[P/D] NIXL Updates (#25844)

Signed-off-by: Sage Moore <sage@neuralmagic.com>
Signed-off-by: simon-mo <simon.mo@hey.com>
Signed-off-by: rentianyue-jk <rentianyue-jk@360shuke.com>
Signed-off-by: Russell Bryant <rbryant@redhat.com>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Signed-off-by: Chenheli Hua <huachenheli@outlook.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Signed-off-by: NickLucche <nlucches@redhat.com>
Signed-off-by: Roger Wang <hey@rogerw.io>
Signed-off-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Sage Moore <sage@neuralmagic.com>
Co-authored-by: Russell Bryant <rbryant@redhat.com>
Co-authored-by: rentianyue-jk <rentianyue-jk@360shuke.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Chenheli Hua <huachenheli@outlook.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>
Co-authored-by: Roger Wang <hey@rogerw.io>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
Signed-off-by: simon-mo <simon.mo@hey.com>

vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py

@@ -474,8 +474,11 @@ class NixlConnectorWorker:
                 "backends", ["UCX"])
         # Agent.
         non_ucx_backends = [b for b in self.nixl_backends if b != "UCX"]
-        config = nixl_agent_config(backends=self.nixl_backends) if len(
-            non_ucx_backends) > 0 and nixl_agent_config is not None else None
+        if nixl_agent_config is None:
+            config = None
+        else:
+            config = nixl_agent_config(backends=self.nixl_backends) if len(
+                non_ucx_backends) > 0 else nixl_agent_config(num_threads=8)
 
         self.nixl_wrapper = NixlWrapper(str(uuid.uuid4()), config)
         # Map of engine_id -> {rank0: agent_name0, rank1: agent_name1..}.

vllm/model_executor/models/qwen3_vl_moe.py

@@ -212,6 +212,8 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
                     # attempted to load as other weights later
                     is_expert_weight = True
                     name_mapped = name.replace(weight_name, param_name)
+                    if is_pp_missing_parameter(name_mapped, self):
+                        continue
                     if is_fused_expert:
                         loaded_weight = loaded_weight.transpose(-1,
                                                                 -2)  # no bias
@@ -230,8 +232,6 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):
                                 name_mapped, params_dict, loaded_weight,
                                 shard_id, num_experts)
                     else:
-                        if is_pp_missing_parameter(name_mapped, self):
-                            continue
                         # Skip loading extra parameters for GPTQ/modelopt models
                         if name_mapped.endswith(
                                 ignore_suffixes

vllm/v1/core/sched/scheduler.py

@@ -1288,4 +1288,9 @@ class Scheduler(SchedulerInterface):
             self.finished_recving_kv_req_ids.add(req_id)
         for req_id in (kv_connector_output.finished_sending or ()):
             logger.debug("Finished sending KV transfer for request %s", req_id)
-            self._free_blocks(self.requests[req_id])
+            if req_id not in self.requests:
+                logger.warning(
+                    "Got finished sending KV transfer for request %s,"
+                    "but the request is already freed.", req_id)
+            else:
+                self._free_blocks(self.requests[req_id])