[TRTLLM-15405][refactor] BREAKING: Remove TRTLLMSampler and sampler_type

[zhaoyangwang-nvidia]

@coderabbitai summary

Description

JIRA: TRTLLM-15405

Removal plan

TRTLLMSampler is removed in two steps so that every user-visible change lands
in one small review and the bulk deletion stays a pure dead-code change:

Step	Scope	Size
1	Python sampler + the sampler_type LLM API option, plus tests and docs	47 files, −1391
2	The C++ decoder stack that step 1 leaves unreachable	131 files, −34126

Step 1 carries all of the API risk and is independently correct: afterwards the
C++ decoder still builds, it simply has no caller. Step 2 then deletes it
without touching behaviour.

Step 2 is not split further because the subsystem is mutually recursive —
layers/ includes runtime/decodingLayerWorkspace.h while
runtime/gptDecoder.cpp and decodingLayerWorkspace.h include
layers/decodingParams.h, so deleting either side alone leaves the other
uncompilable.

This PR is step 1.

TorchSampler becomes the only sampling backend. This has been signposted for a while: sampler_type already carries a status="deprecated" marker and a "will be removed in release 1.4" notice, and TorchSampler has supported a superset of TRTLLMSampler's features for several releases (the docs stated as much). auto already resolved to TorchSampler, so the default path is unchanged.

Removed: the TRTLLMSampler class and its SampleState* types, make_decoding_batch_input_output.py, and the helpers only it reached (get_decoding_mode(), update_sampler_max_seq_len()), plus the now-dead trtllm_sampler feature-conflict rules.

API change. The sampler_type option and its SamplerType enum are removed outright. Once TRTLLMSampler is gone, auto and TorchSampler resolve to the same sampler, so the option no longer selects anything — keeping a single-valued knob would only be a way to get it wrong. auto was the default, so code that never set it is unaffected; code that passed sampler_type= (including sampler_type: in a YAML --extra_llm_api_options file) must drop the line. Both consumers are simplified in step: the AutoDeploy shim constructs TorchSampler directly instead of branching, and validate_feature_combination() drops its sampler_type parameter along with the torch_sampler feature flag, which fed no conflict rule. The telemetry allowlist, golden manifest, API-stability reference and docs are updated to match.

This PR touches no C++ at all, so the C++ build is unaffected; the decoder stack it orphans is deleted in step 2.

Test Coverage

Existing tests. Coverage specific to the removed sampler is deleted; everything else is re-pointed at TorchSampler:

• Deleted: tests/unittest/_torch/sampler/test_trtllm_sampler.py, tests/unittest/auto_deploy/singlegpu/smoke/test_ad_trtllm_sampler.py, TestParameterValidation::test_logprobs_trtllm_sampler (asserted a C++-decoder-only error), and test_llmapi_config_capture.py::test_collect_llm_api_config_captures_sampler_type_categorical (load_format still covers the Union[str, Enum] telemetry path).
• Sampler parametrization is dropped entirely (it now has a single value) across test_beam_search.py, test_logits_logprobs.py, test_overlap_scheduler.py, _test_openai_chat.py, _test_openai_completions.py and test_e2e.py. Two guards in test_beam_search.py had degenerated with the parameter (one skipped unconditionally, one never skipped) and are removed rather than left misleading. Test ids shift as a result, so the affected waives.txt and test-db/l0_a30.yml entries are updated, and the test_e2e.py wrappers now filter on -k logit_bias_effect instead of the removed -k torch_sampler, which would otherwise have matched nothing and silently run zero tests.
• test_beam_search.py::test_beam_search_large_beam_width_regression (nvbugs/6242591) is kept and switched to TorchSampler — it exercises beam-slot reuse, which is not sampler-specific.

Local verification: pre-commit passes on all changed files.

PR Checklist

Please review the following before submitting your PR:

• PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.

• PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

• Test cases are provided for new code paths (see test instructions)

• If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.

• Any new dependencies have been scanned for license and vulnerabilities

• CODEOWNERS updated if ownership changes

• Documentation updated as needed

• Update tava architecture diagram if there is a significant design change in PR.

• The reviewers assigned automatically/manually are appropriate for the PR.

• Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.