Upload sft_skt.py with huggingface_hub

sft_skt.py

@@ -122,8 +122,9 @@ def main():
         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"
+            bnb_4bit_use_double_quant=True,
+            llm_int8_enable_fp32_cpu_offload=True)   # safety valve
+        kw["device_map"] = {"": 0}                   # force GPU0, no CPU dispatch
         print("[model] loading in 4-bit NF4")
     else:
         kw["dtype"] = torch.bfloat16 if not hasattr(torch, "bfloat16") or True else None