weight cache: key daemon paths by GPU UUID (#36101)

Co-authored-by: siyu <liusy58@linux.alibaba.com>
Co-authored-by: Alex Nails <alex.nails@radixark.ai>

python/sglang/srt/model_executor/model_runner.py

@@ -1145,13 +1145,8 @@ class ModelRunner:
             weight_cache_socket=get_model().weight_cache_socket,
         )
 
-        # If the weight cache is enabled, override the load format to IPC_CACHE
-        # and derive the per-rank daemon socket. Idempotent across reloads.
         maybe_enable_ipc_weight_cache(
             load_config=self.load_config,
-            tp_size=self.ps.tp_size,
-            pp_rank=self.ps.pp_rank,
-            tp_rank=self.ps.tp_rank,
         )
         if self.device == "cpu":
             self.model_config = adjust_config_with_unaligned_cpu_tp(

python/sglang/srt/model_executor/model_runner_components/load_model_utils.py

@@ -236,16 +236,13 @@ def build_load_config(
 def maybe_enable_ipc_weight_cache(
     *,
     load_config: LoadConfig,
-    tp_size: int,
-    pp_rank: int,
-    tp_rank: int,
 ) -> None:
     """Switch ``load_config`` onto the IPC weight-cache path, in place.
 
     Overrides the load format to ``IPC_CACHE`` (remembering the original as the
-    disk fallback) and derives the per-rank daemon socket if unset. Idempotent:
-    the format swap is guarded on ``!= IPC_CACHE`` so a second call (e.g. a
-    weight reload) can't overwrite the captured fallback format.
+    disk fallback). Idempotent: the format swap is guarded on ``!= IPC_CACHE``
+    so a second call (e.g. a weight reload) can't overwrite the captured
+    fallback format.
     """
     if get_model().weight_cache_mode == "off":
         return
@@ -254,17 +251,6 @@ def maybe_enable_ipc_weight_cache(
         load_config.fallback_load_format = load_config.load_format
         load_config.load_format = LoadFormat.IPC_CACHE
 
-    # Compute socket path using global rank (tp_size * pp_rank + tp_rank) so
-    # each daemon has a unique socket even across PP stages and nodes.
-    if load_config.weight_cache_socket is None:
-        from sglang.srt.weight_cache.protocol import (
-            compute_global_rank,
-            get_socket_path,
-        )
-
-        global_rank = compute_global_rank(tp_size, pp_rank, tp_rank)
-        load_config.weight_cache_socket = get_socket_path(global_rank=global_rank)
-
 
 def load_model_with_memory_saver(
     *,