[AutoDeploy] Refining AD configurability

[lucaslie]

This PR refines AD configurability in the following ways

• separate out core configs for AD (used in InferenceOptimizer and other core AD classes) into a separate config class
• use multi inheritance to create LlmArgs class for AD to still provide optional full configurability for expert users if needed
• Document use cases for both classes and added improved field validators for some overlapping fields like parallel config
• Limit build_and_run_ad.py CLI to AutoDeployConfig instead of full LlmArgs

New CLI options with reduced configurability:

✗ python build_and_run_ad.py --help                                      

usage: build_and_run_ad.py [-h] [--args [JSON]] [--args.model {str,Path}] [--args.model-factory {AutoModelForCausalLM,AutoModelForImageTextToText}] [--args.model-kwargs Dict[str,Any]] [--args.skip-loading-weights bool]
                           [--args.checkpoint-device {str,null}] [--args.tokenizer {str,Path,null}] [--args.tokenizer-kwargs Dict[str,Any]] [--args.skip-tokenizer-init bool] [--args.disable-overlap-scheduler bool]
                           [--args.mixed-sampler bool] [--args.world-size int] [--args.runtime {demollm,trtllm}] [--args.device str] [--args.kv-cache-dtype str] [--args.attn-backend {flashinfer,triton}]
                           [--args.mla-backend MultiHeadLatentAttention] [--args.free-mem-ratio float] [--args.simple-shard-only bool] [--args.compile-backend {torch-simple,torch-compile,torch-cudagraph,torch-opt}]
                           [--args.cuda-graph-batch-sizes {List[int],null}] [--args.visualize bool] [--args.max-input-len int] [--args.max-num-tokens {int,null}] [--args.max-seq-len int] [--args.max-batch-size int]
                           [--args.attn-page-size int] [--model {str,null}] [--prompt [JSON]] [--prompt.batch-size int] [--prompt.queries {str,List[str]}] [--prompt.sp-kwargs Dict[str,Any]] [--benchmark [JSON]]
                           [--benchmark.enabled bool] [--benchmark.num int] [--benchmark.isl int] [--benchmark.osl int] [--benchmark.bs int] [--benchmark.results-path {str,null}] [--benchmark.store-results bool]
                           [--dry-run | --no-dry-run]

Experiment Configuration for the example script.

    This configuration aggregates all relevant configurations for this example script. It is also
    used to auto-generate the CLI interface.

options:
   [...]

[Copilot]

Pull Request Overview

This PR introduces a dedicated AutoDeployConfig for core AutoDeploy settings, refactors LlmArgs to inherit from it, updates transform and executor components to use the new config, adjusts tests and the example CLI/script to reflect the change, and enhances documentation.

• Introduce AutoDeployConfig and refactor LlmArgs as a subclass
• Update transforms (InferenceOptimizer, KV cache) and executor to accept AutoDeployConfig
• Revise tests, example script, and README to use AutoDeployConfig

Reviewed Changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/.../test_ad_build_small_single.py	Updated test helper to use LlmArgs/AutoDeployConfig mapping
tests/.../test_llm_config.py	Added parallel-config validation tests
tensorrt_llm/.../transformations/transform.py	Switch InferenceOptimizer to accept AutoDeployConfig and return nn.Module
tensorrt_llm/.../transformations/library/kvcache.py	Adjusted typing and signature of insert_cached_attention
tensorrt_llm/.../shim/ad_executor.py	Updated build_from_config to accept AutoDeployConfig
tensorrt_llm/_torch/auto_deploy/llm_args.py	Introduced AutoDeployConfig, reworked LlmArgs inheritance
examples/auto_deploy/build_and_run_ad.py	Switched CLI script to use AutoDeployConfig
examples/auto_deploy/README.md	Updated README to highlight AutoDeployConfig usage

Comments suppressed due to low confidence (3)

tests/unittest/_torch/auto_deploy/unit/singlegpu/test_ad_build_small_single.py:13

• The assertion message still refers to "ad_config" even though the parameter is now named llm_args; please update it for consistency.

def _check_ad_config(experiment_config: ExperimentConfig, llm_args: LlmArgs):

tensorrt_llm/_torch/auto_deploy/transformations/library/kvcache.py:53

• The signature now returns None, but an earlier code path still returns egm; callers expecting a GraphModule may break. Either consistently return the module or update callers to use in-place mutation.

) -> None:

tensorrt_llm/_torch/auto_deploy/llm_args.py:12

• Path is used here but not imported; add from pathlib import Path to avoid a NameError.

PathLike = Union[str, Path]

[nzmora-nvidia]

LGTM

[galagam]

Overall LGTM, but I'm not familiar enough with the code base.