[BugFix] Update AttnFusionPass cache key (vllm-project#21947)

Signed-off-by: Richard Zou <zou3519@gmail.com>

Author: zou3519

vllm/compilation/fusion_attn.py

@@ -164,3 +164,6 @@ def __call__(self, graph: torch.fx.graph.Graph) -> None:
         logger.debug("Fused quantization onto %s attention nodes", count)
         self.dump_graph(graph, "after_attn_fusion")
         self.end_and_log()
+
+    def uuid(self):
+        return VllmInductorPass.hash_source(self, AttentionStaticQuantPattern)

vllm/compilation/inductor_pass.py

@@ -76,9 +76,10 @@ def hash_source(*srcs: Union[str, Any]):
         for src in srcs:
             if isinstance(src, str):
                 src_str = src
-            elif isinstance(src, types.FunctionType):
+            elif isinstance(src, (types.FunctionType, type)):
                 src_str = inspect.getsource(src)
             else:
+                # object instance
                 src_str = inspect.getsource(src.__class__)
             hasher.update(src_str.encode("utf-8"))
         return hasher.hexdigest()