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

Revert "[Performance] Performance improvements in non-blockwise fp8 CUTLASS MoE (#20762) (#21334)

Browse Source 
Signed-off-by: Ming Yang <minos.future@gmail.com>
This commit is contained in:
Ming Yang
2025-07-21 21:49:01 -07:00
committed by GitHub
parent 90f1e55421
commit e7b2042681
6 changed files with 38 additions and 174 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
tests/kernels/moe/test_pplx_cutlass_moe.py
View File

@@ -75,7 +75,6 @@ def pplx_cutlass_moe(
assert torch.cuda.current_device() == pgi.local_rank
num_tokens, hidden_dim = a.shape
intermediate_dim = w2.shape[2]
num_experts = w1.shape[0]
block_size = hidden_dim # TODO support more cases
device = pgi.device
@@ -124,31 +123,10 @@ def pplx_cutlass_moe(
num_local_experts=num_local_experts,
num_dispatchers=num_dispatchers)
ab_strides1 = torch.full((num_local_experts, ),
hidden_dim,
device="cuda",
dtype=torch.int64)
ab_strides2 = torch.full((num_local_experts, ),
intermediate_dim,
device="cuda",
dtype=torch.int64)
c_strides1 = torch.full((num_local_experts, ),
2 * intermediate_dim,
device="cuda",
dtype=torch.int64)
c_strides2 = torch.full((num_local_experts, ),
hidden_dim,
device="cuda",
dtype=torch.int64)
experts = CutlassExpertsFp8(num_local_experts,
out_dtype,
per_act_token,
per_out_ch,
ab_strides1,
ab_strides2,
c_strides1,
c_strides2,
num_dispatchers=num_dispatchers,
use_batched_format=True)