Restore batch_size=8 and grad_ckpt=True for 100% VRAM safety on Dual T4 (10.5GB)

model/configs/train_kaggle_t4.yaml

@@ -13,23 +13,23 @@ curriculum: true
 curriculum_stages:
   - max_step: 50000
     seq_len: 512
-    batch_size: 16
-    grad_accum: 8  # 512 * 16 * 8 = 65,536 tokens/step (On Dual T4: 4 passes per GPU!)
+    batch_size: 8
+    grad_accum: 16  # 512 * 8 * 16 = 65,536 tokens/step (On Dual T4: 8 passes per GPU!)
   - max_step: 200000
     seq_len: 1024
-    batch_size: 8
-    grad_accum: 8  # 1024 * 8 * 8 = 65,536 tokens/step
+    batch_size: 4
+    grad_accum: 16  # 1024 * 4 * 16 = 65,536 tokens/step
   - max_step: null
     seq_len: 2048
-    batch_size: 4
-    grad_accum: 8  # 2048 * 4 * 8 = 65,536 tokens/step
+    batch_size: 2
+    grad_accum: 16  # 2048 * 2 * 16 = 65,536 tokens/step
 
 seq_len: 512
 train_split: 0.95
 
-# Kaggle Dual T4 (2x 16GB VRAM): micro-batch 16 with SDPA, PagedAdamW8bit (~4.5 GB VRAM / 16 GB)
-batch_size: 16
-grad_accum: 8
+# 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
 auto_tune: false
 
@@ -43,13 +43,13 @@ weight_decay: 0.1
 betas: [0.9, 0.95]
 grad_clip: 1.0
 
-# Automatic FP16 on T4 + GradScaler + Muon (grad_ckpt disabled for 242M on 16GB VRAM for 35% speed boost)
+# Automatic FP16 on T4 + GradScaler + Muon + Grad Checkpointing (VRAM safety)
 amp: auto
 optimizer: muon
 muon_lr: 0.02
 muon_momentum: 0.95
 muon_weight_decay: 0.01
-grad_ckpt: false
+grad_ckpt: true
 
 compile: false
 fp8: false