Rainy-W-cy/qwen2-0.5b-instruct-rtn-w8a16

Qwen2-0.5B-Instruct RTN W8A16

This repository contains an INT8 weight-only quantized version of Qwen/Qwen2-0.5B-Instruct. The checkpoint was produced with RTN through llm-compressor and saved in the compressed-tensors format.

这是Qwen2-0.5B-Instruct的RTN W8A16版本。Linear权重以INT8保存，激活、Attention和KV Cache保持浮点；它不是动态W8A8模型。

Model Details

Field	Value
Base model	Qwen/Qwen2-0.5B-Instruct
Base revision	c540970f9e29518b1d8f06ab8b24cba66ad77b6d
Offline algorithm	RTN (round-to-nearest weight quantization)
Quantized targets	Linear
Weight format	Signed INT8, symmetric
Weight granularity	Per-output-channel (strategy=channel, group_size=null)
Activation/KV format	Floating point; not quantized by this checkpoint
Ignored module	lm_head
Resolved model dtype	BF16
Serialization	compressed-tensors, pack-quantized
Export provider	llm-compressor==0.12.0
Compression library	compressed-tensors==0.17.1

RTN was executed without a calibration dataset. No fine-tuning or additional training was performed.

Usage

Use a Transformers environment that supports the serialized compressed-tensors format. The export environment used Transformers 5.10.1 and compressed-tensors==0.17.1.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "<your-hf-account>/Qwen2-0.5B-Instruct-RTN-W8A16"

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

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Explain INT8 weight-only quantization briefly."},
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128, do_sample=False)
response = tokenizer.decode(
    outputs[0][inputs.input_ids.shape[1]:],
    skip_special_tokens=True,
)
print(response)

Quantization Files

• config.json contains the effective compressed-tensors quantization configuration.
• provider_recipe.json records the requested and resolved provider parameters.
• recipe.yaml contains the effective llm-compressor recipe.
• model.safetensors contains the compressed model weights.

Evaluation

No formal benchmark results are claimed in this model card yet. The local artifact manifest records software evaluation as not_checked. Before publishing results, evaluate this checkpoint and the BF16 base model with identical task versions, prompts, sample selections, chat templates, generation parameters, and random seeds.

Example with LM Evaluation Harness:

lm-eval run \
  --model hf \
  --model_args pretrained=<your-hf-account>/Qwen2-0.5B-Instruct-RTN-W8A16,dtype=auto \
  --tasks arc_challenge \
  --num_fewshot 0 \
  --apply_chat_template \
  --batch_size 1 \
  --log_samples \
  --output_path runs/rtn_w8a16/arc_challenge

Intended Use

• Research on INT8 weight-only quantization.
• Software quality comparison against the pinned BF16 base model.
• Input to a separately validated deployment compiler or hardware-runtime workflow.

Limitations

• This is W8A16/weight-only, not end-to-end W8A8 or integer-only inference.
• lm_head, activations, Attention operations, and KV Cache are not INT8-quantized by this checkpoint.
• Runtime speed depends on backend support; reduced weight storage does not guarantee faster inference on every device.
• The model inherits the capabilities, limitations, and potential biases of the base Qwen2 model.
• No formal accuracy, perplexity, latency, power, or hardware-resource result is bundled yet.

License

The base model is released under the Apache 2.0 license. This quantized derivative follows the same license; users remain responsible for complying with the base model terms.

Citation

@article{qwen2,
  title={Qwen2 Technical Report},
  author={Yang, An and others},
  journal={arXiv preprint arXiv:2407.10671},
  year={2024}
}