Cache EnvField reads to avoid per-call os.getenv in the hot path

EnvField.get() re-read os.environ (os.getenv + parse) on every call. The
attention/compose hot path reads many flags per layer / per compose
(nsa_backend ~11/layer, cp_shared_kv_runtime ~12/compose) -- thousands of
os.getenv + Python-method calls per forward for values that are fixed startup
config.

Cache the parsed value on first read; invalidate on every mutation through the
API (set / override / clear). The only env vars mutated at runtime to pass
information are NOT EnvFields (SGLANG_TMP_NCCL_COMM_VALUE is read by C getenv;
SGLANG_RUN_ID / SGLANG_DP_RANK via raw os.environ), so the cache does not affect
them. Convert the two direct os.environ["SGLANG_*"]=... writes in server_args
(SGLANG_VLM_CACHE_SIZE_MB, SGLANG_ENABLE_DETERMINISTIC_INFERENCE) to envs.X.set()
so they invalidate the cache.

Verified: cache + set/override/clear invalidation behavioral test passes;
331 mem_cache/parser/layers unit tests pass (5 pre-existing CUDA-device-assert
failures unrelated).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

python/sglang/srt/environ.py

@@ -43,6 +43,14 @@ class EnvField:
         # NOTE: environ can only accept str values, so we need a flag to indicate
         # whether the env var is explicitly set to None.
         self._set_to_none = False
+        # Env vars are startup config that does not change mid-run, so the parsed
+        # value is cached after the first read (hot paths read flags per-layer /
+        # per-compose).  Every mutation path through this API (set/override/clear)
+        # invalidates the cache.  The only env vars mutated at runtime to pass
+        # information are NOT EnvFields (e.g. SGLANG_TMP_NCCL_COMM_VALUE, read via
+        # raw getenv), so they are unaffected by this cache.
+        self._cache_valid = False
+        self._cache = None
 
     def __set_name__(self, owner, name):
         assert EnvField._allow_set_name, "Usage like `a = envs.A` is not allowed"
@@ -52,6 +60,13 @@ class EnvField:
         raise NotImplementedError()
 
     def get(self) -> Any:
+        if self._cache_valid:
+            return self._cache
+        self._cache = self._compute()
+        self._cache_valid = True
+        return self._cache
+
+    def _compute(self) -> Any:
         value = os.getenv(self.name)
 
         # Explicitly set to None
@@ -77,6 +92,7 @@ class EnvField:
     def set(self, value: Any):
         self._set_to_none = value is None
         os.environ[self.name] = str(value)
+        self._cache_valid = False
 
     @contextmanager
     def override(self, value: Any):
@@ -90,10 +106,12 @@ class EnvField:
         else:
             os.environ.pop(self.name, None)
         self._set_to_none = backup_set_to_none
+        self._cache_valid = False
 
     def clear(self):
         os.environ.pop(self.name, None)
         self._set_to_none = False
+        self._cache_valid = False
 
     def __bool__(self):
         raise RuntimeError(

python/sglang/srt/server_args.py

@@ -3606,9 +3606,9 @@ class ServerArgs:
             self.enable_deterministic_inference = True
 
             # For VLM
-            os.environ["SGLANG_VLM_CACHE_SIZE_MB"] = "0"
+            envs.SGLANG_VLM_CACHE_SIZE_MB.set("0")
             # TODO remove this environment variable as a whole
-            os.environ["SGLANG_ENABLE_DETERMINISTIC_INFERENCE"] = "1"
+            envs.SGLANG_ENABLE_DETERMINISTIC_INFERENCE.set("1")
 
         if self.enable_deterministic_inference:
             if self.enable_aiter_allreduce_fusion: