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

[MoE Refactor] Migrate Unquantized to Full Oracle Flow (#36286)

Browse Source 
Signed-off-by: Yifan Zong <yzong@redhat.com>
Signed-off-by: Robert Shaw <robshaw@redhat.com>
Signed-off-by: yzong-rh <yzong@redhat.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
This commit is contained in:
yzong-rh
2026-03-31 15:43:33 -04:00
committed by GitHub
parent 598190aac3
commit d9b90a07ac
11 changed files with 618 additions and 514 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
tests/kernels/moe/test_moe.py
View File

@@ -1664,7 +1664,7 @@ def test_unquantized_bf16_flashinfer_trtllm_backend(
intermediate_size_per_partition=n,
num_local_experts=e,
num_logical_experts=e,
activation="silu",
activation=MoEActivation.SILU,
device="cuda",
moe_parallel_config=FusedMoEParallelConfig.make_no_parallel(),
in_dtype=dtype,
@@ -1695,13 +1695,25 @@ def test_unquantized_bf16_flashinfer_trtllm_backend(
layer.topk_group = 1
layer.intermediate_size_per_partition = n
layer.ep_rank = 0
layer.activation = "silu"
layer.activation = MoEActivation.SILU
layer.e_score_correction_bias = None
layer.routing_method_type = RoutingMethodType.Renormalize
layer.expert_map = None
layer.apply_router_weight_on_input = False
layer.routed_scaling_factor = None
layer.shared_experts = None
layer._maybe_init_expert_routing_tables = lambda: None
quant_method.process_weights_after_loading(layer)
trtllm_output = quant_method.forward_monolithic_cuda(
assert quant_method.moe_kernel is not None, (
"moe_kernel should be set after process_weights_after_loading"
)
assert quant_method.supports_internal_mk, (
"supports_internal_mk should be True after setup"
)
trtllm_output = quant_method.apply_monolithic(
layer=layer,
x=a,
router_logits=router_logits,