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

[1/N] Elastic EP Milestone 2 (#34861)

Browse Source 
Signed-off-by: Yongji Wu <wuyongji317@gmail.com>
Signed-off-by: Itay Alroy <ialroy@nvidia.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Signed-off-by: Ron Tourgeman <rtourgeman@nvidia.com>
Co-authored-by: Yongji Wu <wuyongji317@gmail.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Ron Tourgeman <rtourgeman@nvidia.com>
This commit is contained in:
Itay Alroy
2026-02-28 06:46:42 +02:00
committed by GitHub
parent 90805ff464
commit dea268336f
53 changed files with 3613 additions and 1016 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm/model_executor/layers/fused_moe/layer.py
View File

@@ -627,6 +627,7 @@ class FusedMoE(CustomOp):
moe_quant_params["intermediate_size_full"] = intermediate_size
self.quant_method.create_weights(layer=self, **moe_quant_params)
self.base_quant_method = self.quant_method
# Disable shared expert overlap if:
# - we are using eplb with non-default backend, because of correctness issues
@@ -683,7 +684,7 @@ class FusedMoE(CustomOp):
# routing_tables only needed for round-robin expert placement with
# DeepEP all2all backend.
routing_tables = self._maybe_init_expert_routing_tables()
prepare_finalize = self.quant_method.maybe_make_prepare_finalize(
prepare_finalize = self.base_quant_method.maybe_make_prepare_finalize(
routing_tables=routing_tables
)
if prepare_finalize is not None:
@@ -693,7 +694,7 @@ class FusedMoE(CustomOp):
self._replace_quant_method(
FusedMoEModularMethod.make(
self,
self.quant_method,
self.base_quant_method,
prepare_finalize,
self.shared_experts,
inplace=not self.moe_config.disable_inplace,