Upload train_glm_qlora_v4.py with huggingface_hub

train_glm_qlora_v4.py

@@ -36,6 +36,10 @@ bnb_config = BitsAndBytesConfig(
     llm_int8_enable_fp32_cpu_offload=True,
 )
 
+import os
+offload_dir = "/tmp/offload"
+os.makedirs(offload_dir, exist_ok=True)
+
 print("Loading model in 4-bit with CPU offload...")
 model = AutoModelForCausalLM.from_pretrained(
     "zai-org/GLM-4.7-Flash",
@@ -43,6 +47,7 @@ model = AutoModelForCausalLM.from_pretrained(
     trust_remote_code=True,
     device_map="auto",
     max_memory={0: "20GiB", "cpu": "30GiB"},
+    offload_folder=offload_dir,
 )
 tokenizer = AutoTokenizer.from_pretrained("zai-org/GLM-4.7-Flash", trust_remote_code=True)
 print("Model loaded.")