[Bugfix][MoE] Tune FlashInfer experts to scheduler token limit (#52989)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Misha Goin
2026-08-20 16:26:24 -04:00
committed by GitHub
co-authored by OpenAI Codex
parent 4f6885fffc
commit bfb6c13499
6 changed files with 12 additions and 1 deletions
+2
View File
@@ -72,6 +72,7 @@ def test_flashinfer_swigluoai_params_are_forwarded(activation, monkeypatch):
moe_parallel_config=FusedMoEParallelConfig.make_no_parallel(),
in_dtype=torch.bfloat16,
routing_method=RoutingMethodType.TopK,
max_num_tokens=16_384,
)
quant_config = FusedMoEQuantConfig.make(
gemm1_alpha=1.702,
@@ -110,6 +111,7 @@ def test_flashinfer_swigluoai_params_are_forwarded(activation, monkeypatch):
assert experts._supports_activation(activation)
assert call_args["activation_type"] == ActivationType.Swiglu
assert call_args["tune_max_num_tokens"] == 16_384
for name, value in (
("swiglu_alpha", 1.702),
("swiglu_beta", 1.0),