[None][feat] AutoDeploy: chunked prefill support

[lucaslie]

Summary by CodeRabbit

• New Features

  • Added optional chunked prefill support, configurable via a new flag. When enabled, request chunking is applied using the attention page size, and scheduling is updated accordingly for improved handling of long contexts.

• Tests

  • Expanded integration tests to parameterize runs with and without chunked prefill across multiple models and environments.
  • Added a unit test validating logits equivalence between chunked and non-chunked prefill paths.
  • Updated test lists to include new parameterized variants and default to non-chunked runs where applicable.

Description

Test Coverage

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.

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.

[lucaslie]

Some long-context prompts to play around with chunked prefill and with some sanity check that the chunked context is correctly read/written:

model: meta-llama/Meta-Llama-3.1-8B-Instruct
args:
  mode: graph
  world_size: 2
  runtime: trtllm
  compile_backend: torch-cudagraph
  attn_backend: flashinfer
  model_factory: AutoModelForCausalLM
  max_batch_size: 32
  attn_page_size: 16
  max_num_tokens: 64
  enable_chunked_prefill: true
  queries:
    - |
      You are an expert assistant designed to provide clear, accurate, and detailed answers across a
      wide range of topics, including technology, science, engineering, and everyday
      problem-solving. Your responses should be structured logically, written in plain and concise
      language, and tailored to the user's level of expertise. Always explain not just what to do
      but also why, offering relevant context, examples, and trade-offs when possible. If a question
      is ambiguous, list possible interpretations and clarify assumptions before answering. Where
      appropriate, provide actionable steps, code snippets, or bullet-point summaries to make your
      answer practical. Use an engaging but professional tone, avoid unnecessary jargon, and ensure
      that your explanations are both technically precise and easy to follow. Your goal is to make
      complex topics accessible while maintaining depth and rigor. Before finalizing an answer,
      verify internal consistency and highlight any uncertainties or limitations clearly so the user
      can make informed decisions based on your response
      Now tell me the answer to the question: "What is the capital of France?"
    - |
      My question is going to be: "What is the capital of France?" Remember that.
      You are an expert assistant designed to provide clear, accurate, and detailed answers across a
      wide range of topics, including technology, science, engineering, and everyday
      problem-solving. Your responses should be structured logically, written in plain and concise
      language, and tailored to the user's level of expertise. Always explain not just what to do
      but also why, offering relevant context, examples, and trade-offs when possible. If a question
      is ambiguous, list possible interpretations and clarify assumptions before answering. Where
      appropriate, provide actionable steps, code snippets, or bullet-point summaries to make your
      answer practical. Use an engaging but professional tone, avoid unnecessary jargon, and ensure
      that your explanations are both technically precise and easy to follow. Your goal is to make
      complex topics accessible while maintaining depth and rigor. Before finalizing an answer,
      verify internal consistency and highlight any uncertainties or limitations clearly so the user
      can make informed decisions based on your response
      Now tell me the answer to the question I asked in the beginning.

[lucaslie]

/bot run

[tensorrt-cicd]

PR_Github #21240 [ run ] triggered by Bot

[coderabbitai]

📝 Walkthrough

Walkthrough

Enables optional chunked prefill in auto-deploy: makes config flag mutable, wires a ContextChunkingConfig into micro-batch scheduling when enabled, truncates cache indices to active KV blocks during context handling, and updates integration/unit tests and test lists to parameterize and validate chunked vs non-chunked behavior.

Changes

Cohort / File(s)	Summary of Changes
Config flag mutability tensorrt_llm/_torch/auto_deploy/llm_args.py	Removed Field(frozen=True) from AutoDeployConfig.enable_chunked_prefill; flag remains default False with same description.
Executor chunked prefill wiring tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py	Added imports ContextChunkingPolicy, ContextChunkingConfig. In create_autodeploy_executor, when enable_chunked_prefill is True, constructs ContextChunkingConfig(policy=FIRST_COME_FIRST_SERVED, unit_size=attn_page_size) and passes to BindMicroBatchScheduler. Updated scheduler ctor calls to keyword args. In _prepare_inputs, truncate page assignments to active KV blocks (end_compute).
Integration test param additions tests/integration/defs/accuracy/test_llm_api_autodeploy.py	Added enable_chunked_prefill parameter to default kwargs and tests. Parameterized TestLlama3_1_8B.test_auto_dtype over False/True; parameterized TestNemotronH.test_auto_dtype and conditionally skip when True. When enabled, set enable_chunked_prefill and max_num_tokens=512.
Test list updates (param variants) tests/integration/test_lists/test-db/l0_b200.yml, .../l0_dgx_h100.yml, .../l0_dgx_h200.yml, .../l0_h100.yml	Updated test entries to include new paramized IDs: replace [N] with [False-N]; expand l0_h100 to include both [False-1] and [True-1] for Llama3.1 and [False]/[True] for Nemotron-H.
Unit test for chunked equivalence tests/unittest/_torch/auto_deploy/unit/singlegpu/shim/test_engine.py	Added dummy KV/resource managers and request types. New test test_ad_engine_chunked_prefill_equivalence(attn_page_size) asserts logits equivalence between single-shot and two-chunk prefill paths (param: attn_page_size={256,2}).

Sequence Diagram(s)

sequenceDiagram
  autonumber
  participant C as Client
  participant AE as create_autodeploy_executor
  participant S1 as BindCapacityScheduler
  participant S2 as BindMicroBatchScheduler
  participant K as KVCacheManager
  rect rgba(230,240,255,0.5)
    note over AE: Initialization
    C->>AE: ad_config, engine, kv_cache_manager, attn_page_size
    AE->>AE: if enable_chunked_prefill: ctx_chunk_config=ContextChunkingConfig(FIFO, unit_size=attn_page_size) else None
    AE->>S1: BindCapacityScheduler(max_num_requests=ad_config.max_batch_size, kv_cache_manager=K.impl, peft_cache_manager=None)
    AE->>S2: BindMicroBatchScheduler(max_batch_size=ad_config.max_batch_size, max_num_tokens=engine.cache_seq_interface.info.max_num_tokens, ctx_chunk_config=ctx_chunk_config)
    AE-->>C: executor
  end

Loading

sequenceDiagram
  autonumber
  participant R as Request (context)
  participant EX as Executor._prepare_inputs
  participant K as KV Cache
  rect rgba(240,255,240,0.5)
    note over EX: Context handling
    R->>EX: page_assignments, end_compute
    EX->>K: query active KV blocks (0..end_compute)
    EX->>EX: truncate page_assignments to active subset
    EX-->>R: prepared inputs (active cache indices)
  end
  note over R,EX: When chunked prefill enabled, scheduling uses ctx_chunk_config during batching

Loading

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~25 minutes

Pre-merge checks and finishing touches

❌ Failed checks (2 warnings)

Check name	Status	Explanation	Resolution
Description Check	⚠️ Warning	The PR description consists solely of the template placeholders and guidance text without any actual summary of the change, explanation of the implementation, or listing of relevant tests that exercise the new chunked prefill behavior.	Please replace the placeholder sections with a concise description of the chunked prefill feature, an overview of the implementation changes, and a clear list of test cases that validate the new code paths.
Docstring Coverage	⚠️ Warning	Docstring coverage is 23.53% which is insufficient. The required threshold is 80.00%.	You can run @coderabbitai generate docstrings to improve docstring coverage.

✅ Passed checks (1 passed)

Check name	Status	Explanation
Title Check	✅ Passed	The title follows the repository’s template and concisely identifies the feature being added, namely support for chunked prefill in the AutoDeploy module. It correctly uses the “[None][feat]” ticket and type format and clearly reflects the primary change in the PR.

✨ 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: 1

Caution

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

⚠️ Outside diff range comments (2)

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

65-69: Fix logging call (formatting bug causes runtime error).

ad_logger.info("...", self.num_blocks) uses %-formatting with no placeholders → TypeError.

Use f-string or add placeholder:

-        ad_logger.info("Using fake cache manager with head_dim=0 and num pages:", self.num_blocks)
+        ad_logger.info("Using fake cache manager with head_dim=0 and num pages: %s", self.num_blocks)

tests/integration/defs/accuracy/test_llm_api_autodeploy.py (1)

44-63: Constraint violation: max_num_tokens must exceed max_batch_size.

The comment on line 62 states that max_num_tokens must be strictly greater than max(attn_page_size, max_batch_size). However, max_num_tokens is set to 512 while max_batch_size is also 512, violating this constraint (512 is not > 512).

Apply this diff to satisfy the constraint:

         if enable_chunked_prefill:
             config["enable_chunked_prefill"] = True
             config[
-                "max_num_tokens"] = 512  # NOTE: must be > max(attn_page_size, max_batch_size)
+                "max_num_tokens"] = 1024  # NOTE: must be > max(attn_page_size, max_batch_size)
         return config

Alternatively, if memory constraints require max_num_tokens to remain at 512, reduce max_batch_size to a value less than 512 when chunked prefill is enabled.

🧹 Nitpick comments (4)

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

139-141: Silence unused-argument lint in tests.

Rename param to _request to address Ruff ARG002.

-    def get_cache_indices(self, request):
+    def get_cache_indices(self, _request):
         # Return many dummy page IDs; ADEngine will truncate as needed
         return list(range(1024))

---

170-218: Optional: guard tests on CUDA.

Tests hardcode cuda. Add skip to avoid local failures.

 @pytest.mark.parametrize("attn_page_size", [256, 2])
 def test_ad_engine_chunked_prefill_equivalence(attn_page_size: int):
+    if not torch.cuda.is_available():
+        pytest.skip("CUDA required")

Also consider adding the same guard to test_engine and test_demo_engine_sampling.

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

208-212: Cache index truncation for context prefill looks correct.

Slicing to num_active_blocks matches end_compute. Please update the preceding comment (“will be used in the future”) since it’s now used.

---

381-388: Surface chunking policy via config (future-proofing).

Policy is hardcoded to FIRST_COME_FIRST_SERVED. Consider adding it to AutoDeployConfig (with sane default) so tests can exercise other policies without code changes.

📜 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 4882815 and c0dcca0.

📒 Files selected for processing (8)

• tensorrt_llm/_torch/auto_deploy/llm_args.py (1 hunks)
• tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py (3 hunks)
• tests/integration/defs/accuracy/test_llm_api_autodeploy.py (6 hunks)
• tests/integration/test_lists/test-db/l0_b200.yml (1 hunks)
• tests/integration/test_lists/test-db/l0_dgx_h100.yml (1 hunks)
• tests/integration/test_lists/test-db/l0_dgx_h200.yml (1 hunks)
• tests/integration/test_lists/test-db/l0_h100.yml (1 hunks)
• tests/unittest/_torch/auto_deploy/unit/singlegpu/shim/test_engine.py (2 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:

• tensorrt_llm/_torch/auto_deploy/llm_args.py
• tests/integration/defs/accuracy/test_llm_api_autodeploy.py
• tests/unittest/_torch/auto_deploy/unit/singlegpu/shim/test_engine.py
• tensorrt_llm/_torch/auto_deploy/shim/ad_executor.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:

• tensorrt_llm/_torch/auto_deploy/llm_args.py
• tests/integration/defs/accuracy/test_llm_api_autodeploy.py
• tests/unittest/_torch/auto_deploy/unit/singlegpu/shim/test_engine.py
• tensorrt_llm/_torch/auto_deploy/shim/ad_executor.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:

• tensorrt_llm/_torch/auto_deploy/llm_args.py
• tests/integration/defs/accuracy/test_llm_api_autodeploy.py
• tests/unittest/_torch/auto_deploy/unit/singlegpu/shim/test_engine.py
• tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py

🧬 Code graph analysis (3)

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

tensorrt_llm/llmapi/llm_args.py (1)

• Field (70-97)

tests/unittest/_torch/auto_deploy/unit/singlegpu/shim/test_engine.py (4)

tensorrt_llm/_torch/auto_deploy/custom_ops/attention_interface.py (3)

• device (218-219)
• SequenceInfo (50-804)
• to (542-550)

tensorrt_llm/_torch/attention_backend/flashinfer.py (1)

• page_size (185-189)

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

• to (49-53)

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

• ADEngine (72-299)
• forward (277-299)

tensorrt_llm/_torch/auto_deploy/shim/ad_executor.py (5)

tensorrt_llm/llmapi/llm_args.py (1)

• ContextChunkingPolicy (917-923)

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

• get_cache_indices (139-141)
• get_num_kv_blocks (143-146)

tensorrt_llm/_torch/pyexecutor/resource_manager.py (2)

• get_cache_indices (647-658)
• get_num_kv_blocks (692-693)

tensorrt_llm/_torch/auto_deploy/custom_ops/attention_interface.py (1)

• page_assignments (420-422)

tensorrt_llm/_torch/pyexecutor/scheduler.py (2)

• BindCapacityScheduler (70-97)
• BindMicroBatchScheduler (169-188)

🪛 Ruff (0.14.0)

tests/unittest/_torch/auto_deploy/unit/singlegpu/shim/test_engine.py

139-139: Unused method argument: request

(ARG002)

⏰ 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 (9)

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

390-399: Scheduler kwargs update LGTM.

Keyword args improve clarity; passing ctx_chunk_config into BindMicroBatchScheduler aligns with chunked prefill.

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

139-139: Field addition LGTM.

enable_chunked_prefill is appropriately added and mutable; integrates with executor setup.

Double-check that attn_page_size yields intended chunk_unit_size across backends (triton/torch set to max_seq_len via validator, flashinfer uses tokens_per_block).

tests/integration/test_lists/test-db/l0_dgx_h200.yml (1)

121-121: No change needed: [False-4] variant exists.

tests/integration/test_lists/test-db/l0_b200.yml (1)

78-78: Test ID [False-1] exists
This suffix correctly maps to enable_chunked_prefill=False and world_size=1 in TestLlama3_1_8B::test_auto_dtype.

tests/integration/test_lists/test-db/l0_h100.yml (1)

119-122: Param variants align with test definitions. IDs [False-1], [True-1] for TestLlama3_1_8B and [False], [True] for TestNemotronH correctly match the pytest parametrization.

tests/integration/test_lists/test-db/l0_dgx_h100.yml (1)

44-44: YAML entry correct
The TestLlama3_1_8B.test_auto_dtype in test_llm_api_autodeploy.py is parametrized over world_size=[1,2,4] and enable_chunked_prefill=[False,True], yielding an item [False-2].

tests/integration/defs/accuracy/test_llm_api_autodeploy.py (3)

73-87: LGTM!

The test parameterization correctly exercises both chunked and non-chunked prefill paths across multiple world sizes, providing good test coverage for the new feature.

---

126-140: LGTM: Skip logic appropriately handles known limitation.

The skip logic correctly prevents test execution when chunked prefill is enabled for NemotronH, with proper issue tracking via GitHub issue #8272. The test parameterization ensures coverage for the non-chunked case.

---

93-116: Constraint satisfied: max_num_tokens (512) > default attn_page_size (64).

[lucaslie]

/bot run

[tensorrt-cicd]

PR_Github #21241 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #21240 [ run ] completed with state ABORTED
LLM/main/L0_MergeRequest_PR #16032 (Blue Ocean) completed with status: ABORTED

[tensorrt-cicd]

PR_Github #21241 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #16033 completed with status: 'FAILURE'

[lucaslie]

/bot run

[tensorrt-cicd]

PR_Github #21254 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #21254 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #16045 completed with status: 'FAILURE'

[lucaslie]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #21364 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #21364 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #16131 completed with status: 'SUCCESS'