[MoE Refactor] Move expert map related code into ExpertMapManager class (#41046)

Signed-off-by: Bill Nell <bnell@redhat.com>
Signed-off-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Signed-off-by: Robert Shaw <robertgshaw2@gmail.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Co-authored-by: Robert Shaw <robertgshaw2@gmail.com>
This commit is contained in:
bnellnm
2026-05-12 09:18:27 -04:00
committed by GitHub
co-authored by Robert Shaw Robert Shaw
parent 8f89381fc6
commit 206eaed08d
19 changed files with 595 additions and 325 deletions
@@ -195,7 +195,7 @@ class CompressedTensorsW4A4Mxfp4MoEMethod(CompressedTensorsMoEMethod):
experts_cls=self.experts_cls,
mxfp4_backend=self.mxfp4_backend,
shared_experts=layer.shared_experts,
routing_tables=layer._maybe_init_expert_routing_tables(),
routing_tables=layer._expert_routing_tables(),
)
def apply(
@@ -236,7 +236,7 @@ class CompressedTensorsW4A4Nvfp4MoEMethod(CompressedTensorsMoEMethod):
moe_config=self.moe,
experts_cls=self.experts_cls,
shared_experts=layer.shared_experts,
routing_tables=layer._maybe_init_expert_routing_tables(),
routing_tables=layer._expert_routing_tables(),
)
self.moe_kernel.fused_experts.process_weights_after_loading(layer)
@@ -336,7 +336,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
moe_config=self.moe,
fp8_backend=self.fp8_backend,
experts_cls=self.experts_cls,
routing_tables=layer._maybe_init_expert_routing_tables(),
routing_tables=layer._expert_routing_tables(),
shared_experts=layer.shared_experts,
)
@@ -147,7 +147,7 @@ class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):
moe_quant_config=self.moe_quant_config,
moe_config=self.moe,
experts_cls=self.experts_cls,
routing_tables=layer._maybe_init_expert_routing_tables(),
routing_tables=layer._expert_routing_tables(),
shared_experts=layer.shared_experts,
)
@@ -138,7 +138,7 @@ class CompressedTensorsW8A8Mxfp8MoEMethod(CompressedTensorsMoEMethod):
moe_config=self.moe,
fp8_backend=self.fp8_backend,
experts_cls=self.experts_cls,
routing_tables=layer._maybe_init_expert_routing_tables(),
routing_tables=layer._expert_routing_tables(),
shared_experts=layer.shared_experts,
)
@@ -786,7 +786,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
moe_config=self.moe,
fp8_backend=self.fp8_backend,
experts_cls=self.experts_cls,
routing_tables=layer._maybe_init_expert_routing_tables(),
routing_tables=layer._expert_routing_tables(),
shared_experts=layer.shared_experts,
)
@@ -769,7 +769,7 @@ class GPTQMarlinMoEMethod(FusedMoEMethodBase):
w2_g_idx=layer.w2_g_idx,
w13_g_idx_sort_indices=layer.w13_g_idx_sort_indices,
w2_g_idx_sort_indices=layer.w2_g_idx_sort_indices,
routing_tables=layer._maybe_init_expert_routing_tables(),
routing_tables=layer._expert_routing_tables(),
shared_experts=layer.shared_experts,
)
@@ -901,7 +901,7 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
moe_config=self.moe,
fp8_backend=self.fp8_backend,
experts_cls=self.experts_cls,
routing_tables=layer._maybe_init_expert_routing_tables(),
routing_tables=layer._expert_routing_tables(),
shared_experts=layer.shared_experts,
)
@@ -1590,7 +1590,7 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
moe_config=self.moe,
experts_cls=self.experts_cls,
shared_experts=layer.shared_experts,
routing_tables=layer._maybe_init_expert_routing_tables(),
routing_tables=layer._expert_routing_tables(),
)
self.moe_kernel.fused_experts.process_weights_after_loading(layer)
@@ -369,7 +369,7 @@ class GptOssMxfp4MoEMethod(FusedMoEMethodBase):
moe_config=self.moe,
mxfp4_backend=self.mxfp4_backend,
experts_cls=self.experts_cls,
routing_tables=layer._maybe_init_expert_routing_tables(),
routing_tables=layer._expert_routing_tables(),
shared_experts=layer.shared_experts,
layer=layer,
)
@@ -702,7 +702,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
moe_config=self.moe,
mxfp4_backend=self.mxfp4_backend,
experts_cls=self.experts_cls,
routing_tables=layer._maybe_init_expert_routing_tables(),
routing_tables=layer._expert_routing_tables(),
shared_experts=layer.shared_experts,
layer=layer,
)
@@ -354,7 +354,7 @@ class _Fp8OnlineMoEBase(OnlineMoEMethodBase):
moe_config=self.moe,
fp8_backend=self.fp8_backend,
experts_cls=self.experts_cls,
routing_tables=layer._maybe_init_expert_routing_tables(),
routing_tables=layer._expert_routing_tables(),
shared_experts=layer.shared_experts,
)
@@ -99,7 +99,7 @@ class Int8OnlineMoEMethod(OnlineMoEMethodBase):
moe_quant_config=self.moe_quant_config,
moe_config=self.moe,
experts_cls=self.experts_cls,
routing_tables=layer._maybe_init_expert_routing_tables(),
routing_tables=layer._expert_routing_tables(),
shared_experts=layer.shared_experts,
)
@@ -199,7 +199,7 @@ class Mxfp8OnlineMoEMethod(OnlineMoEMethodBase):
moe_config=self.moe,
fp8_backend=self.fp8_backend,
experts_cls=self.experts_cls,
routing_tables=layer._maybe_init_expert_routing_tables(),
routing_tables=layer._expert_routing_tables(),
shared_experts=layer.shared_experts,
)
@@ -1340,7 +1340,7 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
moe_config=self.moe,
mxfp4_backend=self.mxfp4_backend,
experts_cls=self.experts_cls,
routing_tables=layer._maybe_init_expert_routing_tables(),
routing_tables=layer._expert_routing_tables(),
shared_experts=layer.shared_experts,
)