[Minor] Fused experts refactor (#15914)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2025-04-03 10:19:38 -07:00
committed by GitHub
parent d2b58ca203
commit 15ba07ef25
8 changed files with 790 additions and 737 deletions
+6 -3
View File
@@ -9,8 +9,11 @@ import torch
from vllm.config import VllmConfig, set_current_vllm_config
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.fused_moe import fused_moe
from vllm.model_executor.layers.fused_moe.fused_moe import (
deep_gemm_moe_fp8, fused_topk, moe_align_block_size)
from vllm.model_executor.layers.fused_moe.deep_gemm_moe import (
deep_gemm_moe_fp8)
from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
from vllm.model_executor.layers.fused_moe.moe_align_block_size import (
moe_align_block_size)
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
per_token_group_quant_fp8, w8a8_block_fp8_matmul)
from vllm.platforms import current_platform
@@ -437,7 +440,7 @@ def test_w8a8_block_fp8_deep_gemm_fused_moe(M, N, K, E, topk, seed):
pytest.skip(
f"Skipping test; bad size m={M}, n={N}, k={K}, topk={topk}, E={E}")
if (N <= 512):
if N <= 512:
pytest.skip("Skipping N <= 512 until performance issues solved.")
vllm_config = VllmConfig()
+8 -8
View File
@@ -4,8 +4,8 @@ import torch
from vllm import _custom_ops as ops
from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config
from vllm.model_executor.layers.fused_moe.fused_moe import (cutlass_moe_fp8,
fused_experts,
from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp8
from vllm.model_executor.layers.fused_moe.fused_moe import (fused_experts,
fused_topk)
from vllm.platforms import current_platform
@@ -131,9 +131,9 @@ def test_cutlass_moe_no_graph(
c_strides2,
a1_scale=a_scale1)
print(triton_output)
print(cutlass_output)
print("*")
#print(triton_output)
#print(cutlass_output)
#print("*")
torch.testing.assert_close(triton_output,
cutlass_output,
@@ -234,9 +234,9 @@ def test_cutlass_moe_cuda_graph(
graph.replay()
torch.cuda.synchronize()
print(triton_output)
print(cutlass_output)
print("*")
#print(triton_output)
#print(cutlass_output)
#print("*")
torch.testing.assert_close(triton_output,
cutlass_output,