Selective inlining of unresolvable JIT submodule calls.
torch_to_nnef's recursive parser identifies the class behind a
prim::CallMethod via importlib.import_module(qualname.module_path) and
isinstance(submod, ref_cls). JIT artifacts shipped without their training
source (e.g. Silero-VAD's silero_vad.jit) carry qualified names like
__torch__.vad.model.vad_annotator.SileroVadBlock that fail to import.
inline_unresolvable_submodules(graph, model) walks the JIT graph and
inlines exactly those calls in place, leaving importable torch.nn.*
calls intact so existing module-level extractors (LSTM, GRU, RNN, ...)
still fire.
Companion to strip_assertion_ifs in this package: a typical JIT-only
preprocessing chain is inline_unresolvable_submodules -> _jit_pass_dce
-> strip_assertion_ifs.
inline_unresolvable_submodules(graph: 'torch._C.Graph', model: nn.Module, max_passes: int = 1024) -> int
Inline every prim::CallMethod whose target class is not importable.
Iterates until fixed point: an inlined body may itself contain further
CallMethods that also need inlining. max_passes bounds the loop
against pathological non-termination on weirdly structured graphs;
raise it for unusually deep nested inlines, lower it to fail fast
when debugging.
Returns the count of inlined calls.