Tune 0.5b v7 job for H200: batch 64, lr 4e-4, 3 epochs, no grad checkpointing

job-0.5b-v7.py

@@ -112,12 +112,20 @@ model = get_peft_model(model, LoraConfig(
     r=16, lora_alpha=32, lora_dropout=0.05, bias="none", task_type="CAUSAL_LM", use_rslora=True,
     target_modules=["q_proj","k_proj","v_proj","o_proj","gate_proj","up_proj","down_proj"]))
 model.print_trainable_parameters()
-args = SFTConfig(output_dir="out-0.5b-v2", num_train_epochs=2, per_device_train_batch_size=8,
-                 gradient_accumulation_steps=2, learning_rate=2e-4,
-                 lr_scheduler_type="cosine", warmup_ratio=0.03, logging_steps=10,
+# Tuned for a single H200 (141 GB): batch 64 in one go instead of 8x2 accumulation.
+# That is a 4x larger effective batch, so the LR is sqrt-scaled 2e-4 -> 4e-4 and
+# warmup widened to 10% (the step count is small enough that 3% was ~3 steps).
+# 3 epochs rather than 2 buys back optimizer steps (66 -> 99) that the big batch
+# costs, and on this hardware the extra epoch is nearly free.
+# Gradient checkpointing is off: with 141 GB there is no reason to trade compute
+# for memory, and it is worth ~30% throughput.
+args = SFTConfig(output_dir="out-0.5b-v2", num_train_epochs=3, per_device_train_batch_size=64,
+                 gradient_accumulation_steps=1, learning_rate=4e-4,
+                 gradient_checkpointing=False,
+                 lr_scheduler_type="cosine", warmup_ratio=0.1, logging_steps=5,
                  save_strategy="no", bf16=True, max_length=MAX_LEN, packing=False,
                  dataset_text_field="text", push_to_hub=False, report_to="none",
-                 run_name="sakthai-0.5b-v2-mlp-rslora-v7")
+                 run_name="sakthai-0.5b-v2-mlp-rslora-v7-h200")
 trainer = SFTTrainer(model=model, args=args, train_dataset=train_ds, processing_class=tokenizer)
 trainer.train()