[feat] Add patch support and required connector interfaces (#1016)

## Purpose
Add patch support for vLLM 0.19.1 with connector interfaces
## Modifications 
- Update the supported vLLM version from 0.19.0 to 0.19.1.
- Add missing connector methods for vllm-ascend:
  - get_finished
  - update_connector_output
## Test

Author: wangwenxin0312

ucm/integration/vllm/patch/apply_patch.py

@@ -114,7 +114,7 @@ def get_vllm_version() -> Optional[str]:
 
 def get_supported_versions() -> list[str]:
     """Get patch-required vLLM versions."""
-    return ["0.11.0", "0.17.0", "0.18.0", "0.19.0"]
+    return ["0.11.0", "0.17.0", "0.18.0", "0.19.1"]
 
 
 def apply_all_patches() -> None:
@@ -150,6 +150,9 @@ def apply_all_patches() -> None:
             case "0.18.0":
                 logger.info("UCM patching vllm for pc...")
                 import ucm.integration.vllm.patch.v0180.vllm.pc_patch
+            case "0.19.1":
+                logger.info("UCM patching vllm for pc...")
+                import ucm.integration.vllm.patch.v0191.vllm.pc_patch
             case _:
                 pass
 
@@ -171,9 +174,12 @@ def apply_all_patches() -> None:
             case "0.18.0":
                 logger.info("UCM patching vllm-ascend for pc...")
                 import ucm.integration.vllm.patch.v0180.vllm_ascend.pc_ascend_patch
-            case "0.17.0" | "0.19.0":
+            case "0.17.0":
                 logger.info(f"UCM patching vllm-ascend {ascend_version} for pc...")
                 import ucm.integration.vllm.patch.v0180.vllm_ascend.ucm_connector_patch
+            case "0.19.1":
+                logger.info(f"UCM patching vllm-ascend {ascend_version} for pc...")
+                import ucm.integration.vllm.patch.v0191.vllm_ascend.pc_ascend_patch
             case _:
                 pass
 

ucm/integration/vllm/patch/v0180/vllm_ascend/pc/distributed/kv_transfer/kv_pool/ucm_connector.py

@@ -74,10 +74,21 @@ def request_finished_all_groups(
     return engine.request_finished(request, [])
 
 
+def get_finished(
+    self: Any,
+    finished_req_ids: set[str],
+) -> tuple[set[str] | None, set[str] | None]:
+    return self._call_ucm_reserved_hook("get_finished", finished_req_ids)
+
+
 def build_connector_worker_meta(self: Any) -> Any:
     return self._call_ucm_reserved_hook("build_connector_worker_meta")
 
 
+def update_connector_output(self: Any, connector_output: Any) -> None:
+    return self._call_ucm_reserved_hook("update_connector_output", connector_output)
+
+
 def take_events(self: Any) -> Iterable[Any]:
     events = self._call_ucm_reserved_hook("take_events")
     return () if events is None else events
@@ -142,11 +153,13 @@ def patch_ucm_connector_v1(mod: Any) -> None:
         "request_finished_all_groups",
         request_finished_all_groups,
     )
+    patch_or_inject(target_cls, "get_finished", get_finished)
     patch_or_inject(
         target_cls,
         "build_connector_worker_meta",
         build_connector_worker_meta,
     )
+    patch_or_inject(target_cls, "update_connector_output", update_connector_output)
     patch_or_inject(target_cls, "take_events", take_events)
     patch_or_inject(target_cls, "get_kv_connector_stats", get_kv_connector_stats)
     patch_or_inject(

ucm/integration/vllm/patch/v0191/__init__.py

@@ -0,0 +1 @@
+

ucm/integration/vllm/patch/v0191/vllm/__init__.py

@@ -0,0 +1 @@
+

ucm/integration/vllm/patch/v0191/vllm/pc_patch.py

@@ -0,0 +1,59 @@
+from ucm.integration.vllm.patch.utils import patch_or_inject, when_imported
+from ucm.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+@when_imported("vllm.config.vllm")
+def patch_vllm_config(mod):
+    logger.debug(f"Patched {mod} called")
+
+    from ucm.integration.vllm.patch.v0180.vllm.pc.config import vllm
+
+    vllm.patch_vllm_config_cls(mod.VllmConfig)
+
+
+@when_imported("vllm.model_executor.models.config")
+def patch_model_config(mod):
+    logger.debug(f"Patched {mod} called")
+
+    from ucm.integration.vllm.patch.v0180.vllm.pc.model_executor.models import (
+        config,
+    )
+
+    config.patch_mamba_model_config_cls(mod.MambaModelConfig)
+
+
+@when_imported("vllm.v1.core.sched.scheduler")
+def patch_core_sched_scheduler(mod):
+    logger.debug(f"Patched {mod} called")
+
+    from ucm.integration.vllm.patch.v0180.vllm.pc.v1.core.sched import scheduler
+
+    patch_or_inject(
+        mod.Scheduler,
+        "_mamba_block_aligned_split",
+        scheduler.Scheduler._mamba_block_aligned_split,
+    )
+
+
+@when_imported("vllm.v1.worker.gpu_model_runner")
+def patch_worker_gpu_model_runner(mod):
+    logger.debug(f"Patched {mod} called")
+
+    from ucm.integration.vllm.patch.v0180.vllm.pc.v1.worker import gpu_model_runner
+
+    patch_or_inject(
+        mod.GPUModelRunner,
+        "may_reinitialize_input_batch",
+        gpu_model_runner.GPUModelRunner.may_reinitialize_input_batch,
+    )
+
+
+@when_imported("vllm.v1.worker.mamba_utils")
+def patch_worker_mamba_utils(mod):
+    logger.debug(f"Patched {mod} called")
+
+    from ucm.integration.vllm.patch.v0180.vllm.pc.v1.worker import mamba_utils
+
+    mamba_utils.patch_preprocess_mamba(mod)

ucm/integration/vllm/patch/v0191/vllm_ascend/__init__.py

@@ -0,0 +1 @@
+

ucm/integration/vllm/patch/v0191/vllm_ascend/pc/__init__.py

@@ -0,0 +1 @@
+

ucm/integration/vllm/patch/v0191/vllm_ascend/pc/platform/__init__.py

@@ -0,0 +1 @@
+

ucm/integration/vllm/patch/v0191/vllm_ascend/pc/platform/patch_mamba_config.py

@@ -0,0 +1,19 @@
+import vllm.model_executor.models.config
+
+
+def patch_hybrid_attention_mamba_model_config() -> None:
+    original = (
+        vllm.model_executor.models.config.HybridAttentionMambaModelConfig.verify_and_update_config
+    )
+
+    @classmethod
+    def verify_and_update_config(cls, vllm_config) -> None:
+        original(vllm_config)
+
+        cache_config = vllm_config.cache_config
+        if cache_config.mamba_cache_mode == "align":
+            cache_config.mamba_block_size = cache_config.block_size
+
+    vllm.model_executor.models.config.HybridAttentionMambaModelConfig.verify_and_update_config = (
+        verify_and_update_config
+    )

ucm/integration/vllm/patch/v0191/vllm_ascend/pc_ascend_patch.py

@@ -0,0 +1,16 @@
+import ucm.integration.vllm.patch.v0191.vllm_ascend.ucm_connector_patch  # noqa: F401
+from ucm.integration.vllm.patch.utils import when_imported
+from ucm.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+@when_imported("vllm_ascend.patch.platform.patch_mamba_config")
+def patch_platform_mamba_config(mod):
+    logger.debug(f"Patched {mod} called")
+
+    from ucm.integration.vllm.patch.v0191.vllm_ascend.pc.platform import (
+        patch_mamba_config,
+    )
+
+    patch_mamba_config.patch_hybrid_attention_mamba_model_config()