Optimize train_kaggle_t4.yaml: micro-batch 8 / accum 16 (8 passes/GPU on Dual T4), log_every 1

model/configs/train_kaggle_t4.yaml

@@ -13,23 +13,23 @@ curriculum: true
 curriculum_stages:
   - max_step: 50000
     seq_len: 512
-    batch_size: 1
-    grad_accum: 128  # 512 * 1 * 128 = 65,536 tokens/step per GPU (x2 GPUs = 131,072 tok/step)
+    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: 1
-    grad_accum: 64   # 1024 * 1 * 64 = 65,536 tokens/step per GPU (x2 GPUs = 131,072 tok/step)
+    batch_size: 4
+    grad_accum: 16  # 1024 * 4 * 16 = 65,536 tokens/step
   - max_step: null
     seq_len: 2048
-    batch_size: 1
-    grad_accum: 32   # 2048 * 1 * 32 = 65,536 tokens/step per GPU (x2 GPUs = 131,072 tok/step)
+    batch_size: 2
+    grad_accum: 16  # 2048 * 2 * 16 = 65,536 tokens/step
 
-seq_len: 1024
+seq_len: 512
 train_split: 0.95
 
-# Kaggle T4 16GB VRAM: micro-batch 1 with SDPA and Gradient Checkpointing (<11GB peak VRAM)
-batch_size: 1
-grad_accum: 64
+# 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
 
@@ -55,17 +55,17 @@ compile: false
 fp8: false
 
 eval_every: 500
-save_every: 1000
-log_every: 10
+save_every: 500
+log_every: 1
 ckpt_dir: ../checkpoints
 log_dir: ../../logs
 
 # Checkpoint auto-push to Hugging Face Hub (background non-blocking thread)
 # Safeguard: Even if Kaggle session terminates at 12 hours, latest checkpoint is safe on HF!
 hf_push: true
-hf_repo: ViuAI/Viu-1.5B-MoE
+hf_repo: ViuAI/ViuMini-MoE-242M
 hf_push_every: 0
-hf_push_full_every: 1000
+hf_push_full_every: 500
 keep_local_full: 2
 keep_remote_full: 3
 resume: auto
@@ -74,7 +74,7 @@ sample_prompt: "bhai kal party me kya scene hai"
 
 # Zero-Disk Streaming from Hub (pure PyArrow memory-mapped buffer, ~300MB RAM)
 hf_dataset: ViuAI/viu-mini-raw-pretrain
-hf_max_blocks: 50000
+hf_max_blocks: 5000
 hf_safety: full
 hf_max_rows: null
 enforce_mix: true