Upload train_glm_qlora_v10.py with huggingface_hub

train_glm_qlora_v10.py

@@ -0,0 +1,128 @@
+# /// script
+# requires-python = ">=3.10"
+# dependencies = [
+#     "trl>=0.12.0",
+#     "peft>=0.7.0",
+#     "transformers @ git+https://github.com/huggingface/transformers.git",
+#     "accelerate @ git+https://github.com/huggingface/accelerate.git",
+#     "bitsandbytes>=0.45.0",
+#     "trackio",
+#     "datasets",
+# ]
+# ///
+
+"""
+Agent Zero SFT: zai-org/GLM-4.7-Flash (30B MoE)
+QLoRA (4-bit) with CPU offload + monkey-patch for Params4bit compat.
+"""
+
+import os
+import torch
+
+# Monkey-patch Params4bit to accept _is_hf_initialized kwarg
+# This fixes accelerate<->bitsandbytes incompatibility where accelerate
+# passes _is_hf_initialized as a kwarg but Params4bit doesn't accept it.
+import bitsandbytes as bnb
+_orig_params4bit_new = bnb.nn.Params4bit.__new__
+def _patched_params4bit_new(cls, *args, **kwargs):
+    kwargs.pop('_is_hf_initialized', None)
+    return _orig_params4bit_new(cls, *args, **kwargs)
+bnb.nn.Params4bit.__new__ = _patched_params4bit_new
+print("Patched Params4bit to accept _is_hf_initialized")
+
+import trackio
+from huggingface_hub import login
+login(token=os.environ["HF_TOKEN"])
+
+from datasets import load_dataset
+from peft import LoraConfig
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
+from trl import SFTTrainer, SFTConfig
+
+print("Loading dataset...")
+train_ds = load_dataset("wheattoast11/agent-zero-sft-v1", data_files="data/train.jsonl", split="train")
+val_ds = load_dataset("wheattoast11/agent-zero-sft-v1", data_files="data/validation.jsonl", split="train")
+print(f"Train: {len(train_ds)}, Val: {len(val_ds)}")
+
+bnb_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_quant_type="nf4",
+    bnb_4bit_compute_dtype=torch.bfloat16,
+    bnb_4bit_use_double_quant=True,
+    llm_int8_enable_fp32_cpu_offload=True,
+)
+
+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",
+    quantization_config=bnb_config,
+    trust_remote_code=True,
+    device_map="auto",
+    max_memory={0: "21GiB", "cpu": "40GiB"},
+    offload_folder=offload_dir,
+    torch_dtype=torch.bfloat16,
+)
+tokenizer = AutoTokenizer.from_pretrained("zai-org/GLM-4.7-Flash", trust_remote_code=True)
+print("Model loaded.")
+
+if hasattr(model, 'hf_device_map'):
+    devices = {}
+    for v in model.hf_device_map.values():
+        devices[str(v)] = devices.get(str(v), 0) + 1
+    print(f"Device distribution: {devices}")
+
+import subprocess
+result = subprocess.run(['nvidia-smi'], capture_output=True, text=True)
+print(result.stdout)
+
+config = SFTConfig(
+    output_dir="agent-zero-glm-4.7-v1",
+    push_to_hub=True,
+    hub_model_id="wheattoast11/agent-zero-glm-4.7-v1",
+    hub_strategy="every_save",
+    hub_private_repo=True,
+    num_train_epochs=2,
+    per_device_train_batch_size=1,
+    gradient_accumulation_steps=16,
+    learning_rate=1e-4,
+    bf16=True,
+    gradient_checkpointing=True,
+    logging_steps=10,
+    save_strategy="steps",
+    save_steps=50,
+    save_total_limit=2,
+    eval_strategy="steps",
+    eval_steps=50,
+    warmup_ratio=0.1,
+    lr_scheduler_type="cosine",
+    report_to="trackio",
+    project="agent-zero-finetune",
+    run_name="glm-4.7-flash-qlora-v1",
+)
+
+peft_config = LoraConfig(
+    r=16, lora_alpha=32, lora_dropout=0.05,
+    bias="none", task_type="CAUSAL_LM",
+    target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
+)
+
+print("Initializing trainer...")
+trainer = SFTTrainer(
+    model=model,
+    tokenizer=tokenizer,
+    train_dataset=train_ds,
+    eval_dataset=val_ds,
+    args=config,
+    peft_config=peft_config,
+)
+
+print("Starting training...")
+trainer.train()
+
+print("Pushing to Hub...")
+trainer.push_to_hub()
+trackio.finish()
+print("Done!")