boboliu/Qwen2.5-14B-Instruct-FP8-KV-CN

🤗 Hugging Face sourcetext-generationapache-2.014.8B params16 GBsafetensors✓ 5 checksumsupdated today
Submit in one command

Run it next to your model folder. It makes the torrent, checks your files against Hugging Face, and submits it. You just start seeding and paste your key from your account. It only reads your files and never changes them. Read the script first if you like.

curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo boboliu/Qwen2.5-14B-Instruct-FP8-KV-CN ./model-folder
Needs a seeder →

Qwen2.5-14B-Instruct FP8 Static Quantized with KVCache

使用 m-a-p/COIG-CQIA zhihu subset 校准静态量化的 Qwen/Qwen2.5-14B-Instruct 模型,可启用 fp8 kv cache。

subset 中主要为中文通常语言逻辑任务,为 vLLM 准备。

使用

参数加入 kv_cache_dtype="fp8"

评估

使用 lm-evaluation-harness + vLLM 进行评估(A100 with no FP8 support):

lm_eval --model vllm --model_args pretrained=MODEL_DIR,kv_cache_dtype=fp8,add_bos_token=True --tasks gsm8k --num_fewshot 5 --batch_size auto

得分

项目 Qwen2.5-14B-Instruct 此项目 Recovery
GSM8K 5shots 83.78 83.47 99.63%

Prefill 速度

项目 Qwen2.5-14B-Instruct 此项目 Improvement
GSM8K 5shots 1240.53 1395.89 12.52%

Decode 速度

项目 Qwen2.5-14B-Instruct 此项目 Improvement
GSM8K 5shots 287.38 316.10 10%