Downloads · 30 days
15
19% of all-time downloads
hangxian/Qwen3.5-27B-FP8-PerChannel
Qwen3.5-27B-FP8-PerChannel is a text generation model from hangxian. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as apache-2.0.
- 来源模型: Qwen/Qwen3.5-27B(官方 BF16 原版) - 权重未做任何修改: 本模型是纯量化版本,未经过任何微调、蒸馏或去审查处理,权重数值与官方原版完全一致 - 量化方式: per-channel FP8(每列一个 scale)+ dynamic activation(每次推理在线量化激活) - 量化工具: llmcompressor,scheme=FP8DYNAMIC
Downloads · 30 days
15
19% of all-time downloads
All-time downloads
78
Public
Parameters
26.9B
29.5 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors29.4 GB · 100%
How the weights are stored.
F8_E4M324.4B · 91%
From the Hugging Face model README
Qwen/Qwen3.5-27B(官方 BF16 原版){
"quant_method": "compressed-tensors",
"config_groups": {
"group_0": {
"weights": {
"strategy": "channel",
"num_bits": 8,
"type": "float",
"dynamic": false
},
"input_activations": {
"strategy": "token",
"num_bits": 8,
"type": "float",
"dynamic": true
}
}
}
}
| 本模型 | Qwen/Qwen3.5-27B-FP8(官方) | |
|---|---|---|
| 权重来源 | 官方原版,未修改 | 官方原版,未修改 |
| 权重量化粒度 | per-channel | block 128×128 |
| 激活量化 | dynamic | dynamic |
| SGLang GEMM 后端 | sgl_kernel.fp8_scaled_mm (CUTLASS) | DeepGEMM / Triton |
| 实测性能 (H20, 4k/4k, c=1) | ~160 tok/s | ~108 tok/s (DeepGEMM) |
per-channel 版本在 decode 阶段(batch=1)比 block-wise 版本快约 50%,因为 per-channel 的 scale 处理更简单,GEMM kernel 开销更小。
python3 -m sglang.launch_server \
--model-path hangxian/Qwen3.5-27B-FP8-PerChannel \
--tp 1 \
--host 0.0.0.0 \
--port 30000
python3 -m sglang.launch_server \
--model-path Qwen/Qwen3.5-27B \
--quantization fp8 \
--tp 1 \
--host 0.0.0.0 \
--port 30000
两种方式在推理时效果完全一致,区别仅在于本模型预先将 FP8 权重存储在磁盘上,方式二在加载时在线量化。
Apache 2.0(与原模型一致)