[#8245][feat] Autodeploy: Guided Decoding Support

[govind-ramnarayan]

Summary by CodeRabbit

• New Features

  • Added guided decoding support to AutoDeploy with JSON schema and regex constraint options.

• Tests

  • Added unit and integration tests for guided decoding functionality.
  • Added end-to-end tests for regex-constrained and JSON-formatted generation.

Description

Implements guided decoding in AutoDeploy. Main code change (thanks to @2ez4bz ) is to simply pass the guided decoding backend already in the TRT-LLM backend in the PyExecutor created by the AutoDeploy backend.

Test Coverage

Testing Plan:

• Unit test for create_autodeploy_executor with Mocking to test that the guided decoding config makes it into the PyExecutor class when a guided_decoding_backend is in the input LlmArgs.
• Lightweight integration test for guided decoding with regex generation. The test checks that if the output must follow some regex, then a prefix of the output is matched by the network. Since the network is untrained, it is possible that it will only be able to generate a prefix of some regexes so we check that a prefix is matched.
• Full integration test that loads a model checkpoint. Currently in the unit test folder, should be moved to be with integration tests in test_llm_examples.py

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)

• Any new dependencies have been scanned for license and vulnerabilities

• CODEOWNERS updated if ownership changes

• Documentation updated as needed

• 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

/bot [-h] ['run', 'kill', 'skip', 'reuse-pipeline'] ...

Provide a user friendly way for developers to interact with a Jenkins server.

Run /bot [-h|--help] to print this help message.

See details below for each supported subcommand.

Details

run [--reuse-test (optional)pipeline-id --disable-fail-fast --skip-test --stage-list "A10-PyTorch-1, xxx" --gpu-type "A30, H100_PCIe" --test-backend "pytorch, cpp" --add-multi-gpu-test --only-multi-gpu-test --disable-multi-gpu-test --post-merge --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" --detailed-log --debug(experimental)]

Launch build/test pipelines. All previously running jobs will be killed.

--reuse-test (optional)pipeline-id (OPTIONAL) : Allow the new pipeline to reuse build artifacts and skip successful test stages from a specified pipeline or the last pipeline if no pipeline-id is indicated. If the Git commit ID has changed, this option will be always ignored. The DEFAULT behavior of the bot is to reuse build artifacts and successful test results from the last pipeline.

--disable-reuse-test (OPTIONAL) : Explicitly prevent the pipeline from reusing build artifacts and skipping successful test stages from a previous pipeline. Ensure that all builds and tests are run regardless of previous successes.

--disable-fail-fast (OPTIONAL) : Disable fail fast on build/tests/infra failures.

--skip-test (OPTIONAL) : Skip all test stages, but still run build stages, package stages and sanity check stages. Note: Does NOT update GitHub check status.

--stage-list "A10-PyTorch-1, xxx" (OPTIONAL) : Only run the specified test stages. Examples: "A10-PyTorch-1, xxx". Note: Does NOT update GitHub check status.

--gpu-type "A30, H100_PCIe" (OPTIONAL) : Only run the test stages on the specified GPU types. Examples: "A30, H100_PCIe". Note: Does NOT update GitHub check status.

--test-backend "pytorch, cpp" (OPTIONAL) : Skip test stages which don't match the specified backends. Only support [pytorch, cpp, tensorrt, triton]. Examples: "pytorch, cpp" (does not run test stages with tensorrt or triton backend). Note: Does NOT update GitHub pipeline status.

--only-multi-gpu-test (OPTIONAL) : Only run the multi-GPU tests. Note: Does NOT update GitHub check status.

--disable-multi-gpu-test (OPTIONAL) : Disable the multi-GPU tests. Note: Does NOT update GitHub check status.

--add-multi-gpu-test (OPTIONAL) : Force run the multi-GPU tests in addition to running L0 pre-merge pipeline.

--post-merge (OPTIONAL) : Run the L0 post-merge pipeline instead of the ordinary L0 pre-merge pipeline.

--extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" (OPTIONAL) : Run the ordinary L0 pre-merge pipeline and specified test stages. Examples: --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx".

--detailed-log (OPTIONAL) : Enable flushing out all logs to the Jenkins console. This will significantly increase the log volume and may slow down the job.

--debug (OPTIONAL) : Experimental feature. Enable access to the CI container for debugging purpose. Note: Specify exactly one stage in the stage-list parameter to access the appropriate container environment. Note: Does NOT update GitHub check status.

For guidance on mapping tests to stage names, see docs/source/reference/ci-overview.md
and the scripts/test_to_stage_mapping.py helper.

kill

kill

Kill all running builds associated with pull request.

skip

skip --comment COMMENT

Skip testing for latest commit on pull request. --comment "Reason for skipping build/test" is required. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.

reuse-pipeline

reuse-pipeline

Reuse a previous pipeline to validate current commit. This action will also kill all currently running builds associated with the pull request. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.

[coderabbitai]

📝 Walkthrough

Walkthrough

The pull request adds guided decoding support to AutoDeploy by propagating vocab_size_padded through the model factory and engine, integrating tokenizer and GuidedDecoder into the executor pipeline, and introducing unit and integration tests for regex and JSON-guided generation.

Changes

Cohort / File(s)	Summary
Model Configuration tensorrt_llm/_torch/auto_deploy/models/hf.py	Adds vocab_size_padded property to AutoModelForCausalLMFactory, retrieving and exposing padded vocabulary size from model config.
Executor Infrastructure tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py	Extends ADEngine.__init__ to accept optional vocab_size_padded parameter and exposes via property. Updates create_autodeploy_executor to accept optional tokenizer, conditionally instantiates GuidedDecoder when guided decoding is enabled, and validates availability of vocab_size_padded for guided decoding.
Worker Integration tensorrt_llm/executor/base_worker.py	Passes tokenizer to autodeploy executor via args["tokenizer"] = self._tokenizer in the autodeploy backend path.
Unit Test - Guided Decoding tests/unittest/_torch/auto_deploy/unit/singlegpu/shim/test_guided_decoding.py	Introduces unit test for create_autodeploy_executor with mock tokenizer, guided decoding config, and decoder. Validates proper instantiation of GuidedDecoder and PyExecutor with expected state.
Integration Tests tests/unittest/_torch/auto_deploy/unit/singlegpu/test_ad_guided_decoding_integration.py, test_ad_guided_decoding_regex.py	Two integration test modules: JSON-guided generation validates outputs against schema; regex-guided generation validates outputs against regex constraints across parameterized backends.

Sequence Diagram(s)

sequenceDiagram
    participant Factory as AutoModelForCausalLMFactory
    participant Engine as ADEngine
    participant Executor as create_autodeploy_executor
    participant Decoder as GuidedDecoder
    participant PyEx as PyExecutor

    Factory->>Engine: vocab_size_padded
    Executor->>Engine: build_from_config(ad_config)
    Engine-->>Executor: vocab_size_padded
    
    rect rgb(200, 220, 240)
    note over Executor: Guided Decoding Path
    Executor->>Executor: guided_decoding_backend set?
    alt Yes & on_last_pp_rank
        Executor->>Decoder: new GuidedDecoder(tokenizer, config)
        Decoder->>Decoder: store vocab_size_padded
        Executor->>PyEx: __init__(..., guided_decoder)
    else No
        Executor->>PyEx: __init__(..., guided_decoder=None)
    end
    end
    
    Executor-->>PyEx: return instance

Loading

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

The changes span multiple files with distinct responsibilities: simple property addition in model factory, moderately complex guided decoding integration logic in executor with conditional instantiation and validation, straightforward tokenizer propagation in worker, and three test files with comprehensive coverage. The heterogeneity of changes across configuration, logic, and testing requires careful verification of each component, but the logic density is moderate and patterns are relatively consistent.

Pre-merge checks and finishing touches

❌ Failed checks (2 warnings)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 66.67% which is insufficient. The required threshold is 80.00%.	You can run @coderabbitai generate docstrings to improve docstring coverage.
Description Check	⚠️ Warning	The PR description still contains the original instruction comment block for writing the PR title instead of providing a formatted title or using the @coderabbitai summary header as required by the repository template. While it includes Description, Test Coverage, and Checklist sections, the Test Coverage section does not clearly enumerate the relevant tests under the specified heading and the summary placeholder remains unpopulated. Consequently, the description does not conform to the prescribed structure and formatting outlined in the repository template.	Please remove the instruction comment block and include a properly formatted PR title and summary at the top following the template (for example, using @coderabbitai summary or an explicit title line). Populate the Test Coverage section with a clear list of the specific test cases that safeguard the changes. Verify that all template sections (Description, Test Coverage, PR Checklist) are fully completed and adhere to the repository’s required structure.

✅ Passed checks (1 passed)

Check name	Status	Explanation
Title Check	✅ Passed	The pull request title "[#8245][feat] Autodeploy: Guided Decoding Support" clearly and specifically communicates the primary change in the changeset. It follows the repository's naming conventions with the GitHub issue reference and type tag, and directly summarizes the main feature being added: guided decoding support for the AutoDeploy backend. The title is concise, meaningful, and would be immediately clear to a teammate reviewing the git history.

✨ Finishing touches

• 📝 Generate docstrings

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 4

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py (1)

1-4: Add NVIDIA Apache-2.0 header (2025).

This file is missing the required copyright header. Please prepend it.

Apply this diff at the top of the file:

+# Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#     http://www.apache.org/licenses/LICENSE-2.0
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

🧹 Nitpick comments (9)

tensorrt_llm/_torch/auto_deploy/models/hf.py (2)

127-130: Avoid redundant config fetch in init (potential perf hit).

Calling _get_model_config() here can duplicate work done later in build/load paths. Consider deferring until first access or caching the result to avoid repeated HF config downloads/parsing.

---

135-138: Verify “padded” semantics vs raw vocab size.

vocab_size_padded currently mirrors config.vocab_size. In many stacks the padded size differs (e.g., multiple-of-8/64). Please confirm this is truly padded; otherwise either compute the padded value where known or rename to avoid confusion.

tests/unittest/_torch/auto_deploy/unit/singlegpu/test_ad_guided_decoding_integration.py (2)

41-46: Gate the test for CUDA and consider marking as integration/slow.

This loads a real checkpoint and uses the AutoDeploy backend; it will fail or be flaky on CPU-only CI or constrained runners. Add a CUDA skip and, ideally, move to an integration suite.

Apply this diff near the imports:

+ import torch
+ 
+ pytestmark = pytest.mark.skipif(
+     not torch.cuda.is_available(),
+     reason="CUDA-required AutoDeploy integration test",
+ )

As per coding guidelines.

---

46-51: Stabilize output for CI.

For deterministic behavior, set temperature=0.0. Guided decoding enforces structure, but lower entropy reduces flakes.

-        temperature=0.1,
+        temperature=0.0,

tests/unittest/_torch/auto_deploy/unit/singlegpu/shim/test_guided_decoding.py (1)

61-62: Instantiate the mock tokenizer.

Use an instance to better reflect production usage and avoid passing a type.

-    mock_tokenizer = MockTokenizer
+    mock_tokenizer = MockTokenizer()

tests/unittest/_torch/auto_deploy/unit/singlegpu/test_ad_guided_decoding_regex.py (2)

24-32: Optionally skip when guided decoding backend isn’t available.

If xgrammar/llguidance backends are optional builds, add a skip to prevent false CI failures when not installed.

+ import importlib
+ 
+ if guided_decoding_backend == "xgrammar":
+     pytest.skipif(importlib.util.find_spec("xgrammar") is None, reason="xgrammar not installed")
+ elif guided_decoding_backend == "llguidance":
+     pytest.skipif(importlib.util.find_spec("llguidance") is None, reason="llguidance not installed")

---

89-103: Also validate against the regex (not just prefixes).

Complement the prefix check with a regex match to align with the test intent.

-                is_valid_prefix = any(
+                is_valid_prefix = any(
                     response.startswith(generated_text) or generated_text.startswith(response)
                     for response in valid_responses
                 )
 
-                assert is_valid_prefix, (
+                import re
+                regex_ok = re.match(regex_pattern, generated_text) is not None
+                assert is_valid_prefix or regex_ok, (
                     f"Test case {test_idx + 1}, Output {i} is not a valid prefix of '{valid_responses}'\n"
                     f"Generated text: '{generated_text}'"
                 )

tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py (2)

8-12: Prefer module-namespace imports per guidelines.

Import modules and reference symbols via the module to keep namespaces clear.

Apply this diff and update call sites below:

-from tensorrt_llm._torch.pyexecutor.guided_decoder import GuidedDecoder
-from tensorrt_llm._torch.pyexecutor.py_executor_creator import get_guided_decoding_config
+from tensorrt_llm._torch.pyexecutor import guided_decoder as guided_decoder_mod
+from tensorrt_llm._torch.pyexecutor import py_executor_creator as pyexec_creator

And later replace:

• GuidedDecoder( → guided_decoder_mod.GuidedDecoder(
• get_guided_decoding_config( → pyexec_creator.get_guided_decoding_config(

---

170-175: Expose a brief docstring for the new property.

Add a one-line docstring to clarify units/meaning (padded vocab size used for bitmasking).

Example:

 @property
 def vocab_size_padded(self) -> Optional[int]:
-    return self._vocab_size_padded
+    """Model vocabulary size padded to kernel alignment; required for guided decoding."""
+    return self._vocab_size_padded

📜 Review details

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

📥 Commits

Reviewing files that changed from the base of the PR and between 8dc4aac and 1694ac3.

📒 Files selected for processing (6)

• tensorrt_llm/_torch/auto_deploy/models/hf.py (1 hunks)
• tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py (7 hunks)
• tensorrt_llm/executor/base_worker.py (1 hunks)
• tests/unittest/_torch/auto_deploy/unit/singlegpu/shim/test_guided_decoding.py (1 hunks)
• tests/unittest/_torch/auto_deploy/unit/singlegpu/test_ad_guided_decoding_integration.py (1 hunks)
• tests/unittest/_torch/auto_deploy/unit/singlegpu/test_ad_guided_decoding_regex.py (1 hunks)

🧰 Additional context used

📓 Path-based instructions (3)

**/*.{h,hpp,hh,hxx,cpp,cxx,cc,cu,cuh,py}

📄 CodeRabbit inference engine (CODING_GUIDELINES.md)

Use only spaces, no tabs; indent with 4 spaces.

Files:

• tests/unittest/_torch/auto_deploy/unit/singlegpu/shim/test_guided_decoding.py
• tests/unittest/_torch/auto_deploy/unit/singlegpu/test_ad_guided_decoding_integration.py
• tensorrt_llm/_torch/auto_deploy/models/hf.py
• tensorrt_llm/executor/base_worker.py
• tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py
• tests/unittest/_torch/auto_deploy/unit/singlegpu/test_ad_guided_decoding_regex.py

**/*.py

📄 CodeRabbit inference engine (CODING_GUIDELINES.md)

**/*.py: Python code must target Python 3.8+.
Indent Python code with 4 spaces; do not use tabs.
Maintain module namespace when importing; prefer 'from package.subpackage import foo' then 'foo.SomeClass()' instead of importing the class directly.
Python filenames should be snake_case (e.g., some_file.py).
Python classes use PascalCase names.
Functions and methods use snake_case names.
Local variables use snake_case; prefix 'k' for variables that start with a number (e.g., k_99th_percentile).
Global variables use upper SNAKE_CASE prefixed with 'G' (e.g., G_MY_GLOBAL).
Constants use upper SNAKE_CASE (e.g., MY_CONSTANT).
Avoid shadowing variables from an outer scope.
Initialize all externally visible members of a class in the constructor.
Prefer docstrings for interfaces that may be used outside a file; comments for in-function or file-local interfaces.
Use Google-style docstrings for classes and functions (Sphinx-parsable).
Document attributes and variables inline so they render under the class/function docstring.
Avoid reflection when a simpler, explicit approach suffices (e.g., avoid dict(**locals()) patterns).
In try/except, catch the most specific exceptions possible.
For duck-typing try/except, keep the try body minimal and use else for the main logic.

Files:

• tests/unittest/_torch/auto_deploy/unit/singlegpu/shim/test_guided_decoding.py
• tests/unittest/_torch/auto_deploy/unit/singlegpu/test_ad_guided_decoding_integration.py
• tensorrt_llm/_torch/auto_deploy/models/hf.py
• tensorrt_llm/executor/base_worker.py
• tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py
• tests/unittest/_torch/auto_deploy/unit/singlegpu/test_ad_guided_decoding_regex.py

**/*.{cpp,cxx,cc,h,hpp,hh,hxx,cu,cuh,py}

📄 CodeRabbit inference engine (CODING_GUIDELINES.md)

Prepend the NVIDIA Apache-2.0 copyright header with current year to the top of all source files (e.g., .cpp, .h, .cu, .py).

Files:

• tests/unittest/_torch/auto_deploy/unit/singlegpu/shim/test_guided_decoding.py
• tests/unittest/_torch/auto_deploy/unit/singlegpu/test_ad_guided_decoding_integration.py
• tensorrt_llm/_torch/auto_deploy/models/hf.py
• tensorrt_llm/executor/base_worker.py
• tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py
• tests/unittest/_torch/auto_deploy/unit/singlegpu/test_ad_guided_decoding_regex.py

🧬 Code graph analysis (5)

tests/unittest/_torch/auto_deploy/unit/singlegpu/shim/test_guided_decoding.py (2)

tensorrt_llm/_torch/auto_deploy/llm_args.py (1)

• LlmArgs (303-417)

tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py (1)

• create_autodeploy_executor (314-457)

tests/unittest/_torch/auto_deploy/unit/singlegpu/test_ad_guided_decoding_integration.py (1)

tensorrt_llm/sampling_params.py (1)

• GuidedDecodingParams (15-37)

tensorrt_llm/_torch/auto_deploy/models/hf.py (1)

tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py (1)

• vocab_size_padded (173-174)

tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py (4)

tensorrt_llm/_torch/pyexecutor/guided_decoder.py (1)

• GuidedDecoder (138-402)

tensorrt_llm/_torch/pyexecutor/py_executor_creator.py (1)

• get_guided_decoding_config (183-199)

tensorrt_llm/llmapi/tokenizer.py (1)

• TokenizerBase (24-25)

tensorrt_llm/_torch/auto_deploy/models/hf.py (1)

• vocab_size_padded (136-137)

tests/unittest/_torch/auto_deploy/unit/singlegpu/test_ad_guided_decoding_regex.py (2)

tests/unittest/_torch/auto_deploy/_utils_test/_model_test_utils.py (1)

• get_small_model_config (508-547)

tensorrt_llm/sampling_params.py (1)

• GuidedDecodingParams (15-37)

🪛 Ruff (0.14.1)

tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py

429-431: Avoid specifying long messages outside the exception class

(TRY003)

⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)

• GitHub Check: Pre-commit Check

🔇 Additional comments (3)

tensorrt_llm/executor/base_worker.py (1)

131-134: Passing tokenizer into AutoDeploy executor looks correct.

Matches the PyTorch path and aligns with create_autodeploy_executor(..., tokenizer=...). Please just confirm that tokenizer=None is handled gracefully when guided decoding is disabled.

tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py (2)

455-455: PyExecutor wiring looks correct.

Passing guided_decoder into PyExecutor completes the integration.

---

120-126: Review comment is incorrect. The suggested defensive callable check is unnecessary because all vocab_size_padded definitions across the codebase are consistently @property-decorated, not plain methods. When accessed via getattr(), a property automatically returns its value (int or None), not a function object.

Likely an incorrect or invalid review comment.

[coderabbitai]

Caution

Docstrings generation - FAILED

An unexpected error occurred while opening a pull request: Reference update failed - https://docs.github.com/rest/git/refs#create-a-reference

[govind-ramnarayan]

/bot run

[tensorrt-cicd]

PR_Github #22328 [ run ] triggered by Bot. Commit: af359e2

[tensorrt-cicd]

PR_Github #22328 [ run ] completed with state SUCCESS. Commit: af359e2
/LLM/main/L0_MergeRequest_PR pipeline #16833 completed with status: 'FAILURE'

[govind-ramnarayan]

/bot run

[tensorrt-cicd]

PR_Github #22465 [ run ] triggered by Bot. Commit: 14a94a2

[tensorrt-cicd]

PR_Github #22470 [ kill ] triggered by Bot. Commit: 25dd086

[tensorrt-cicd]

PR_Github #22469 [ run ] completed with state ABORTED. Commit: 25dd086

[tensorrt-cicd]

PR_Github #22470 [ kill ] completed with state SUCCESS. Commit: 25dd086
Successfully killed previous jobs for commit 25dd086

[govind-ramnarayan]

/bot run

[tensorrt-cicd]

PR_Github #22472 [ run ] triggered by Bot. Commit: d62eb7a

[tensorrt-cicd]

PR_Github #22472 [ run ] completed with state SUCCESS. Commit: d62eb7a
/LLM/main/L0_MergeRequest_PR pipeline #16937 completed with status: 'FAILURE'

[govind-ramnarayan]

/bot run

[tensorrt-cicd]

PR_Github #22481 [ run ] triggered by Bot. Commit: 371bb95

[tensorrt-cicd]

PR_Github #22481 [ run ] completed with state SUCCESS. Commit: 371bb95
/LLM/main/L0_MergeRequest_PR pipeline #16941 completed with status: 'FAILURE'

[govind-ramnarayan]

/bot run

[tensorrt-cicd]

PR_Github #22673 [ run ] triggered by Bot. Commit: 371bb95

[tensorrt-cicd]

PR_Github #22673 [ run ] completed with state SUCCESS. Commit: 371bb95
/LLM/main/L0_MergeRequest_PR pipeline #17093 completed with status: 'SUCCESS'

[Superjomn]

LGTM on the llmapi changes

[2ez4bz]

🎉