Qwen3-Next-80B-A3B-Thinking-FP8-Dynamic

Qwen3-80B FP8 Dynamic Quantization with LLMCompressor

Introduction

---

Environment Requirements

• Python 3.10+
• NVIDIA GPU (Hopper architecture supporting FP8, e.g., H100/A100)
• CUDA 12.x
• PyTorch 2.6
• Dependencies installation:

  uv pip install llmcompressor torch
  uv pip install git+https://github.com/huggingface/transformers.git@main
  

---

Usage Steps

1. Save the following script as quantize.py:

   from llmcompressor.transformers import SparseAutoModelForCausalLM
   from transformers import AutoTokenizer
   model_name = "Qwen/Qwen3-Next-80B-A3B-Thinking"
   # Load tokenizer and model
   tokenizer = AutoTokenizer.from_pretrained(model_name)
   model = SparseAutoModelForCausalLM.from_pretrained(
       model_name,
       dtype="auto",
       device_map="auto"
   )
   from llmcompressor.transformers import oneshot
   from llmcompressor.modifiers.quantization import QuantizationModifier
   # Configure simple PTQ quantization
   recipe = QuantizationModifier(
       targets="Linear",
       scheme="FP8_DYNAMIC",
       ignore=[
           "lm_head",
           "re:.*mlp.gate$",           # Ignore standard gate layers
           "re:.*shared_expert_gate$",  # Ignore shared expert gate layers
           "re:.*router$"               # Ignore router layers
       ]
   )
   # Apply quantization algorithm
   oneshot(model=model, recipe=recipe)
   # Save model
   SAVE_DIR = model_name.split("/")[1] + "-FP8-Dynamic"
   model.save_pretrained(SAVE_DIR)
   tokenizer.save_pretrained(SAVE_DIR)
   

2. Run the script:

   python quantize.py
   

3. The quantized model will be saved in the Qwen3-Next-80B-A3B-Thinking-FP8-Dynamic directory.

VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 vllm serve Qwen3-Next-80B-A3B-Thinking-FP8-Dynamic --port 8080 --tensor-parallel-size 2 --api-key 123 --gpu-memory-utilization 0.95 --max_num_seqs 2 --max-model-len 131072 --enable-auto-tool-choice --tool-call-parser hermes --reasoning-parser deepseek_r1 # --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'

---

---

Notes

1. There is compatibility issues between the quantized version and MTP

---

References

• LLMCompressor Official Documentation