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

[Kernels] Remove BatchedTritonOrDeepGemmExperts and default fallback to Triton (#29929)

Browse Source 
Signed-off-by: Bill Nell <bnell@redhat.com>
Signed-off-by: bnellnm <49004751+bnellnm@users.noreply.github.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
This commit is contained in:
bnellnm
2025-12-03 15:49:00 -05:00
committed by GitHub
parent ac1886588f
commit 2902c34826
5 changed files with 46 additions and 217 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
tests/kernels/moe/modular_kernel_tools/mk_objects.py
View File

@@ -13,9 +13,6 @@ from vllm.model_executor.layers.fused_moe.all2all_utils import (
from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
BatchedDeepGemmExperts,
)
from vllm.model_executor.layers.fused_moe.batched_triton_or_deep_gemm_moe import (
BatchedTritonOrDeepGemmExperts,
)
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
FusedMoEQuantConfig,
@@ -286,16 +283,6 @@ if has_deep_gemm() and is_deep_gemm_supported():
needs_matching_quant=False,
needs_deep_gemm=True,
)
register_experts(
BatchedTritonOrDeepGemmExperts,
batched_format,
common_float_and_int_types,
blocked_quantization_support=True,
supports_chunking=False,
supports_expert_map=False,
needs_matching_quant=True,
needs_deep_gemm=True,
)
register_experts(
TritonOrDeepGemmExperts,
standard_format,
@@ -457,10 +444,6 @@ def make_fused_experts(
kwargs = batch_kwargs | quant_kwargs
print(f"Making BatchedTritonExperts {kwargs} ...")
experts = BatchedTritonExperts(**kwargs)
elif fused_experts_type == BatchedTritonOrDeepGemmExperts:
kwargs = batch_kwargs | quant_kwargs | deepgemm_kwargs
print(f"Making BatchedTritonOrDeepGemmExperts {kwargs} ...")
experts = BatchedTritonOrDeepGemmExperts(**kwargs)
elif fused_experts_type == DeepGemmExperts:
print(f"Making DeepGemmExperts {quant_config} ...")
experts = DeepGemmExperts(quant_config)