Set grad_accum=8: 2x step speed, eff_batch=64 (32,768 tokens/step)

model/configs/train_kaggle_t4.yaml

@@ -14,23 +14,23 @@ curriculum_stages:
   - max_step: 50000
     seq_len: 512
     batch_size: 8
-    grad_accum: 16  # 512 * 8 * 16 = 65,536 tokens/step (On Dual T4: 8 passes per GPU!)
+    grad_accum: 8   # 512 * 8 * 8 = 32,768 tokens/step (On Dual T4: 4 passes per GPU!)
   - max_step: 200000
     seq_len: 1024
     batch_size: 4
-    grad_accum: 16  # 1024 * 4 * 16 = 65,536 tokens/step
+    grad_accum: 8   # 1024 * 4 * 8 = 32,768 tokens/step
   - max_step: null
     seq_len: 2048
     batch_size: 2
-    grad_accum: 16  # 2048 * 2 * 16 = 65,536 tokens/step
+    grad_accum: 8   # 2048 * 2 * 8 = 32,768 tokens/step
 
 seq_len: 512
 train_split: 0.95
 
 # Kaggle Dual T4 (2x 16GB VRAM): micro-batch 8 with SDPA, PagedAdamW8bit, and Grad Checkpointing (~10.5 GB VRAM)
 batch_size: 8
-grad_accum: 16
-target_global_batch: 128
+grad_accum: 8
+target_global_batch: 64
 auto_tune: false
 
 # FULL PRETRAINING: Dynamic Auto-Steps from 125B+ Tokens (2,109 Parquet files on Hub)