fix(dynamo): handle dynamic (ITensor) negative start in impl.slice.sl… - #4486
Open
apbose wants to merge 1 commit into
Open
fix(dynamo): handle dynamic (ITensor) negative start in impl.slice.sl…#4486apbose wants to merge 1 commit into
apbose wants to merge 1 commit into
Conversation
github-actions
Bot
added
component: tests
…ice_op
slice_op's negative-index wraparound (start -> start + dim_size) only
fired for a literal Python int start known at trace time. When start is a
TRTTensor instead - a value computed at conversion time by some other
converter's own implementation code, not something that can come from the
FX graph itself - its sign isn't known until the engine runs, so the
wraparound was silently skipped and the raw (possibly negative) value was
fed straight to ISliceLayer, which TensorRT rejects as an out-of-bounds
start.
No currently-registered converter passes a TRTTensor start (aten_ops_slice,
set_item, and embedding.py's slice_op call sites all use literal ints), so
this has no effect on any shipped converter today. It matters for
converter authors: slice_op is a public, reusable impl-layer helper (see
tutorials/extensibility/converters), and the reporter hit this while
building a converter of their own that computes its start dynamically,
resorting to a hand-rolled workaround because the wraparound wasn't
generic.
A model-level regression test isn't possible here: torch.export resolves
the sign of a slice bound via guards at trace time rather than carrying it
as a truly runtime-only-known value. E.g.
class M(torch.nn.Module):
def forward(self, x):
return x[-(x.shape[0] // 2):]
torch.export.export(
M(), (x,),
dynamic_shapes={"x": {0: torch.export.Dim("n", min=2, max=64)}},
)
raises ConstraintViolationError ("Not all values of n ... satisfy the
generated guard (-1)*(x.size()[0] // 2) < 0 ...") - export refuses to
generalize the trace across the declared dynamic range once the sign
isn't fixed for every value in it. So this case can never reach slice_op
through torch_tensorrt.compile() with a compiled torch.nn.Module; it is
only reachable by calling impl.slice.slice_op directly, the same way a
converter author would. The added test (TestSliceOpDynamicStart) does
exactly that: it builds a small TensorRT network by hand and calls
slice_op with a start TRTTensor whose value (-2) is only computed at
runtime, verifying the build fails before this fix and produces the
correct wrapped-around result after it.
Fixes #4186
apbose
force-pushed
the
abose/slice_op_dynamic_negative_start
branch
from
August 13, 2026 21:54
b76b23f to
9c57a30
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes #4186