[misc] Make NaN-logit sanitization opt-in (default off) (#28829)

Co-authored-by: hnyls2002 <lsyincs@gmail.com>

python/sglang/srt/environ.py

@@ -653,7 +653,7 @@ class Envs:
     SGLANG_ENABLE_ASYNC_ASSERT = EnvBool(False)
     # Sanitize NaN logits before sampling kernels and log a throttled warning
     # (see sanitize_nan_logits).
-    SGLANG_SANITIZE_NAN_LOGITS = EnvBool(True)
+    SGLANG_SANITIZE_NAN_LOGITS = EnvBool(False)
 
     # VLM
     SGLANG_VLM_CACHE_SIZE_MB = EnvInt(100)

scripts/ci/amd/amd_ci_exec.sh

@@ -21,10 +21,6 @@ declare -A ENV_MAP=(
   # the MXFP4 EAGLE-MTP decode path and abort the queue with an HSA hardware
   # exception.
   [SGLANG_ENABLE_ASYNC_ASSERT]=0
-  # Disabled on AMD: the per-step NaN-logit probe/sanitize kernels (#27883) in
-  # the sampler hot path lower single-batch decode fwd_occupancy below the
-  # base-a sanity threshold.
-  [SGLANG_SANITIZE_NAN_LOGITS]=0
   [SGLANG_USE_AITER]=1
 )
 

test/registered/4-gpu-models/test_deepseek_v3_cutedsl_4gpu.py

@@ -146,6 +146,7 @@ class TestDummyWithSBO(CustomTestCase):
                 # exception, crash-time coredump) so NaN is sanitized with a
                 # warning instead of killing the scheduler.
                 "SGLANG_ENABLE_ASYNC_ASSERT": "0",
+                "SGLANG_SANITIZE_NAN_LOGITS": "1",
                 "SGLANG_CUDA_COREDUMP": "0",
                 # Already injected into os.environ by the test process when
                 # SGLANG_CUDA_COREDUMP=1, so it must be overridden explicitly.