EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

fix: resolve flashinfer 0.4.1 import (#11940)

Browse Source
This commit is contained in:
Yineng Zhang
2025-10-21 17:19:57 -07:00
committed by GitHub
parent 87a92e459a
commit 704160017d
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/layers/quantization/modelopt_quant.py
View File

@@ -1061,8 +1061,8 @@ class ModelOptNvFp4FusedMoEMethod(FusedMoEMethodBase):
):
from flashinfer import nvfp4_block_scale_interleave
from flashinfer.fused_moe.core import (
_maybe_get_cached_w2_permute_indices,
_maybe_get_cached_w3_w1_permute_indices,
get_w2_permute_indices_with_cache,
)
"""Prepare quantized weights for kernel (done offline with weights)."""
@@ -1123,7 +1123,7 @@ class ModelOptNvFp4FusedMoEMethod(FusedMoEMethodBase):
)
)
permute_indices = _maybe_get_cached_w2_permute_indices(
permute_indices = get_w2_permute_indices_with_cache(
self._cache_permute_indices,
gemm2_weights_fp4[i].view(torch.uint8),
epilogue_tile_m,
@@ -1134,7 +1134,7 @@ class ModelOptNvFp4FusedMoEMethod(FusedMoEMethodBase):
.contiguous()
)
permute_sf_indices = _maybe_get_cached_w2_permute_indices(
permute_sf_indices = get_w2_permute_indices_with_cache(
self._cache_permute_indices,
gemm2_scales_linear_fp4[i].view(torch.uint8),
epilogue_tile_m,