yujiepan/mistral-small-4-tiny-random

This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from mistralai/Mistral-Small-4-119B-2603.

File path	Size
model.safetensors	11.8MB

Example usage:

import torch
from transformers import AutoProcessor, Mistral3ForConditionalGeneration

# Load model and tokenizer
model_id = "yujiepan/mistral-small-4-tiny-random"
model = Mistral3ForConditionalGeneration.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="bfloat16",
    trust_remote_code=True,
)
processor = AutoProcessor.from_pretrained(model_id)
image_url = "https://static.wikia.nocookie.net/essentialsdocs/images/7/70/Battle.png/revision/latest?cb=20220523172438"
messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": "What is this?",
            },
            {"type": "image_url", "image_url": {"url": image_url}},
        ],
    },
]
inputs = processor.apply_chat_template(
    messages,
    return_tensors="pt",
    tokenize=True,
    return_dict=True,
    reasoning_effort="high",
)
inputs = inputs.to(model.device)

output = model.generate(
    **inputs,
    max_new_tokens=32,
    do_sample=True,
    temperature=0.7,
)[0]
decoded_output = processor.decode(output, skip_special_tokens=False).replace(
    "[IMG]", "I"
)
print(decoded_output)

Codes to create this repo:

Click to expand

import json
from pathlib import Path

import accelerate
import torch
from huggingface_hub import file_exists, hf_hub_download
from transformers import (
    AutoConfig,
    AutoModelForCausalLM,
    AutoProcessor,
    GenerationConfig,
    Mistral3ForConditionalGeneration,
    MistralCommonBackend,
    set_seed,
)

source_model_id = "mistralai/Mistral-Small-4-119B-2603"
save_folder = "/tmp/yujiepan/mistral-small-4-tiny-random"

processor = AutoProcessor.from_pretrained(source_model_id, trust_remote_code=True)
processor.save_pretrained(save_folder)
processor = MistralCommonBackend.from_pretrained(
    source_model_id, trust_remote_code=True
)
processor.save_pretrained(save_folder)

with open(
    hf_hub_download(source_model_id, filename="config.json", repo_type="model"),
    "r",
    encoding="utf-8",
) as f:
    config_json = json.load(f)

config_json["text_config"].update(
    {
        "hidden_size": 8,
        "intermediate_size": 32,
        "moe_intermediate_size": 32,
        "num_hidden_layers": 2,
        "q_lora_rank": 32,
    }
)
# config_json['tie_word_embeddings'] = True
config_json["vision_config"].update(
    {
        "head_dim": 32,
        "hidden_size": 64,
        "intermediate_size": 64,
        "num_attention_heads": 2,
        "num_hidden_layers": 2,
    }
)
del config_json["quantization_config"]
with open(f"{save_folder}/config.json", "w", encoding="utf-8") as f:
    json.dump(config_json, f, indent=2)

config = AutoConfig.from_pretrained(
    save_folder,
    trust_remote_code=True,
)
print(config)
torch.set_default_dtype(torch.bfloat16)
model = Mistral3ForConditionalGeneration(config)
torch.set_default_dtype(torch.float32)
if file_exists(
    filename="generation_config.json", repo_id=source_model_id, repo_type="model"
):
    model.generation_config = GenerationConfig.from_pretrained(
        source_model_id,
        trust_remote_code=True,
    )
    model.generation_config.do_sample = True
    print(model.generation_config)
model = model.cpu()
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        torch.nn.init.normal_(p, 0, 0.2)
        print(name, p.shape)
model.save_pretrained(save_folder)
print(model)

Printing the model:

Click to expand

Mistral3ForConditionalGeneration(
  (model): Mistral3Model(
    (vision_tower): PixtralVisionModel(
      (patch_conv): Conv2d(3, 64, kernel_size=(14, 14), stride=(14, 14), bias=False)
      (ln_pre): PixtralRMSNorm((64,), eps=1e-05)
      (transformer): PixtralTransformer(
        (layers): ModuleList(
          (0-1): 2 x PixtralAttentionLayer(
            (attention_norm): PixtralRMSNorm((64,), eps=1e-05)
            (feed_forward): PixtralMLP(
              (gate_proj): Linear(in_features=64, out_features=64, bias=False)
              (up_proj): Linear(in_features=64, out_features=64, bias=False)
              (down_proj): Linear(in_features=64, out_features=64, bias=False)
              (act_fn): SiLUActivation()
            )
            (attention): PixtralAttention(
              (k_proj): Linear(in_features=64, out_features=64, bias=False)
              (v_proj): Linear(in_features=64, out_features=64, bias=False)
              (q_proj): Linear(in_features=64, out_features=64, bias=False)
              (o_proj): Linear(in_features=64, out_features=64, bias=False)
            )
            (ffn_norm): PixtralRMSNorm((64,), eps=1e-05)
          )
        )
      )
      (patch_positional_embedding): PixtralRotaryEmbedding()
    )
    (multi_modal_projector): Mistral3MultiModalProjector(
      (norm): Mistral3RMSNorm((64,), eps=1e-06)
      (patch_merger): Mistral3PatchMerger(
        (merging_layer): Linear(in_features=256, out_features=64, bias=False)
      )
      (linear_1): Linear(in_features=64, out_features=8, bias=False)
      (act): GELUActivation()
      (linear_2): Linear(in_features=8, out_features=8, bias=False)
    )
    (language_model): Mistral4Model(
      (embed_tokens): Embedding(131072, 8, padding_idx=11)
      (layers): ModuleList(
        (0-1): 2 x Mistral4DecoderLayer(
          (self_attn): Mistral4Attention(
            (q_a_proj): Linear(in_features=8, out_features=32, bias=False)
            (q_a_layernorm): Mistral4RMSNorm((32,), eps=1e-06)
            (q_b_proj): Linear(in_features=32, out_features=4096, bias=False)
            (kv_a_proj_with_mqa): Linear(in_features=8, out_features=320, bias=False)
            (kv_a_layernorm): Mistral4RMSNorm((256,), eps=1e-06)
            (kv_b_proj): Linear(in_features=256, out_features=6144, bias=False)
            (o_proj): Linear(in_features=4096, out_features=8, bias=False)
          )
          (mlp): Mistral4MoE(
            (experts): Mistral4NaiveMoe(
              (act_fn): SiLUActivation()
            )
            (gate): Mistral4TopkRouter()
            (shared_experts): Mistral4MLP(
              (gate_proj): Linear(in_features=8, out_features=32, bias=False)
              (up_proj): Linear(in_features=8, out_features=32, bias=False)
              (down_proj): Linear(in_features=32, out_features=8, bias=False)
              (act_fn): SiLUActivation()
            )
          )
          (input_layernorm): Mistral4RMSNorm((8,), eps=1e-06)
          (post_attention_layernorm): Mistral4RMSNorm((8,), eps=1e-06)
        )
      )
      (norm): Mistral4RMSNorm((8,), eps=1e-06)
      (rotary_emb): Mistral4RotaryEmbedding()
    )
  )
  (lm_head): Linear(in_features=8, out_features=131072, bias=False)
)