[BugFix] Add patch for local_cache_hit calculation in vllm 0.18.0 met… (#941)

## Purpose
Fix negative `local_cache_hit` calculation that can crash Prometheus
counters.
Related vLLM issue: vllm-project/vllm#36755
When preemption occurs with async scheduling, there's a race condition
where:
- `schedule(N+1)` can preempt a request and reset its state
- `update_from_output(N)` reads the already-mutated request state
This can result in `num_external_computed_tokens` exceeding
`num_cached_tokens + recomputed`, causing `local_cache_hit` to become
negative. Prometheus counters then crash with:
ValueError: Counters can only be incremented by non-negative amounts.
## Modifications
- Patch `vllm.v1.metrics.stats.PromptTokenStats.update_from_output` to
wrap `local_cache_hit` calculation with `max(0, ...)`
- Add `v0180/vllm/pc/metrics/stats.py` with the patched method
- Add `v0180/vllm/pc_patch.py` to register the patch
- Update `apply_patch.py` to support vllm 0.18.0 version

Author: sumingZero

ucm/integration/vllm/patch/apply_patch.py

@@ -136,6 +136,9 @@ def apply_all_patches() -> None:
                 if ENABLE_SPARSE:
                     logger.info("UCM patching vllm for sparse...")
                     import ucm.integration.vllm.patch.v0110.vllm.sparse_patch
+            case "0.18.0":
+                logger.info("UCM patching vllm for pc...")
+                import ucm.integration.vllm.patch.v0180.vllm.pc_patch
             case _:
                 pass
 

…integration/vllm/patch/v0180/__init__.py …ration/vllm/patch/v0180/vllm/__init__.pyucm/integration/vllm/patch/v0180/__init__.py renamed to ucm/integration/vllm/patch/v0180/vllm/__init__.py ucm/integration/vllm/patch/v0180/__init__.py renamed to ucm/integration/vllm/patch/v0180/vllm/__init__.py

@@ -0,0 +1,17 @@
+def update_from_output(
+    self,
+    num_cached_tokens: int,
+    num_external_computed_tokens: int,
+    prompt_len: int,
+) -> None:
+    """Update stats from a prefill output."""
+    recomputed = 1 if (num_cached_tokens + 1 == prompt_len) else 0
+
+    self.computed += prompt_len - num_cached_tokens
+    self.external_kv_transfer += num_external_computed_tokens
+    self.local_cache_hit += max(
+        0, num_cached_tokens + recomputed - num_external_computed_tokens
+    )
+    self.cached_tokens += num_cached_tokens
+    self.recomputed_tokens += recomputed
+    self.total += prompt_len

ucm/integration/vllm/patch/v0180/vllm/pc/__init__.py

@@ -0,0 +1,17 @@
+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.v1.metrics.stats")
+def patch_stats(mod):
+    logger.debug(f"Patched {mod} called")
+
+    from ucm.integration.vllm.patch.v0180.vllm.pc.metrics import stats
+
+    patch_or_inject(
+        mod.PromptTokenStats,
+        "update_from_output",
+        stats.update_from_output,
+    )