train_glm_qlora.py

# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "unsloth[cu124-ampere]",
#     "trackio",
#     "datasets",
# ]
# ///

"""
Agent Zero SFT: zai-org/GLM-4.7-Flash (30B MoE)
QLoRA (4-bit) fine-tuning with Unsloth optimizations.
Router layers frozen - only attention layers trained.
"""

import trackio
from datasets import load_dataset
from unsloth import FastLanguageModel

# Load dataset
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)}")

# Load model in 4-bit with Unsloth
print("Loading model with Unsloth (4-bit QLoRA)...")
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="zai-org/GLM-4.7-Flash",
    max_seq_length=2048,
    load_in_4bit=True,
    dtype=None,  # auto-detect
)

# Apply LoRA adapters via Unsloth
model = FastLanguageModel.get_peft_model(
    model,
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    target_modules=["q_proj", "v_proj", "k_proj", "o_proj"],
    bias="none",
    use_gradient_checkpointing="unsloth",  # Unsloth optimized
)

from trl import SFTTrainer, SFTConfig

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,

    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",
)

print("Initializing trainer...")
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=train_ds,
    eval_dataset=val_ds,
    args=config,
)

print("Starting training...")
trainer.train()

print("Pushing to Hub...")
trainer.push_to_hub()

trackio.finish()
print("Done! Model at: https://huggingface.co/wheattoast11/agent-zero-glm-4.7-v1")