[Bugfix][Test] Fix monolithic routing replay test buffer capacity (#50640)
Amir Samani committed
Jul 31, 2026 at 19:37 UTC
c4a4a42d80a50901baec34c39e7c3e8c68e7890b
1 file changed
+3
-3
tests/kernels/moe/test_routed_experts_capture_monolithic.py
+3
-3
@@ -100,7 +100,7 @@ def _make_bf16_monolithic_experts(
100
activation=MoEActivation.SILU,
101
device=device,
102
routing_method=routing_method,
103
- max_num_tokens=max(8, 1),
103
+ max_num_tokens=16,
104
)
105
quant_config = FusedMoEQuantConfig.make(
106
quant_dtype=None,
@@ -518,7 +518,7 @@ def _make_fp8_block_scale_monolithic_experts(
518
activation=MoEActivation.SILU,
519
device=device,
520
routing_method=RoutingMethodType.DeepSeekV3,
521
- max_num_tokens=max(8, 1),
521
+ max_num_tokens=16,
522
)
523
524
# Random fp8 weights + ones-block scales (the kernel decoder only cares
@@ -703,7 +703,7 @@ def _make_nvfp4_monolithic_experts(
703
activation=MoEActivation.SILU,
704
device=device,
705
routing_method=RoutingMethodType.DeepSeekV3,
706
- max_num_tokens=max(8, 1),
706
+ max_num_tokens=16,
707
)
708
709
gemm1 = torch.randn(