Qwen3-4B-Instruct-2507-heretic-W8A16
Overview
This is an 8-bit weight-only quantized version of Qwen3-4B-Instruct-2507-heretic using the W8A16 scheme, quantized with LLMCompressor.
| Property | Value |
|---|---|
| Base Model | Qwen3-4B-Instruct-2507-heretic |
| Quantization | W8A16 (8-bit weights, 16-bit activations) |
| Quant Method | compressed-tensors (AWQ-compatible) |
| Model Size | ~4.1 GB |
| Context Length | 262,144 tokens |
Quantization Details
- Framework: LLMCompressor
- Quantization Scheme: W8A16 (channel-wise symmetric int8)
- Format:
pack-quantized(compressed-tensors) - Target Layers: All
Linearlayers (exceptlm_head)
Performance
Perplexity (WikiText-2-raw-v1, test split)
| Model | Perplexity | Degradation |
|---|---|---|
| Original FP16 | 17.4634 | - |
| This Model (W8A16) | 17.5556 | +0.53% |
Nearly lossless quantization with minimal perplexity degradation!
Usage
With vLLM
vllm serve groxaxo/Qwen3-4B-Instruct-2507-heretic-W8A16 \
--quantization compressed-tensors \
--trust-remote-code
With Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"groxaxo/Qwen3-4B-Instruct-2507-heretic-W8A16",
device_map="auto",
trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(
"groxaxo/Qwen3-4B-Instruct-2507-heretic-W8A16"
)
messages = [{"role": "user", "content": "Hello, how are you?"}]
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=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Python API (vLLM)
from vllm import LLM, SamplingParams
llm = LLM(
model="groxaxo/Qwen3-4B-Instruct-2507-heretic-W8A16",
quantization="compressed-tensors",
trust_remote_code=True
)
sampling_params = SamplingParams(temperature=0.7, top_p=0.9, max_tokens=256)
outputs = llm.generate(["Hello, how are you?"], sampling_params)
for output in outputs:
print(output.outputs[0].text)
Model Architecture
- Architecture: Qwen3ForCausalLM
- Hidden Size: 2,560
- Intermediate Size: 9,728
- Attention Heads: 32
- KV Heads: 8 (GQA)
- Layers: 36
- Vocab Size: 151,936
- RoPE Theta: 5,000,000
Acknowledgements
- Qwen Team for the original Qwen3 model
- heretic-org for the fine-tuned variant
- vLLM Team for LLMCompressor and vLLM
License
Apache 2.0 (inherited from base model)
- Downloads last month
- 11
Model tree for groxaxo/Qwen3-4B-Instruct-2507-heretic-W8A16
Base model
Qwen/Qwen3-4B-Instruct-2507