config: set Stage 1 micro 32 accum 4 for peak Tensor Core speed

model/configs/train_kaggle_t4.yaml

@@ -14,23 +14,23 @@ curriculum: true
 curriculum_stages:
   - max_step: 50000
     seq_len: 512
-    batch_size: 16
-    grad_accum: 16  # 512 * 16 * 16 = 131,072 tokens/step (On Dual T4: micro 16, accum 8 per GPU!)
+    batch_size: 32
+    grad_accum: 8   # 512 * 32 * 8 = 131,072 tokens/step (On Dual T4: micro 32, accum 4 per GPU!)
   - max_step: 200000
     seq_len: 1024
-    batch_size: 8
-    grad_accum: 16  # 1024 * 8 * 16 = 131,072 tokens/step (On Dual T4: micro 8, accum 8 per GPU!)
+    batch_size: 16
+    grad_accum: 8   # 1024 * 16 * 8 = 131,072 tokens/step (On Dual T4: micro 16, accum 4 per GPU!)
   - max_step: null
     seq_len: 2048
-    batch_size: 4
-    grad_accum: 16  # 2048 * 4 * 16 = 131,072 tokens/step (On Dual T4: micro 4, accum 8 per GPU!)
+    batch_size: 8
+    grad_accum: 8   # 2048 * 8 * 8 = 131,072 tokens/step (On Dual T4: micro 8, accum 4 per GPU!)
 
 seq_len: 512
 train_split: 0.95
 
-# Kaggle Dual T4 (2x 16GB VRAM): micro-batch 16 with SDPA, PagedAdamW8bit, and Grad Checkpointing (~5.8 GB VRAM, 100% Zero-OOM)
-batch_size: 16
-grad_accum: 16
+# Kaggle Dual T4 (2x 16GB VRAM): micro-batch 32 with SDPA, PagedAdamW8bit, and Grad Checkpointing (~7.5 GB VRAM, 100% Zero-OOM)
+batch_size: 32
+grad_accum: 8
 target_global_batch: 256
 auto_tune: false