biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Kernels] MoE refactor (#19636)

Browse Source 
Signed-off-by: Bill Nell <bnell@redhat.com>
Signed-off-by: ElizaWszola <ewszola@redhat.com>
Co-authored-by: ElizaWszola <ewszola@redhat.com>
This commit is contained in:
bnellnm
2025-07-02 09:08:27 -04:00
committed by GitHub
parent b95877509b
commit c1909e7e8c
36 changed files with 2698 additions and 1584 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
tests/kernels/moe/test_pplx_cutlass_moe.py
View File

@@ -15,7 +15,7 @@ from vllm.model_executor.layers.fused_moe.modular_kernel import (
FusedMoEModularKernel)
from vllm.platforms import current_platform
from .utils import ProcessGroupInfo, parallel_launch
from .parallel_utils import ProcessGroupInfo, parallel_launch
try:
from pplx_kernels import AllToAll
@@ -93,7 +93,7 @@ def pplx_cutlass_moe(
num_experts=num_experts,
experts_per_token=topk,
rank=rank,
world_size=pgi.world_size,
world_size=world_size,
dp_size=dp_size,
hidden_dim=hidden_dim,
hidden_dim_bytes=hidden_dim, # because a.dtype.itemsize == 1
@@ -118,8 +118,6 @@ def pplx_cutlass_moe(
pgi.world_size,
rank,
dp_size,
quant_dtype=torch.float8_e4m3fn,
per_act_token=per_act_token,
)
experts = CutlassExpertsFp8((num_experts + world_size - 1) // world_size,