test: re-enable FlashInfer per-token NVFP4 coverage (#36985)

test/registered/backends/test_flashinfer_nvfp4_online_moe_backend.py

@@ -77,14 +77,6 @@ class FlashinferNvFp4OnlineMoeBackendBase:
             )
 
 
-# Only this class is affected, but the file runs with failfast, so leaving it
-# enabled also cuts off the class that sorts after it.
-@unittest.skip(
-    "flashinfer-ai/flashinfer#4486: on SM100/SM103 the TRTLLM_GEN tile-192 BMM "
-    "path returns non-finite MoE output from FlashInfer 0.6.16.post4 on, so the "
-    "first real prefill trips the sampler NaN assert and gsm8k scores 0.0. "
-    "See #34629 for the package bisect."
-)
 class TestFlashinferTrtllmGenMoeBackendNvFp4Online(
     FlashinferNvFp4OnlineMoeBackendBase, CustomTestCase
 ):

test/registered/backends/test_flashinfer_trtllm_gen_moe_backend.py

@@ -273,12 +273,6 @@ class TestFlashinferTrtllmGenMoeBackendBF16Routed(
     backend = "flashinfer_trtllm_routed"
 
 
-@unittest.skip(
-    "flashinfer-ai/flashinfer#4486: on SM100/SM103 the TRTLLM_GEN tile-192 BMM "
-    "path returns non-finite MoE output from FlashInfer 0.6.16.post4 on, so the "
-    "first real prefill trips the sampler NaN assert and gsm8k scores 0.0. "
-    "See #34629 for the package bisect."
-)
 class TestFlashinferTrtllmGenMoeBackendNvFp4PerTokenActivationRouted(
     FlashinferTrtllmGenMoeBackendNVFP4Base, CustomTestCase
 ):