pytorch - 💡(How to fix) Fix insert_deferred_runtime_asserts shall not prune all torch checks on inputs [1 participants]

insert_deferred_runtime_asserts (in torch/fx/passes/runtime_assert.py) currently prunes the inline _assert_scalar for input-bound checks because the ShapeEnv range info is considered to already cover them. This is wrong: user-written torch._check calls should never be dropped from the graph body — they're part of the model's semantics, not just hints to the tracer. The pruning is especially harmful under torch.compile (no ep.module() wrapper) where the check then has no runtime effect at all.

Fix should ensure user-emitted asserts are preserved while still allowing dedup of compiler-internal redundant ones.

at least do not prone ones that checks more than >=0?

cc @chauhang @penguinwu @ezyang @bobrenjc93 @aditvenk