_k
157
+ ^blk\.17\.attn_gate\.weight$=q3_k
158
+ ^blk\.17\.attn_qkv\.weight$=q3_k
159
+ ^blk\.17\.ffn_down\.weight$=q3_k
160
+ ^blk\.17\.ffn_gate\.weight$=q3_k
161
+ ^blk\.17\.ffn_up\.weight$=q3_k
162
+ ^blk\.17\.ssm_out\.weight$=q5_k
163
+ ^blk\.18\.attn_gate\.weight$=q3_k
164
+ ^blk\.18\.attn_qkv\.weight$=q3_k
165
+ ^blk\.18\.ffn_down\.weight$=q3_k
166
+ ^blk\.18\.ffn_gate\.weight$=q3_k
167
+ ^blk\.18\.ffn_up\.weight$=q3_k
168
+ ^blk\.18\.ssm_out\.weight$=q5_k
169
+ ^blk\.19\.attn_k\.weight$=q5_k
170
+ ^blk\.19\.attn_output\.weight$=q5_k
171
+ ^blk\.19\.attn_q\.weight$=q3_k
172
+ ^blk\.19\.attn_v\.weight$=q6_k
173
+ ^blk\.19\.ffn_down\.weight$=q3_k
174
+ ^blk\.19\.ffn_gate\.weight$=q3_k
175
+ ^blk\.19\.ffn_up\.weight$=q3_k
176
+ ^blk\.20\.attn_gate\.weight$=q3_k
177
+ ^blk\.20\.attn_qkv\.weight$=q3_k
178
+ ^blk\.20\.ffn_down\.weight$=q3_k
179
+ ^blk\.20\.ffn_gate\.weight$=q3_k
180
+ ^blk\.20\.ffn_up\.weight$=q3_k
181
+ ^blk\.20\.ssm_out\.weight$=q5_k
182
+ ^blk\.21\.attn_gate\.weight$=q3_k
183
+ ^blk\.21\.attn_qkv\.weight$=q3_k
184
+ ^blk\.21\.ffn_down\.weight$=q3_k
185
+ ^blk\.21\.ffn_gate\.weight$=q3_k
186
+ ^blk\.21\.ffn_up\.weight$=q3_k
187
+ ^blk\.21\.ssm_out\.weight$=q5_k
188
+ ^blk\.22\.attn_gate\.weight$=q3_k
189
+ ^blk\.22\.attn_qkv\.weight$=q3_k
190
+ ^blk\.22\.ffn_down\.weight$=q3_k
191
+ ^blk\.22\.ffn_gate\.weight$=q3_k
192
+ ^blk\.22\.ffn_up\.weight$=q3_k
193
+ ^blk\.22\.ssm_out\.weight$=q5_k
194
+ ^blk\.23\.attn_k\.weight$=q5_k
195
+ ^blk\.23\.attn_output\.weight$=q5_k
196
+ ^blk\.23\.attn_q\.weight$=q3_k
197
+ ^blk\.23\.attn_v\.weight$=q6_k
198
+ ^blk\.23\.ffn_down\.weight$=q3_k
199
+ ^blk\.23\.ffn_gate\.weight$=q3_k
200
+ ^blk\.23\.ffn_up\.weight$=q3_k
201
+ ^blk\.24\.attn_gate\.weight$=q4_k
202
+ ^blk\.24\.attn_qkv\.weight$=q3_k
203
+ ^blk\.24\.ffn_down\.weight$=q3_k
204
+ ^blk\.24\.ffn_gate\.weight$=q3_k
205
+ ^blk\.24\.ffn_up\.weight$=q3_k
206
+ ^blk\.24\.ssm_out\.weight$=q5_k
207
+ ^blk\.25\.attn_gate\.weight$=q3_k
208
+ ^blk\.25\.attn_qkv\.weight$=q3_k
209
+ ^blk\.25\.ffn_down\.weight$=q3_k
210
+ ^blk\.25\.ffn_gate\.weight$=q3_k
211
+ ^blk\.25\.ffn_up\.weight$=q3_k
212
+ ^blk\.25\.ssm_out\.weight$=q5_k
213
+ ^blk\.26\.attn_gate\.weight$=q3_k
214
+ ^blk\.26\.attn_qkv\.weight$=q3_k
215
+ ^blk\.26\.ffn_down\.weight$=q3_k
216
+ ^blk\.26\.ffn_gate\.weight$=q3_k
217
+ ^blk\.26\.ffn_up\.weight$=q3_k
218
+ ^blk\.26\.ssm_out\.weight$=q5_k
219
+ ^blk\.27\.attn_k\.weight$=q6_k
220
+ ^blk\.27\.attn_output\.weight$=q5_k
221
+ ^blk\.27\.attn_q\.weight$=q3_k
222
+ ^blk\.27\.attn_v\.weight$=q8_0
223
+ ^blk\.27\.ffn_down\.weight$=q4_k
224
+ ^blk\.27\.ffn_gate\.weight$=q3_k
225
+ ^blk\.27\.ffn_up\.weight$=q3_k
226
+ ^blk\.28\.attn_gate\.weight$=q3_k
227
+ ^blk\.28\.attn_qkv\.weight$=q3_k
228
+ ^blk\.28\.ffn_down\.weight$=q3_k
229
+ ^blk\.28\.ffn_gate\.weight$=q3_k
230
+ ^blk\.28\.ffn_up\.weight$=q3_k
231
+ ^blk\.28\.ssm_out\.weight$=q5_k
232
+ ^blk\.29\.attn_gate\.weight$=q3_k
233
+ ^blk\.29\.attn_qkv\.weight$=q3_k
234
+ ^blk\.29\.ffn_down\.weight$=q3_k
235
+ ^blk\.29\.ffn_gate\.weight$=q3_k
236
+ ^blk\.29\.ffn_up\.weight$=q3_k
237
+ ^blk\.29\.ssm_out\.weight$=q5_k
238
+ ^blk\.30\.attn_gate\.weight$=q3_k
239
+ ^blk\.30\.attn_qkv\.weight$=q3_k
240
+ ^blk\.30\.ffn_down\.weight$=q3_k
241
+ ^blk\.30\.ffn_gate\.weight$=q3_k
242
+ ^blk\.30\.ffn_up\.weight$=q3_k
243
+ ^blk\.30\.ssm_out\.weight$=q5_k
244
+ ^blk\.31\.attn_k\.weight$=q5_k
245
+ ^blk\.31\.attn_output\.weight$=q6_k
246
+ ^blk\.31\.attn_q\.weight$=q4_k
247
+ ^blk\.31\.attn_v\.weight$=q6_k
248
+ ^blk\.31\.ffn_down\.weight$=q5_k
249
+ ^blk\.31\.ffn_gate\.weight$=q5_k
250
+ ^blk\.31\.ffn_up\.weight$=q5_k
layouts/LensVLM-9B-Q3_K_S.layout.json ADDED
@@ -0,0 +1,193 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "Q3_K_S",
5
+ "base_type": "q3_k",
6
+ "family": "k",
7
+ "recipe": "ident",
8
+ "variant": "shr0.90",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "S",
12
+ "rung_share": 0.9,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.1791,
17
+ "type": "q4_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.0938,
21
+ "type": "q4_k"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q4_k",
131
+ "token_embd.weight": "q4_k"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 4628938752,
135
+ "map_body+embd": 4219928576,
136
+ "ratio": 0.9116,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.9,
140
+ "body_bpw": 3.56,
141
+ "body_bpw_heuristic": 3.886,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 4249339904,
144
+ "file_bpw_pred": 3.797,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "q3_k": 23,
148
+ "q5_k": 1
149
+ },
150
+ "attn_k": {
151
+ "q5_k": 8
152
+ },
153
+ "attn_output": {
154
+ "q5_k": 8
155
+ },
156
+ "attn_q": {
157
+ "q3_k": 8
158
+ },
159
+ "attn_qkv": {
160
+ "q3_k": 23,
161
+ "q5_k": 1
162
+ },
163
+ "attn_v": {
164
+ "q5_k": 6,
165
+ "q6_k": 2
166
+ },
167
+ "ffn_down": {
168
+ "q3_k": 31,
169
+ "q5_k": 1
170
+ },
171
+ "ffn_gate": {
172
+ "q3_k": 32
173
+ },
174
+ "ffn_up": {
175
+ "q3_k": 31,
176
+ "q4_k": 1
177
+ },
178
+ "ssm_out": {
179
+ "q3_k": 17,
180
+ "q4_k": 6,
181
+ "q6_k": 1
182
+ }
183
+ },
184
+ "prior": "layout/prior.json",
185
+ "pin_rule": "floor",
186
+ "embd_cap": "heur",
187
+ "generator": "auto_quant_v2.layout",
188
+ "generator_key": "93ecba14028d5e4b",
189
+ "quant_type": "Q3_K_S",
190
+ "quantize_ftype": "Q3_K_S",
191
+ "llama_cpp_version": "b11081",
192
+ "commit": "d7cfc9bb610f62821944853c508648097409f4a7"
193
+ }
layouts/LensVLM-9B-Q3_K_S.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q4_k
50
+ ^token_embd\.weight$=q4_k
51
+ ^blk\.0\.attn_gate\.weight$=q5_k
52
+ ^blk\.0\.attn_qkv\.weight$=q5_k
53
+ ^blk\.0\.ffn_down\.weight$=q3_k
54
+ ^blk\.0\.ffn_gate\.weight$=q3_k
55
+ ^blk\.0\.ffn_up\.weight$=q3_k
56
+ ^blk\.0\.ssm_out\.weight$=q6_k
57
+ ^blk\.1\.attn_gate\.weight$=q3_k
58
+ ^blk\.1\.attn_qkv\.weight$=q3_k
59
+ ^blk\.1\.ffn_down\.weight$=q3_k
60
+ ^blk\.1\.ffn_gate\.weight$=q3_k
61
+ ^blk\.1\.ffn_up\.weight$=q3_k
62
+ ^blk\.1\.ssm_out\.weight$=q4_k
63
+ ^blk\.2\.attn_gate\.weight$=q3_k
64
+ ^blk\.2\.attn_qkv\.weight$=q3_k
65
+ ^blk\.2\.ffn_down\.weight$=q3_k
66
+ ^blk\.2\.ffn_gate\.weight$=q3_k
67
+ ^blk\.2\.ffn_up\.weight$=q3_k
68
+ ^blk\.2\.ssm_out\.weight$=q4_k
69
+ ^blk\.3\.attn_k\.weight$=q5_k
70
+ ^blk\.3\.attn_output\.weight$=q5_k
71
+ ^blk\.3\.attn_q\.weight$=q3_k
72
+ ^blk\.3\.attn_v\.weight$=q5_k
73
+ ^blk\.3\.ffn_down\.weight$=q3_k
74
+ ^blk\.3\.ffn_gate\.weight$=q3_k
75
+ ^blk\.3\.ffn_up\.weight$=q3_k
76
+ ^blk\.4\.attn_gate\.weight$=q3_k
77
+ ^blk\.4\.attn_qkv\.weight$=q3_k
78
+ ^blk\.4\.ffn_down\.weight$=q3_k
79
+ ^blk\.4\.ffn_gate\.weight$=q3_k
80
+ ^blk\.4\.ffn_up\.weight$=q3_k
81
+ ^blk\.4\.ssm_out\.weight$=q4_k
82
+ ^blk\.5\.attn_gate\.weight$=q3_k
83
+ ^blk\.5\.attn_qkv\.weight$=q3_k
84
+ ^blk\.5\.ffn_down\.weight$=q3_k
85
+ ^blk\.5\.ffn_gate\.weight$=q3_k
86
+ ^blk\.5\.ffn_up\.weight$=q3_k
87
+ ^blk\.5\.ssm_out\.weight$=q4_k
88
+ ^blk\.6\.attn_gate\.weight$=q3_k
89
+ ^blk\.6\.attn_qkv\.weight$=q3_k
90
+ ^blk\.6\.ffn_down\.weight$=q3_k
91
+ ^blk\.6\.ffn_gate\.weight$=q3_k
92
+ ^blk\.6\.ffn_up\.weight$=q3_k
93
+ ^blk\.6\.ssm_out\.weight$=q4_k
94
+ ^blk\.7\.attn_k\.weight$=q5_k
95
+ ^blk\.7\.attn_output\.weight$=q5_k
96
+ ^blk\.7\.attn_q\.weight$=q3_k
97
+ ^blk\.7\.attn_v\.weight$=q5_k
98
+ ^blk\.7\.ffn_down\.weight$=q3_k
99
+ ^blk\.7\.ffn_gate\.weight$=q3_k
100
+ ^blk\.7\.ffn_up\.weight$=q3_k
101
+ ^blk\.8\.attn_gate\.weight$=q3_k
102
+ ^blk\.8\.attn_qkv\.weight$=q3_k
103
+ ^blk\.8\.ffn_down\.weight$=q3_k
104
+ ^blk\.8\.ffn_gate\.weight$=q3_k
105
+ ^blk\.8\.ffn_up\.weight$=q3_k
106
+ ^blk\.8\.ssm_out\.weight$=q3_k
107
+ ^blk\.9\.attn_gate\.weight$=q3_k
108
+ ^blk\.9\.attn_qkv\.weight$=q3_k
109
+ ^blk\.9\.ffn_down\.weight$=q3_k
110
+ ^blk\.9\.ffn_gate\.weight$=q3_k
111
+ ^blk\.9\.ffn_up\.weight$=q3_k
112
+ ^blk\.9\.ssm_out\.weight$=q3_k
113
+ ^blk\.10\.attn_gate\.weight$=q3_k
114
+ ^blk\.10\.attn_qkv\.weight$=q3_k
115
+ ^blk\.10\.ffn_down\.weight$=q3_k
116
+ ^blk\.10\.ffn_gate\.weight$=q3_k
117
+ ^blk\.10\.ffn_up\.weight$=q3_k
118
+ ^blk\.10\.ssm_out\.weight$=q3_k
119
+ ^blk\.11\.attn_k\.weight$=q5_k
120
+ ^blk\.11\.attn_output\.weight$=q5_k
121
+ ^blk\.11\.attn_q\.weight$=q3_k
122
+ ^blk\.11\.attn_v\.weight$=q5_k
123
+ ^blk\.11\.ffn_down\.weight$=q3_k
124
+ ^blk\.11\.ffn_gate\.weight$=q3_k
125
+ ^blk\.11\.ffn_up\.weight$=q3_k
126
+ ^blk\.12\.attn_gate\.weight$=q3_k
127
+ ^blk\.12\.attn_qkv\.weight$=q3_k
128
+ ^blk\.12\.ffn_down\.weight$=q3_k
129
+ ^blk\.12\.ffn_gate\.weight$=q3_k
130
+ ^blk\.12\.ffn_up\.weight$=q3_k
131
+ ^blk\.12\.ssm_out\.weight$=q3_k
132
+ ^blk\.13\.attn_gate\.weight$=q3_k
133
+ ^blk\.13\.attn_qkv\.weight$=q3_k
134
+ ^blk\.13\.ffn_down\.weight$=q3_k
135
+ ^blk\.13\.ffn_gate\.weight$=q3_k
136
+ ^blk\.13\.ffn_up\.weight$=q3_k
137
+ ^blk\.13\.ssm_out\.weight$=q3_k
138
+ ^blk\.14\.attn_gate\.weight$=q3_k
139
+ ^blk\.14\.attn_qkv\.weight$=q3_k
140
+ ^blk\.14\.ffn_down\.weight$=q3_k
141
+ ^blk\.14\.ffn_gate\.weight$=q3_k
142
+ ^blk\.14\.ffn_up\.weight$=q3_k
143
+ ^blk\.14\.ssm_out\.weight$=q3_k
144
+ ^blk\.15\.attn_k\.weight$=q5_k
145
+ ^blk\.15\.attn_output\.weight$=q5_k
146
+ ^blk\.15\.attn_q\.weight$=q3_k
147
+ ^blk\.15\.attn_v\.weight$=q5_k
148
+ ^blk\.15\.ffn_down\.weight$=q3_k
149
+ ^blk\.15\.ffn_gate\.weight$=q3_k
150
+ ^blk\.15\.ffn_up\.weight$=q3_k
151
+ ^blk\.16\.attn_gate\.weight$=q3_k
152
+ ^blk\.16\.attn_qkv\.weight$=q3_k
153
+ ^blk\.16\.ffn_down\.weight$=q3_k
154
+ ^blk\.16\.ffn_gate\.weight$=q3_k
155
+ ^blk\.16\.ffn_up\.weight$=q3_k
156
+ ^blk\.16\.ssm_out\.weight$=q3_k
157
+ ^blk\.17\.attn_gate\.weight$=q3_k
158
+ ^blk\.17\.attn_qkv\.weight$=q3_k
159
+ ^blk\.17\.ffn_down\.weight$=q3_k
160
+ ^blk\.17\.ffn_gate\.weight$=q3_k
161
+ ^blk\.17\.ffn_up\.weight$=q3_k
162
+ ^blk\.17\.ssm_out\.weight$=q3_k
163
+ ^blk\.18\.attn_gate\.weight$=q3_k
164
+ ^blk\.18\.attn_qkv\.weight$=q3_k
165
+ ^blk\.18\.ffn_down\.weight$=q3_k
166
+ ^blk\.18\.ffn_gate\.weight$=q3_k
167
+ ^blk\.18\.ffn_up\.weight$=q3_k
168
+ ^blk\.18\.ssm_out\.weight$=q3_k
169
+ ^blk\.19\.attn_k\.weight$=q5_k
170
+ ^blk\.19\.attn_output\.weight$=q5_k
171
+ ^blk\.19\.attn_q\.weight$=q3_k
172
+ ^blk\.19\.attn_v\.weight$=q5_k
173
+ ^blk\.19\.ffn_down\.weight$=q3_k
174
+ ^blk\.19\.ffn_gate\.weight$=q3_k
175
+ ^blk\.19\.ffn_up\.weight$=q3_k
176
+ ^blk\.20\.attn_gate\.weight$=q3_k
177
+ ^blk\.20\.attn_qkv\.weight$=q3_k
178
+ ^blk\.20\.ffn_down\.weight$=q3_k
179
+ ^blk\.20\.ffn_gate\.weight$=q3_k
180
+ ^blk\.20\.ffn_up\.weight$=q3_k
181
+ ^blk\.20\.ssm_out\.weight$=q3_k
182
+ ^blk\.21\.attn_gate\.weight$=q3_k
183
+ ^blk\.21\.attn_qkv\.weight$=q3_k
184
+ ^blk\.21\.ffn_down\.weight$=q3_k
185
+ ^blk\.21\.ffn_gate\.weight$=q3_k
186
+ ^blk\.21\.ffn_up\.weight$=q3_k
187
+ ^blk\.21\.ssm_out\.weight$=q3_k
188
+ ^blk\.22\.attn_gate\.weight$=q3_k
189
+ ^blk\.22\.attn_qkv\.weight$=q3_k
190
+ ^blk\.22\.ffn_down\.weight$=q3_k
191
+ ^blk\.22\.ffn_gate\.weight$=q3_k
192
+ ^blk\.22\.ffn_up\.weight$=q3_k
193
+ ^blk\.22\.ssm_out\.weight$=q3_k
194
+ ^blk\.23\.attn_k\.weight$=q5_k
195
+ ^blk\.23\.attn_output\.weight$=q5_k
196
+ ^blk\.23\.attn_q\.weight$=q3_k
197
+ ^blk\.23\.attn_v\.weight$=q5_k
198
+ ^blk\.23\.ffn_down\.weight$=q3_k
199
+ ^blk\.23\.ffn_gate\.weight$=q3_k
200
+ ^blk\.23\.ffn_up\.weight$=q3_k
201
+ ^blk\.24\.attn_gate\.weight$=q3_k
202
+ ^blk\.24\.attn_qkv\.weight$=q3_k
203
+ ^blk\.24\.ffn_down\.weight$=q3_k
204
+ ^blk\.24\.ffn_gate\.weight$=q3_k
205
+ ^blk\.24\.ffn_up\.weight$=q3_k
206
+ ^blk\.24\.ssm_out\.weight$=q4_k
207
+ ^blk\.25\.attn_gate\.weight$=q3_k
208
+ ^blk\.25\.attn_qkv\.weight$=q3_k
209
+ ^blk\.25\.ffn_down\.weight$=q3_k
210
+ ^blk\.25\.ffn_gate\.weight$=q3_k
211
+ ^blk\.25\.ffn_up\.weight$=q3_k
212
+ ^blk\.25\.ssm_out\.weight$=q3_k
213
+ ^blk\.26\.attn_gate\.weight$=q3_k
214
+ ^blk\.26\.attn_qkv\.weight$=q3_k
215
+ ^blk\.26\.ffn_down\.weight$=q3_k
216
+ ^blk\.26\.ffn_gate\.weight$=q3_k
217
+ ^blk\.26\.ffn_up\.weight$=q3_k
218
+ ^blk\.26\.ssm_out\.weight$=q3_k
219
+ ^blk\.27\.attn_k\.weight$=q5_k
220
+ ^blk\.27\.attn_output\.weight$=q5_k
221
+ ^blk\.27\.attn_q\.weight$=q3_k
222
+ ^blk\.27\.attn_v\.weight$=q6_k
223
+ ^blk\.27\.ffn_down\.weight$=q3_k
224
+ ^blk\.27\.ffn_gate\.weight$=q3_k
225
+ ^blk\.27\.ffn_up\.weight$=q3_k
226
+ ^blk\.28\.attn_gate\.weight$=q3_k
227
+ ^blk\.28\.attn_qkv\.weight$=q3_k
228
+ ^blk\.28\.ffn_down\.weight$=q3_k
229
+ ^blk\.28\.ffn_gate\.weight$=q3_k
230
+ ^blk\.28\.ffn_up\.weight$=q3_k
231
+ ^blk\.28\.ssm_out\.weight$=q3_k
232
+ ^blk\.29\.attn_gate\.weight$=q3_k
233
+ ^blk\.29\.attn_qkv\.weight$=q3_k
234
+ ^blk\.29\.ffn_down\.weight$=q3_k
235
+ ^blk\.29\.ffn_gate\.weight$=q3_k
236
+ ^blk\.29\.ffn_up\.weight$=q3_k
237
+ ^blk\.29\.ssm_out\.weight$=q3_k
238
+ ^blk\.30\.attn_gate\.weight$=q3_k
239
+ ^blk\.30\.attn_qkv\.weight$=q3_k
240
+ ^blk\.30\.ffn_down\.weight$=q3_k
241
+ ^blk\.30\.ffn_gate\.weight$=q3_k
242
+ ^blk\.30\.ffn_up\.weight$=q3_k
243
+ ^blk\.30\.ssm_out\.weight$=q3_k
244
+ ^blk\.31\.attn_k\.weight$=q5_k
245
+ ^blk\.31\.attn_output\.weight$=q5_k
246
+ ^blk\.31\.attn_q\.weight$=q3_k
247
+ ^blk\.31\.attn_v\.weight$=q6_k
248
+ ^blk\.31\.ffn_down\.weight$=q5_k
249
+ ^blk\.31\.ffn_gate\.weight$=q3_k
250
+ ^blk\.31\.ffn_up\.weight$=q4_k
layouts/LensVLM-9B-Q4_K_L.layout.json ADDED
@@ -0,0 +1,198 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "Q4_K_L",
5
+ "base_type": "q4_k",
6
+ "family": "k",
7
+ "recipe": "ident",
8
+ "variant": "shr0.50",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "L",
12
+ "rung_share": 0.5,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.1414,
17
+ "type": "q6_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.097,
21
+ "type": "q4_k"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q6_k",
131
+ "token_embd.weight": "q4_k"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 5870403584,
135
+ "map_body+embd": 6161645568,
136
+ "ratio": 1.0496,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.5,
140
+ "body_bpw": 5.503,
141
+ "body_bpw_heuristic": 5.166,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 6191056896,
144
+ "file_bpw_pred": 5.532,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "q6_k": 11,
148
+ "q4_k": 11,
149
+ "q8_0": 1,
150
+ "q5_k": 1
151
+ },
152
+ "attn_k": {
153
+ "q8_0": 8
154
+ },
155
+ "attn_output": {
156
+ "q8_0": 8
157
+ },
158
+ "attn_q": {
159
+ "q4_k": 6,
160
+ "q6_k": 2
161
+ },
162
+ "attn_qkv": {
163
+ "q4_k": 11,
164
+ "q6_k": 8,
165
+ "q5_k": 4,
166
+ "q8_0": 1
167
+ },
168
+ "attn_v": {
169
+ "q8_0": 8
170
+ },
171
+ "ffn_down": {
172
+ "q4_k": 16,
173
+ "q6_k": 15,
174
+ "q8_0": 1
175
+ },
176
+ "ffn_gate": {
177
+ "q4_k": 31,
178
+ "q8_0": 1
179
+ },
180
+ "ffn_up": {
181
+ "q4_k": 22,
182
+ "q6_k": 9,
183
+ "q8_0": 1
184
+ },
185
+ "ssm_out": {
186
+ "q8_0": 24
187
+ }
188
+ },
189
+ "prior": "layout/prior.json",
190
+ "pin_rule": "floor",
191
+ "embd_cap": "heur",
192
+ "generator": "auto_quant_v2.layout",
193
+ "generator_key": "93ecba14028d5e4b",
194
+ "quant_type": "Q4_K_L",
195
+ "quantize_ftype": "Q4_K_M",
196
+ "llama_cpp_version": "b11081",
197
+ "commit": "e6c963412b6904665bc6346cbd7a27c10a0a6abc"
198
+ }
layouts/LensVLM-9B-Q4_K_L.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q6_k
50
+ ^token_embd\.weight$=q4_k
51
+ ^blk\.0\.attn_gate\.weight$=q8_0
52
+ ^blk\.0\.attn_qkv\.weight$=q8_0
53
+ ^blk\.0\.ffn_down\.weight$=q6_k
54
+ ^blk\.0\.ffn_gate\.weight$=q4_k
55
+ ^blk\.0\.ffn_up\.weight$=q6_k
56
+ ^blk\.0\.ssm_out\.weight$=q8_0
57
+ ^blk\.1\.attn_gate\.weight$=q6_k
58
+ ^blk\.1\.attn_qkv\.weight$=q6_k
59
+ ^blk\.1\.ffn_down\.weight$=q6_k
60
+ ^blk\.1\.ffn_gate\.weight$=q4_k
61
+ ^blk\.1\.ffn_up\.weight$=q6_k
62
+ ^blk\.1\.ssm_out\.weight$=q8_0
63
+ ^blk\.2\.attn_gate\.weight$=q6_k
64
+ ^blk\.2\.attn_qkv\.weight$=q6_k
65
+ ^blk\.2\.ffn_down\.weight$=q6_k
66
+ ^blk\.2\.ffn_gate\.weight$=q4_k
67
+ ^blk\.2\.ffn_up\.weight$=q6_k
68
+ ^blk\.2\.ssm_out\.weight$=q8_0
69
+ ^blk\.3\.attn_k\.weight$=q8_0
70
+ ^blk\.3\.attn_output\.weight$=q8_0
71
+ ^blk\.3\.attn_q\.weight$=q4_k
72
+ ^blk\.3\.attn_v\.weight$=q8_0
73
+ ^blk\.3\.ffn_down\.weight$=q6_k
74
+ ^blk\.3\.ffn_gate\.weight$=q4_k
75
+ ^blk\.3\.ffn_up\.weight$=q6_k
76
+ ^blk\.4\.attn_gate\.weight$=q6_k
77
+ ^blk\.4\.attn_qkv\.weight$=q6_k
78
+ ^blk\.4\.ffn_down\.weight$=q6_k
79
+ ^blk\.4\.ffn_gate\.weight$=q4_k
80
+ ^blk\.4\.ffn_up\.weight$=q6_k
81
+ ^blk\.4\.ssm_out\.weight$=q8_0
82
+ ^blk\.5\.attn_gate\.weight$=q6_k
83
+ ^blk\.5\.attn_qkv\.weight$=q6_k
84
+ ^blk\.5\.ffn_down\.weight$=q6_k
85
+ ^blk\.5\.ffn_gate\.weight$=q4_k
86
+ ^blk\.5\.ffn_up\.weight$=q6_k
87
+ ^blk\.5\.ssm_out\.weight$=q8_0
88
+ ^blk\.6\.attn_gate\.weight$=q6_k
89
+ ^blk\.6\.attn_qkv\.weight$=q6_k
90
+ ^blk\.6\.ffn_down\.weight$=q6_k
91
+ ^blk\.6\.ffn_gate\.weight$=q4_k
92
+ ^blk\.6\.ffn_up\.weight$=q6_k
93
+ ^blk\.6\.ssm_out\.weight$=q8_0
94
+ ^blk\.7\.attn_k\.weight$=q8_0
95
+ ^blk\.7\.attn_output\.weight$=q8_0
96
+ ^blk\.7\.attn_q\.weight$=q4_k
97
+ ^blk\.7\.attn_v\.weight$=q8_0
98
+ ^blk\.7\.ffn_down\.weight$=q6_k
99
+ ^blk\.7\.ffn_gate\.weight$=q4_k
100
+ ^blk\.7\.ffn_up\.weight$=q6_k
101
+ ^blk\.8\.attn_gate\.weight$=q5_k
102
+ ^blk\.8\.attn_qkv\.weight$=q6_k
103
+ ^blk\.8\.ffn_down\.weight$=q4_k
104
+ ^blk\.8\.ffn_gate\.weight$=q4_k
105
+ ^blk\.8\.ffn_up\.weight$=q4_k
106
+ ^blk\.8\.ssm_out\.weight$=q8_0
107
+ ^blk\.9\.attn_gate\.weight$=q4_k
108
+ ^blk\.9\.attn_qkv\.weight$=q6_k
109
+ ^blk\.9\.ffn_down\.weight$=q4_k
110
+ ^blk\.9\.ffn_gate\.weight$=q4_k
111
+ ^blk\.9\.ffn_up\.weight$=q4_k
112
+ ^blk\.9\.ssm_out\.weight$=q8_0
113
+ ^blk\.10\.attn_gate\.weight$=q4_k
114
+ ^blk\.10\.attn_qkv\.weight$=q6_k
115
+ ^blk\.10\.ffn_down\.weight$=q4_k
116
+ ^blk\.10\.ffn_gate\.weight$=q4_k
117
+ ^blk\.10\.ffn_up\.weight$=q4_k
118
+ ^blk\.10\.ssm_out\.weight$=q8_0
119
+ ^blk\.11\.attn_k\.weight$=q8_0
120
+ ^blk\.11\.attn_output\.weight$=q8_0
121
+ ^blk\.11\.attn_q\.weight$=q4_k
122
+ ^blk\.11\.attn_v\.weight$=q8_0
123
+ ^blk\.11\.ffn_down\.weight$=q4_k
124
+ ^blk\.11\.ffn_gate\.weight$=q4_k
125
+ ^blk\.11\.ffn_up\.weight$=q4_k
126
+ ^blk\.12\.attn_gate\.weight$=q4_k
127
+ ^blk\.12\.attn_qkv\.weight$=q5_k
128
+ ^blk\.12\.ffn_down\.weight$=q4_k
129
+ ^blk\.12\.ffn_gate\.weight$=q4_k
130
+ ^blk\.12\.ffn_up\.weight$=q4_k
131
+ ^blk\.12\.ssm_out\.weight$=q8_0
132
+ ^blk\.13\.attn_gate\.weight$=q4_k
133
+ ^blk\.13\.attn_qkv\.weight$=q5_k
134
+ ^blk\.13\.ffn_down\.weight$=q4_k
135
+ ^blk\.13\.ffn_gate\.weight$=q4_k
136
+ ^blk\.13\.ffn_up\.weight$=q4_k
137
+ ^blk\.13\.ssm_out\.weight$=q8_0
138
+ ^blk\.14\.attn_gate\.weight$=q4_k
139
+ ^blk\.14\.attn_qkv\.weight$=q5_k
140
+ ^blk\.14\.ffn_down\.weight$=q4_k
141
+ ^blk\.14\.ffn_gate\.weight$=q4_k
142
+ ^blk\.14\.ffn_up\.weight$=q4_k
143
+ ^blk\.14\.ssm_out\.weight$=q8_0
144
+ ^blk\.15\.attn_k\.weight$=q8_0
145
+ ^blk\.15\.attn_output\.weight$=q8_0
146
+ ^blk\.15\.attn_q\.weight$=q4_k
147
+ ^blk\.15\.attn_v\.weight$=q8_0
148
+ ^blk\.15\.ffn_down\.weight$=q4_k
149
+ ^blk\.15\.ffn_gate\.weight$=q4_k
150
+ ^blk\.15\.ffn_up\.weight$=q4_k
151
+ ^blk\.16\.attn_gate\.weight$=q4_k
152
+ ^blk\.16\.attn_qkv\.weight$=q5_k
153
+ ^blk\.16\.ffn_down\.weight$=q4_k
154
+ ^blk\.16\.ffn_gate\.weight$=q4_k
155
+ ^blk\.16\.ffn_up\.weight$=q4_k
156
+ ^blk\.16\.ssm_out\.weight$=q8_0
157
+ ^blk\.17\.attn_gate\.weight$=q4_k
158
+ ^blk\.17\.attn_qkv\.weight$=q4_k
159
+ ^blk\.17\.ffn_down\.weight$=q4_k
160
+ ^blk\.17\.ffn_gate\.weight$=q4_k
161
+ ^blk\.17\.ffn_up\.weight$=q4_k
162
+ ^blk\.17\.ssm_out\.weight$=q8_0
163
+ ^blk\.18\.attn_gate\.weight$=q4_k
164
+ ^blk\.18\.attn_qkv\.weight$=q4_k
165
+ ^blk\.18\.ffn_down\.weight$=q4_k
166
+ ^blk\.18\.ffn_gate\.weight$=q4_k
167
+ ^blk\.18\.ffn_up\.weight$=q4_k
168
+ ^blk\.18\.ssm_out\.weight$=q8_0
169
+ ^blk\.19\.attn_k\.weight$=q8_0
170
+ ^blk\.19\.attn_output\.weight$=q8_0
171
+ ^blk\.19\.attn_q\.weight$=q4_k
172
+ ^blk\.19\.attn_v\.weight$=q8_0
173
+ ^blk\.19\.ffn_down\.weight$=q4_k
174
+ ^blk\.19\.ffn_gate\.weight$=q4_k
175
+ ^blk\.19\.ffn_up\.weight$=q4_k
176
+ ^blk\.20\.attn_gate\.weight$=q4_k
177
+ ^blk\.20\.attn_qkv\.weight$=q4_k
178
+ ^blk\.20\.ffn_down\.weight$=q4_k
179
+ ^blk\.20\.ffn_gate\.weight$=q4_k
180
+ ^blk\.20\.ffn_up\.weight$=q4_k
181
+ ^blk\.20\.ssm_out\.weight$=q8_0
182
+ ^blk\.21\.attn_gate\.weight$=q4_k
183
+ ^blk\.21\.attn_qkv\.weight$=q4_k
184
+ ^blk\.21\.ffn_down\.weight$=q4_k
185
+ ^blk\.21\.ffn_gate\.weight$=q4_k
186
+ ^blk\.21\.ffn_up\.weight$=q4_k
187
+ ^blk\.21\.ssm_out\.weight$=q8_0
188
+ ^blk\.22\.attn_gate\.weight$=q4_k
189
+ ^blk\.22\.attn_qkv\.weight$=q4_k
190
+ ^blk\.22\.ffn_down\.weight$=q4_k
191
+ ^blk\.22\.ffn_gate\.weight$=q4_k
192
+ ^blk\.22\.ffn_up\.weight$=q4_k
193
+ ^blk\.22\.ssm_out\.weight$=q8_0
194
+ ^blk\.23\.attn_k\.weight$=q8_0
195
+ ^blk\.23\.attn_output\.weight$=q8_0
196
+ ^blk\.23\.attn_q\.weight$=q4_k
197
+ ^blk\.23\.attn_v\.weight$=q8_0
198
+ ^blk\.23\.ffn_down\.weight$=q4_k
199
+ ^blk\.23\.ffn_gate\.weight$=q4_k
200
+ ^blk\.23\.ffn_up\.weight$=q4_k
201
+ ^blk\.24\.attn_gate\.weight$=q6_k
202
+ ^blk\.24\.attn_qkv\.weight$=q4_k
203
+ ^blk\.24\.ffn_down\.weight$=q6_k
204
+ ^blk\.24\.ffn_gate\.weight$=q4_k
205
+ ^blk\.24\.ffn_up\.weight$=q4_k
206
+ ^blk\.24\.ssm_out\.weight$=q8_0
207
+ ^blk\.25\.attn_gate\.weight$=q6_k
208
+ ^blk\.25\.attn_qkv\.weight$=q4_k
209
+ ^blk\.25\.ffn_down\.weight$=q6_k
210
+ ^blk\.25\.ffn_gate\.weight$=q4_k
211
+ ^blk\.25\.ffn_up\.weight$=q4_k
212
+ ^blk\.25\.ssm_out\.weight$=q8_0
213
+ ^blk\.26\.attn_gate\.weight$=q6_k
214
+ ^blk\.26\.attn_qkv\.weight$=q4_k
215
+ ^blk\.26\.ffn_down\.weight$=q6_k
216
+ ^blk\.26\.ffn_gate\.weight$=q4_k
217
+ ^blk\.26\.ffn_up\.weight$=q4_k
218
+ ^blk\.26\.ssm_out\.weight$=q8_0
219
+ ^blk\.27\.attn_k\.weight$=q8_0
220
+ ^blk\.27\.attn_output\.weight$=q8_0
221
+ ^blk\.27\.attn_q\.weight$=q6_k
222
+ ^blk\.27\.attn_v\.weight$=q8_0
223
+ ^blk\.27\.ffn_down\.weight$=q6_k
224
+ ^blk\.27\.ffn_gate\.weight$=q4_k
225
+ ^blk\.27\.ffn_up\.weight$=q6_k
226
+ ^blk\.28\.attn_gate\.weight$=q6_k
227
+ ^blk\.28\.attn_qkv\.weight$=q4_k
228
+ ^blk\.28\.ffn_down\.weight$=q6_k
229
+ ^blk\.28\.ffn_gate\.weight$=q4_k
230
+ ^blk\.28\.ffn_up\.weight$=q4_k
231
+ ^blk\.28\.ssm_out\.weight$=q8_0
232
+ ^blk\.29\.attn_gate\.weight$=q6_k
233
+ ^blk\.29\.attn_qkv\.weight$=q4_k
234
+ ^blk\.29\.ffn_down\.weight$=q6_k
235
+ ^blk\.29\.ffn_gate\.weight$=q4_k
236
+ ^blk\.29\.ffn_up\.weight$=q4_k
237
+ ^blk\.29\.ssm_out\.weight$=q8_0
238
+ ^blk\.30\.attn_gate\.weight$=q6_k
239
+ ^blk\.30\.attn_qkv\.weight$=q4_k
240
+ ^blk\.30\.ffn_down\.weight$=q6_k
241
+ ^blk\.30\.ffn_gate\.weight$=q4_k
242
+ ^blk\.30\.ffn_up\.weight$=q4_k
243
+ ^blk\.30\.ssm_out\.weight$=q8_0
244
+ ^blk\.31\.attn_k\.weight$=q8_0
245
+ ^blk\.31\.attn_output\.weight$=q8_0
246
+ ^blk\.31\.attn_q\.weight$=q6_k
247
+ ^blk\.31\.attn_v\.weight$=q8_0
248
+ ^blk\.31\.ffn_down\.weight$=q8_0
249
+ ^blk\.31\.ffn_gate\.weight$=q8_0
250
+ ^blk\.31\.ffn_up\.weight$=q8_0
layouts/LensVLM-9B-Q4_K_M.layout.json ADDED
@@ -0,0 +1,196 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "Q4_K_M",
5
+ "base_type": "q4_k",
6
+ "family": "k",
7
+ "recipe": "ident",
8
+ "variant": "shr0.70",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "M",
12
+ "rung_share": 0.7,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.1414,
17
+ "type": "q6_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.097,
21
+ "type": "q4_k"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q6_k",
131
+ "token_embd.weight": "q4_k"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 5870403584,
135
+ "map_body+embd": 5800673280,
136
+ "ratio": 0.9881,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.7,
140
+ "body_bpw": 5.086,
141
+ "body_bpw_heuristic": 5.166,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 5830084608,
144
+ "file_bpw_pred": 5.209,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "q4_k": 17,
148
+ "q6_k": 6,
149
+ "q8_0": 1
150
+ },
151
+ "attn_k": {
152
+ "q8_0": 8
153
+ },
154
+ "attn_output": {
155
+ "q8_0": 8
156
+ },
157
+ "attn_q": {
158
+ "q4_k": 7,
159
+ "q6_k": 1
160
+ },
161
+ "attn_qkv": {
162
+ "q4_k": 23,
163
+ "q8_0": 1
164
+ },
165
+ "attn_v": {
166
+ "q8_0": 8
167
+ },
168
+ "ffn_down": {
169
+ "q4_k": 29,
170
+ "q6_k": 2,
171
+ "q8_0": 1
172
+ },
173
+ "ffn_gate": {
174
+ "q4_k": 31,
175
+ "q6_k": 1
176
+ },
177
+ "ffn_up": {
178
+ "q4_k": 23,
179
+ "q6_k": 8,
180
+ "q8_0": 1
181
+ },
182
+ "ssm_out": {
183
+ "q8_0": 12,
184
+ "q6_k": 12
185
+ }
186
+ },
187
+ "prior": "layout/prior.json",
188
+ "pin_rule": "floor",
189
+ "embd_cap": "heur",
190
+ "generator": "auto_quant_v2.layout",
191
+ "generator_key": "93ecba14028d5e4b",
192
+ "quant_type": "Q4_K_M",
193
+ "quantize_ftype": "Q4_K_M",
194
+ "llama_cpp_version": "b11081",
195
+ "commit": "d7cfc9bb610f62821944853c508648097409f4a7"
196
+ }
layouts/LensVLM-9B-Q4_K_M.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q6_k
50
+ ^token_embd\.weight$=q4_k
51
+ ^blk\.0\.attn_gate\.weight$=q8_0
52
+ ^blk\.0\.attn_qkv\.weight$=q8_0
53
+ ^blk\.0\.ffn_down\.weight$=q6_k
54
+ ^blk\.0\.ffn_gate\.weight$=q4_k
55
+ ^blk\.0\.ffn_up\.weight$=q6_k
56
+ ^blk\.0\.ssm_out\.weight$=q8_0
57
+ ^blk\.1\.attn_gate\.weight$=q6_k
58
+ ^blk\.1\.attn_qkv\.weight$=q4_k
59
+ ^blk\.1\.ffn_down\.weight$=q4_k
60
+ ^blk\.1\.ffn_gate\.weight$=q4_k
61
+ ^blk\.1\.ffn_up\.weight$=q6_k
62
+ ^blk\.1\.ssm_out\.weight$=q8_0
63
+ ^blk\.2\.attn_gate\.weight$=q6_k
64
+ ^blk\.2\.attn_qkv\.weight$=q4_k
65
+ ^blk\.2\.ffn_down\.weight$=q4_k
66
+ ^blk\.2\.ffn_gate\.weight$=q4_k
67
+ ^blk\.2\.ffn_up\.weight$=q6_k
68
+ ^blk\.2\.ssm_out\.weight$=q8_0
69
+ ^blk\.3\.attn_k\.weight$=q8_0
70
+ ^blk\.3\.attn_output\.weight$=q8_0
71
+ ^blk\.3\.attn_q\.weight$=q4_k
72
+ ^blk\.3\.attn_v\.weight$=q8_0
73
+ ^blk\.3\.ffn_down\.weight$=q4_k
74
+ ^blk\.3\.ffn_gate\.weight$=q4_k
75
+ ^blk\.3\.ffn_up\.weight$=q6_k
76
+ ^blk\.4\.attn_gate\.weight$=q6_k
77
+ ^blk\.4\.attn_qkv\.weight$=q4_k
78
+ ^blk\.4\.ffn_down\.weight$=q4_k
79
+ ^blk\.4\.ffn_gate\.weight$=q4_k
80
+ ^blk\.4\.ffn_up\.weight$=q6_k
81
+ ^blk\.4\.ssm_out\.weight$=q8_0
82
+ ^blk\.5\.attn_gate\.weight$=q6_k
83
+ ^blk\.5\.attn_qkv\.weight$=q4_k
84
+ ^blk\.5\.ffn_down\.weight$=q4_k
85
+ ^blk\.5\.ffn_gate\.weight$=q4_k
86
+ ^blk\.5\.ffn_up\.weight$=q6_k
87
+ ^blk\.5\.ssm_out\.weight$=q8_0
88
+ ^blk\.6\.attn_gate\.weight$=q6_k
89
+ ^blk\.6\.attn_qkv\.weight$=q4_k
90
+ ^blk\.6\.ffn_down\.weight$=q4_k
91
+ ^blk\.6\.ffn_gate\.weight$=q4_k
92
+ ^blk\.6\.ffn_up\.weight$=q6_k
93
+ ^blk\.6\.ssm_out\.weight$=q8_0
94
+ ^blk\.7\.attn_k\.weight$=q8_0
95
+ ^blk\.7\.attn_output\.weight$=q8_0
96
+ ^blk\.7\.attn_q\.weight$=q4_k
97
+ ^blk\.7\.attn_v\.weight$=q8_0
98
+ ^blk\.7\.ffn_down\.weight$=q4_k
99
+ ^blk\.7\.ffn_gate\.weight$=q4_k
100
+ ^blk\.7\.ffn_up\.weight$=q6_k
101
+ ^blk\.8\.attn_gate\.weight$=q4_k
102
+ ^blk\.8\.attn_qkv\.weight$=q4_k
103
+ ^blk\.8\.ffn_down\.weight$=q4_k
104
+ ^blk\.8\.ffn_gate\.weight$=q4_k
105
+ ^blk\.8\.ffn_up\.weight$=q4_k
106
+ ^blk\.8\.ssm_out\.weight$=q6_k
107
+ ^blk\.9\.attn_gate\.weight$=q4_k
108
+ ^blk\.9\.attn_qkv\.weight$=q4_k
109
+ ^blk\.9\.ffn_down\.weight$=q4_k
110
+ ^blk\.9\.ffn_gate\.weight$=q4_k
111
+ ^blk\.9\.ffn_up\.weight$=q4_k
112
+ ^blk\.9\.ssm_out\.weight$=q6_k
113
+ ^blk\.10\.attn_gate\.weight$=q4_k
114
+ ^blk\.10\.attn_qkv\.weight$=q4_k
115
+ ^blk\.10\.ffn_down\.weight$=q4_k
116
+ ^blk\.10\.ffn_gate\.weight$=q4_k
117
+ ^blk\.10\.ffn_up\.weight$=q4_k
118
+ ^blk\.10\.ssm_out\.weight$=q6_k
119
+ ^blk\.11\.attn_k\.weight$=q8_0
120
+ ^blk\.11\.attn_output\.weight$=q8_0
121
+ ^blk\.11\.attn_q\.weight$=q4_k
122
+ ^blk\.11\.attn_v\.weight$=q8_0
123
+ ^blk\.11\.ffn_down\.weight$=q4_k
124
+ ^blk\.11\.ffn_gate\.weight$=q4_k
125
+ ^blk\.11\.ffn_up\.weight$=q4_k
126
+ ^blk\.12\.attn_gate\.weight$=q4_k
127
+ ^blk\.12\.attn_qkv\.weight$=q4_k
128
+ ^blk\.12\.ffn_down\.weight$=q4_k
129
+ ^blk\.12\.ffn_gate\.weight$=q4_k
130
+ ^blk\.12\.ffn_up\.weight$=q4_k
131
+ ^blk\.12\.ssm_out\.weight$=q6_k
132
+ ^blk\.13\.attn_gate\.weight$=q4_k
133
+ ^blk\.13\.attn_qkv\.weight$=q4_k
134
+ ^blk\.13\.ffn_down\.weight$=q4_k
135
+ ^blk\.13\.ffn_gate\.weight$=q4_k
136
+ ^blk\.13\.ffn_up\.weight$=q4_k
137
+ ^blk\.13\.ssm_out\.weight$=q6_k
138
+ ^blk\.14\.attn_gate\.weight$=q4_k
139
+ ^blk\.14\.attn_qkv\.weight$=q4_k
140
+ ^blk\.14\.ffn_down\.weight$=q4_k
141
+ ^blk\.14\.ffn_gate\.weight$=q4_k
142
+ ^blk\.14\.ffn_up\.weight$=q4_k
143
+ ^blk\.14\.ssm_out\.weight$=q6_k
144
+ ^blk\.15\.attn_k\.weight$=q8_0
145
+ ^blk\.15\.attn_output\.weight$=q8_0
146
+ ^blk\.15\.attn_q\.weight$=q4_k
147
+ ^blk\.15\.attn_v\.weight$=q8_0
148
+ ^blk\.15\.ffn_down\.weight$=q4_k
149
+ ^blk\.15\.ffn_gate\.weight$=q4_k
150
+ ^blk\.15\.ffn_up\.weight$=q4_k
151
+ ^blk\.16\.attn_gate\.weight$=q4_k
152
+ ^blk\.16\.attn_qkv\.weight$=q4_k
153
+ ^blk\.16\.ffn_down\.weight$=q4_k
154
+ ^blk\.16\.ffn_gate\.weight$=q4_k
155
+ ^blk\.16\.ffn_up\.weight$=q4_k
156
+ ^blk\.16\.ssm_out\.weight$=q6_k
157
+ ^blk\.17\.attn_gate\.weight$=q4_k
158
+ ^blk\.17\.attn_qkv\.weight$=q4_k
159
+ ^blk\.17\.ffn_down\.weight$=q4_k
160
+ ^blk\.17\.ffn_gate\.weight$=q4_k
161
+ ^blk\.17\.ffn_up\.weight$=q4_k
162
+ ^blk\.17\.ssm_out\.weight$=q6_k
163
+ ^blk\.18\.attn_gate\.weight$=q4_k
164
+ ^blk\.18\.attn_qkv\.weight$=q4_k
165
+ ^blk\.18\.ffn_down\.weight$=q4_k
166
+ ^blk\.18\.ffn_gate\.weight$=q4_k
167
+ ^blk\.18\.ffn_up\.weight$=q4_k
168
+ ^blk\.18\.ssm_out\.weight$=q6_k
169
+ ^blk\.19\.attn_k\.weight$=q8_0
170
+ ^blk\.19\.attn_output\.weight$=q8_0
171
+ ^blk\.19\.attn_q\.weight$=q4_k
172
+ ^blk\.19\.attn_v\.weight$=q8_0
173
+ ^blk\.19\.ffn_down\.weight$=q4_k
174
+ ^blk\.19\.ffn_gate\.weight$=q4_k
175
+ ^blk\.19\.ffn_up\.weight$=q4_k
176
+ ^blk\.20\.attn_gate\.weight$=q4_k
177
+ ^blk\.20\.attn_qkv\.weight$=q4_k
178
+ ^blk\.20\.ffn_down\.weight$=q4_k
179
+ ^blk\.20\.ffn_gate\.weight$=q4_k
180
+ ^blk\.20\.ffn_up\.weight$=q4_k
181
+ ^blk\.20\.ssm_out\.weight$=q6_k
182
+ ^blk\.21\.attn_gate\.weight$=q4_k
183
+ ^blk\.21\.attn_qkv\.weight$=q4_k
184
+ ^blk\.21\.ffn_down\.weight$=q4_k
185
+ ^blk\.21\.ffn_gate\.weight$=q4_k
186
+ ^blk\.21\.ffn_up\.weight$=q4_k
187
+ ^blk\.21\.ssm_out\.weight$=q6_k
188
+ ^blk\.22\.attn_gate\.weight$=q4_k
189
+ ^blk\.22\.attn_qkv\.weight$=q4_k
190
+ ^blk\.22\.ffn_down\.weight$=q4_k
191
+ ^blk\.22\.ffn_gate\.weight$=q4_k
192
+ ^blk\.22\.ffn_up\.weight$=q4_k
193
+ ^blk\.22\.ssm_out\.weight$=q6_k
194
+ ^blk\.23\.attn_k\.weight$=q8_0
195
+ ^blk\.23\.attn_output\.weight$=q8_0
196
+ ^blk\.23\.attn_q\.weight$=q4_k
197
+ ^blk\.23\.attn_v\.weight$=q8_0
198
+ ^blk\.23\.ffn_down\.weight$=q4_k
199
+ ^blk\.23\.ffn_gate\.weight$=q4_k
200
+ ^blk\.23\.ffn_up\.weight$=q4_k
201
+ ^blk\.24\.attn_gate\.weight$=q6_k
202
+ ^blk\.24\.attn_qkv\.weight$=q4_k
203
+ ^blk\.24\.ffn_down\.weight$=q4_k
204
+ ^blk\.24\.ffn_gate\.weight$=q4_k
205
+ ^blk\.24\.ffn_up\.weight$=q4_k
206
+ ^blk\.24\.ssm_out\.weight$=q8_0
207
+ ^blk\.25\.attn_gate\.weight$=q4_k
208
+ ^blk\.25\.attn_qkv\.weight$=q4_k
209
+ ^blk\.25\.ffn_down\.weight$=q4_k
210
+ ^blk\.25\.ffn_gate\.weight$=q4_k
211
+ ^blk\.25\.ffn_up\.weight$=q4_k
212
+ ^blk\.25\.ssm_out\.weight$=q8_0
213
+ ^blk\.26\.attn_gate\.weight$=q4_k
214
+ ^blk\.26\.attn_qkv\.weight$=q4_k
215
+ ^blk\.26\.ffn_down\.weight$=q4_k
216
+ ^blk\.26\.ffn_gate\.weight$=q4_k
217
+ ^blk\.26\.ffn_up\.weight$=q4_k
218
+ ^blk\.26\.ssm_out\.weight$=q8_0
219
+ ^blk\.27\.attn_k\.weight$=q8_0
220
+ ^blk\.27\.attn_output\.weight$=q8_0
221
+ ^blk\.27\.attn_q\.weight$=q4_k
222
+ ^blk\.27\.attn_v\.weight$=q8_0
223
+ ^blk\.27\.ffn_down\.weight$=q6_k
224
+ ^blk\.27\.ffn_gate\.weight$=q4_k
225
+ ^blk\.27\.ffn_up\.weight$=q4_k
226
+ ^blk\.28\.attn_gate\.weight$=q4_k
227
+ ^blk\.28\.attn_qkv\.weight$=q4_k
228
+ ^blk\.28\.ffn_down\.weight$=q4_k
229
+ ^blk\.28\.ffn_gate\.weight$=q4_k
230
+ ^blk\.28\.ffn_up\.weight$=q4_k
231
+ ^blk\.28\.ssm_out\.weight$=q8_0
232
+ ^blk\.29\.attn_gate\.weight$=q4_k
233
+ ^blk\.29\.attn_qkv\.weight$=q4_k
234
+ ^blk\.29\.ffn_down\.weight$=q4_k
235
+ ^blk\.29\.ffn_gate\.weight$=q4_k
236
+ ^blk\.29\.ffn_up\.weight$=q4_k
237
+ ^blk\.29\.ssm_out\.weight$=q8_0
238
+ ^blk\.30\.attn_gate\.weight$=q4_k
239
+ ^blk\.30\.attn_qkv\.weight$=q4_k
240
+ ^blk\.30\.ffn_down\.weight$=q4_k
241
+ ^blk\.30\.ffn_gate\.weight$=q4_k
242
+ ^blk\.30\.ffn_up\.weight$=q4_k
243
+ ^blk\.30\.ssm_out\.weight$=q8_0
244
+ ^blk\.31\.attn_k\.weight$=q8_0
245
+ ^blk\.31\.attn_output\.weight$=q8_0
246
+ ^blk\.31\.attn_q\.weight$=q6_k
247
+ ^blk\.31\.attn_v\.weight$=q8_0
248
+ ^blk\.31\.ffn_down\.weight$=q8_0
249
+ ^blk\.31\.ffn_gate\.weight$=q6_k
250
+ ^blk\.31\.ffn_up\.weight$=q8_0
layouts/LensVLM-9B-Q4_K_S.layout.json ADDED
@@ -0,0 +1,194 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "Q4_K_S",
5
+ "base_type": "q4_k",
6
+ "family": "k",
7
+ "recipe": "ident",
8
+ "variant": "shr0.90",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "S",
12
+ "rung_share": 0.9,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.1493,
17
+ "type": "q6_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.1024,
21
+ "type": "q4_k"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q6_k",
131
+ "token_embd.weight": "q4_k"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 5557796864,
135
+ "map_body+embd": 5442879488,
136
+ "ratio": 0.9793,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.9,
140
+ "body_bpw": 4.672,
141
+ "body_bpw_heuristic": 4.805,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 5472290816,
144
+ "file_bpw_pred": 4.889,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "q4_k": 23,
148
+ "q8_0": 1
149
+ },
150
+ "attn_k": {
151
+ "q6_k": 5,
152
+ "q8_0": 3
153
+ },
154
+ "attn_output": {
155
+ "q6_k": 5,
156
+ "q8_0": 3
157
+ },
158
+ "attn_q": {
159
+ "q4_k": 8
160
+ },
161
+ "attn_qkv": {
162
+ "q4_k": 23,
163
+ "q6_k": 1
164
+ },
165
+ "attn_v": {
166
+ "q8_0": 8
167
+ },
168
+ "ffn_down": {
169
+ "q4_k": 31,
170
+ "q6_k": 1
171
+ },
172
+ "ffn_gate": {
173
+ "q4_k": 32
174
+ },
175
+ "ffn_up": {
176
+ "q4_k": 31,
177
+ "q6_k": 1
178
+ },
179
+ "ssm_out": {
180
+ "q4_k": 18,
181
+ "q6_k": 5,
182
+ "q8_0": 1
183
+ }
184
+ },
185
+ "prior": "layout/prior.json",
186
+ "pin_rule": "floor",
187
+ "embd_cap": "heur",
188
+ "generator": "auto_quant_v2.layout",
189
+ "generator_key": "93ecba14028d5e4b",
190
+ "quant_type": "Q4_K_S",
191
+ "quantize_ftype": "Q4_K_S",
192
+ "llama_cpp_version": "b11081",
193
+ "commit": "d7cfc9bb610f62821944853c508648097409f4a7"
194
+ }
layouts/LensVLM-9B-Q4_K_S.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q6_k
50
+ ^token_embd\.weight$=q4_k
51
+ ^blk\.0\.attn_gate\.weight$=q8_0
52
+ ^blk\.0\.attn_qkv\.weight$=q6_k
53
+ ^blk\.0\.ffn_down\.weight$=q4_k
54
+ ^blk\.0\.ffn_gate\.weight$=q4_k
55
+ ^blk\.0\.ffn_up\.weight$=q4_k
56
+ ^blk\.0\.ssm_out\.weight$=q8_0
57
+ ^blk\.1\.attn_gate\.weight$=q4_k
58
+ ^blk\.1\.attn_qkv\.weight$=q4_k
59
+ ^blk\.1\.ffn_down\.weight$=q4_k
60
+ ^blk\.1\.ffn_gate\.weight$=q4_k
61
+ ^blk\.1\.ffn_up\.weight$=q4_k
62
+ ^blk\.1\.ssm_out\.weight$=q6_k
63
+ ^blk\.2\.attn_gate\.weight$=q4_k
64
+ ^blk\.2\.attn_qkv\.weight$=q4_k
65
+ ^blk\.2\.ffn_down\.weight$=q4_k
66
+ ^blk\.2\.ffn_gate\.weight$=q4_k
67
+ ^blk\.2\.ffn_up\.weight$=q4_k
68
+ ^blk\.2\.ssm_out\.weight$=q6_k
69
+ ^blk\.3\.attn_k\.weight$=q6_k
70
+ ^blk\.3\.attn_output\.weight$=q6_k
71
+ ^blk\.3\.attn_q\.weight$=q4_k
72
+ ^blk\.3\.attn_v\.weight$=q8_0
73
+ ^blk\.3\.ffn_down\.weight$=q4_k
74
+ ^blk\.3\.ffn_gate\.weight$=q4_k
75
+ ^blk\.3\.ffn_up\.weight$=q4_k
76
+ ^blk\.4\.attn_gate\.weight$=q4_k
77
+ ^blk\.4\.attn_qkv\.weight$=q4_k
78
+ ^blk\.4\.ffn_down\.weight$=q4_k
79
+ ^blk\.4\.ffn_gate\.weight$=q4_k
80
+ ^blk\.4\.ffn_up\.weight$=q4_k
81
+ ^blk\.4\.ssm_out\.weight$=q6_k
82
+ ^blk\.5\.attn_gate\.weight$=q4_k
83
+ ^blk\.5\.attn_qkv\.weight$=q4_k
84
+ ^blk\.5\.ffn_down\.weight$=q4_k
85
+ ^blk\.5\.ffn_gate\.weight$=q4_k
86
+ ^blk\.5\.ffn_up\.weight$=q4_k
87
+ ^blk\.5\.ssm_out\.weight$=q6_k
88
+ ^blk\.6\.attn_gate\.weight$=q4_k
89
+ ^blk\.6\.attn_qkv\.weight$=q4_k
90
+ ^blk\.6\.ffn_down\.weight$=q4_k
91
+ ^blk\.6\.ffn_gate\.weight$=q4_k
92
+ ^blk\.6\.ffn_up\.weight$=q4_k
93
+ ^blk\.6\.ssm_out\.weight$=q6_k
94
+ ^blk\.7\.attn_k\.weight$=q6_k
95
+ ^blk\.7\.attn_output\.weight$=q6_k
96
+ ^blk\.7\.attn_q\.weight$=q4_k
97
+ ^blk\.7\.attn_v\.weight$=q8_0
98
+ ^blk\.7\.ffn_down\.weight$=q4_k
99
+ ^blk\.7\.ffn_gate\.weight$=q4_k
100
+ ^blk\.7\.ffn_up\.weight$=q4_k
101
+ ^blk\.8\.attn_gate\.weight$=q4_k
102
+ ^blk\.8\.attn_qkv\.weight$=q4_k
103
+ ^blk\.8\.ffn_down\.weight$=q4_k
104
+ ^blk\.8\.ffn_gate\.weight$=q4_k
105
+ ^blk\.8\.ffn_up\.weight$=q4_k
106
+ ^blk\.8\.ssm_out\.weight$=q4_k
107
+ ^blk\.9\.attn_gate\.weight$=q4_k
108
+ ^blk\.9\.attn_qkv\.weight$=q4_k
109
+ ^blk\.9\.ffn_down\.weight$=q4_k
110
+ ^blk\.9\.ffn_gate\.weight$=q4_k
111
+ ^blk\.9\.ffn_up\.weight$=q4_k
112
+ ^blk\.9\.ssm_out\.weight$=q4_k
113
+ ^blk\.10\.attn_gate\.weight$=q4_k
114
+ ^blk\.10\.attn_qkv\.weight$=q4_k
115
+ ^blk\.10\.ffn_down\.weight$=q4_k
116
+ ^blk\.10\.ffn_gate\.weight$=q4_k
117
+ ^blk\.10\.ffn_up\.weight$=q4_k
118
+ ^blk\.10\.ssm_out\.weight$=q4_k
119
+ ^blk\.11\.attn_k\.weight$=q8_0
120
+ ^blk\.11\.attn_output\.weight$=q8_0
121
+ ^blk\.11\.attn_q\.weight$=q4_k
122
+ ^blk\.11\.attn_v\.weight$=q8_0
123
+ ^blk\.11\.ffn_down\.weight$=q4_k
124
+ ^blk\.11\.ffn_gate\.weight$=q4_k
125
+ ^blk\.11\.ffn_up\.weight$=q4_k
126
+ ^blk\.12\.attn_gate\.weight$=q4_k
127
+ ^blk\.12\.attn_qkv\.weight$=q4_k
128
+ ^blk\.12\.ffn_down\.weight$=q4_k
129
+ ^blk\.12\.ffn_gate\.weight$=q4_k
130
+ ^blk\.12\.ffn_up\.weight$=q4_k
131
+ ^blk\.12\.ssm_out\.weight$=q4_k
132
+ ^blk\.13\.attn_gate\.weight$=q4_k
133
+ ^blk\.13\.attn_qkv\.weight$=q4_k
134
+ ^blk\.13\.ffn_down\.weight$=q4_k
135
+ ^blk\.13\.ffn_gate\.weight$=q4_k
136
+ ^blk\.13\.ffn_up\.weight$=q4_k
137
+ ^blk\.13\.ssm_out\.weight$=q4_k
138
+ ^blk\.14\.attn_gate\.weight$=q4_k
139
+ ^blk\.14\.attn_qkv\.weight$=q4_k
140
+ ^blk\.14\.ffn_down\.weight$=q4_k
141
+ ^blk\.14\.ffn_gate\.weight$=q4_k
142
+ ^blk\.14\.ffn_up\.weight$=q4_k
143
+ ^blk\.14\.ssm_out\.weight$=q4_k
144
+ ^blk\.15\.attn_k\.weight$=q8_0
145
+ ^blk\.15\.attn_output\.weight$=q8_0
146
+ ^blk\.15\.attn_q\.weight$=q4_k
147
+ ^blk\.15\.attn_v\.weight$=q8_0
148
+ ^blk\.15\.ffn_down\.weight$=q4_k
149
+ ^blk\.15\.ffn_gate\.weight$=q4_k
150
+ ^blk\.15\.ffn_up\.weight$=q4_k
151
+ ^blk\.16\.attn_gate\.weight$=q4_k
152
+ ^blk\.16\.attn_qkv\.weight$=q4_k
153
+ ^blk\.16\.ffn_down\.weight$=q4_k
154
+ ^blk\.16\.ffn_gate\.weight$=q4_k
155
+ ^blk\.16\.ffn_up\.weight$=q4_k
156
+ ^blk\.16\.ssm_out\.weight$=q4_k
157
+ ^blk\.17\.attn_gate\.weight$=q4_k
158
+ ^blk\.17\.attn_qkv\.weight$=q4_k
159
+ ^blk\.17\.ffn_down\.weight$=q4_k
160
+ ^blk\.17\.ffn_gate\.weight$=q4_k
161
+ ^blk\.17\.ffn_up\.weight$=q4_k
162
+ ^blk\.17\.ssm_out\.weight$=q4_k
163
+ ^blk\.18\.attn_gate\.weight$=q4_k
164
+ ^blk\.18\.attn_qkv\.weight$=q4_k
165
+ ^blk\.18\.ffn_down\.weight$=q4_k
166
+ ^blk\.18\.ffn_gate\.weight$=q4_k
167
+ ^blk\.18\.ffn_up\.weight$=q4_k
168
+ ^blk\.18\.ssm_out\.weight$=q4_k
169
+ ^blk\.19\.attn_k\.weight$=q6_k
170
+ ^blk\.19\.attn_output\.weight$=q6_k
171
+ ^blk\.19\.attn_q\.weight$=q4_k
172
+ ^blk\.19\.attn_v\.weight$=q8_0
173
+ ^blk\.19\.ffn_down\.weight$=q4_k
174
+ ^blk\.19\.ffn_gate\.weight$=q4_k
175
+ ^blk\.19\.ffn_up\.weight$=q4_k
176
+ ^blk\.20\.attn_gate\.weight$=q4_k
177
+ ^blk\.20\.attn_qkv\.weight$=q4_k
178
+ ^blk\.20\.ffn_down\.weight$=q4_k
179
+ ^blk\.20\.ffn_gate\.weight$=q4_k
180
+ ^blk\.20\.ffn_up\.weight$=q4_k
181
+ ^blk\.20\.ssm_out\.weight$=q4_k
182
+ ^blk\.21\.attn_gate\.weight$=q4_k
183
+ ^blk\.21\.attn_qkv\.weight$=q4_k
184
+ ^blk\.21\.ffn_down\.weight$=q4_k
185
+ ^blk\.21\.ffn_gate\.weight$=q4_k
186
+ ^blk\.21\.ffn_up\.weight$=q4_k
187
+ ^blk\.21\.ssm_out\.weight$=q4_k
188
+ ^blk\.22\.attn_gate\.weight$=q4_k
189
+ ^blk\.22\.attn_qkv\.weight$=q4_k
190
+ ^blk\.22\.ffn_down\.weight$=q4_k
191
+ ^blk\.22\.ffn_gate\.weight$=q4_k
192
+ ^blk\.22\.ffn_up\.weight$=q4_k
193
+ ^blk\.22\.ssm_out\.weight$=q4_k
194
+ ^blk\.23\.attn_k\.weight$=q6_k
195
+ ^blk\.23\.attn_output\.weight$=q6_k
196
+ ^blk\.23\.attn_q\.weight$=q4_k
197
+ ^blk\.23\.attn_v\.weight$=q8_0
198
+ ^blk\.23\.ffn_down\.weight$=q4_k
199
+ ^blk\.23\.ffn_gate\.weight$=q4_k
200
+ ^blk\.23\.ffn_up\.weight$=q4_k
201
+ ^blk\.24\.attn_gate\.weight$=q4_k
202
+ ^blk\.24\.attn_qkv\.weight$=q4_k
203
+ ^blk\.24\.ffn_down\.weight$=q4_k
204
+ ^blk\.24\.ffn_gate\.weight$=q4_k
205
+ ^blk\.24\.ffn_up\.weight$=q4_k
206
+ ^blk\.24\.ssm_out\.weight$=q4_k
207
+ ^blk\.25\.attn_gate\.weight$=q4_k
208
+ ^blk\.25\.attn_qkv\.weight$=q4_k
209
+ ^blk\.25\.ffn_down\.weight$=q4_k
210
+ ^blk\.25\.ffn_gate\.weight$=q4_k
211
+ ^blk\.25\.ffn_up\.weight$=q4_k
212
+ ^blk\.25\.ssm_out\.weight$=q4_k
213
+ ^blk\.26\.attn_gate\.weight$=q4_k
214
+ ^blk\.26\.attn_qkv\.weight$=q4_k
215
+ ^blk\.26\.ffn_down\.weight$=q4_k
216
+ ^blk\.26\.ffn_gate\.weight$=q4_k
217
+ ^blk\.26\.ffn_up\.weight$=q4_k
218
+ ^blk\.26\.ssm_out\.weight$=q4_k
219
+ ^blk\.27\.attn_k\.weight$=q8_0
220
+ ^blk\.27\.attn_output\.weight$=q6_k
221
+ ^blk\.27\.attn_q\.weight$=q4_k
222
+ ^blk\.27\.attn_v\.weight$=q8_0
223
+ ^blk\.27\.ffn_down\.weight$=q4_k
224
+ ^blk\.27\.ffn_gate\.weight$=q4_k
225
+ ^blk\.27\.ffn_up\.weight$=q4_k
226
+ ^blk\.28\.attn_gate\.weight$=q4_k
227
+ ^blk\.28\.attn_qkv\.weight$=q4_k
228
+ ^blk\.28\.ffn_down\.weight$=q4_k
229
+ ^blk\.28\.ffn_gate\.weight$=q4_k
230
+ ^blk\.28\.ffn_up\.weight$=q4_k
231
+ ^blk\.28\.ssm_out\.weight$=q4_k
232
+ ^blk\.29\.attn_gate\.weight$=q4_k
233
+ ^blk\.29\.attn_qkv\.weight$=q4_k
234
+ ^blk\.29\.ffn_down\.weight$=q4_k
235
+ ^blk\.29\.ffn_gate\.weight$=q4_k
236
+ ^blk\.29\.ffn_up\.weight$=q4_k
237
+ ^blk\.29\.ssm_out\.weight$=q4_k
238
+ ^blk\.30\.attn_gate\.weight$=q4_k
239
+ ^blk\.30\.attn_qkv\.weight$=q4_k
240
+ ^blk\.30\.ffn_down\.weight$=q4_k
241
+ ^blk\.30\.ffn_gate\.weight$=q4_k
242
+ ^blk\.30\.ffn_up\.weight$=q4_k
243
+ ^blk\.30\.ssm_out\.weight$=q4_k
244
+ ^blk\.31\.attn_k\.weight$=q6_k
245
+ ^blk\.31\.attn_output\.weight$=q8_0
246
+ ^blk\.31\.attn_q\.weight$=q4_k
247
+ ^blk\.31\.attn_v\.weight$=q8_0
248
+ ^blk\.31\.ffn_down\.weight$=q6_k
249
+ ^blk\.31\.ffn_gate\.weight$=q4_k
250
+ ^blk\.31\.ffn_up\.weight$=q6_k
layouts/LensVLM-9B-Q5_K_M.layout.json ADDED
@@ -0,0 +1,193 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "Q5_K_M",
5
+ "base_type": "q5_k",
6
+ "family": "k",
7
+ "recipe": "ident",
8
+ "variant": "shr0.70",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "M",
12
+ "rung_share": 0.7,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.1219,
17
+ "type": "q6_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.1022,
21
+ "type": "q5_k"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q6_k",
131
+ "token_embd.weight": "q5_k"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 6812549120,
135
+ "map_body+embd": 6835748864,
136
+ "ratio": 1.0034,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.7,
140
+ "body_bpw": 6.137,
141
+ "body_bpw_heuristic": 6.11,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 6865160192,
144
+ "file_bpw_pred": 6.134,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "q5_k": 17,
148
+ "q8_0": 7
149
+ },
150
+ "attn_k": {
151
+ "q8_0": 8
152
+ },
153
+ "attn_output": {
154
+ "q8_0": 8
155
+ },
156
+ "attn_q": {
157
+ "q5_k": 7,
158
+ "q8_0": 1
159
+ },
160
+ "attn_qkv": {
161
+ "q5_k": 21,
162
+ "q6_k": 2,
163
+ "q8_0": 1
164
+ },
165
+ "attn_v": {
166
+ "q8_0": 8
167
+ },
168
+ "ffn_down": {
169
+ "q5_k": 29,
170
+ "q8_0": 3
171
+ },
172
+ "ffn_gate": {
173
+ "q5_k": 31,
174
+ "q8_0": 1
175
+ },
176
+ "ffn_up": {
177
+ "q5_k": 23,
178
+ "q8_0": 9
179
+ },
180
+ "ssm_out": {
181
+ "q8_0": 24
182
+ }
183
+ },
184
+ "prior": "layout/prior.json",
185
+ "pin_rule": "floor",
186
+ "embd_cap": "heur",
187
+ "generator": "auto_quant_v2.layout",
188
+ "generator_key": "93ecba14028d5e4b",
189
+ "quant_type": "Q5_K_M",
190
+ "quantize_ftype": "Q5_K_M",
191
+ "llama_cpp_version": "b11081",
192
+ "commit": "9bd4ca72e6b98d9c35d19c58991bf5d41945daa9"
193
+ }
layouts/LensVLM-9B-Q5_K_M.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q6_k
50
+ ^token_embd\.weight$=q5_k
51
+ ^blk\.0\.attn_gate\.weight$=q8_0
52
+ ^blk\.0\.attn_qkv\.weight$=q8_0
53
+ ^blk\.0\.ffn_down\.weight$=q8_0
54
+ ^blk\.0\.ffn_gate\.weight$=q5_k
55
+ ^blk\.0\.ffn_up\.weight$=q8_0
56
+ ^blk\.0\.ssm_out\.weight$=q8_0
57
+ ^blk\.1\.attn_gate\.weight$=q8_0
58
+ ^blk\.1\.attn_qkv\.weight$=q6_k
59
+ ^blk\.1\.ffn_down\.weight$=q5_k
60
+ ^blk\.1\.ffn_gate\.weight$=q5_k
61
+ ^blk\.1\.ffn_up\.weight$=q8_0
62
+ ^blk\.1\.ssm_out\.weight$=q8_0
63
+ ^blk\.2\.attn_gate\.weight$=q8_0
64
+ ^blk\.2\.attn_qkv\.weight$=q6_k
65
+ ^blk\.2\.ffn_down\.weight$=q5_k
66
+ ^blk\.2\.ffn_gate\.weight$=q5_k
67
+ ^blk\.2\.ffn_up\.weight$=q8_0
68
+ ^blk\.2\.ssm_out\.weight$=q8_0
69
+ ^blk\.3\.attn_k\.weight$=q8_0
70
+ ^blk\.3\.attn_output\.weight$=q8_0
71
+ ^blk\.3\.attn_q\.weight$=q5_k
72
+ ^blk\.3\.attn_v\.weight$=q8_0
73
+ ^blk\.3\.ffn_down\.weight$=q5_k
74
+ ^blk\.3\.ffn_gate\.weight$=q5_k
75
+ ^blk\.3\.ffn_up\.weight$=q8_0
76
+ ^blk\.4\.attn_gate\.weight$=q8_0
77
+ ^blk\.4\.attn_qkv\.weight$=q5_k
78
+ ^blk\.4\.ffn_down\.weight$=q5_k
79
+ ^blk\.4\.ffn_gate\.weight$=q5_k
80
+ ^blk\.4\.ffn_up\.weight$=q8_0
81
+ ^blk\.4\.ssm_out\.weight$=q8_0
82
+ ^blk\.5\.attn_gate\.weight$=q8_0
83
+ ^blk\.5\.attn_qkv\.weight$=q5_k
84
+ ^blk\.5\.ffn_down\.weight$=q5_k
85
+ ^blk\.5\.ffn_gate\.weight$=q5_k
86
+ ^blk\.5\.ffn_up\.weight$=q8_0
87
+ ^blk\.5\.ssm_out\.weight$=q8_0
88
+ ^blk\.6\.attn_gate\.weight$=q8_0
89
+ ^blk\.6\.attn_qkv\.weight$=q5_k
90
+ ^blk\.6\.ffn_down\.weight$=q5_k
91
+ ^blk\.6\.ffn_gate\.weight$=q5_k
92
+ ^blk\.6\.ffn_up\.weight$=q8_0
93
+ ^blk\.6\.ssm_out\.weight$=q8_0
94
+ ^blk\.7\.attn_k\.weight$=q8_0
95
+ ^blk\.7\.attn_output\.weight$=q8_0
96
+ ^blk\.7\.attn_q\.weight$=q5_k
97
+ ^blk\.7\.attn_v\.weight$=q8_0
98
+ ^blk\.7\.ffn_down\.weight$=q5_k
99
+ ^blk\.7\.ffn_gate\.weight$=q5_k
100
+ ^blk\.7\.ffn_up\.weight$=q8_0
101
+ ^blk\.8\.attn_gate\.weight$=q5_k
102
+ ^blk\.8\.attn_qkv\.weight$=q5_k
103
+ ^blk\.8\.ffn_down\.weight$=q5_k
104
+ ^blk\.8\.ffn_gate\.weight$=q5_k
105
+ ^blk\.8\.ffn_up\.weight$=q5_k
106
+ ^blk\.8\.ssm_out\.weight$=q8_0
107
+ ^blk\.9\.attn_gate\.weight$=q5_k
108
+ ^blk\.9\.attn_qkv\.weight$=q5_k
109
+ ^blk\.9\.ffn_down\.weight$=q5_k
110
+ ^blk\.9\.ffn_gate\.weight$=q5_k
111
+ ^blk\.9\.ffn_up\.weight$=q5_k
112
+ ^blk\.9\.ssm_out\.weight$=q8_0
113
+ ^blk\.10\.attn_gate\.weight$=q5_k
114
+ ^blk\.10\.attn_qkv\.weight$=q5_k
115
+ ^blk\.10\.ffn_down\.weight$=q5_k
116
+ ^blk\.10\.ffn_gate\.weight$=q5_k
117
+ ^blk\.10\.ffn_up\.weight$=q5_k
118
+ ^blk\.10\.ssm_out\.weight$=q8_0
119
+ ^blk\.11\.attn_k\.weight$=q8_0
120
+ ^blk\.11\.attn_output\.weight$=q8_0
121
+ ^blk\.11\.attn_q\.weight$=q5_k
122
+ ^blk\.11\.attn_v\.weight$=q8_0
123
+ ^blk\.11\.ffn_down\.weight$=q5_k
124
+ ^blk\.11\.ffn_gate\.weight$=q5_k
125
+ ^blk\.11\.ffn_up\.weight$=q5_k
126
+ ^blk\.12\.attn_gate\.weight$=q5_k
127
+ ^blk\.12\.attn_qkv\.weight$=q5_k
128
+ ^blk\.12\.ffn_down\.weight$=q5_k
129
+ ^blk\.12\.ffn_gate\.weight$=q5_k
130
+ ^blk\.12\.ffn_up\.weight$=q5_k
131
+ ^blk\.12\.ssm_out\.weight$=q8_0
132
+ ^blk\.13\.attn_gate\.weight$=q5_k
133
+ ^blk\.13\.attn_qkv\.weight$=q5_k
134
+ ^blk\.13\.ffn_down\.weight$=q5_k
135
+ ^blk\.13\.ffn_gate\.weight$=q5_k
136
+ ^blk\.13\.ffn_up\.weight$=q5_k
137
+ ^blk\.13\.ssm_out\.weight$=q8_0
138
+ ^blk\.14\.attn_gate\.weight$=q5_k
139
+ ^blk\.14\.attn_qkv\.weight$=q5_k
140
+ ^blk\.14\.ffn_down\.weight$=q5_k
141
+ ^blk\.14\.ffn_gate\.weight$=q5_k
142
+ ^blk\.14\.ffn_up\.weight$=q5_k
143
+ ^blk\.14\.ssm_out\.weight$=q8_0
144
+ ^blk\.15\.attn_k\.weight$=q8_0
145
+ ^blk\.15\.attn_output\.weight$=q8_0
146
+ ^blk\.15\.attn_q\.weight$=q5_k
147
+ ^blk\.15\.attn_v\.weight$=q8_0
148
+ ^blk\.15\.ffn_down\.weight$=q5_k
149
+ ^blk\.15\.ffn_gate\.weight$=q5_k
150
+ ^blk\.15\.ffn_up\.weight$=q5_k
151
+ ^blk\.16\.attn_gate\.weight$=q5_k
152
+ ^blk\.16\.attn_qkv\.weight$=q5_k
153
+ ^blk\.16\.ffn_down\.weight$=q5_k
154
+ ^blk\.16\.ffn_gate\.weight$=q5_k
155
+ ^blk\.16\.ffn_up\.weight$=q5_k
156
+ ^blk\.16\.ssm_out\.weight$=q8_0
157
+ ^blk\.17\.attn_gate\.weight$=q5_k
158
+ ^blk\.17\.attn_qkv\.weight$=q5_k
159
+ ^blk\.17\.ffn_down\.weight$=q5_k
160
+ ^blk\.17\.ffn_gate\.weight$=q5_k
161
+ ^blk\.17\.ffn_up\.weight$=q5_k
162
+ ^blk\.17\.ssm_out\.weight$=q8_0
163
+ ^blk\.18\.attn_gate\.weight$=q5_k
164
+ ^blk\.18\.attn_qkv\.weight$=q5_k
165
+ ^blk\.18\.ffn_down\.weight$=q5_k
166
+ ^blk\.18\.ffn_gate\.weight$=q5_k
167
+ ^blk\.18\.ffn_up\.weight$=q5_k
168
+ ^blk\.18\.ssm_out\.weight$=q8_0
169
+ ^blk\.19\.attn_k\.weight$=q8_0
170
+ ^blk\.19\.attn_output\.weight$=q8_0
171
+ ^blk\.19\.attn_q\.weight$=q5_k
172
+ ^blk\.19\.attn_v\.weight$=q8_0
173
+ ^blk\.19\.ffn_down\.weight$=q5_k
174
+ ^blk\.19\.ffn_gate\.weight$=q5_k
175
+ ^blk\.19\.ffn_up\.weight$=q5_k
176
+ ^blk\.20\.attn_gate\.weight$=q5_k
177
+ ^blk\.20\.attn_qkv\.weight$=q5_k
178
+ ^blk\.20\.ffn_down\.weight$=q5_k
179
+ ^blk\.20\.ffn_gate\.weight$=q5_k
180
+ ^blk\.20\.ffn_up\.weight$=q5_k
181
+ ^blk\.20\.ssm_out\.weight$=q8_0
182
+ ^blk\.21\.attn_gate\.weight$=q5_k
183
+ ^blk\.21\.attn_qkv\.weight$=q5_k
184
+ ^blk\.21\.ffn_down\.weight$=q5_k
185
+ ^blk\.21\.ffn_gate\.weight$=q5_k
186
+ ^blk\.21\.ffn_up\.weight$=q5_k
187
+ ^blk\.21\.ssm_out\.weight$=q8_0
188
+ ^blk\.22\.attn_gate\.weight$=q5_k
189
+ ^blk\.22\.attn_qkv\.weight$=q5_k
190
+ ^blk\.22\.ffn_down\.weight$=q5_k
191
+ ^blk\.22\.ffn_gate\.weight$=q5_k
192
+ ^blk\.22\.ffn_up\.weight$=q5_k
193
+ ^blk\.22\.ssm_out\.weight$=q8_0
194
+ ^blk\.23\.attn_k\.weight$=q8_0
195
+ ^blk\.23\.attn_output\.weight$=q8_0
196
+ ^blk\.23\.attn_q\.weight$=q5_k
197
+ ^blk\.23\.attn_v\.weight$=q8_0
198
+ ^blk\.23\.ffn_down\.weight$=q5_k
199
+ ^blk\.23\.ffn_gate\.weight$=q5_k
200
+ ^blk\.23\.ffn_up\.weight$=q5_k
201
+ ^blk\.24\.attn_gate\.weight$=q8_0
202
+ ^blk\.24\.attn_qkv\.weight$=q5_k
203
+ ^blk\.24\.ffn_down\.weight$=q5_k
204
+ ^blk\.24\.ffn_gate\.weight$=q5_k
205
+ ^blk\.24\.ffn_up\.weight$=q5_k
206
+ ^blk\.24\.ssm_out\.weight$=q8_0
207
+ ^blk\.25\.attn_gate\.weight$=q5_k
208
+ ^blk\.25\.attn_qkv\.weight$=q5_k
209
+ ^blk\.25\.ffn_down\.weight$=q5_k
210
+ ^blk\.25\.ffn_gate\.weight$=q5_k
211
+ ^blk\.25\.ffn_up\.weight$=q5_k
212
+ ^blk\.25\.ssm_out\.weight$=q8_0
213
+ ^blk\.26\.attn_gate\.weight$=q5_k
214
+ ^blk\.26\.attn_qkv\.weight$=q5_k
215
+ ^blk\.26\.ffn_down\.weight$=q5_k
216
+ ^blk\.26\.ffn_gate\.weight$=q5_k
217
+ ^blk\.26\.ffn_up\.weight$=q5_k
218
+ ^blk\.26\.ssm_out\.weight$=q8_0
219
+ ^blk\.27\.attn_k\.weight$=q8_0
220
+ ^blk\.27\.attn_output\.weight$=q8_0
221
+ ^blk\.27\.attn_q\.weight$=q5_k
222
+ ^blk\.27\.attn_v\.weight$=q8_0
223
+ ^blk\.27\.ffn_down\.weight$=q8_0
224
+ ^blk\.27\.ffn_gate\.weight$=q5_k
225
+ ^blk\.27\.ffn_up\.weight$=q5_k
226
+ ^blk\.28\.attn_gate\.weight$=q5_k
227
+ ^blk\.28\.attn_qkv\.weight$=q5_k
228
+ ^blk\.28\.ffn_down\.weight$=q5_k
229
+ ^blk\.28\.ffn_gate\.weight$=q5_k
230
+ ^blk\.28\.ffn_up\.weight$=q5_k
231
+ ^blk\.28\.ssm_out\.weight$=q8_0
232
+ ^blk\.29\.attn_gate\.weight$=q5_k
233
+ ^blk\.29\.attn_qkv\.weight$=q5_k
234
+ ^blk\.29\.ffn_down\.weight$=q5_k
235
+ ^blk\.29\.ffn_gate\.weight$=q5_k
236
+ ^blk\.29\.ffn_up\.weight$=q5_k
237
+ ^blk\.29\.ssm_out\.weight$=q8_0
238
+ ^blk\.30\.attn_gate\.weight$=q5_k
239
+ ^blk\.30\.attn_qkv\.weight$=q5_k
240
+ ^blk\.30\.ffn_down\.weight$=q5_k
241
+ ^blk\.30\.ffn_gate\.weight$=q5_k
242
+ ^blk\.30\.ffn_up\.weight$=q5_k
243
+ ^blk\.30\.ssm_out\.weight$=q8_0
244
+ ^blk\.31\.attn_k\.weight$=q8_0
245
+ ^blk\.31\.attn_output\.weight$=q8_0
246
+ ^blk\.31\.attn_q\.weight$=q8_0
247
+ ^blk\.31\.attn_v\.weight$=q8_0
248
+ ^blk\.31\.ffn_down\.weight$=q8_0
249
+ ^blk\.31\.ffn_gate\.weight$=q8_0
250
+ ^blk\.31\.ffn_up\.weight$=q8_0
layouts/LensVLM-9B-Q5_K_S.layout.json ADDED
@@ -0,0 +1,191 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "Q5_K_S",
5
+ "base_type": "q5_k",
6
+ "family": "k",
7
+ "recipe": "ident",
8
+ "variant": "shr0.90",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "S",
12
+ "rung_share": 0.9,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.1281,
17
+ "type": "q6_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.1073,
21
+ "type": "q5_k"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q6_k",
131
+ "token_embd.weight": "q5_k"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 6486048768,
135
+ "map_body+embd": 6455640064,
136
+ "ratio": 0.9953,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.902,
140
+ "body_bpw": 5.697,
141
+ "body_bpw_heuristic": 5.732,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 6485051392,
144
+ "file_bpw_pred": 5.794,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "q5_k": 23,
148
+ "q8_0": 1
149
+ },
150
+ "attn_k": {
151
+ "q8_0": 8
152
+ },
153
+ "attn_output": {
154
+ "q8_0": 8
155
+ },
156
+ "attn_q": {
157
+ "q5_k": 8
158
+ },
159
+ "attn_qkv": {
160
+ "q5_k": 23,
161
+ "q8_0": 1
162
+ },
163
+ "attn_v": {
164
+ "q8_0": 8
165
+ },
166
+ "ffn_down": {
167
+ "q5_k": 31,
168
+ "q8_0": 1
169
+ },
170
+ "ffn_gate": {
171
+ "q5_k": 32
172
+ },
173
+ "ffn_up": {
174
+ "q5_k": 31,
175
+ "q8_0": 1
176
+ },
177
+ "ssm_out": {
178
+ "q5_k": 18,
179
+ "q8_0": 6
180
+ }
181
+ },
182
+ "prior": "layout/prior.json",
183
+ "pin_rule": "floor",
184
+ "embd_cap": "heur",
185
+ "generator": "auto_quant_v2.layout",
186
+ "generator_key": "93ecba14028d5e4b",
187
+ "quant_type": "Q5_K_S",
188
+ "quantize_ftype": "Q5_K_S",
189
+ "llama_cpp_version": "b11081",
190
+ "commit": "9bd4ca72e6b98d9c35d19c58991bf5d41945daa9"
191
+ }
layouts/LensVLM-9B-Q5_K_S.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q6_k
50
+ ^token_embd\.weight$=q5_k
51
+ ^blk\.0\.attn_gate\.weight$=q8_0
52
+ ^blk\.0\.attn_qkv\.weight$=q8_0
53
+ ^blk\.0\.ffn_down\.weight$=q5_k
54
+ ^blk\.0\.ffn_gate\.weight$=q5_k
55
+ ^blk\.0\.ffn_up\.weight$=q5_k
56
+ ^blk\.0\.ssm_out\.weight$=q8_0
57
+ ^blk\.1\.attn_gate\.weight$=q5_k
58
+ ^blk\.1\.attn_qkv\.weight$=q5_k
59
+ ^blk\.1\.ffn_down\.weight$=q5_k
60
+ ^blk\.1\.ffn_gate\.weight$=q5_k
61
+ ^blk\.1\.ffn_up\.weight$=q5_k
62
+ ^blk\.1\.ssm_out\.weight$=q8_0
63
+ ^blk\.2\.attn_gate\.weight$=q5_k
64
+ ^blk\.2\.attn_qkv\.weight$=q5_k
65
+ ^blk\.2\.ffn_down\.weight$=q5_k
66
+ ^blk\.2\.ffn_gate\.weight$=q5_k
67
+ ^blk\.2\.ffn_up\.weight$=q5_k
68
+ ^blk\.2\.ssm_out\.weight$=q8_0
69
+ ^blk\.3\.attn_k\.weight$=q8_0
70
+ ^blk\.3\.attn_output\.weight$=q8_0
71
+ ^blk\.3\.attn_q\.weight$=q5_k
72
+ ^blk\.3\.attn_v\.weight$=q8_0
73
+ ^blk\.3\.ffn_down\.weight$=q5_k
74
+ ^blk\.3\.ffn_gate\.weight$=q5_k
75
+ ^blk\.3\.ffn_up\.weight$=q5_k
76
+ ^blk\.4\.attn_gate\.weight$=q5_k
77
+ ^blk\.4\.attn_qkv\.weight$=q5_k
78
+ ^blk\.4\.ffn_down\.weight$=q5_k
79
+ ^blk\.4\.ffn_gate\.weight$=q5_k
80
+ ^blk\.4\.ffn_up\.weight$=q5_k
81
+ ^blk\.4\.ssm_out\.weight$=q8_0
82
+ ^blk\.5\.attn_gate\.weight$=q5_k
83
+ ^blk\.5\.attn_qkv\.weight$=q5_k
84
+ ^blk\.5\.ffn_down\.weight$=q5_k
85
+ ^blk\.5\.ffn_gate\.weight$=q5_k
86
+ ^blk\.5\.ffn_up\.weight$=q5_k
87
+ ^blk\.5\.ssm_out\.weight$=q8_0
88
+ ^blk\.6\.attn_gate\.weight$=q5_k
89
+ ^blk\.6\.attn_qkv\.weight$=q5_k
90
+ ^blk\.6\.ffn_down\.weight$=q5_k
91
+ ^blk\.6\.ffn_gate\.weight$=q5_k
92
+ ^blk\.6\.ffn_up\.weight$=q5_k
93
+ ^blk\.6\.ssm_out\.weight$=q8_0
94
+ ^blk\.7\.attn_k\.weight$=q8_0
95
+ ^blk\.7\.attn_output\.weight$=q8_0
96
+ ^blk\.7\.attn_q\.weight$=q5_k
97
+ ^blk\.7\.attn_v\.weight$=q8_0
98
+ ^blk\.7\.ffn_down\.weight$=q5_k
99
+ ^blk\.7\.ffn_gate\.weight$=q5_k
100
+ ^blk\.7\.ffn_up\.weight$=q5_k
101
+ ^blk\.8\.attn_gate\.weight$=q5_k
102
+ ^blk\.8\.attn_qkv\.weight$=q5_k
103
+ ^blk\.8\.ffn_down\.weight$=q5_k
104
+ ^blk\.8\.ffn_gate\.weight$=q5_k
105
+ ^blk\.8\.ffn_up\.weight$=q5_k
106
+ ^blk\.8\.ssm_out\.weight$=q5_k
107
+ ^blk\.9\.attn_gate\.weight$=q5_k
108
+ ^blk\.9\.attn_qkv\.weight$=q5_k
109
+ ^blk\.9\.ffn_down\.weight$=q5_k
110
+ ^blk\.9\.ffn_gate\.weight$=q5_k
111
+ ^blk\.9\.ffn_up\.weight$=q5_k
112
+ ^blk\.9\.ssm_out\.weight$=q5_k
113
+ ^blk\.10\.attn_gate\.weight$=q5_k
114
+ ^blk\.10\.attn_qkv\.weight$=q5_k
115
+ ^blk\.10\.ffn_down\.weight$=q5_k
116
+ ^blk\.10\.ffn_gate\.weight$=q5_k
117
+ ^blk\.10\.ffn_up\.weight$=q5_k
118
+ ^blk\.10\.ssm_out\.weight$=q5_k
119
+ ^blk\.11\.attn_k\.weight$=q8_0
120
+ ^blk\.11\.attn_output\.weight$=q8_0
121
+ ^blk\.11\.attn_q\.weight$=q5_k
122
+ ^blk\.11\.attn_v\.weight$=q8_0
123
+ ^blk\.11\.ffn_down\.weight$=q5_k
124
+ ^blk\.11\.ffn_gate\.weight$=q5_k
125
+ ^blk\.11\.ffn_up\.weight$=q5_k
126
+ ^blk\.12\.attn_gate\.weight$=q5_k
127
+ ^blk\.12\.attn_qkv\.weight$=q5_k
128
+ ^blk\.12\.ffn_down\.weight$=q5_k
129
+ ^blk\.12\.ffn_gate\.weight$=q5_k
130
+ ^blk\.12\.ffn_up\.weight$=q5_k
131
+ ^blk\.12\.ssm_out\.weight$=q5_k
132
+ ^blk\.13\.attn_gate\.weight$=q5_k
133
+ ^blk\.13\.attn_qkv\.weight$=q5_k
134
+ ^blk\.13\.ffn_down\.weight$=q5_k
135
+ ^blk\.13\.ffn_gate\.weight$=q5_k
136
+ ^blk\.13\.ffn_up\.weight$=q5_k
137
+ ^blk\.13\.ssm_out\.weight$=q5_k
138
+ ^blk\.14\.attn_gate\.weight$=q5_k
139
+ ^blk\.14\.attn_qkv\.weight$=q5_k
140
+ ^blk\.14\.ffn_down\.weight$=q5_k
141
+ ^blk\.14\.ffn_gate\.weight$=q5_k
142
+ ^blk\.14\.ffn_up\.weight$=q5_k
143
+ ^blk\.14\.ssm_out\.weight$=q5_k
144
+ ^blk\.15\.attn_k\.weight$=q8_0
145
+ ^blk\.15\.attn_output\.weight$=q8_0
146
+ ^blk\.15\.attn_q\.weight$=q5_k
147
+ ^blk\.15\.attn_v\.weight$=q8_0
148
+ ^blk\.15\.ffn_down\.weight$=q5_k
149
+ ^blk\.15\.ffn_gate\.weight$=q5_k
150
+ ^blk\.15\.ffn_up\.weight$=q5_k
151
+ ^blk\.16\.attn_gate\.weight$=q5_k
152
+ ^blk\.16\.attn_qkv\.weight$=q5_k
153
+ ^blk\.16\.ffn_down\.weight$=q5_k
154
+ ^blk\.16\.ffn_gate\.weight$=q5_k
155
+ ^blk\.16\.ffn_up\.weight$=q5_k
156
+ ^blk\.16\.ssm_out\.weight$=q5_k
157
+ ^blk\.17\.attn_gate\.weight$=q5_k
158
+ ^blk\.17\.attn_qkv\.weight$=q5_k
159
+ ^blk\.17\.ffn_down\.weight$=q5_k
160
+ ^blk\.17\.ffn_gate\.weight$=q5_k
161
+ ^blk\.17\.ffn_up\.weight$=q5_k
162
+ ^blk\.17\.ssm_out\.weight$=q5_k
163
+ ^blk\.18\.attn_gate\.weight$=q5_k
164
+ ^blk\.18\.attn_qkv\.weight$=q5_k
165
+ ^blk\.18\.ffn_down\.weight$=q5_k
166
+ ^blk\.18\.ffn_gate\.weight$=q5_k
167
+ ^blk\.18\.ffn_up\.weight$=q5_k
168
+ ^blk\.18\.ssm_out\.weight$=q5_k
169
+ ^blk\.19\.attn_k\.weight$=q8_0
170
+ ^blk\.19\.attn_output\.weight$=q8_0
171
+ ^blk\.19\.attn_q\.weight$=q5_k
172
+ ^blk\.19\.attn_v\.weight$=q8_0
173
+ ^blk\.19\.ffn_down\.weight$=q5_k
174
+ ^blk\.19\.ffn_gate\.weight$=q5_k
175
+ ^blk\.19\.ffn_up\.weight$=q5_k
176
+ ^blk\.20\.attn_gate\.weight$=q5_k
177
+ ^blk\.20\.attn_qkv\.weight$=q5_k
178
+ ^blk\.20\.ffn_down\.weight$=q5_k
179
+ ^blk\.20\.ffn_gate\.weight$=q5_k
180
+ ^blk\.20\.ffn_up\.weight$=q5_k
181
+ ^blk\.20\.ssm_out\.weight$=q5_k
182
+ ^blk\.21\.attn_gate\.weight$=q5_k
183
+ ^blk\.21\.attn_qkv\.weight$=q5_k
184
+ ^blk\.21\.ffn_down\.weight$=q5_k
185
+ ^blk\.21\.ffn_gate\.weight$=q5_k
186
+ ^blk\.21\.ffn_up\.weight$=q5_k
187
+ ^blk\.21\.ssm_out\.weight$=q5_k
188
+ ^blk\.22\.attn_gate\.weight$=q5_k
189
+ ^blk\.22\.attn_qkv\.weight$=q5_k
190
+ ^blk\.22\.ffn_down\.weight$=q5_k
191
+ ^blk\.22\.ffn_gate\.weight$=q5_k
192
+ ^blk\.22\.ffn_up\.weight$=q5_k
193
+ ^blk\.22\.ssm_out\.weight$=q5_k
194
+ ^blk\.23\.attn_k\.weight$=q8_0
195
+ ^blk\.23\.attn_output\.weight$=q8_0
196
+ ^blk\.23\.attn_q\.weight$=q5_k
197
+ ^blk\.23\.attn_v\.weight$=q8_0
198
+ ^blk\.23\.ffn_down\.weight$=q5_k
199
+ ^blk\.23\.ffn_gate\.weight$=q5_k
200
+ ^blk\.23\.ffn_up\.weight$=q5_k
201
+ ^blk\.24\.attn_gate\.weight$=q5_k
202
+ ^blk\.24\.attn_qkv\.weight$=q5_k
203
+ ^blk\.24\.ffn_down\.weight$=q5_k
204
+ ^blk\.24\.ffn_gate\.weight$=q5_k
205
+ ^blk\.24\.ffn_up\.weight$=q5_k
206
+ ^blk\.24\.ssm_out\.weight$=q5_k
207
+ ^blk\.25\.attn_gate\.weight$=q5_k
208
+ ^blk\.25\.attn_qkv\.weight$=q5_k
209
+ ^blk\.25\.ffn_down\.weight$=q5_k
210
+ ^blk\.25\.ffn_gate\.weight$=q5_k
211
+ ^blk\.25\.ffn_up\.weight$=q5_k
212
+ ^blk\.25\.ssm_out\.weight$=q5_k
213
+ ^blk\.26\.attn_gate\.weight$=q5_k
214
+ ^blk\.26\.attn_qkv\.weight$=q5_k
215
+ ^blk\.26\.ffn_down\.weight$=q5_k
216
+ ^blk\.26\.ffn_gate\.weight$=q5_k
217
+ ^blk\.26\.ffn_up\.weight$=q5_k
218
+ ^blk\.26\.ssm_out\.weight$=q5_k
219
+ ^blk\.27\.attn_k\.weight$=q8_0
220
+ ^blk\.27\.attn_output\.weight$=q8_0
221
+ ^blk\.27\.attn_q\.weight$=q5_k
222
+ ^blk\.27\.attn_v\.weight$=q8_0
223
+ ^blk\.27\.ffn_down\.weight$=q5_k
224
+ ^blk\.27\.ffn_gate\.weight$=q5_k
225
+ ^blk\.27\.ffn_up\.weight$=q5_k
226
+ ^blk\.28\.attn_gate\.weight$=q5_k
227
+ ^blk\.28\.attn_qkv\.weight$=q5_k
228
+ ^blk\.28\.ffn_down\.weight$=q5_k
229
+ ^blk\.28\.ffn_gate\.weight$=q5_k
230
+ ^blk\.28\.ffn_up\.weight$=q5_k
231
+ ^blk\.28\.ssm_out\.weight$=q5_k
232
+ ^blk\.29\.attn_gate\.weight$=q5_k
233
+ ^blk\.29\.attn_qkv\.weight$=q5_k
234
+ ^blk\.29\.ffn_down\.weight$=q5_k
235
+ ^blk\.29\.ffn_gate\.weight$=q5_k
236
+ ^blk\.29\.ffn_up\.weight$=q5_k
237
+ ^blk\.29\.ssm_out\.weight$=q5_k
238
+ ^blk\.30\.attn_gate\.weight$=q5_k
239
+ ^blk\.30\.attn_qkv\.weight$=q5_k
240
+ ^blk\.30\.ffn_down\.weight$=q5_k
241
+ ^blk\.30\.ffn_gate\.weight$=q5_k
242
+ ^blk\.30\.ffn_up\.weight$=q5_k
243
+ ^blk\.30\.ssm_out\.weight$=q5_k
244
+ ^blk\.31\.attn_k\.weight$=q8_0
245
+ ^blk\.31\.attn_output\.weight$=q8_0
246
+ ^blk\.31\.attn_q\.weight$=q5_k
247
+ ^blk\.31\.attn_v\.weight$=q8_0
248
+ ^blk\.31\.ffn_down\.weight$=q8_0
249
+ ^blk\.31\.ffn_gate\.weight$=q5_k
250
+ ^blk\.31\.ffn_up\.weight$=q8_0
layouts/LensVLM-9B-Q6_K.layout.json ADDED
@@ -0,0 +1,192 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "Q6_K",
5
+ "base_type": "q6_k",
6
+ "family": "k",
7
+ "recipe": "ident",
8
+ "variant": "shr0.70",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "M",
12
+ "rung_share": 0.7,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.1085,
17
+ "type": "q6_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.1085,
21
+ "type": "q6_k"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q6_k",
131
+ "token_embd.weight": "q6_k"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 7659880448,
135
+ "map_body+embd": 7753334784,
136
+ "ratio": 1.0122,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.701,
140
+ "body_bpw": 7.042,
141
+ "body_bpw_heuristic": 6.934,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 7782746112,
144
+ "file_bpw_pred": 6.954,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "q6_k": 17,
148
+ "q8_0": 7
149
+ },
150
+ "attn_k": {
151
+ "q8_0": 8
152
+ },
153
+ "attn_output": {
154
+ "q8_0": 8
155
+ },
156
+ "attn_q": {
157
+ "q6_k": 7,
158
+ "q8_0": 1
159
+ },
160
+ "attn_qkv": {
161
+ "q6_k": 18,
162
+ "q8_0": 6
163
+ },
164
+ "attn_v": {
165
+ "q8_0": 8
166
+ },
167
+ "ffn_down": {
168
+ "q6_k": 27,
169
+ "q8_0": 5
170
+ },
171
+ "ffn_gate": {
172
+ "q6_k": 31,
173
+ "q8_0": 1
174
+ },
175
+ "ffn_up": {
176
+ "q6_k": 23,
177
+ "q8_0": 9
178
+ },
179
+ "ssm_out": {
180
+ "q8_0": 24
181
+ }
182
+ },
183
+ "prior": "layout/prior.json",
184
+ "pin_rule": "floor",
185
+ "embd_cap": "heur",
186
+ "generator": "auto_quant_v2.layout",
187
+ "generator_key": "93ecba14028d5e4b",
188
+ "quant_type": "Q6_K",
189
+ "quantize_ftype": "Q6_K",
190
+ "llama_cpp_version": "b11081",
191
+ "commit": "9bd4ca72e6b98d9c35d19c58991bf5d41945daa9"
192
+ }
layouts/LensVLM-9B-Q6_K.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q6_k
50
+ ^token_embd\.weight$=q6_k
51
+ ^blk\.0\.attn_gate\.weight$=q8_0
52
+ ^blk\.0\.attn_qkv\.weight$=q8_0
53
+ ^blk\.0\.ffn_down\.weight$=q8_0
54
+ ^blk\.0\.ffn_gate\.weight$=q6_k
55
+ ^blk\.0\.ffn_up\.weight$=q8_0
56
+ ^blk\.0\.ssm_out\.weight$=q8_0
57
+ ^blk\.1\.attn_gate\.weight$=q8_0
58
+ ^blk\.1\.attn_qkv\.weight$=q8_0
59
+ ^blk\.1\.ffn_down\.weight$=q6_k
60
+ ^blk\.1\.ffn_gate\.weight$=q6_k
61
+ ^blk\.1\.ffn_up\.weight$=q8_0
62
+ ^blk\.1\.ssm_out\.weight$=q8_0
63
+ ^blk\.2\.attn_gate\.weight$=q8_0
64
+ ^blk\.2\.attn_qkv\.weight$=q8_0
65
+ ^blk\.2\.ffn_down\.weight$=q6_k
66
+ ^blk\.2\.ffn_gate\.weight$=q6_k
67
+ ^blk\.2\.ffn_up\.weight$=q8_0
68
+ ^blk\.2\.ssm_out\.weight$=q8_0
69
+ ^blk\.3\.attn_k\.weight$=q8_0
70
+ ^blk\.3\.attn_output\.weight$=q8_0
71
+ ^blk\.3\.attn_q\.weight$=q6_k
72
+ ^blk\.3\.attn_v\.weight$=q8_0
73
+ ^blk\.3\.ffn_down\.weight$=q8_0
74
+ ^blk\.3\.ffn_gate\.weight$=q6_k
75
+ ^blk\.3\.ffn_up\.weight$=q8_0
76
+ ^blk\.4\.attn_gate\.weight$=q8_0
77
+ ^blk\.4\.attn_qkv\.weight$=q8_0
78
+ ^blk\.4\.ffn_down\.weight$=q6_k
79
+ ^blk\.4\.ffn_gate\.weight$=q6_k
80
+ ^blk\.4\.ffn_up\.weight$=q8_0
81
+ ^blk\.4\.ssm_out\.weight$=q8_0
82
+ ^blk\.5\.attn_gate\.weight$=q8_0
83
+ ^blk\.5\.attn_qkv\.weight$=q8_0
84
+ ^blk\.5\.ffn_down\.weight$=q6_k
85
+ ^blk\.5\.ffn_gate\.weight$=q6_k
86
+ ^blk\.5\.ffn_up\.weight$=q8_0
87
+ ^blk\.5\.ssm_out\.weight$=q8_0
88
+ ^blk\.6\.attn_gate\.weight$=q8_0
89
+ ^blk\.6\.attn_qkv\.weight$=q8_0
90
+ ^blk\.6\.ffn_down\.weight$=q6_k
91
+ ^blk\.6\.ffn_gate\.weight$=q6_k
92
+ ^blk\.6\.ffn_up\.weight$=q8_0
93
+ ^blk\.6\.ssm_out\.weight$=q8_0
94
+ ^blk\.7\.attn_k\.weight$=q8_0
95
+ ^blk\.7\.attn_output\.weight$=q8_0
96
+ ^blk\.7\.attn_q\.weight$=q6_k
97
+ ^blk\.7\.attn_v\.weight$=q8_0
98
+ ^blk\.7\.ffn_down\.weight$=q8_0
99
+ ^blk\.7\.ffn_gate\.weight$=q6_k
100
+ ^blk\.7\.ffn_up\.weight$=q8_0
101
+ ^blk\.8\.attn_gate\.weight$=q6_k
102
+ ^blk\.8\.attn_qkv\.weight$=q6_k
103
+ ^blk\.8\.ffn_down\.weight$=q6_k
104
+ ^blk\.8\.ffn_gate\.weight$=q6_k
105
+ ^blk\.8\.ffn_up\.weight$=q6_k
106
+ ^blk\.8\.ssm_out\.weight$=q8_0
107
+ ^blk\.9\.attn_gate\.weight$=q6_k
108
+ ^blk\.9\.attn_qkv\.weight$=q6_k
109
+ ^blk\.9\.ffn_down\.weight$=q6_k
110
+ ^blk\.9\.ffn_gate\.weight$=q6_k
111
+ ^blk\.9\.ffn_up\.weight$=q6_k
112
+ ^blk\.9\.ssm_out\.weight$=q8_0
113
+ ^blk\.10\.attn_gate\.weight$=q6_k
114
+ ^blk\.10\.attn_qkv\.weight$=q6_k
115
+ ^blk\.10\.ffn_down\.weight$=q6_k
116
+ ^blk\.10\.ffn_gate\.weight$=q6_k
117
+ ^blk\.10\.ffn_up\.weight$=q6_k
118
+ ^blk\.10\.ssm_out\.weight$=q8_0
119
+ ^blk\.11\.attn_k\.weight$=q8_0
120
+ ^blk\.11\.attn_output\.weight$=q8_0
121
+ ^blk\.11\.attn_q\.weight$=q6_k
122
+ ^blk\.11\.attn_v\.weight$=q8_0
123
+ ^blk\.11\.ffn_down\.weight$=q6_k
124
+ ^blk\.11\.ffn_gate\.weight$=q6_k
125
+ ^blk\.11\.ffn_up\.weight$=q6_k
126
+ ^blk\.12\.attn_gate\.weight$=q6_k
127
+ ^blk\.12\.attn_qkv\.weight$=q6_k
128
+ ^blk\.12\.ffn_down\.weight$=q6_k
129
+ ^blk\.12\.ffn_gate\.weight$=q6_k
130
+ ^blk\.12\.ffn_up\.weight$=q6_k
131
+ ^blk\.12\.ssm_out\.weight$=q8_0
132
+ ^blk\.13\.attn_gate\.weight$=q6_k
133
+ ^blk\.13\.attn_qkv\.weight$=q6_k
134
+ ^blk\.13\.ffn_down\.weight$=q6_k
135
+ ^blk\.13\.ffn_gate\.weight$=q6_k
136
+ ^blk\.13\.ffn_up\.weight$=q6_k
137
+ ^blk\.13\.ssm_out\.weight$=q8_0
138
+ ^blk\.14\.attn_gate\.weight$=q6_k
139
+ ^blk\.14\.attn_qkv\.weight$=q6_k
140
+ ^blk\.14\.ffn_down\.weight$=q6_k
141
+ ^blk\.14\.ffn_gate\.weight$=q6_k
142
+ ^blk\.14\.ffn_up\.weight$=q6_k
143
+ ^blk\.14\.ssm_out\.weight$=q8_0
144
+ ^blk\.15\.attn_k\.weight$=q8_0
145
+ ^blk\.15\.attn_output\.weight$=q8_0
146
+ ^blk\.15\.attn_q\.weight$=q6_k
147
+ ^blk\.15\.attn_v\.weight$=q8_0
148
+ ^blk\.15\.ffn_down\.weight$=q6_k
149
+ ^blk\.15\.ffn_gate\.weight$=q6_k
150
+ ^blk\.15\.ffn_up\.weight$=q6_k
151
+ ^blk\.16\.attn_gate\.weight$=q6_k
152
+ ^blk\.16\.attn_qkv\.weight$=q6_k
153
+ ^blk\.16\.ffn_down\.weight$=q6_k
154
+ ^blk\.16\.ffn_gate\.weight$=q6_k
155
+ ^blk\.16\.ffn_up\.weight$=q6_k
156
+ ^blk\.16\.ssm_out\.weight$=q8_0
157
+ ^blk\.17\.attn_gate\.weight$=q6_k
158
+ ^blk\.17\.attn_qkv\.weight$=q6_k
159
+ ^blk\.17\.ffn_down\.weight$=q6_k
160
+ ^blk\.17\.ffn_gate\.weight$=q6_k
161
+ ^blk\.17\.ffn_up\.weight$=q6_k
162
+ ^blk\.17\.ssm_out\.weight$=q8_0
163
+ ^blk\.18\.attn_gate\.weight$=q6_k
164
+ ^blk\.18\.attn_qkv\.weight$=q6_k
165
+ ^blk\.18\.ffn_down\.weight$=q6_k
166
+ ^blk\.18\.ffn_gate\.weight$=q6_k
167
+ ^blk\.18\.ffn_up\.weight$=q6_k
168
+ ^blk\.18\.ssm_out\.weight$=q8_0
169
+ ^blk\.19\.attn_k\.weight$=q8_0
170
+ ^blk\.19\.attn_output\.weight$=q8_0
171
+ ^blk\.19\.attn_q\.weight$=q6_k
172
+ ^blk\.19\.attn_v\.weight$=q8_0
173
+ ^blk\.19\.ffn_down\.weight$=q6_k
174
+ ^blk\.19\.ffn_gate\.weight$=q6_k
175
+ ^blk\.19\.ffn_up\.weight$=q6_k
176
+ ^blk\.20\.attn_gate\.weight$=q6_k
177
+ ^blk\.20\.attn_qkv\.weight$=q6_k
178
+ ^blk\.20\.ffn_down\.weight$=q6_k
179
+ ^blk\.20\.ffn_gate\.weight$=q6_k
180
+ ^blk\.20\.ffn_up\.weight$=q6_k
181
+ ^blk\.20\.ssm_out\.weight$=q8_0
182
+ ^blk\.21\.attn_gate\.weight$=q6_k
183
+ ^blk\.21\.attn_qkv\.weight$=q6_k
184
+ ^blk\.21\.ffn_down\.weight$=q6_k
185
+ ^blk\.21\.ffn_gate\.weight$=q6_k
186
+ ^blk\.21\.ffn_up\.weight$=q6_k
187
+ ^blk\.21\.ssm_out\.weight$=q8_0
188
+ ^blk\.22\.attn_gate\.weight$=q6_k
189
+ ^blk\.22\.attn_qkv\.weight$=q6_k
190
+ ^blk\.22\.ffn_down\.weight$=q6_k
191
+ ^blk\.22\.ffn_gate\.weight$=q6_k
192
+ ^blk\.22\.ffn_up\.weight$=q6_k
193
+ ^blk\.22\.ssm_out\.weight$=q8_0
194
+ ^blk\.23\.attn_k\.weight$=q8_0
195
+ ^blk\.23\.attn_output\.weight$=q8_0
196
+ ^blk\.23\.attn_q\.weight$=q6_k
197
+ ^blk\.23\.attn_v\.weight$=q8_0
198
+ ^blk\.23\.ffn_down\.weight$=q6_k
199
+ ^blk\.23\.ffn_gate\.weight$=q6_k
200
+ ^blk\.23\.ffn_up\.weight$=q6_k
201
+ ^blk\.24\.attn_gate\.weight$=q8_0
202
+ ^blk\.24\.attn_qkv\.weight$=q6_k
203
+ ^blk\.24\.ffn_down\.weight$=q6_k
204
+ ^blk\.24\.ffn_gate\.weight$=q6_k
205
+ ^blk\.24\.ffn_up\.weight$=q6_k
206
+ ^blk\.24\.ssm_out\.weight$=q8_0
207
+ ^blk\.25\.attn_gate\.weight$=q6_k
208
+ ^blk\.25\.attn_qkv\.weight$=q6_k
209
+ ^blk\.25\.ffn_down\.weight$=q6_k
210
+ ^blk\.25\.ffn_gate\.weight$=q6_k
211
+ ^blk\.25\.ffn_up\.weight$=q6_k
212
+ ^blk\.25\.ssm_out\.weight$=q8_0
213
+ ^blk\.26\.attn_gate\.weight$=q6_k
214
+ ^blk\.26\.attn_qkv\.weight$=q6_k
215
+ ^blk\.26\.ffn_down\.weight$=q6_k
216
+ ^blk\.26\.ffn_gate\.weight$=q6_k
217
+ ^blk\.26\.ffn_up\.weight$=q6_k
218
+ ^blk\.26\.ssm_out\.weight$=q8_0
219
+ ^blk\.27\.attn_k\.weight$=q8_0
220
+ ^blk\.27\.attn_output\.weight$=q8_0
221
+ ^blk\.27\.attn_q\.weight$=q6_k
222
+ ^blk\.27\.attn_v\.weight$=q8_0
223
+ ^blk\.27\.ffn_down\.weight$=q8_0
224
+ ^blk\.27\.ffn_gate\.weight$=q6_k
225
+ ^blk\.27\.ffn_up\.weight$=q6_k
226
+ ^blk\.28\.attn_gate\.weight$=q6_k
227
+ ^blk\.28\.attn_qkv\.weight$=q6_k
228
+ ^blk\.28\.ffn_down\.weight$=q6_k
229
+ ^blk\.28\.ffn_gate\.weight$=q6_k
230
+ ^blk\.28\.ffn_up\.weight$=q6_k
231
+ ^blk\.28\.ssm_out\.weight$=q8_0
232
+ ^blk\.29\.attn_gate\.weight$=q6_k
233
+ ^blk\.29\.attn_qkv\.weight$=q6_k
234
+ ^blk\.29\.ffn_down\.weight$=q6_k
235
+ ^blk\.29\.ffn_gate\.weight$=q6_k
236
+ ^blk\.29\.ffn_up\.weight$=q6_k
237
+ ^blk\.29\.ssm_out\.weight$=q8_0
238
+ ^blk\.30\.attn_gate\.weight$=q6_k
239
+ ^blk\.30\.attn_qkv\.weight$=q6_k
240
+ ^blk\.30\.ffn_down\.weight$=q6_k
241
+ ^blk\.30\.ffn_gate\.weight$=q6_k
242
+ ^blk\.30\.ffn_up\.weight$=q6_k
243
+ ^blk\.30\.ssm_out\.weight$=q8_0
244
+ ^blk\.31\.attn_k\.weight$=q8_0
245
+ ^blk\.31\.attn_output\.weight$=q8_0
246
+ ^blk\.31\.attn_q\.weight$=q8_0
247
+ ^blk\.31\.attn_v\.weight$=q8_0
248
+ ^blk\.31\.ffn_down\.weight$=q8_0
249
+ ^blk\.31\.ffn_gate\.weight$=q8_0
250
+ ^blk\.31\.ffn_up\.weight$=q8_0
layouts/LensVLM-9B-Q6_K_L.layout.json ADDED
@@ -0,0 +1,192 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "Q6_K_L",
5
+ "base_type": "q6_k",
6
+ "family": "k",
7
+ "recipe": "ident",
8
+ "variant": "shr0.50",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "L",
12
+ "rung_share": 0.5,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.1085,
17
+ "type": "q6_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.1085,
21
+ "type": "q6_k"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q6_k",
131
+ "token_embd.weight": "q6_k"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 7659880448,
135
+ "map_body+embd": 8066203648,
136
+ "ratio": 1.053,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.501,
140
+ "body_bpw": 7.404,
141
+ "body_bpw_heuristic": 6.934,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 8095614976,
144
+ "file_bpw_pred": 7.233,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "q8_0": 12,
148
+ "q6_k": 12
149
+ },
150
+ "attn_k": {
151
+ "q8_0": 8
152
+ },
153
+ "attn_output": {
154
+ "q8_0": 8
155
+ },
156
+ "attn_q": {
157
+ "q8_0": 4,
158
+ "q6_k": 4
159
+ },
160
+ "attn_qkv": {
161
+ "q8_0": 18,
162
+ "q6_k": 6
163
+ },
164
+ "attn_v": {
165
+ "q8_0": 8
166
+ },
167
+ "ffn_down": {
168
+ "q8_0": 16,
169
+ "q6_k": 16
170
+ },
171
+ "ffn_gate": {
172
+ "q6_k": 31,
173
+ "q8_0": 1
174
+ },
175
+ "ffn_up": {
176
+ "q6_k": 20,
177
+ "q8_0": 12
178
+ },
179
+ "ssm_out": {
180
+ "q8_0": 24
181
+ }
182
+ },
183
+ "prior": "layout/prior.json",
184
+ "pin_rule": "floor",
185
+ "embd_cap": "heur",
186
+ "generator": "auto_quant_v2.layout",
187
+ "generator_key": "93ecba14028d5e4b",
188
+ "quant_type": "Q6_K_L",
189
+ "quantize_ftype": "Q6_K",
190
+ "llama_cpp_version": "b11081",
191
+ "commit": "9bd4ca72e6b98d9c35d19c58991bf5d41945daa9"
192
+ }
layouts/LensVLM-9B-Q6_K_L.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q6_k
50
+ ^token_embd\.weight$=q6_k
51
+ ^blk\.0\.attn_gate\.weight$=q8_0
52
+ ^blk\.0\.attn_qkv\.weight$=q8_0
53
+ ^blk\.0\.ffn_down\.weight$=q8_0
54
+ ^blk\.0\.ffn_gate\.weight$=q6_k
55
+ ^blk\.0\.ffn_up\.weight$=q8_0
56
+ ^blk\.0\.ssm_out\.weight$=q8_0
57
+ ^blk\.1\.attn_gate\.weight$=q8_0
58
+ ^blk\.1\.attn_qkv\.weight$=q8_0
59
+ ^blk\.1\.ffn_down\.weight$=q8_0
60
+ ^blk\.1\.ffn_gate\.weight$=q6_k
61
+ ^blk\.1\.ffn_up\.weight$=q8_0
62
+ ^blk\.1\.ssm_out\.weight$=q8_0
63
+ ^blk\.2\.attn_gate\.weight$=q8_0
64
+ ^blk\.2\.attn_qkv\.weight$=q8_0
65
+ ^blk\.2\.ffn_down\.weight$=q8_0
66
+ ^blk\.2\.ffn_gate\.weight$=q6_k
67
+ ^blk\.2\.ffn_up\.weight$=q8_0
68
+ ^blk\.2\.ssm_out\.weight$=q8_0
69
+ ^blk\.3\.attn_k\.weight$=q8_0
70
+ ^blk\.3\.attn_output\.weight$=q8_0
71
+ ^blk\.3\.attn_q\.weight$=q8_0
72
+ ^blk\.3\.attn_v\.weight$=q8_0
73
+ ^blk\.3\.ffn_down\.weight$=q8_0
74
+ ^blk\.3\.ffn_gate\.weight$=q6_k
75
+ ^blk\.3\.ffn_up\.weight$=q8_0
76
+ ^blk\.4\.attn_gate\.weight$=q8_0
77
+ ^blk\.4\.attn_qkv\.weight$=q8_0
78
+ ^blk\.4\.ffn_down\.weight$=q8_0
79
+ ^blk\.4\.ffn_gate\.weight$=q6_k
80
+ ^blk\.4\.ffn_up\.weight$=q8_0
81
+ ^blk\.4\.ssm_out\.weight$=q8_0
82
+ ^blk\.5\.attn_gate\.weight$=q8_0
83
+ ^blk\.5\.attn_qkv\.weight$=q8_0
84
+ ^blk\.5\.ffn_down\.weight$=q8_0
85
+ ^blk\.5\.ffn_gate\.weight$=q6_k
86
+ ^blk\.5\.ffn_up\.weight$=q8_0
87
+ ^blk\.5\.ssm_out\.weight$=q8_0
88
+ ^blk\.6\.attn_gate\.weight$=q8_0
89
+ ^blk\.6\.attn_qkv\.weight$=q8_0
90
+ ^blk\.6\.ffn_down\.weight$=q8_0
91
+ ^blk\.6\.ffn_gate\.weight$=q6_k
92
+ ^blk\.6\.ffn_up\.weight$=q8_0
93
+ ^blk\.6\.ssm_out\.weight$=q8_0
94
+ ^blk\.7\.attn_k\.weight$=q8_0
95
+ ^blk\.7\.attn_output\.weight$=q8_0
96
+ ^blk\.7\.attn_q\.weight$=q8_0
97
+ ^blk\.7\.attn_v\.weight$=q8_0
98
+ ^blk\.7\.ffn_down\.weight$=q8_0
99
+ ^blk\.7\.ffn_gate\.weight$=q6_k
100
+ ^blk\.7\.ffn_up\.weight$=q8_0
101
+ ^blk\.8\.attn_gate\.weight$=q6_k
102
+ ^blk\.8\.attn_qkv\.weight$=q8_0
103
+ ^blk\.8\.ffn_down\.weight$=q6_k
104
+ ^blk\.8\.ffn_gate\.weight$=q6_k
105
+ ^blk\.8\.ffn_up\.weight$=q6_k
106
+ ^blk\.8\.ssm_out\.weight$=q8_0
107
+ ^blk\.9\.attn_gate\.weight$=q6_k
108
+ ^blk\.9\.attn_qkv\.weight$=q8_0
109
+ ^blk\.9\.ffn_down\.weight$=q6_k
110
+ ^blk\.9\.ffn_gate\.weight$=q6_k
111
+ ^blk\.9\.ffn_up\.weight$=q6_k
112
+ ^blk\.9\.ssm_out\.weight$=q8_0
113
+ ^blk\.10\.attn_gate\.weight$=q6_k
114
+ ^blk\.10\.attn_qkv\.weight$=q8_0
115
+ ^blk\.10\.ffn_down\.weight$=q6_k
116
+ ^blk\.10\.ffn_gate\.weight$=q6_k
117
+ ^blk\.10\.ffn_up\.weight$=q6_k
118
+ ^blk\.10\.ssm_out\.weight$=q8_0
119
+ ^blk\.11\.attn_k\.weight$=q8_0
120
+ ^blk\.11\.attn_output\.weight$=q8_0
121
+ ^blk\.11\.attn_q\.weight$=q6_k
122
+ ^blk\.11\.attn_v\.weight$=q8_0
123
+ ^blk\.11\.ffn_down\.weight$=q6_k
124
+ ^blk\.11\.ffn_gate\.weight$=q6_k
125
+ ^blk\.11\.ffn_up\.weight$=q6_k
126
+ ^blk\.12\.attn_gate\.weight$=q6_k
127
+ ^blk\.12\.attn_qkv\.weight$=q8_0
128
+ ^blk\.12\.ffn_down\.weight$=q6_k
129
+ ^blk\.12\.ffn_gate\.weight$=q6_k
130
+ ^blk\.12\.ffn_up\.weight$=q6_k
131
+ ^blk\.12\.ssm_out\.weight$=q8_0
132
+ ^blk\.13\.attn_gate\.weight$=q6_k
133
+ ^blk\.13\.attn_qkv\.weight$=q8_0
134
+ ^blk\.13\.ffn_down\.weight$=q6_k
135
+ ^blk\.13\.ffn_gate\.weight$=q6_k
136
+ ^blk\.13\.ffn_up\.weight$=q6_k
137
+ ^blk\.13\.ssm_out\.weight$=q8_0
138
+ ^blk\.14\.attn_gate\.weight$=q6_k
139
+ ^blk\.14\.attn_qkv\.weight$=q8_0
140
+ ^blk\.14\.ffn_down\.weight$=q6_k
141
+ ^blk\.14\.ffn_gate\.weight$=q6_k
142
+ ^blk\.14\.ffn_up\.weight$=q6_k
143
+ ^blk\.14\.ssm_out\.weight$=q8_0
144
+ ^blk\.15\.attn_k\.weight$=q8_0
145
+ ^blk\.15\.attn_output\.weight$=q8_0
146
+ ^blk\.15\.attn_q\.weight$=q6_k
147
+ ^blk\.15\.attn_v\.weight$=q8_0
148
+ ^blk\.15\.ffn_down\.weight$=q6_k
149
+ ^blk\.15\.ffn_gate\.weight$=q6_k
150
+ ^blk\.15\.ffn_up\.weight$=q6_k
151
+ ^blk\.16\.attn_gate\.weight$=q6_k
152
+ ^blk\.16\.attn_qkv\.weight$=q8_0
153
+ ^blk\.16\.ffn_down\.weight$=q6_k
154
+ ^blk\.16\.ffn_gate\.weight$=q6_k
155
+ ^blk\.16\.ffn_up\.weight$=q6_k
156
+ ^blk\.16\.ssm_out\.weight$=q8_0
157
+ ^blk\.17\.attn_gate\.weight$=q6_k
158
+ ^blk\.17\.attn_qkv\.weight$=q8_0
159
+ ^blk\.17\.ffn_down\.weight$=q6_k
160
+ ^blk\.17\.ffn_gate\.weight$=q6_k
161
+ ^blk\.17\.ffn_up\.weight$=q6_k
162
+ ^blk\.17\.ssm_out\.weight$=q8_0
163
+ ^blk\.18\.attn_gate\.weight$=q6_k
164
+ ^blk\.18\.attn_qkv\.weight$=q8_0
165
+ ^blk\.18\.ffn_down\.weight$=q6_k
166
+ ^blk\.18\.ffn_gate\.weight$=q6_k
167
+ ^blk\.18\.ffn_up\.weight$=q6_k
168
+ ^blk\.18\.ssm_out\.weight$=q8_0
169
+ ^blk\.19\.attn_k\.weight$=q8_0
170
+ ^blk\.19\.attn_output\.weight$=q8_0
171
+ ^blk\.19\.attn_q\.weight$=q6_k
172
+ ^blk\.19\.attn_v\.weight$=q8_0
173
+ ^blk\.19\.ffn_down\.weight$=q6_k
174
+ ^blk\.19\.ffn_gate\.weight$=q6_k
175
+ ^blk\.19\.ffn_up\.weight$=q6_k
176
+ ^blk\.20\.attn_gate\.weight$=q6_k
177
+ ^blk\.20\.attn_qkv\.weight$=q8_0
178
+ ^blk\.20\.ffn_down\.weight$=q6_k
179
+ ^blk\.20\.ffn_gate\.weight$=q6_k
180
+ ^blk\.20\.ffn_up\.weight$=q6_k
181
+ ^blk\.20\.ssm_out\.weight$=q8_0
182
+ ^blk\.21\.attn_gate\.weight$=q6_k
183
+ ^blk\.21\.attn_qkv\.weight$=q8_0
184
+ ^blk\.21\.ffn_down\.weight$=q6_k
185
+ ^blk\.21\.ffn_gate\.weight$=q6_k
186
+ ^blk\.21\.ffn_up\.weight$=q6_k
187
+ ^blk\.21\.ssm_out\.weight$=q8_0
188
+ ^blk\.22\.attn_gate\.weight$=q6_k
189
+ ^blk\.22\.attn_qkv\.weight$=q8_0
190
+ ^blk\.22\.ffn_down\.weight$=q6_k
191
+ ^blk\.22\.ffn_gate\.weight$=q6_k
192
+ ^blk\.22\.ffn_up\.weight$=q6_k
193
+ ^blk\.22\.ssm_out\.weight$=q8_0
194
+ ^blk\.23\.attn_k\.weight$=q8_0
195
+ ^blk\.23\.attn_output\.weight$=q8_0
196
+ ^blk\.23\.attn_q\.weight$=q6_k
197
+ ^blk\.23\.attn_v\.weight$=q8_0
198
+ ^blk\.23\.ffn_down\.weight$=q6_k
199
+ ^blk\.23\.ffn_gate\.weight$=q6_k
200
+ ^blk\.23\.ffn_up\.weight$=q6_k
201
+ ^blk\.24\.attn_gate\.weight$=q8_0
202
+ ^blk\.24\.attn_qkv\.weight$=q6_k
203
+ ^blk\.24\.ffn_down\.weight$=q8_0
204
+ ^blk\.24\.ffn_gate\.weight$=q6_k
205
+ ^blk\.24\.ffn_up\.weight$=q8_0
206
+ ^blk\.24\.ssm_out\.weight$=q8_0
207
+ ^blk\.25\.attn_gate\.weight$=q8_0
208
+ ^blk\.25\.attn_qkv\.weight$=q6_k
209
+ ^blk\.25\.ffn_down\.weight$=q8_0
210
+ ^blk\.25\.ffn_gate\.weight$=q6_k
211
+ ^blk\.25\.ffn_up\.weight$=q8_0
212
+ ^blk\.25\.ssm_out\.weight$=q8_0
213
+ ^blk\.26\.attn_gate\.weight$=q8_0
214
+ ^blk\.26\.attn_qkv\.weight$=q6_k
215
+ ^blk\.26\.ffn_down\.weight$=q8_0
216
+ ^blk\.26\.ffn_gate\.weight$=q6_k
217
+ ^blk\.26\.ffn_up\.weight$=q6_k
218
+ ^blk\.26\.ssm_out\.weight$=q8_0
219
+ ^blk\.27\.attn_k\.weight$=q8_0
220
+ ^blk\.27\.attn_output\.weight$=q8_0
221
+ ^blk\.27\.attn_q\.weight$=q8_0
222
+ ^blk\.27\.attn_v\.weight$=q8_0
223
+ ^blk\.27\.ffn_down\.weight$=q8_0
224
+ ^blk\.27\.ffn_gate\.weight$=q6_k
225
+ ^blk\.27\.ffn_up\.weight$=q8_0
226
+ ^blk\.28\.attn_gate\.weight$=q8_0
227
+ ^blk\.28\.attn_qkv\.weight$=q6_k
228
+ ^blk\.28\.ffn_down\.weight$=q8_0
229
+ ^blk\.28\.ffn_gate\.weight$=q6_k
230
+ ^blk\.28\.ffn_up\.weight$=q6_k
231
+ ^blk\.28\.ssm_out\.weight$=q8_0
232
+ ^blk\.29\.attn_gate\.weight$=q8_0
233
+ ^blk\.29\.attn_qkv\.weight$=q6_k
234
+ ^blk\.29\.ffn_down\.weight$=q8_0
235
+ ^blk\.29\.ffn_gate\.weight$=q6_k
236
+ ^blk\.29\.ffn_up\.weight$=q6_k
237
+ ^blk\.29\.ssm_out\.weight$=q8_0
238
+ ^blk\.30\.attn_gate\.weight$=q8_0
239
+ ^blk\.30\.attn_qkv\.weight$=q6_k
240
+ ^blk\.30\.ffn_down\.weight$=q8_0
241
+ ^blk\.30\.ffn_gate\.weight$=q6_k
242
+ ^blk\.30\.ffn_up\.weight$=q6_k
243
+ ^blk\.30\.ssm_out\.weight$=q8_0
244
+ ^blk\.31\.attn_k\.weight$=q8_0
245
+ ^blk\.31\.attn_output\.weight$=q8_0
246
+ ^blk\.31\.attn_q\.weight$=q8_0
247
+ ^blk\.31\.attn_v\.weight$=q8_0
248
+ ^blk\.31\.ffn_down\.weight$=q8_0
249
+ ^blk\.31\.ffn_gate\.weight$=q8_0
250
+ ^blk\.31\.ffn_up\.weight$=q8_0
layouts/LensVLM-9B-Q6_K_S.layout.json ADDED
@@ -0,0 +1,191 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-bf16.gguf",
3
+ "imatrix": "/media/m2teamgroup/models/LensVLM-9B-GGUF/LensVLM-9B-imatrix.gguf",
4
+ "ftype": "Q6_K_S",
5
+ "base_type": "q6_k",
6
+ "family": "k",
7
+ "recipe": "ident",
8
+ "variant": "shr0.90",
9
+ "auto_crush": false,
10
+ "rung_mode": "share",
11
+ "tier": "S",
12
+ "rung_share": 0.9,
13
+ "crush_pass_bytes": null,
14
+ "embd_shares": {
15
+ "output.weight": {
16
+ "share": 0.1085,
17
+ "type": "q6_k"
18
+ },
19
+ "token_embd.weight": {
20
+ "share": 0.1085,
21
+ "type": "q6_k"
22
+ }
23
+ },
24
+ "arch": "hybrid",
25
+ "tied": false,
26
+ "n_layer": 32,
27
+ "body_rule": true,
28
+ "pins": {
29
+ "^blk\\.0\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
30
+ "^blk\\.0\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
31
+ "^blk\\.1\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
32
+ "^blk\\.1\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
33
+ "^blk\\.2\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
34
+ "^blk\\.2\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
35
+ "^blk\\.4\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
36
+ "^blk\\.4\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
37
+ "^blk\\.5\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
38
+ "^blk\\.5\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
39
+ "^blk\\.6\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
40
+ "^blk\\.6\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
41
+ "^blk\\.8\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
42
+ "^blk\\.8\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
43
+ "^blk\\.9\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
44
+ "^blk\\.9\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
45
+ "^blk\\.10\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
46
+ "^blk\\.10\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
47
+ "^blk\\.12\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
48
+ "^blk\\.12\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
49
+ "^blk\\.13\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
50
+ "^blk\\.13\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
51
+ "^blk\\.14\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
52
+ "^blk\\.14\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
53
+ "^blk\\.16\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
54
+ "^blk\\.16\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
55
+ "^blk\\.17\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
56
+ "^blk\\.17\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
57
+ "^blk\\.18\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
58
+ "^blk\\.18\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
59
+ "^blk\\.20\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
60
+ "^blk\\.20\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
61
+ "^blk\\.21\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
62
+ "^blk\\.21\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
63
+ "^blk\\.22\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
64
+ "^blk\\.22\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
65
+ "^blk\\.24\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
66
+ "^blk\\.24\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
67
+ "^blk\\.25\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
68
+ "^blk\\.25\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
69
+ "^blk\\.26\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
70
+ "^blk\\.26\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
71
+ "^blk\\.28\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
72
+ "^blk\\.28\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
73
+ "^blk\\.29\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
74
+ "^blk\\.29\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32",
75
+ "^blk\\.30\\.ssm_alpha\\.weight$": "tiny: 131,072 params, kept at f32",
76
+ "^blk\\.30\\.ssm_beta\\.weight$": "tiny: 131,072 params, kept at f32"
77
+ },
78
+ "shape_pins": {},
79
+ "fixed_pins": {
80
+ "blk.0.ssm_alpha.weight": "f32",
81
+ "blk.0.ssm_beta.weight": "f32",
82
+ "blk.1.ssm_alpha.weight": "f32",
83
+ "blk.1.ssm_beta.weight": "f32",
84
+ "blk.2.ssm_alpha.weight": "f32",
85
+ "blk.2.ssm_beta.weight": "f32",
86
+ "blk.4.ssm_alpha.weight": "f32",
87
+ "blk.4.ssm_beta.weight": "f32",
88
+ "blk.5.ssm_alpha.weight": "f32",
89
+ "blk.5.ssm_beta.weight": "f32",
90
+ "blk.6.ssm_alpha.weight": "f32",
91
+ "blk.6.ssm_beta.weight": "f32",
92
+ "blk.8.ssm_alpha.weight": "f32",
93
+ "blk.8.ssm_beta.weight": "f32",
94
+ "blk.9.ssm_alpha.weight": "f32",
95
+ "blk.9.ssm_beta.weight": "f32",
96
+ "blk.10.ssm_alpha.weight": "f32",
97
+ "blk.10.ssm_beta.weight": "f32",
98
+ "blk.12.ssm_alpha.weight": "f32",
99
+ "blk.12.ssm_beta.weight": "f32",
100
+ "blk.13.ssm_alpha.weight": "f32",
101
+ "blk.13.ssm_beta.weight": "f32",
102
+ "blk.14.ssm_alpha.weight": "f32",
103
+ "blk.14.ssm_beta.weight": "f32",
104
+ "blk.16.ssm_alpha.weight": "f32",
105
+ "blk.16.ssm_beta.weight": "f32",
106
+ "blk.17.ssm_alpha.weight": "f32",
107
+ "blk.17.ssm_beta.weight": "f32",
108
+ "blk.18.ssm_alpha.weight": "f32",
109
+ "blk.18.ssm_beta.weight": "f32",
110
+ "blk.20.ssm_alpha.weight": "f32",
111
+ "blk.20.ssm_beta.weight": "f32",
112
+ "blk.21.ssm_alpha.weight": "f32",
113
+ "blk.21.ssm_beta.weight": "f32",
114
+ "blk.22.ssm_alpha.weight": "f32",
115
+ "blk.22.ssm_beta.weight": "f32",
116
+ "blk.24.ssm_alpha.weight": "f32",
117
+ "blk.24.ssm_beta.weight": "f32",
118
+ "blk.25.ssm_alpha.weight": "f32",
119
+ "blk.25.ssm_beta.weight": "f32",
120
+ "blk.26.ssm_alpha.weight": "f32",
121
+ "blk.26.ssm_beta.weight": "f32",
122
+ "blk.28.ssm_alpha.weight": "f32",
123
+ "blk.28.ssm_beta.weight": "f32",
124
+ "blk.29.ssm_alpha.weight": "f32",
125
+ "blk.29.ssm_beta.weight": "f32",
126
+ "blk.30.ssm_alpha.weight": "f32",
127
+ "blk.30.ssm_beta.weight": "f32"
128
+ },
129
+ "embeddings": {
130
+ "output.weight": "q6_k",
131
+ "token_embd.weight": "q6_k"
132
+ },
133
+ "bytes": {
134
+ "heuristic_body+embd": 7659880448,
135
+ "map_body+embd": 7468908544,
136
+ "ratio": 0.9751,
137
+ "note": "quantizable, unpinned tensors only; pinned/unquantized tensors are identical in both"
138
+ },
139
+ "body_base_share": 0.902,
140
+ "body_bpw": 6.713,
141
+ "body_bpw_heuristic": 6.934,
142
+ "total_params": 8953803264,
143
+ "file_bytes_pred": 7498319872,
144
+ "file_bpw_pred": 6.7,
145
+ "histogram": {
146
+ "attn_gate": {
147
+ "q6_k": 23,
148
+ "q8_0": 1
149
+ },
150
+ "attn_k": {
151
+ "q8_0": 8
152
+ },
153
+ "attn_output": {
154
+ "q8_0": 8
155
+ },
156
+ "attn_q": {
157
+ "q6_k": 8
158
+ },
159
+ "attn_qkv": {
160
+ "q6_k": 23,
161
+ "q8_0": 1
162
+ },
163
+ "attn_v": {
164
+ "q8_0": 8
165
+ },
166
+ "ffn_down": {
167
+ "q6_k": 31,
168
+ "q8_0": 1
169
+ },
170
+ "ffn_gate": {
171
+ "q6_k": 32
172
+ },
173
+ "ffn_up": {
174
+ "q6_k": 31,
175
+ "q8_0": 1
176
+ },
177
+ "ssm_out": {
178
+ "q6_k": 13,
179
+ "q8_0": 11
180
+ }
181
+ },
182
+ "prior": "layout/prior.json",
183
+ "pin_rule": "floor",
184
+ "embd_cap": "heur",
185
+ "generator": "auto_quant_v2.layout",
186
+ "generator_key": "93ecba14028d5e4b",
187
+ "quant_type": "Q6_K_S",
188
+ "quantize_ftype": "Q6_K",
189
+ "llama_cpp_version": "b11081",
190
+ "commit": "9bd4ca72e6b98d9c35d19c58991bf5d41945daa9"
191
+ }
layouts/LensVLM-9B-Q6_K_S.tensor-types.txt ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ^blk\.0\.ssm_alpha\.weight$=f32
2
+ ^blk\.0\.ssm_beta\.weight$=f32
3
+ ^blk\.1\.ssm_alpha\.weight$=f32
4
+ ^blk\.1\.ssm_beta\.weight$=f32
5
+ ^blk\.2\.ssm_alpha\.weight$=f32
6
+ ^blk\.2\.ssm_beta\.weight$=f32
7
+ ^blk\.4\.ssm_alpha\.weight$=f32
8
+ ^blk\.4\.ssm_beta\.weight$=f32
9
+ ^blk\.5\.ssm_alpha\.weight$=f32
10
+ ^blk\.5\.ssm_beta\.weight$=f32
11
+ ^blk\.6\.ssm_alpha\.weight$=f32
12
+ ^blk\.6\.ssm_beta\.weight$=f32
13
+ ^blk\.8\.ssm_alpha\.weight$=f32
14
+ ^blk\.8\.ssm_beta\.weight$=f32
15
+ ^blk\.9\.ssm_alpha\.weight$=f32
16
+ ^blk\.9\.ssm_beta\.weight$=f32
17
+ ^blk\.10\.ssm_alpha\.weight$=f32
18
+ ^blk\.10\.ssm_beta\.weight$=f32
19
+ ^blk\.12\.ssm_alpha\.weight$=f32
20
+ ^blk\.12\.ssm_beta\.weight$=f32
21
+ ^blk\.13\.ssm_alpha\.weight$=f32
22
+ ^blk\.13\.ssm_beta\.weight$=f32
23
+ ^blk\.14\.ssm_alpha\.weight$=f32
24
+ ^blk\.14\.ssm_beta\.weight$=f32
25
+ ^blk\.16\.ssm_alpha\.weight$=f32
26
+ ^blk\.16\.ssm_beta\.weight$=f32
27
+ ^blk\.17\.ssm_alpha\.weight$=f32
28
+ ^blk\.17\.ssm_beta\.weight$=f32
29
+ ^blk\.18\.ssm_alpha\.weight$=f32
30
+ ^blk\.18\.ssm_beta\.weight$=f32
31
+ ^blk\.20\.ssm_alpha\.weight$=f32
32
+ ^blk\.20\.ssm_beta\.weight$=f32
33
+ ^blk\.21\.ssm_alpha\.weight$=f32
34
+ ^blk\.21\.ssm_beta\.weight$=f32
35
+ ^blk\.22\.ssm_alpha\.weight$=f32
36
+ ^blk\.22\.ssm_beta\.weight$=f32
37
+ ^blk\.24\.ssm_alpha\.weight$=f32
38
+ ^blk\.24\.ssm_beta\.weight$=f32
39
+ ^blk\.25\.ssm_alpha\.weight$=f32
40
+ ^blk\.25\.ssm_beta\.weight$=f32
41
+ ^blk\.26\.ssm_alpha\.weight$=f32
42
+ ^blk\.26\.ssm_beta\.weight$=f32
43
+ ^blk\.28\.ssm_alpha\.weight$=f32
44
+ ^blk\.28\.ssm_beta\.weight$=f32
45
+ ^blk\.29\.ssm_alpha\.weight$=f32
46
+ ^blk\.29\.ssm_beta\.weight$=f32
47
+ ^blk\.30\.ssm_alpha\.weight$=f32
48
+ ^blk\.30\.ssm_beta\.weight$=f32
49
+ ^output\.weight$=q6_k
50
+ ^token_embd\.weight$=q6_k
51
+ ^blk\.0\.attn_gate\.weight$=q8_0
52
+ ^blk\.0\.attn_qkv\.weight$=q8_0
53
+ ^blk\.0\.ffn_down\.weight$=q6_k
54
+ ^blk\.0\.ffn_gate\.weight$=q6_k
55
+ ^blk\.0\.ffn_up\.weight$=q6_k
56
+ ^blk\.0\.ssm_out\.weight$=q8_0
57
+ ^blk\.1\.attn_gate\.weight$=q6_k
58
+ ^blk\.1\.attn_qkv\.weight$=q6_k
59
+ ^blk\.1\.ffn_down\.weight$=q6_k
60
+ ^blk\.1\.ffn_gate\.weight$=q6_k
61
+ ^blk\.1\.ffn_up\.weight$=q6_k
62
+ ^blk\.1\.ssm_out\.weight$=q8_0
63
+ ^blk\.2\.attn_gate\.weight$=q6_k
64
+ ^blk\.2\.attn_qkv\.weight$=q6_k
65
+ ^blk\.2\.ffn_down\.weight$=q6_k
66
+ ^blk\.2\.ffn_gate\.weight$=q6_k
67
+ ^blk\.2\.ffn_up\.weight$=q6_k
68
+ ^blk\.2\.ssm_out\.weight$=q8_0
69
+ ^blk\.3\.attn_k\.weight$=q8_0
70
+ ^blk\.3\.attn_output\.weight$=q8_0
71
+ ^blk\.3\.attn_q\.weight$=q6_k
72
+ ^blk\.3\.attn_v\.weight$=q8_0
73
+ ^blk\.3\.ffn_down\.weight$=q6_k
74
+ ^blk\.3\.ffn_gate\.weight$=q6_k
75
+ ^blk\.3\.ffn_up\.weight$=q6_k
76
+ ^blk\.4\.attn_gate\.weight$=q6_k
77
+ ^blk\.4\.attn_qkv\.weight$=q6_k
78
+ ^blk\.4\.ffn_down\.weight$=q6_k
79
+ ^blk\.4\.ffn_gate\.weight$=q6_k
80
+ ^blk\.4\.ffn_up\.weight$=q6_k
81
+ ^blk\.4\.ssm_out\.weight$=q8_0
82
+ ^blk\.5\.attn_gate\.weight$=q6_k
83
+ ^blk\.5\.attn_qkv\.weight$=q6_k
84
+ ^blk\.5\.ffn_down\.weight$=q6_k
85
+ ^blk\.5\.ffn_gate\.weight$=q6_k
86
+ ^blk\.5\.ffn_up\.weight$=q6_k
87
+ ^blk\.5\.ssm_out\.weight$=q8_0
88
+ ^blk\.6\.attn_gate\.weight$=q6_k
89
+ ^blk\.6\.attn_qkv\.weight$=q6_k
90
+ ^blk\.6\.ffn_down\.weight$=q6_k
91
+ ^blk\.6\.ffn_gate\.weight$=q6_k
92
+ ^blk\.6\.ffn_up\.weight$=q6_k
93
+ ^blk\.6\.ssm_out\.weight$=q8_0
94
+ ^blk\.7\.attn_k\.weight$=q8_0
95
+ ^blk\.7\.attn_output\.weight$=q8_0
96
+ ^blk\.7\.attn_q\.weight$=q6_k
97
+ ^blk\.7\.attn_v\.weight$=q8_0
98
+ ^blk\.7\.ffn_down\.weight$=q6_k
99
+ ^blk\.7\.ffn_gate\.weight$=q6_k
100
+ ^blk\.7\.ffn_up\.weight$=q6_k
101
+ ^blk\.8\.attn_gate\.weight$=q6_k
102
+ ^blk\.8\.attn_qkv\.weight$=q6_k
103
+ ^blk\.8\.ffn_down\.weight$=q6_k
104
+ ^blk\.8\.ffn_gate\.weight$=q6_k
105
+ ^blk\.8\.ffn_up\.weight$=q6_k
106
+ ^blk\.8\.ssm_out\.weight$=q6_k
107
+ ^blk\.9\.attn_gate\.weight$=q6_k
108
+ ^blk\.9\.attn_qkv\.weight$=q6_k
109
+ ^blk\.9\.ffn_down\.weight$=q6_k
110
+ ^blk\.9\.ffn_gate\.weight$=q6_k
111
+ ^blk\.9\.ffn_up\.weight$=q6_k
112
+ ^blk\.9\.ssm_out\.weight$=q6_k
113
+ ^blk\.10\.attn_gate\.weight$=q6_k
114
+ ^blk\.10\.attn_qkv\.weight$=q6_k
115
+ ^blk\.10\.ffn_down\.weight$=q6_k
116
+ ^blk\.10\.ffn_gate\.weight$=q6_k
117
+ ^blk\.10\.ffn_up\.weight$=q6_k
118
+ ^blk\.10\.ssm_out\.weight$=q6_k
119
+ ^blk\.11\.attn_k\.weight$=q8_0
120
+ ^blk\.11\.attn_output\.weight$=q8_0
121
+ ^blk\.11\.attn_q\.weight$=q6_k
122
+ ^blk\.11\.attn_v\.weight$=q8_0
123
+ ^blk\.11\.ffn_down\.weight$=q6_k
124
+ ^blk\.11\.ffn_gate\.weight$=q6_k
125
+ ^blk\.11\.ffn_up\.weight$=q6_k
126
+ ^blk\.12\.attn_gate\.weight$=q6_k
127
+ ^blk\.12\.attn_qkv\.weight$=q6_k
128
+ ^blk\.12\.ffn_down\.weight$=q6_k
129
+ ^blk\.12\.ffn_gate\.weight$=q6_k
130
+ ^blk\.12\.ffn_up\.weight$=q6_k
131
+ ^blk\.12\.ssm_out\.weight$=q6_k
132
+ ^blk\.13\.attn_gate\.weight$=q6_k
133
+ ^blk\.13\.attn_qkv\.weight$=q6_k
134
+ ^blk\.13\.ffn_down\.weight$=q6_k
135
+ ^blk\.13\.ffn_gate\.weight$=q6_k
136
+ ^blk\.13\.ffn_up\.weight$=q6_k
137
+ ^blk\.13\.ssm_out\.weight$=q6_k
138
+ ^blk\.14\.attn_gate\.weight$=q6_k
139
+ ^blk\.14\.attn_qkv\.weight$=q6_k
140
+ ^blk\.14\.ffn_down\.weight$=q6_k
141
+ ^blk\.14\.ffn_gate\.weight$=q6_k
142
+ ^blk\.14\.ffn_up\.weight$=q6_k
143
+ ^blk\.14\.ssm_out\.weight$=q6_k
144
+ ^blk\.15\.attn_k\.weight$=q8_0
145
+ ^blk\.15\.attn_output\.weight$=q8_0
146
+ ^blk\.15\.attn_q\.weight$=q6_k
147
+ ^blk\.15\.attn_v\.weight$=q8_0
148
+ ^blk\.15\.ffn_down\.weight$=q6_k
149
+ ^blk\.15\.ffn_gate\.weight$=q6_k
150
+ ^blk\.15\.ffn_up\.weight$=q6_k
151
+ ^blk\.16\.attn_gate\.weight$=q6_k
152
+ ^blk\.16\.attn_qkv\.weight$=q6_k
153
+ ^blk\.16\.ffn_down\.weight$=q6_k
154
+ ^blk\.16\.ffn_gate\.weight$=q6_k
155
+ ^blk\.16\.ffn_up\.weight$=q6_k
156
+ ^blk\.16\.ssm_out\.weight$=q6_k
157
+ ^blk\.17\.attn_gate\.weight$=q6_k
158
+ ^blk\.17\.attn_qkv\.weight$=q6_k
159
+ ^blk\.17\.ffn_down\.weight$=q6_k
160
+ ^blk\.17\.ffn_gate\.weight$=q6_k
161
+ ^blk\.17\.ffn_up\.weight$=q6_k
162
+ ^blk\.17\.ssm_out\.weight$=q6_k
163
+ ^blk\.18\.attn_gate\.weight$=q6_k
164
+ ^blk\.18\.attn_qkv\.weight$=q6_k
165
+ ^blk\.18\.ffn_down\.weight$=q6_k
166
+ ^blk\.18\.ffn_gate\.weight$=q6_k
167
+ ^blk\.18\.ffn_up\.weight$=q6_k
168
+ ^blk\.18\.ssm_out\.weight$=q6_k
169
+ ^blk\.19\.attn_k\.weight$=q8_0
170
+ ^blk\.19\.attn_output\.weight$=q8_0
171
+ ^blk\.19\.attn_q\.weight$=q6_k
172
+ ^blk\.19\.attn_v\.weight$=q8_0
173
+ ^blk\.19\.ffn_down\.weight$=q6_k
174
+ ^blk\.19\.ffn_gate\.weight$=q6_k
175
+ ^blk\.19\.ffn_up\.weight$=q6_k
176
+ ^blk\.20\.attn_gate\.weight$=q6_k
177
+ ^blk\.20\.attn_qkv\.weight$=q6_k
178
+ ^blk\.20\.ffn_down\.weight$=q6_k
179
+ ^blk\.20\.ffn_gate\.weight$=q6_k
180
+ ^blk\.20\.ffn_up\.weight$=q6_k
181
+ ^blk\.20\.ssm_out\.weight$=q6_k
182
+ ^blk\.21\.attn_gate\.weight$=q6_k
183
+ ^blk\.21\.attn_qkv\.weight$=q6_k
184
+ ^blk\.21\.ffn_down\.weight$=q6_k
185
+ ^blk\.21\.ffn_gate\.weight$=q6_k
186
+ ^blk\.21\.ffn_up\.weight$=q6_k
187
+ ^blk\.21\.ssm_out\.weight$=q6_k
188
+ ^blk\.22\.attn_gate\.weight$=q6_k
189
+ ^blk\.22\.attn_qkv\.weight$=q6_k
190
+ ^blk\.22\.ffn_down\.weight$=q6_k
191
+ ^blk\.22\.ffn_gate\.weight$=q6_k
192
+ ^blk\.22\.ffn_up\.weight$=q6_k
193
+ ^blk\.22\.ssm_out\.weight$=q6_k
194
+ ^blk\.23\.attn_k\.weight$=q8_0
195
+ ^blk\.23\.attn_output\.weight$=q8_0
196
+ ^blk\.23\.attn_q\.weight$=q6_k
197
+ ^blk\.23\.attn_v\.weight$=q8_0
198
+ ^blk\.23\.ffn_down\.weight$=q6_k
199
+ ^blk\.23\.ffn_gate\.weight$=q6_k
200
+ ^blk\.23\.ffn_up\.weight$=q6_k
201
+ ^blk\.24\.attn_gate\.weight$=q6_k
202
+ ^blk\.24\.attn_qkv\.weight$=q6_k
203
+ ^blk\.24\.ffn_down\.weight$=q6_k
204
+ ^blk\.24\.ffn_gate\.weight$=q6_k
205
+ ^blk\.24\.ffn_up\.weight$=q6_k
206
+ ^blk\.24\.ssm_out\.weight$=q8_0
207
+ ^blk\.25\.attn_gate\.weight$=q6_k
208
+ ^blk\.25\.attn_qkv\.weight$=q6_k
209
+ ^blk\.25\.ffn_down\.weight$=q6_k
210
+ ^blk\.25\.ffn_gate\.weight$=q6_k
211
+ ^blk\.25\.ffn_up\.weight$=q6_k
212
+ ^blk\.25\.ssm_out\.weight$=q8_0
213
+ ^blk\.26\.attn_gate\.weight$=q6_k
214
+ ^blk\.26\.attn_qkv\.weight$=q6_k
215
+ ^blk\.26\.ffn_down\.weight$=q6_k
216
+ ^blk\.26\.ffn_gate\.weight$=q6_k
217
+ ^blk\.26\.ffn_up\.weight$=q6_k
218
+ ^blk\.26\.ssm_out\.weight$=q8_0
219
+ ^blk\.27\.attn_k\.weight$=q8_0
220
+ ^blk\.27\.attn_output\.weight$=q8_0
221
+ ^blk\.27\.attn_q\.weight$=q6_k
222
+ ^blk\.27\.attn_v\.weight$=q8_0
223
+ ^blk\.27\.ffn_down\.weight$=q6_k
224
+ ^blk\.27\.ffn_gate\.weight$=q6_k
225
+ ^blk\.27\.ffn_up\.weight$=q6_k
226
+ ^blk\.28\.attn_gate\.weight$=q6_k
227
+ ^blk\.28\.attn_qkv\.weight$=q6_k
228
+ ^blk\.28\.ffn_down\.weight$=q6_k
229
+ ^blk\.28\.ffn_gate\.weight$=q6_k
230
+ ^blk\.28\.ffn_up\.weight$=q6_k
231
+ ^blk\.28\.ssm_out\.weight$=q8_0
232
+ ^blk\.29\.attn_gate\.weight$=q6_k
233
+ ^blk\.29\.attn_qkv\.weight$=q6_k
234
+ ^blk\.29\.ffn_down\.weight$=q6_k
235
+ ^blk\.29\.ffn_gate\.weight$=q6_k
236
+ ^blk\.29\.ffn_up\.weight$=q6_k
237
+ ^blk\.29\.ssm_out\.weight$=q8_0
238
+ ^blk\.30\.attn_gate\.weight$=q6_k
239
+ ^blk\.30\.attn_qkv\.weight$=q6_k
240
+ ^blk\.30\.ffn_down\.weight$=q6_k
241
+ ^blk\.30\.ffn_gate\.weight$=q6_k
242
+ ^blk\.30\.ffn_up\.weight$=q6_k
243
+ ^blk\.30\.ssm_out\.weight$=q6_k
244
+ ^blk\.31\.attn_k\.weight$=q8_0
245
+ ^blk\.31\.attn_output\.weight$=q8_0
246
+ ^blk\.31\.attn_q\.weight$=q6_k
247
+ ^blk\.31\.attn_v\.weight$=q8_0
248
+ ^blk\.31\.ffn_down\.weight$=q8_0
249
+ ^blk\.31\.ffn_gate\.weight$=q6_k
250
+ ^blk\.31\.ffn_up\.weight$=q8_0