Qwen2.5-7B-Instruct-FP4-W4A4

Qwen2.5-7B-Instruct-FP4-W4A4

Model Description

This is an NVFP4 (NVIDIA FP4) quantized version of Qwen/Qwen2.5-7B-Instruct using the compressed-tensors quantization method.

• Base Model: Qwen/Qwen2.5-7B-Instruct
• Quantization Method: compressed-tensors
• Quantization Type: NVFP4 W4A4 (4-bit Weight and Activation)
• Model Size: ~5.5GB (compared to ~15GB for BF16)
• Compression Ratio: ~2.7x

Quantization Configuration

This model uses NVFP4 (NVIDIA FP4) quantization with grouped quantization for both weights and activations:

Weights

• Precision: NVFP4 (4-bit floating point)
• Strategy: Tensor-group (grouped quantization)
• Group Size: 16
• Symmetric: Yes
• Dynamic: No (static quantization)
• Observer: MinMax

Activations

• Precision: NVFP4 (4-bit floating point)
• Strategy: Tensor-group (grouped quantization)
• Group Size: 16
• Symmetric: Yes
• Dynamic: Local (dynamic quantization with local calibration)
• Observer: MinMax

Other Details

• Format: nvfp4-pack-quantized (packed 4-bit format)
• KV Cache: Not quantized
• Ignored Layers: lm_head
• Target Layers: Linear layers
• Quantization Version: 0.11.0

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "JongYeop/Qwen2.5-7B-Instruct-FP4-W4A4"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="auto"
)

messages = [
    {"role": "user", "content": "What is machine learning?"}
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(
    input_ids,
    max_new_tokens=256,
    do_sample=True,
    temperature=0.7,
    top_p=0.9,
)

response = tokenizer.decode(outputs[0][input_ids.shape[-1]:], skip_special_tokens=True)
print(response)

Model Architecture

• Architecture: Qwen2ForCausalLM
• Hidden Size: 3584
• Intermediate Size: 18944
• Number of Layers: 28
• Number of Attention Heads: 28
• Number of KV Heads: 4
• Vocabulary Size: 152064
• Max Position Embeddings: 32768

Intended Use

This quantized model is intended for efficient inference with significantly reduced memory footprint while maintaining reasonable performance. It is suitable for:

• Resource-constrained environments
• Edge deployment
• Applications requiring minimal memory usage
• High throughput scenarios
• GPU inference with FP4 support

Limitations

• FP4 quantization may result in more accuracy loss compared to FP8 or INT8 quantization
• Best performance is achieved on hardware with native FP4 support (e.g., NVIDIA H100, Ada Lovelace, Blackwell GPUs)
• Dynamic activation quantization may introduce additional runtime overhead
• Grouped quantization requires compatible inference engines

Performance Notes

• Memory Usage: ~2.7x reduction compared to BF16
• Speed: Requires hardware with FP4 tensor core support for optimal performance
• Accuracy: May experience some degradation compared to higher precision formats

Citation

If you use this model, please cite the original Qwen2.5 paper and the compressed-tensors library.

License

Same as the base model: Apache 2.0