Upload sft_skt.py with huggingface_hub

sft_skt.py

@@ -1,5 +1,15 @@
 #!/usr/bin/env python3
+"""ROCm/MI300X-safe SFT for SKT-ST-X-10B-MoE  (TRL + PEFT, bf16, no Unsloth).
 
+Budget-tuned: 2,500 steps @ batch 16 ~= 2.3h train + ~15 min overhead = ~$5.06
+on MI300X @ $1.99/hr. Optimizer-step checkpointing every 500 steps so a crash
+costs at most ~28 min.
+
+Usage:
+    python sft_skt.py --model workingplace/SKT-ST-X-10B-MoE \
+                      --data train_500k.jsonl.gz \
+                      --out ./skt-sft --steps 2500
+"""
 import argparse, glob, gzip, json, os, sys, time
 
 import torch
@@ -114,10 +124,8 @@ def main():
             bnb_4bit_compute_dtype=torch.float16,
             bnb_4bit_use_double_quant=True,
             llm_int8_enable_fp32_cpu_offload=True)
-        kw["device_map"] = "auto"
-        kw["max_memory"] = {0: "11GiB", "cpu": "16GiB"}
-        kw["offload_folder"] = "/kaggle/working/offload"
-        print("[model] 4-bit NF4, fp16 compute, GPU cap 11GiB")
+        kw["device_map"] = "auto"          # splits across BOTH T4s (no max_memory)
+        print("[model] 4-bit NF4, fp16 compute, auto device_map")
     else:
         kw["torch_dtype"] = torch.bfloat16
     model = AutoModelForCausalLM.from_pretrained(args.model, **kw)