| --- |
| tags: |
| - fp4 |
| - vllm |
| language: |
| - en |
| - de |
| - fr |
| - it |
| - pt |
| - hi |
| - es |
| - th |
| pipeline_tag: text-generation |
| license: apache-2.0 |
| base_model: Qwen/Qwen3-32B |
| --- |
| |
| # Qwen3-32B-NVFP4 |
|
|
| ## Model Overview |
| - **Model Architecture:** Qwen/Qwen3-32B |
| - **Input:** Text |
| - **Output:** Text |
| - **Model Optimizations:** |
| - **Weight quantization:** FP4 |
| - **Activation quantization:** FP4 |
| - **Out-of-scope:** Use in any manner that violates applicable laws or regulations (including trade compliance laws). Use in languages other than English. |
| - **Release Date:** 6/25/2025 |
| - **Version:** 1.0 |
| - **Model Developers:** RedHatAI |
|
|
| This model is a quantized version of [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B). |
| It was evaluated on a several tasks to assess the its quality in comparison to the unquatized model. |
|
|
| ### Model Optimizations |
|
|
| This model was obtained by quantizing the weights and activations of [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) to FP4 data type, ready for inference with vLLM>=0.9.1 |
| This optimization reduces the number of bits per parameter from 16 to 4, reducing the disk size and GPU memory requirements by approximately 75%. |
|
|
| Only the weights and activations of the linear operators within transformers blocks are quantized using [LLM Compressor](https://github.com/vllm-project/llm-compressor). |
|
|
| ## Deployment |
|
|
| ### Use with vLLM |
|
|
| This model can be deployed efficiently using the [vLLM](https://docs.vllm.ai/en/latest/) backend, as shown in the example below. |
|
|
| ```python |
| from vllm import LLM, SamplingParams |
| from transformers import AutoTokenizer |
| |
| model_id = "RedHatAI/Qwen3-32B-NVFP4" |
| number_gpus = 2 |
| |
| sampling_params = SamplingParams(temperature=0.6, top_p=0.9, max_tokens=256) |
| |
| tokenizer = AutoTokenizer.from_pretrained(model_id) |
| |
| messages = [ |
| {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"}, |
| {"role": "user", "content": "Who are you?"}, |
| ] |
| |
| prompts = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False) |
| |
| llm = LLM(model=model_id, tensor_parallel_size=number_gpus) |
| |
| outputs = llm.generate(prompts, sampling_params) |
| |
| generated_text = outputs[0].outputs[0].text |
| print(generated_text) |
| ``` |
|
|
| vLLM aslo supports OpenAI-compatible serving. See the [documentation](https://docs.vllm.ai/en/latest/) for more details. |
|
|
| ## Creation |
|
|
| This model was created by applying [LLM Compressor with calibration samples from UltraChat](https://github.com/vllm-project/llm-compressor/blob/main/examples/quantization_w4a4_fp4/llama3_example.py), as presented in the code snipet below. |
|
|
| <details> |
| |
| ```python |
| 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 |
| |
| MODEL_ID = "Qwen/Qwen3-32B" |
| |
| # Load model. |
| model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype="auto") |
| tokenizer = AutoTokenizer.from_pretrained(MODEL_ID) |
| |
| DATASET_ID = "HuggingFaceH4/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 = 512 |
| 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) |
| |
| # Configure the quantization algorithm and scheme. |
| # In this case, we: |
| # * quantize the weights to fp4 with per group 16 via ptq |
| # * calibrate a global_scale for activations, which will be used to |
| # quantize activations to fp4 on the fly |
| recipe = QuantizationModifier(targets="Linear", scheme="NVFP4", ignore=["lm_head"]) |
| |
| # Save to disk in compressed-tensors format. |
| SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-NVFP4" |
| |
| # Apply quantization. |
| oneshot( |
| model=model, |
| dataset=ds, |
| recipe=recipe, |
| max_seq_length=MAX_SEQUENCE_LENGTH, |
| num_calibration_samples=NUM_CALIBRATION_SAMPLES, |
| output_dir=SAVE_DIR, |
| ) |
| |
| print("\n\n") |
| print("========== SAMPLE GENERATION ==============") |
| dispatch_for_generation(model) |
| input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to("cuda") |
| output = model.generate(input_ids, max_new_tokens=100) |
| print(tokenizer.decode(output[0])) |
| print("==========================================\n\n") |
| |
| model.save_pretrained(SAVE_DIR, save_compressed=True) |
| tokenizer.save_pretrained(SAVE_DIR) |
| |
| ``` |
| </details> |
|
|
| ## Evaluation |
|
|
| This model was evaluated on the well-known OpenLLM v1, OpenLLM v2 and HumanEval_64 benchmarks using [lm-evaluation-harness](https://github.com/neuralmagic/lm-evaluation-harness). The Reasoning evals were done using [ligheval](https://github.com/neuralmagic/lighteval). |
| |
| ### Accuracy |
| |
| <table> |
| <thead> |
| <tr> |
| <th>Category</th> |
| <th>Metric</th> |
| <th>Qwen/Qwen3-32B</th> |
| <th>RedHatAI/Qwen3-32B-NVFP4 (this model)</th> |
| <th>Recovery (%)</th> |
| </tr> |
| </thead> |
| <tbody> |
| <!-- OpenLLM V1 (Core) --> |
| <tr> |
| <td rowspan="7"><b>OpenLLM V1</b></td> |
| <td>arc_challenge</td> |
| <td>70.65</td> |
| <td>70.22</td> |
| <td>99.39</td> |
| </tr> |
| <tr> |
| <td>gsm8k</td> |
| <td>74.15</td> |
| <td>74.68</td> |
| <td>100.71</td> |
| </tr> |
| <tr> |
| <td>hellaswag</td> |
| <td>84.00</td> |
| <td>83.33</td> |
| <td>99.20</td> |
| </tr> |
| <tr> |
| <td>mmlu</td> |
| <td>81.84</td> |
| <td>81.23</td> |
| <td>99.25</td> |
| </tr> |
| <tr> |
| <td>truthfulqa_mc2</td> |
| <td>59.36</td> |
| <td>58.92</td> |
| <td>99.26</td> |
| </tr> |
| <tr> |
| <td>winogrande</td> |
| <td>75.93</td> |
| <td>76.80</td> |
| <td>101.15</td> |
| </tr> |
| <tr> |
| <td><b>Average</b></td> |
| <td><b>74.32</b></td> |
| <td><b>74.20</b></td> |
| <td><b>99.83</b></td> |
| </tr> |
| <tr> |
| <td rowspan="7"><b>OpenLLM V2</b></td> |
| <td>BBH (3-shot)</td> |
| <td>62.35</td> |
| <td>60.72</td> |
| <td>97.39</td> |
| </tr> |
| <tr> |
| <td>MMLU-Pro (5-shot)</td> |
| <td>54.39</td> |
| <td>51.13</td> |
| <td>94.01</td> |
| </tr> |
| <tr> |
| <td>MuSR (0-shot)</td> |
| <td>39.29</td> |
| <td>41.01</td> |
| <td>104.38</td> |
| </tr> |
| <tr> |
| <td>IFEval (0-shot)</td> |
| <td>88.97</td> |
| <td>87.29</td> |
| <td>98.11</td> |
| </tr> |
| <tr> |
| <td>GPQA (0-shot)</td> |
| <td>30.12</td> |
| <td>30.29</td> |
| <td>100.56</td> |
| </tr> |
| <tr> |
| <td>Math-|v|-5 (4-shot)</td> |
| <td>58.99</td> |
| <td>56.27</td> |
| <td>95.39</td> |
| </tr> |
| <tr> |
| <td><b>Average</b></td> |
| <td><b>55.69</b></td> |
| <td><b>54.45</b></td> |
| <td><b>97.79</b></td> |
| </tr> |
| <tr> |
| <td><b>Coding</b></td> |
| <td>HumanEval_64 pass@2</td> |
| <td>90.14</td> |
| <td>90.40</td> |
| <td>100.29</td> |
| </tr> |
| <tr> |
| <td rowspan="4"><b>Reasoning</b></td> |
| <td>AIME24 (0-shot)</td> |
| <td>75.86</td> |
| <td>68.97</td> |
| <td>90.93</td> |
| </tr> |
| <tr> |
| <td>AIME25 (0-shot)</td> |
| <td>72.41</td> |
| <td>65.52</td> |
| <td>90.52</td> |
| </tr> |
| <tr> |
| <td>GPQA (Diamond, 0-shot)</td> |
| <td>62.94</td> |
| <td>64.47</td> |
| <td>102.43</td> |
| </tr> |
| <tr> |
| <td><b>Average</b></td> |
| <td><b>70.40</b></td> |
| <td><b>66.32</b></td> |
| <td><b>94.21</b></td> |
| </tr> |
| </tbody> |
| </table> |
| |
|
|
| ### Reproduction |
|
|
| The results were obtained using the following commands: |
|
|
| <details> |
|
|
| #### OpenLLM v1 |
| ``` |
| lm_eval \ |
| --model vllm \ |
| --model_args pretrained="RedHatAI/Qwen3-32B-NVFP4",dtype=auto,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True,enforce_eager=True\ |
| --apply_chat_template \ |
| --fewshot_as_multiturn \ |
| --tasks openllm \ |
| --batch_size auto |
| ``` |
|
|
|
|
| #### OpenLLM v2 |
| ``` |
| lm_eval \ |
| --model vllm \ |
| --model_args pretrained="RedHatAI/Qwen3-32B-NVFP4",dtype=auto,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True,enforce_eager=True\ |
| --apply_chat_template \ |
| --fewshot_as_multiturn \ |
| --tasks leaderboard \ |
| --batch_size auto |
| ``` |
|
|
| #### HumanEval_64 |
| ``` |
| lm_eval \ |
| --model vllm \ |
| --model_args pretrained="RedHatAI/Qwen3-32B-NVFP4",dtype=auto,max_model_len=4096,tensor_parallel_size=2,enable_chunked_prefill=True,enforce_eager=True\ |
| --apply_chat_template \ |
| --fewshot_as_multiturn \ |
| --tasks humaneval_64_instruct \ |
| --batch_size auto |
| ``` |
| |
| #### LightEval |
| |
| ``` |
| # --- model_args.yaml --- |
| cat > model_args.yaml <<'YAML' |
| model_parameters: |
| model_name: "RedHatAI/Qwen3-32B-NVFP4" |
| dtype: auto |
| gpu_memory_utilization: 0.9 |
| tensor_parallel_size: 2 |
| max_model_length: 40960 |
| generation_parameters: |
| seed: 42 |
| temperature: 0.6 |
| top_k: 20 |
| top_p: 0.95 |
| min_p: 0.0 |
| max_new_tokens: 32768 |
| YAML |
| |
| lighteval vllm model_args.yaml \ |
| "lighteval|aime24|0,lighteval|aime25|0,lighteval|gpqa:diamond|0" \ |
| --max-samples -1 \ |
| --output-dir out_dir |
|
|
| ``` |
| |
| </details> |
