Skip to content

[Bug][Relax] ReorderTakeAfterMatmul fails on vector dot products #20595

Description

@Yuhx141

Description

ReorderTakeAfterMatmul rewrites a valid vector dot product into a scalar MatMul followed by take(axis=-1). A scalar has no valid axis, so the pass raises while constructing the replacement.

Environment

  • TVM 0.26.dev0, executed at e269315c90e3a061c9e1c77b370ce883b1b223f4
  • Ubuntu 24.04 x86-64, LLVM 18.1.3 CPU target
  • Current upstream main checked at d55759ec1ebed554d9782a52e43e5730b8a96656
  • src/relax/transform/reorder_take_after_matmul.cc is byte-identical at both revisions (SHA-256 80324bd216442834d9c181f2466fc6ff8aa5538b93e78251b59f863ce5a6e6ba)
  • I did not have a binary built from current main, so the runtime reproduction below is from the executed revision

Minimal reproduction

from tvm import relax
from tvm.script import ir as I
from tvm.script import relax as R

@I.ir_module
class Module:
    @R.function
    def main(
        lhs: R.Tensor((3,), "float32"),
        weight: R.Tensor((3,), "float32"),
        indices: R.Tensor((3,), "int64"),
    ):
        selected = R.take(weight, indices, axis=0, mode="clip")
        return R.matmul(lhs, selected)

assert relax.analysis.check_well_formed(Module, check_ty=True)
relax.transform.ReorderTakeAfterMatmul()(Module)

Expected behavior

The pass should preserve the result or leave this unsupported rank pattern unchanged.

Actual behavior

The source is well formed, executes on LLVM, and returns a scalar. The pass raises:

ValueError: the input axis -1 is out of range. The input tensor has 0 dimensions

The adjacent rank-two weight control is rewritten successfully and its source and target outputs match.

Likely cause

The rewrite guard accepts a rank-one weight and indices tensor. The replacement MatMul therefore produces a scalar, but the pass constructs take(out_table, indices, matmul_ty->ndim - 1, mode), yielding axis -1 for a rank-zero value.

Duplicate check

I searched open and closed TVM issues and PRs for ReorderTakeAfterMatmul, vector weights, scalar outputs, and axis -1. #20198/#20331 concern out_dtype, and #20201/#20206 concern take mode; I did not find this rank-guard failure.

I can send a fix PR and help with follow-up testing if this diagnosis looks right.

Triage

  • needs-triage
  • type: bug

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions