[Paddle TensorRT] Fix int64 shape tensor in stack/arange converters breaking TRT>=10.8 engine build

[WheelWell9876]

PR Category

Inference

PR Types

Bug fixes

Description

Problem. On TensorRT ≥ 10.8 the Shape op returns int64 (it returned int32 up to 10.0.1).
The pd_op.stack converter builds its output-shape subgraph from a raw network.add_shape(...) and
then concatenates that int64 tensor with the int32 add_1D_constant_layer(network, 1), which breaks the
TensorRT engine build. Paddle already handles this elsewhere via converter_utils.py::trt_shape() (which
casts the Shape result back to int32 on TRT ≥ 10, per its docstring); stack_converter simply wasn't
routed through it. arange_converter has the analogous gap — only its float branch casts the quotient.

Fix (2 files):

• python/paddle/tensorrt/impls/manipulation.py — stack_converter: replace raw
  network.add_shape(...) with trt_shape(...) so the shape tensor is int32 before the
  add_concatenation with the int32 constant.
• python/paddle/tensorrt/impls/creation.py — arange_converter: cast the integer-branch quotient to
  int32 (mirroring the existing float branch).

Verification — the repo's own tests go FAIL → PASS, on Turing and Blackwell. Built the released
paddlepaddle-gpu==3.3.0 (cu129) + tensorrt==10.15.1.29 and ran test/tensorrt/test_converter_*.py:

[Tesla T4, sm_75]  and  [RTX PRO 6000, sm_120]  — identical:

BEFORE this fix:
  test_converter_manipulation.py::TestStackTRTPattern::test_trt_result       FAILED
  test_converter_manipulation.py::TestStackCase2TRTPattern::test_trt_result  FAILED
    [TRT] [E] Error Code 2: graphShapeAnalyzer ... Assertion !isInFlight(p.second.symbolicRep) failed
    E  AttributeError: 'NoneType' object has no attribute 'axis'  (manipulation.py:999)
  => 2 failed, 2 passed

AFTER this fix (same tests):
  test_converter_manipulation.py::TestStackTRTPattern::test_trt_result       PASSED
  test_converter_manipulation.py::TestStackCase2TRTPattern::test_trt_result  PASSED
  => 4 passed

The existing TestStackTRTPattern / TestStackCase2TRTPattern already cover the path, so no new test
is needed — they fail on TRT ≥ 10.8 without this change and pass with it (and pass on TRT < 10.8 either
way, where Shape still returns int32, which is why this regressed silently on CI). In larger graphs the
same int64 shape tensor instead trips Error Code 4: ... incompatible types Int32 and Int64 (observed
with PP-FormulaNet_plus-L).

Note on arange: the existing arange tests use constant inputs and pass before and after, so they
don't independently exercise this; the integer-branch cast is included as the obvious consistency fix
matching the float branch (the same issue surfaces at the model level, e.g. PP-DocLayout-L). Happy to
split it into a follow-up if preferred.

Regression sources: pd_op.stack (#68839), pd_op.arange (#68757). Good cherry-pick candidate for
release/3.3 / release/3.4.

Fixes #79319.

是否引起精度变化 (Does this change precision?)

否 (No). Shape/index tensors are bounded well within int32 range; the change only makes the TensorRT
operand dtype consistent so the engine builds, and is numerically equivalent.

[CLAassistant]

All committers have signed the CLA.

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | pr_review | 2026-06-16 07:21:05

📋 Review 摘要

PR 概述：修复 TensorRT 10.8+ 下 stack/arange converter 中 int64 shape tensor 与 int32 shape subgraph 混用导致的 engine build 问题。
变更范围：python/paddle/tensorrt/impls/creation.py、python/paddle/tensorrt/impls/manipulation.py
影响面 Tag：[Inference]

问题

级别	文件	概述
🟡 建议	python/paddle/tensorrt/impls/creation.py:134	arange 的 runtime tensor integer 分支缺少对应回归测试

📝 PR 规范检查

符合规范。

总体评价

stack_converter 改用 trt_shape() 与现有 shape helper 语义一致，能把 TRT>=10 的 shape 输出统一回 int32；arange_converter 的整数分支 cast 也与 float 分支和后续 zero_tensor 的 int32 约束对齐。主要剩余风险是 arange 现有测试只覆盖常量输入，建议补一个 runtime tensor 输入用例锁住这次修复。

[PaddlePaddle-bot]

🟡 建议 这个 cast 修的是 start/end/step 作为 TensorRT 输入 tensor 时的非 float 分支，但现有 arange converter 测试只用常量输入（test/tensorrt/test_converter_creation.py:109-114 的 feed_list 为空）。常量折叠路径不会稳定覆盖 f_quotient_tensor 在 TRT>=10 上产生 int64 后再和 zero_tensor 做 trt_sub 的运行时 subgraph；后续删除或改坏这行 cast 时，现有测试仍可能通过。

建议修复方式：
新增一个 arange TensorRT 用例，把 start、end、step 放进 feed_list（例如三个 int64、shape 为 [1] 的输入），并设置对应的输入 shape 数据，使 pd_op.arange 的 integer 分支以 runtime tensor 进入 converter；该用例应在 TRT>=10.8 删除这行 cast 时 engine build 失败。