From c4a4a42d80a50901baec34c39e7c3e8c68e7890b Mon Sep 17 00:00:00 2001 From: Amir Samani Date: Fri, 31 Jul 2026 19:37:39 -0700 Subject: [PATCH] [Bugfix][Test] Fix monolithic routing replay test buffer capacity (#50640) --- tests/kernels/moe/test_routed_experts_capture_monolithic.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/tests/kernels/moe/test_routed_experts_capture_monolithic.py b/tests/kernels/moe/test_routed_experts_capture_monolithic.py index 047438762c..d4a95750ae 100644 --- a/tests/kernels/moe/test_routed_experts_capture_monolithic.py +++ b/tests/kernels/moe/test_routed_experts_capture_monolithic.py @@ -100,7 +100,7 @@ def _make_bf16_monolithic_experts( activation=MoEActivation.SILU, device=device, routing_method=routing_method, - max_num_tokens=max(8, 1), + max_num_tokens=16, ) quant_config = FusedMoEQuantConfig.make( quant_dtype=None, @@ -518,7 +518,7 @@ def _make_fp8_block_scale_monolithic_experts( activation=MoEActivation.SILU, device=device, routing_method=RoutingMethodType.DeepSeekV3, - max_num_tokens=max(8, 1), + max_num_tokens=16, ) # Random fp8 weights + ones-block scales (the kernel decoder only cares @@ -703,7 +703,7 @@ def _make_nvfp4_monolithic_experts( activation=MoEActivation.SILU, device=device, routing_method=RoutingMethodType.DeepSeekV3, - max_num_tokens=max(8, 1), + max_num_tokens=16, ) gemm1 = torch.randn(