xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-06-01 21:11:19 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Perf] Reuse workspace for FP8+FP4 Marlin MoE (#20500)

Browse Source 
Signed-off-by: mgoin <mgoin64@gmail.com>
Signed-off-by: Michael Goin <mgoin64@gmail.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
This commit is contained in:
Michael Goin 2025-09-16 17:45:10 -04:00 committed by GitHub
parent 3053a22b33
commit dbebb7f812
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 8 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
View File

@ -398,7 +398,8 @@ class CompressedTensorsW4A4MoeMethod(CompressedTensorsMoEMethod):
quant_type_id=scalar_types.float4_e2m1f.id, quant_type_id=scalar_types.float4_e2m1f.id,
apply_router_weight_on_input=apply_router_weight_on_input, apply_router_weight_on_input=apply_router_weight_on_input,
global_num_experts=global_num_experts, global_num_experts=global_num_experts,
expert_map=expert_map) expert_map=expert_map,
workspace=layer.workspace)
# FlashInfer fused experts path # FlashInfer fused experts path
if self.fused_experts is not None: if self.fused_experts is not None:
@ -940,7 +941,8 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
quant_type_id=scalar_types.float8_e4m3fn.id, quant_type_id=scalar_types.float8_e4m3fn.id,
apply_router_weight_on_input=apply_router_weight_on_input, apply_router_weight_on_input=apply_router_weight_on_input,
global_num_experts=global_num_experts, global_num_experts=global_num_experts,
expert_map=expert_map) expert_map=expert_map,
workspace=layer.workspace)
assert self.fused_experts_func is not None assert self.fused_experts_func is not None

3
vllm/model_executor/layers/quantization/fp8.py
View File

@ -1103,7 +1103,8 @@ class Fp8MoEMethod(FusedMoEMethodBase):
quant_type_id=scalar_types.float8_e4m3fn.id, quant_type_id=scalar_types.float8_e4m3fn.id,
apply_router_weight_on_input=apply_router_weight_on_input, apply_router_weight_on_input=apply_router_weight_on_input,
global_num_experts=global_num_experts, global_num_experts=global_num_experts,
expert_map=expert_map) expert_map=expert_map,
workspace=layer.workspace)
elif self.flashinfer_moe_backend == FlashinferMoeBackend.CUTLASS: elif self.flashinfer_moe_backend == FlashinferMoeBackend.CUTLASS:
assert self.block_quant is None assert self.block_quant is None
assert (not renormalize and custom_routing_function is not None) assert (not renormalize and custom_routing_function is not None)

3
vllm/model_executor/layers/quantization/modelopt.py
View File

@ -1474,7 +1474,8 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
quant_type_id=scalar_types.float4_e2m1f.id, quant_type_id=scalar_types.float4_e2m1f.id,
apply_router_weight_on_input=apply_router_weight_on_input, apply_router_weight_on_input=apply_router_weight_on_input,
global_num_experts=global_num_experts, global_num_experts=global_num_experts,
expert_map=expert_map) expert_map=expert_map,
workspace=layer.workspace)
if self.fused_experts is not None: if self.fused_experts is not None:
assert self.allow_flashinfer and \ assert self.allow_flashinfer and \