[AutoDeploy][2/n] Modular InferenceOptimizer

[lucaslie]

moving more transforms into the modular inference optimizer

[Copilot]

Pull Request Overview

This PR refactors the modular inference optimizer by moving low-level cleanup transforms out of the export module and into configurable stages, introduces stricter config typing, updates the InferenceOptimizer API, and adjusts tests to use a shared FakeFactory helper and the new optimizer in the attention matcher tests.

• Remove inline cleanup functions (no-op slice/add/check removals) from export.py, drive them via optimizer config
• Introduce StrictInferenceOptimizerConfig, allow supplying an existing GraphModule to InferenceOptimizer
• Centralize FakeFactory in a test helper and update test_attention_matcher to use InferenceOptimizer

Reviewed Changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/unittest/_torch/auto_deploy/unit/singlegpu/transformations/library/test_kv_cache.py	Remove local FakeFactory and import shared helper
tests/unittest/_torch/auto_deploy/unit/singlegpu/transformations/library/test_attention_matcher.py	Add _get_match_repeat_kv_optimizer using InferenceOptimizer
tests/unittest/_torch/auto_deploy/_utils_test/_graph_test_helpers.py	Add shared FakeFactory
tensorrt_llm/_torch/auto_deploy/transformations/export.py	Remove cleanup transforms and update export signatures
tensorrt_llm/_torch/auto_deploy/transform/optimizer.py	Import strict config type, make gm optional in __call__
tensorrt_llm/_torch/auto_deploy/transform/interface.py	Define StrictInferenceOptimizerConfig and refine config type aliases
tensorrt_llm/_torch/auto_deploy/config/default.yaml	Add cleanup transforms under post_export stage

Comments suppressed due to low confidence (2)

tensorrt_llm/_torch/auto_deploy/transform/optimizer.py:48

• [nitpick] The signature now accepts an optional gm, but the docstring above still only mentions cm. Update the docstring to explain the new gm parameter.

        self, cm: CachedSequenceInterface, gm: Optional[GraphModule] = None

tests/unittest/_torch/auto_deploy/_utils_test/_graph_test_helpers.py:14

• The helper file refers to nn.Module but nn is not imported. Add import torch.nn as nn at the top to avoid a NameError.

    def __init__(self, model: nn.Module):

[Fridah-nv]

LGTM