GPTQ quantization

#2
by ArtemSultanov - opened

Hi! Can you tell me how possible it is to independently prepare the GPTQ modification of this model? We're exploring production deployment of Huihui-Qwen3-Coder-Next-abliterated via vLLM with 4-bit quantization. Due to the custom qwen3_next architecture and MoE setup, standard tools (AutoGPTQ/AutoAWQ) fail to recognize the model type. Any guidance on quantization workflows or pre-quantized weights would be highly appreciated.

  1. Architecture & Quantization Support
    Does Qwen3NextForCausalLM (custom qwen3_next architecture with mixed linear_attention/full_attention layers and 512-expert MoE) support standard 4-bit quantization frameworks like AutoGPTQ or AutoAWQ?
    If not officially supported, are there known workarounds or custom quantization scripts you recommend?

  2. Pre-quantized Versions
    Do you provide official GPTQ or AWQ 4-bit quantized versions of this model?
    If yes, please share the Hugging Face Hub link or conversion script used.

  3. MoE-Specific Quantization Guidance
    Given the extreme MoE configuration (num_experts=512, num_experts_per_tok=10):
    Should all 512 experts be quantized, or only the top-k activated ones?
    Are there known accuracy drops when quantizing MoE layers vs. dense layers?
    Any recommended group_size or desc_act settings for MoE stability?

  4. Critical Parameters for Quantization
    For optimal quantization quality, please confirm:
    Recommended group_size (128 for Marlin compatibility?)
    desc_act=True or False?
    Required calibration dataset (e.g., c4-new, domain-specific code data)?
    Minimum VRAM needed for quantization (e.g., 24GB for 7B-class models?)

Since our space on hf.co is now limited, there are no plans to switch to other formats of storage.

Maybe you can try llm-compressor

Creation details

import torch

from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer

from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.utils import dispatch_for_generation
from llmcompressor.modifiers.quantization import GPTQModifier

# NOTE: Requires a minimum of transformers 4.57.0

MODEL_ID = "/data/model-cache/Huihui-Qwen3-Coder-Next-abliterated"

# Load model.
model = AutoModelForCausalLM.from_pretrained(MODEL_ID, dtype="auto")
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

# Select calibration dataset.
DATASET_ID = "/data/model-cache/ultrachat_200k"
DATASET_SPLIT = "train_sft"

# Select number of samples. 512 samples is a good place to start.
# Increasing the number of samples can improve accuracy.
NUM_CALIBRATION_SAMPLES = 1024
MAX_SEQUENCE_LENGTH = 4096

# 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)

# Configure the quantization algorithm to run.
#   * quantize the weights to 4 bit with GPTQ with a group size 128
recipe = GPTQModifier(
    targets="Linear",
    scheme="W4A16",
    weight_observer="mse",
    ignore=[
        "lm_head",
        "re:.*mlp.gate$",
        "re:.*mlp.shared_expert_gate$",
        "re:.*linear_attn.*",
    ],
)

# Apply algorithms.
oneshot(
    model=model,
    dataset=ds,
    recipe=recipe,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
)

# Save to disk compressed.
SAVE_DIR = MODEL_ID + "-quantized.w4a16"
model.save_pretrained(SAVE_DIR, save_compressed=True)
tokenizer.save_pretrained(SAVE_DIR)

I will try later

Sign up or log in to comment