Upload sft_skt.py with huggingface_hub

sft_skt.py

@@ -32,6 +32,10 @@ def parse_args():
     p.add_argument("--warmup", type=int, default=150)
     p.add_argument("--save-steps", type=int, default=500)
     p.add_argument("--seed", type=int, default=42)
+    p.add_argument("--hf-token", default=os.environ.get("HF_TOKEN", ""),
+                   help="HF token for private model repos")
+    p.add_argument("--load-4bit", action="store_true",
+                   help="load in 4-bit NF4 (needed for <=24GB GPUs, e.g. T4)")
     p.add_argument("--resume", action="store_true",
                    help="resume from the newest checkpoint in --out")
     return p.parse_args()
@@ -110,9 +114,21 @@ def main():
     if tok.pad_token is None:
         tok.pad_token = tok.eos_token
 
-    model = AutoModelForCausalLM.from_pretrained(
-        args.model, torch_dtype=torch.bfloat16,
-        attn_implementation="sdpa", trust_remote_code=True)
+    kw = dict(attn_implementation="sdpa", trust_remote_code=True)
+    if args.hf_token:
+        kw["token"] = args.hf_token
+    if args.load_4bit:
+        from transformers import BitsAndBytesConfig
+        kw["quantization_config"] = BitsAndBytesConfig(
+            load_in_4bit=True, bnb_4bit_quant_type="nf4",
+            bnb_4bit_compute_dtype=torch.bfloat16,
+            bnb_4bit_use_double_quant=True)
+        kw["device_map"] = "auto"
+        print("[model] loading in 4-bit NF4")
+    else:
+        kw["dtype"] = torch.bfloat16 if not hasattr(torch, "bfloat16") or True else None
+        kw["torch_dtype"] = torch.bfloat16
+    model = AutoModelForCausalLM.from_pretrained(args.model, **kw)
     model.config.use_cache = False                   # needed for grad ckpt
 
     targets = ["q_proj", "k_proj", "v_proj", "o_proj"]
@@ -121,6 +137,9 @@ def main():
     peft_cfg = LoraConfig(r=args.lora_r, lora_alpha=args.lora_alpha,
                           lora_dropout=0.05, bias="none",
                           task_type="CAUSAL_LM", target_modules=targets)
+    if args.load_4bit:
+        from peft import prepare_model_for_kbit_training
+        model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=True)
     model = get_peft_model(model, peft_cfg)
     model.print_trainable_parameters()