Handle coredump dirs and cache hit updates (#30897)

Co-authored-by: Yongji Wu <30348494+libertyeagle@users.noreply.github.com>

python/sglang/srt/managers/scheduler.py

@@ -2900,9 +2900,9 @@ class Scheduler(
                     # skip staging requests that are ongoing prefetch
                     continue
                 # Pop the number of tokens loaded from storage (L3 hits)
-                req.storage_hit_length = self.tree_cache.pop_prefetch_loaded_tokens(
+                loaded_tokens = self.tree_cache.pop_prefetch_loaded_tokens(req.rid)
-                    req.rid
+                if loaded_tokens > 0:
-                )
+                    req.storage_hit_length = loaded_tokens
 
             req.init_next_round_input(self.tree_cache)
             res = adder.add_one_req(

python/sglang/srt/server_args.py

@@ -6425,14 +6425,26 @@ class ServerArgs:
                 os.environ[key] = value
                 logger.info("Auto-set %s=%s (from --crash-dump-folder)", key, value)
 
-                if key == "CUDA_COREDUMP_FILE":
+        coredump_dir = os.path.dirname(
-                    # cuda curedump cannot write to a folder that does not exist,
+            os.environ["CUDA_COREDUMP_FILE"].replace("%h", socket.gethostname())
-                    # so we have to create the folder first.
+        )
-                    hostname = socket.gethostname()
+        if "%" in coredump_dir:
-                    os.makedirs(
+            logger.warning(
-                        os.path.join(self.crash_dump_folder, hostname),
+                "Cannot pre-create CUDA coredump directory %s: only %%h is "
-                        exist_ok=True,
+                "supported in the directory part of CUDA_COREDUMP_FILE; "
-                    )
+                "coredumps may fail to write.",
+                coredump_dir,
+            )
+        elif coredump_dir:
+            try:
+                os.makedirs(coredump_dir, exist_ok=True)
+            except OSError as e:
+                logger.warning(
+                    "Failed to create CUDA coredump directory %s: %s; "
+                    "coredumps may fail to write.",
+                    coredump_dir,
+                    e,
+                )
 
     def _handle_debug_utils(self):
         if is_in_ci() and self.soft_watchdog_timeout is None:

test/registered/unit/server_args/test_server_args.py

@@ -1,6 +1,7 @@
 import importlib
 import json
 import os
+import socket
 import tempfile
 import unittest
 from types import SimpleNamespace
@@ -1294,6 +1295,48 @@ class TestSamplingBackendTokenOracleEnvGate(CustomTestCase):
         self.assertEqual(parsed.sampling_backend, "token_oracle")
 
 
+class TestHandleCrashDumpEnv(CustomTestCase):
+    _COREDUMP_ENV_KEYS = (
+        "CUDA_ENABLE_COREDUMP_ON_EXCEPTION",
+        "CUDA_ENABLE_USER_TRIGGERED_COREDUMP",
+        "CUDA_COREDUMP_SHOW_PROGRESS",
+        "CUDA_COREDUMP_GENERATION_FLAGS",
+        "CUDA_COREDUMP_FILE",
+        "CUDA_COREDUMP_PIPE",
+    )
+
+    def _run_handler(self, crash_dump_folder, preset_env=None):
+        server_args = ServerArgs.__new__(ServerArgs)
+        server_args.crash_dump_folder = crash_dump_folder
+        with patch.dict(os.environ, preset_env or {}):
+            for key in self._COREDUMP_ENV_KEYS:
+                if key not in (preset_env or {}):
+                    os.environ.pop(key, None)
+            ServerArgs._handle_crash_dump_env(server_args)
+
+    def test_creates_coredump_dir_when_auto_set(self):
+        with tempfile.TemporaryDirectory() as tmp:
+            self._run_handler(tmp)
+            self.assertTrue(
+                os.path.isdir(os.path.join(tmp, socket.gethostname())),
+                "coredump dir not created for auto-set CUDA_COREDUMP_FILE",
+            )
+
+    def test_creates_coredump_dir_when_env_preset(self):
+        # Regression test: when CUDA_COREDUMP_FILE is preset, the coredump
+        # directory must still be created up front.
+        with tempfile.TemporaryDirectory() as tmp:
+            preset_dir = os.path.join(tmp, "preset-location")
+            self._run_handler(
+                tmp,
+                preset_env={"CUDA_COREDUMP_FILE": f"{preset_dir}/%h/core.cuda.%t.%p"},
+            )
+            self.assertTrue(
+                os.path.isdir(os.path.join(preset_dir, socket.gethostname())),
+                "coredump dir not created for preset CUDA_COREDUMP_FILE",
+            )
+
+
 class TestGrpcServerArgs(CustomTestCase):
     """Native gRPC is enabled by --grpc-port (or SGLANG_GRPC_PORT) and runs
     alongside HTTP; --smg-grpc-mode (and the deprecated --grpc-mode) select the