chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[quant] Ignore FP8 quantization layers (#20340)

Browse Source 
Co-authored-by: qiuxuan.lzw <qiuxuan.lzw@alibaba-inc.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
This commit is contained in:
StonyPort
2026-03-13 13:59:39 +08:00
committed by GitHub
co-authored by qiuxuan.lzw gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
parent e00328d1e5
commit d4e68ead1d
3 changed files with 11 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/references/environment_variables.md
+2 -1
View File
@@ -126,7 +126,8 @@ SGLang supports various environment variables that can be used to configure its
| `SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE` | Quantize moe of nextn layer from BF16 to FP8 when launching DeepSeek NVFP4 checkpoint | `false` | | `SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE` | Quantize moe of nextn layer from BF16 to FP8 when launching DeepSeek NVFP4 checkpoint | `false` |
| `SGLANG_ENABLE_FLASHINFER_FP8_GEMM` (deprecated) | Use flashinfer kernels when running blockwise fp8 GEMM on Blackwell GPUs. **DEPRECATED**: Please use `--fp8-gemm-backend=flashinfer_trtllm` (SM100/SM103) or `--fp8-gemm-backend=flashinfer_cutlass` (SM120/SM121 and newer) instead. | `false` | | `SGLANG_ENABLE_FLASHINFER_FP8_GEMM` (deprecated) | Use flashinfer kernels when running blockwise fp8 GEMM on Blackwell GPUs. **DEPRECATED**: Please use `--fp8-gemm-backend=flashinfer_trtllm` (SM100/SM103) or `--fp8-gemm-backend=flashinfer_cutlass` (SM120/SM121 and newer) instead. | `false` |
| `SGLANG_SUPPORT_CUTLASS_BLOCK_FP8` (deprecated) | Use Cutlass kernels when running blockwise fp8 GEMM on Hopper or Blackwell GPUs. **DEPRECATED**: Please use `--fp8-gemm-backend=cutlass` instead. | `false` | | `SGLANG_SUPPORT_CUTLASS_BLOCK_FP8` (deprecated) | Use Cutlass kernels when running blockwise fp8 GEMM on Hopper or Blackwell GPUs. **DEPRECATED**: Please use `--fp8-gemm-backend=cutlass` instead. | `false` |
| `SGLANG_QUANT_ALLOW_DOWNCASTING` | Allow weights downcasting | `false` | | `SGLANG_QUANT_ALLOW_DOWNCASTING` | Allow weight dtype downcasting during loading (e.g., fp32 → fp16). By default, SGLang rejects this kind of downcasting when using quantization. | `false` |
| `SGLANG_FP8_IGNORED_LAYERS` | A comma-separated list of layer names to ignore during FP8 quantization. For example: `model.layers.0,model.layers.1.,qkv_proj`. | `""` |
## Distributed Computing ## Distributed Computing
python/sglang/srt/environ.py
+1
View File
@@ -333,6 +333,7 @@ class Envs:
SGLANG_PER_TOKEN_GROUP_QUANT_8BIT_V2 = EnvBool(False) SGLANG_PER_TOKEN_GROUP_QUANT_8BIT_V2 = EnvBool(False)
SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE = EnvBool(False) SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE = EnvBool(False)
SGLANG_QUANT_ALLOW_DOWNCASTING = EnvBool(False) SGLANG_QUANT_ALLOW_DOWNCASTING = EnvBool(False)
SGLANG_FP8_IGNORED_LAYERS = EnvStr("")
# Flashinfer # Flashinfer
SGLANG_IS_FLASHINFER_AVAILABLE = EnvBool(True) SGLANG_IS_FLASHINFER_AVAILABLE = EnvBool(True)
python/sglang/srt/layers/quantization/fp8.py
+8
View File
@@ -131,6 +131,14 @@ class Fp8Config(QuantizationConfig):
raise ValueError(f"Unsupported activation scheme {activation_scheme}") raise ValueError(f"Unsupported activation scheme {activation_scheme}")
self.activation_scheme = activation_scheme self.activation_scheme = activation_scheme
self.ignored_layers = ignored_layers or [] self.ignored_layers = ignored_layers or []
if ignored_layers_str := envs.SGLANG_FP8_IGNORED_LAYERS.get():
self.ignored_layers.extend(
[
layer.strip()
for layer in ignored_layers_str.split(",")
if layer.strip()
]
)
self.packed_modules_mapping = packed_modules_mapping or {} self.packed_modules_mapping = packed_modules_mapping or {}
self.use_mxfp8 = use_mxfp8 self.use_mxfp8 = use_mxfp8
if weight_block_size is not None: if weight_block_size is not None: