nvfp4_quantize.py

from transformers import AutoModelForCausalLM, AutoTokenizer
from datasets import load_dataset

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import QuantizationModifier
from compressed_tensors.offload import dispatch_model

MODEL_ID = "Qwen/Qwen3-Coder-Next"

# Load model.
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype="auto",
    low_cpu_mem_usage=True,
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)

DATASET_ID = "HuggingFaceH4/ultrachat_200k"
DATASET_SPLIT = "train_sft"

# Select number of samples
NUM_CALIBRATION_SAMPLES = 20
MAX_SEQUENCE_LENGTH = 2048

# Load dataset and preprocess.
ds = load_dataset(DATASET_ID, split=f"{DATASET_SPLIT}[:{NUM_CALIBRATION_SAMPLES}]")
ds = ds.shuffle(seed=42)


def preprocess(example):
    return {
        "text": tokenizer.apply_chat_template(
            example["messages"],
            tokenize=False,
        )
    }


ds = ds.map(preprocess)


# Tokenize inputs.
def tokenize(sample):
    return tokenizer(
        sample["text"],
        padding=False,
        max_length=MAX_SEQUENCE_LENGTH,
        truncation=True,
        add_special_tokens=False,
    )


ds = ds.map(tokenize, remove_columns=ds.column_names)



recipe = QuantizationModifier(
    targets="Linear",
    scheme="NVFP4",
    weight_observer="mse",
    ignore= ['re:.*lm_head', 're:.*mlp.gate$', 're:.*mlp.shared_expert_gate$', 're:.*linear_attn.*'],
)


oneshot(
    model=model,
    dataset=ds,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    moe_calibrate_all_experts=True,
)


print("\n\n")
print("========== SAMPLE GENERATION ==============")

dispatch_model(model)

input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to(
    model.device
)
output = model.generate(input_ids, max_new_tokens=100)
print(tokenizer.decode(output[0]))
print("==========================================\n\n")


# Save to disk in compressed-tensors format.
SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-NVFP4"
model.save_pretrained(SAVE_DIR, save_compressed=True)
tokenizer.save_pretrained(SAVE_DIR)