fix(fx): drop the dynamic_range argument from the 8 remaining mark_as_int8_layer calls

[Anai-Guo]

Problem

13873bd "remove dynamic_range in fx" (#4221) changed

# py/torch_tensorrt/fx/converters/converter_utils.py:710
-def mark_as_int8_layer(layer, dynamic_range):
+def mark_as_int8_layer(layer):

and cleaned up adaptive_avgpool.py, impl/activation.py, maxpool.py,
quantization.py, transformation.py and one of the three sites in
linear.py. Eight call sites were missed and still pass two positional
arguments:

TypeError: mark_as_int8_layer() takes 1 positional argument but 2 were given

Replaying every mark_as_int8_layer(...) call site in fx/converters/ against
the signature parsed out of converter_utils.py on main:

main signature: (layer)

  TypeError add.py:47   (2 positional) -> mark_as_int8_layer() takes 1 positional argument but 2 were given
  TypeError add.py:64   (2 positional) -> ...
  TypeError add.py:70   (2 positional) -> ...
  TypeError batchnorm.py:21  (2 positional) -> ...
  OK        linear.py:27  (1 positional)
  TypeError linear.py:43  (2 positional) -> ...
  TypeError linear.py:51  (2 positional) -> ...
  TypeError mul.py:46   (2 positional) -> ...
  TypeError impl/convolution.py:130  (2 positional) -> ...

8 broken / 1 correct

Each one sits behind an is_quantized / scale is not None guard, so it is
reached by quantized add, add_relu, mul, BatchNorm, Linear and
Conv conversion through the FX path.

Fix

Same idiom #4221 used for linear.py:27: keep the call, drop the argument.
Where the argument was the only consumer of a get_dyn_range(...) result, the
computation and the now-unused import go too.

get_dyn_range itself is kept — quantization.py still uses it.

Note the two shapes #4221 dealt with: sites guarded by if input_val.dynamic_range:
were deleted whole (that attribute is gone in TRT 10), while sites guarded by
if is_quantized: kept the call. All eight sites here are the second kind, so
none of them are deleted.

Checks

black (26.3.1, the pinned pre-commit rev) reports all five files unchanged.
pyflakes is clean on the patched files — the one remaining note,
'typing.Any' imported but unused in impl/convolution.py, is pre-existing on
main and untouched here.

🤖 Generated with Claude Code

[micwill755]

After this change, get_dyn_range has no remaining callers. #4221 already removed the last use in quantization.py and left only the import.

Could you also delete get_dyn_range from converter_utils.py and the unused import in quantization.py? It's leftover from the old ITensor.dynamic_range path which TRT 10 no longer exposes.

[Anai-Guo]

Thanks @micwill755 — done in 367eff1.

• py/torch_tensorrt/fx/converters/converter_utils.py: removed get_dyn_range.
• py/torch_tensorrt/fx/converters/quantization.py: dropped it from the from .converter_utils import ... line (the other name on that line, get_inputs_from_args_and_kwargs, is still used).

Confirmed there are no other references left in the repo — the only remaining hits for get_dyn_range before this commit were the eight mark_as_int8_layer call sites this PR already fixed, the definition itself, and that import.

py/torch_tensorrt/fx is excluded from both black (force-exclude) and isort (skip) in pyproject.toml, so the deletions are formatter-neutral.

[Anai-Guo]

Small note on the red checks here, since they are unrelated to this PR.

executorch-runtime-test / test / executorch-runtime-test--3.10-cu132 fails on a single assertion in tests/py/dynamo/executorch/test_api.py (1 failed, 220 passed):

FAILED executorch/test_api.py::test_runtime_wheel_pins_cuda_13_native_dependencies
  assert 'TENSORRT_DISTRIBUTION = "tensorrt-cu13"' in setup_source

That literal stopped existing on main in #4635 ("bring back cu126 support", 2026-09-02), which replaced the hard-coded constant with TENSORRT_DISTRIBUTION = tensorrt_distribution() in py/torch-tensorrt-executorch-runtime/setup.py. The test still greps for the old string, so it is red for every PR based on current main; #4659 already covers it.

This PR only touches py/torch_tensorrt/fx/converters/{converter_utils,quantization}.py and the fx tests, none of which the ExecuTorch runtime test imports. The gate failure is just the aggregate of that job.

Everything you asked for is in 367eff1; happy to rebase once #4659 lands if you would like a green run before merging.