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(
-                    req.rid
-                )
+                loaded_tokens = self.tree_cache.pop_prefetch_loaded_tokens(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":
-                    # cuda curedump cannot write to a folder that does not exist,
-                    # so we have to create the folder first.
-                    hostname = socket.gethostname()
-                    os.makedirs(
-                        os.path.join(self.crash_dump_folder, hostname),
-                        exist_ok=True,
-                    )
+        coredump_dir = os.path.dirname(
+            os.environ["CUDA_COREDUMP_FILE"].replace("%h", socket.gethostname())
+        )
+        if "%" in coredump_dir:
+            logger.warning(
+                "Cannot pre-create CUDA coredump directory %s: only %%h is "
+                "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: