frozenleaves/vllm
1
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2025-10-20 14:53:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

[ModelOpt] Remove NVFP4 MoE K%16==0 constraint (#26891)

Browse Source 
Signed-off-by: XiaobingSuper <xiaobingzhangupc@gmail.com>
This commit is contained in:
XiaobingZhang
2025-10-16 01:06:17 +08:00
committed by GitHub
parent 14f8456344
commit d796375258
1 changed files with 0 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/model_executor/layers/quantization/modelopt.py
View File

@ -1542,23 +1542,11 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
del layer.w2_input_scale_quant
else:
# Non-TRT-LLM processing (Cutlass or non-flashinfer)
assert layer.w13_weight_scale.shape[2] % 16 == 0, (
"Expected weight_scale.dim(1) to be divisible by 16"
)
assert layer.w13_weight_scale.dtype == torch.float8_e4m3fn, (
"Weight Blockscale must be represented as FP8-E4M3"
)
w13_blockscale_swizzled = swizzle_blockscale(layer.w13_weight_scale)
layer.w13_weight_scale = Parameter(
w13_blockscale_swizzled, requires_grad=False
)
assert layer.w2_weight_scale.shape[2] % 16 == 0, (
"Expected weight_scale.dim(1) to be divisible by 16"
)
assert layer.w2_weight_scale.dtype == torch.float8_e4m3fn, (
"Weight Blockscale must be represented as FP8-E4M3"
)
w2_blockscale_swizzled = swizzle_blockscale(layer.w2_weight_scale)
layer.w2_weight_scale = Parameter(
w2_blockscale_swizzled, requires_grad=False