A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 22:35 UTC
[BugFix] Update AttnFusionPass cache key (#21947)
Signed-off-by: Richard Zou <zou3519@gmail.com>
Richard Zou committed
Aug 1, 2025 at 10:11 UTC
8026a335a135af2e53c7d89652863312d7a3c936
2 files changed
+5
-1
vllm/compilation/fusion_attn.py
+3
@@ -164,3 +164,6 @@ class AttnFusionPass(VllmInductorPass):
164
logger.debug("Fused quantization onto %s attention nodes", count)
165
self.dump_graph(graph, "after_attn_fusion")
166
self.end_and_log()
167
+
168
+ def uuid(self):
169
+ return VllmInductorPass.hash_source(self, AttentionStaticQuantPattern)
vllm/compilation/inductor_pass.py
+2
-1
@@ -76,9 +76,10 @@ class InductorPass(CustomGraphPass):
76
for src in srcs:
77
if isinstance(src, str):
78
src_str = src
79
- elif isinstance(src, types.FunctionType):
79
+ elif isinstance(src, (types.FunctionType, type)):
80
src_str = inspect.getsource(src)
81
else:
82
+ # object instance
83
src_str = inspect.getsource(src.__class__)
84
hasher.update(src_str.encode("utf-8"))
85
return hasher.hexdigest()