Upload folder using huggingface_hub

src/mathcompose/train/train.py

@@ -62,6 +62,20 @@ def build_and_train(args) -> str:
         model_kwargs["quantization_config"] = build_bnb_config(cfg, compute_dtype=compute_dtype)
     model = AutoModelForCausalLM.from_pretrained(base_id, **model_kwargs)
 
+    if compute_dtype == torch.float16:
+        # On T4 the fp16 GradScaler can't unscale bf16 grads. transformers may load
+        # the non-quantized weights in the model's default bf16 regardless of dtype=,
+        # so LoRA inherits bf16. Force any remaining bf16 params to fp16 and pin the
+        # config dtype so PEFT doesn't recreate bf16 adapters from the stale config.
+        model.config.torch_dtype = torch.float16
+        cast = 0
+        for p in model.parameters():
+            if p.dtype == torch.bfloat16:
+                p.data = p.data.to(torch.float16)
+                cast += 1
+        if cast:
+            print(f"[dtype] cast {cast} bf16 param tensors -> fp16 (T4 AMP compatibility)")
+
     data_files = {"train": train_file}
     if val_file and Path(val_file).exists():
         data_files["validation"] = val_file