remove resolve_op_overloads and use splitting_ops directly (#28081)

Signed-off-by: Boyuan Feng <boyuan@meta.com>
This commit is contained in:
Boyuan Feng
2025-11-08 01:13:13 +00:00
committed by GitHub
parent 1aaecda078
commit b158df2813
3 changed files with 89 additions and 65 deletions
+5 -10
View File
@@ -19,7 +19,7 @@ import vllm.envs as envs
from vllm.compilation.inductor_pass import pass_context
from vllm.compilation.partition_rules import (
inductor_partition_rule_context,
resolve_defined_ops,
should_split,
)
from vllm.config import CompilationConfig, CUDAGraphMode, VllmConfig
from vllm.logger import init_logger
@@ -303,7 +303,7 @@ class SplitItem:
def split_graph(
graph: fx.GraphModule, resolved_ops: list[torch._ops.OpOverload]
graph: fx.GraphModule, splitting_ops: list[str]
) -> tuple[fx.GraphModule, list[SplitItem]]:
# split graph by ops
subgraph_id = 0
@@ -312,12 +312,8 @@ def split_graph(
for node in graph.graph.nodes:
if node.op in ("output", "placeholder"):
continue
# Match node.target against resolved_ops
# node.target can be OpOverloadPacket, need to check .default
if node.op == "call_function" and (
node.target in resolved_ops
or (hasattr(node.target, "default") and node.target.default in resolved_ops)
):
if should_split(node, splitting_ops):
subgraph_id += 1
node_to_subgraph_id[node] = subgraph_id
split_op_graphs.append(subgraph_id)
@@ -653,8 +649,7 @@ class VllmBackend:
else:
fx_split_ops = self.compilation_config.splitting_ops or []
resolved_split_ops = resolve_defined_ops(fx_split_ops)
self.split_gm, self.piecewise_graphs = split_graph(graph, resolved_split_ops)
self.split_gm, self.piecewise_graphs = split_graph(graph, fx_split_ops)
from torch._dynamo.utils import lazy_format_graph_code