57 lines
1.3 KiB
Markdown
57 lines
1.3 KiB
Markdown
---
|
|
license: apache-2.0
|
|
base_model: Qwen/Qwen3-14B
|
|
tags:
|
|
- quantized
|
|
- 4-bit
|
|
- int4
|
|
- qwen3
|
|
language:
|
|
- en
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
# Qwen3-14B-AWQ-INT4
|
|
|
|
INT4 quantization of [`Qwen/Qwen3-14B`](https://huggingface.co/Qwen/Qwen3-14B). Built to run on a single consumer GPU (≥10 GB VRAM).
|
|
|
|
## Footprint
|
|
|
|
| | |
|
|
|---|---|
|
|
| Source params | 14B |
|
|
| Quantized weights | ~9.4 GB on disk |
|
|
| Inference VRAM (incl. KV cache @ 32K context) | ~16 GB |
|
|
|
|
Fits any 16 GB+ consumer card. No homelab needed.
|
|
|
|
## Bench
|
|
|
|
Scored on [`drawais/needle-1M-bench-mvp`](https://huggingface.co/datasets/drawais/needle-1M-bench-mvp) (50K-token haystack, real arxiv text):
|
|
|
|
| Metric | Score |
|
|
|---|---|
|
|
| Overall recall | **90.0%** |
|
|
| Paper-anchored | 80.0% |
|
|
| Synthetic codes | 100.0% |
|
|
|
|
## Quick start
|
|
|
|
```bash
|
|
vllm serve drawais/Qwen3-14B-AWQ-INT4 --quantization awq_marlin --max-model-len 32768
|
|
```
|
|
|
|
```python
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM
|
|
tok = AutoTokenizer.from_pretrained("drawais/Qwen3-14B-AWQ-INT4")
|
|
model = AutoModelForCausalLM.from_pretrained("drawais/Qwen3-14B-AWQ-INT4", device_map="auto")
|
|
```
|
|
|
|
## Context length
|
|
|
|
Native: 40,960 tokens (inherits from base model). For longer contexts, enable YaRN rope-scaling per the base model's config — supported by vLLM and Transformers.
|
|
|
|
## License
|
|
|
|
Apache 2.0 (inherits from base model).
|