mirror of
https://github.com/langchain-ai/langgraph.git
synced 2026-08-22 07:32:25 +02:00
1e44bda48f
Fixes langchain-ai/langgraph#8559 `find_subgraph_pregel` runs once per node at build time and recovers each node function's reachable values with `inspect.getsource` + `ast.parse`; langchain-ai/langgraph#8559 measures that source parsing at 80% of `StateGraph.compile()`. This replaces it with a `dis` walk over `func.__code__`, which reads the same information already in memory. Closure cells and the globals named in `co_names` supply the values directly, and a walk over the instruction stream recovers the attribute paths the function actually takes. The closure alone cannot express those: a captured `holder` whose graph lives at `holder.graph` is reachable only if something records that `graph` is loaded off `holder`. Both implementations over-declare — a node can reference a graph it never invokes — and this one over-declares a different set. It no longer reports a graph named only along an attribute path in code the compiler removed; such a path cannot execute, so that entry was always a phantom. Nothing that can actually run stopped being detected. `find_subgraph_pregel` returns the first `PregelProtocol` it finds, so the remaining extra candidates only widen detection. ## Release note Subgraph auto-detection now reads node functions' bytecode instead of parsing their source. Graph builds with many function-backed nodes are substantially faster, and subgraphs are now detected inside functions with no retrievable source — defined in a REPL or notebook cell, or via `exec` — where detection previously failed silently and returned nothing. ## Performance Against [langgraph-build-bench](<https://github.com/soarez/langgraph-build-bench>) — 713 nodes, 500 state fields, 264 tools; CPython 3.12.8, Apple silicon: ``` before build 1412 ms subgraph detection 1132 ms (80%) after build 286 ms subgraph detection 4 ms (2%) ``` **4.9x on total build, \~266x on detection.** The bench prints whole ms; at full precision detection is 1139 ms -> 4.23 ms. Co-authored-by: Elior Nataf Lackritz <elior.nataflackritz@langchain.dev>