[https://nvbugs/5517404][fix] Use the correct cuda graph for dynamic spec dec #7728
Conversation
ziyixiong-nv
force-pushed
the
dev-fxiong-bug5517404
branch
from
0dc46f8 to
a9189c8
Compare
📝 WalkthroughWalkthroughRefactors speculative decoding to derive per-request token sizing from draft_len, updates CUDA graph runner and sampler logic accordingly, and revises a unit test to run in a single-process environment with deterministic spec-decode gating based on draft token length. No public APIs are changed. Changes
Sequence Diagram(s)sequenceDiagram
autonumber
actor Tester
participant Env as Env Var
participant Test as test_dynamic_spec_decode
participant Sampler as TorchSampler
participant Runner as CUDAGraphRunner
participant Engine as Engine
Tester->>Env: Save TLLM_WORKER_USE_SINGLE_PROCESS
Tester->>Env: Set to "1"
Note over Test,Sampler: Run speculative decode path
Test->>Sampler: update_requests(batch)
alt draft_len(req) > 0
Sampler->>Sampler: set py_num_accepted_draft_tokens, py_rewind_len
else zero draft
Sampler->>Sampler: set both fields to 0
end
Test->>Engine: should_use_spec_decode?(batch)
alt any active req has draft_len > 0
Engine-->>Test: False
else none have drafts
Engine-->>Test: True
end
Test->>Runner: capture()
Runner->>Runner: token_per_request = draft_len + 1
Runner->>Engine: enable_spec_decode
Test-->>Env: Restore original env var
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~25 minutes Pre-merge checks and finishing touches❌ Failed checks (1 warning, 1 inconclusive)
✅ Passed checks (1 passed)
✨ Finishing touches
🧪 Generate unit tests
Tip 👮 Agentic pre-merge checks are now available in preview!Pro plan users can now enable pre-merge checks in their settings to enforce checklists before merging PRs.
Please see the documentation for more information. Example: reviews:
pre_merge_checks:
custom_checks:
- name: "Undocumented Breaking Changes"
mode: "warning"
instructions: |
Pass/fail criteria: All breaking changes to public APIs, CLI flags, environment variables, configuration keys, database schemas, or HTTP/GraphQL endpoints must be documented in the "Breaking Change" section of the PR description and in CHANGELOG.md. Exclude purely internal or private changes (e.g., code not exported from package entry points or explicitly marked as internal).Please share your feedback with us on this Discord post. 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. Comment |
|
/bot run |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 0
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (4)
tensorrt_llm/_torch/pyexecutor/sampler.py (1)
1-1: Missing NVIDIA Apache‑2.0 header.Per coding guidelines, prepend the NVIDIA Apache‑2.0 copyright header (current year) to this source file.
tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (2)
58-69: Static buffers sized by current draft_len break when spec decode toggles on later. Allocate for worst‑case (max_draft_len).
_create_shared_static_tensors()usesself.draft_len + 1, which is 0 when spec decode is off at construction. If spec decode is later enabled (draft_len > 0), capture slices exceed the preallocated size. Allocate usingspec_config.max_draft_len + 1instead.Apply:
- token_per_request = self.draft_len + 1 + # Allocate for worst-case so dynamic spec-decode toggling never outgrows buffers + token_per_request = ( + (self.spec_config.max_draft_len if self.spec_config is not None else 0) + + 1 + )Optionally, add a runtime guard in
capture()to assert the static tensors are large enough fornum_tokens_for_capture.
1-1: Missing NVIDIA Apache‑2.0 header.Add the required header at the top of this file.
tests/unittest/_torch/speculative/test_dynamic_spec_decode.py (1)
1-1: Missing NVIDIA Apache‑2.0 header.Add the required header to this test file as well.
🧹 Nitpick comments (5)
tests/unittest/_torch/speculative/test_dynamic_spec_decode.py (5)
21-26: Prefer pytest’s monkeypatch for env var handling.Cleaner and auto‑restores without try/finally.
Example:
-@pytest.mark.high_cuda_memory -def test_dynamic_spec_decode(): +@pytest.mark.high_cuda_memory +def test_dynamic_spec_decode(monkeypatch): - original_value = os.environ.get("TLLM_WORKER_USE_SINGLE_PROCESS") - os.environ["TLLM_WORKER_USE_SINGLE_PROCESS"] = "1" + monkeypatch.setenv("TLLM_WORKER_USE_SINGLE_PROCESS", "1")
73-76: Patch the attribute via patch.object to avoid fragile import paths.Import the class and patch the bound attribute.
- with patch( - 'tensorrt_llm._torch.speculative.model_drafter.ModelDrafter.should_use_spec_decode', - side_effect=mock_should_use_spec_decode): + from tensorrt_llm._torch.speculative.model_drafter import ModelDrafter + with patch.object(ModelDrafter, 'should_use_spec_decode', + side_effect=mock_should_use_spec_decode):
77-78: Remove redundant SamplingParams assignment.
sampling_paramsis immediately reassigned below; drop the first assignment.- sampling_params = SamplingParams(max_tokens=128, temperature=0) ... - sampling_params = SamplingParams(max_tokens=10, temperature=0) + sampling_params = SamplingParams(max_tokens=10, temperature=0)Also applies to: 84-85
64-65: Static analysis: unused parameters in mock (ARG001).Switching to
*argsabove resolves this; otherwise, prefix with_to silence linters.
63-72: Make mock robust to bound/unbound call signatures; drop unused params.File: tests/unittest/_torch/speculative/test_dynamic_spec_decode.py:63-72
Apply:
- def mock_should_use_spec_decode(requests, max_batch_size, - max_num_tokens, max_draft_len): + def mock_should_use_spec_decode(*args, **_): + # Support both (self, requests, ...) and (requests, ...) signatures + requests = args[1] if len(args) == 5 else args[0] for req in requests: if req.state != LlmRequestState.GENERATION_IN_PROGRESS: continue if get_draft_token_length(req) > 0: return False return True
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (3)
tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py(3 hunks)tensorrt_llm/_torch/pyexecutor/sampler.py(1 hunks)tests/unittest/_torch/speculative/test_dynamic_spec_decode.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/pyexecutor/sampler.pytensorrt_llm/_torch/pyexecutor/cuda_graph_runner.pytests/unittest/_torch/speculative/test_dynamic_spec_decode.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/pyexecutor/sampler.pytensorrt_llm/_torch/pyexecutor/cuda_graph_runner.pytests/unittest/_torch/speculative/test_dynamic_spec_decode.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/pyexecutor/sampler.pytensorrt_llm/_torch/pyexecutor/cuda_graph_runner.pytests/unittest/_torch/speculative/test_dynamic_spec_decode.py
🧠 Learnings (4)
📚 Learning: 2025-08-28T10:25:22.370Z
Learnt from: ixlmar
PR: NVIDIA/TensorRT-LLM#7294
File: tensorrt_llm/_torch/pyexecutor/sampler.py:887-891
Timestamp: 2025-08-28T10:25:22.370Z
Learning: In tensorrt_llm/_torch/pyexecutor/sampler.py, the draft_probs and target_probs tensors have shapes [1, steps] not [steps, vocab_size] as might be expected, making the .squeeze(0) operations appropriate for removing the batch dimension of size 1.
Applied to files:
tensorrt_llm/_torch/pyexecutor/sampler.py
📚 Learning: 2025-08-01T15:14:45.673Z
Learnt from: yibinl-nvidia
PR: NVIDIA/TensorRT-LLM#6506
File: examples/models/core/mixtral/requirements.txt:3-3
Timestamp: 2025-08-01T15:14:45.673Z
Learning: In TensorRT-LLM, examples directory can have different dependency versions than the root requirements.txt file. Version conflicts between root and examples dependencies are acceptable because examples are designed to be standalone and self-contained.
Applied to files:
tests/unittest/_torch/speculative/test_dynamic_spec_decode.py
📚 Learning: 2025-08-26T09:37:10.463Z
Learnt from: jiaganc
PR: NVIDIA/TensorRT-LLM#7031
File: tensorrt_llm/bench/dataclasses/configuration.py:90-104
Timestamp: 2025-08-26T09:37:10.463Z
Learning: In TensorRT-LLM, the `get_pytorch_perf_config()` method returns `self.pytorch_config` which can contain default `cuda_graph_config` values, so `llm_args` may already have this config before the extra options processing.
Applied to files:
tests/unittest/_torch/speculative/test_dynamic_spec_decode.py
📚 Learning: 2025-07-28T17:06:08.621Z
Learnt from: moraxu
PR: NVIDIA/TensorRT-LLM#6303
File: tests/integration/test_lists/qa/examples_test_list.txt:494-494
Timestamp: 2025-07-28T17:06:08.621Z
Learning: In TensorRT-LLM testing, it's common to have both CLI flow tests (test_cli_flow.py) and PyTorch API tests (test_llm_api_pytorch.py) for the same model. These serve different purposes: CLI flow tests validate the traditional command-line workflow, while PyTorch API tests validate the newer LLM API backend. Both are legitimate and should coexist.
Applied to files:
tests/unittest/_torch/speculative/test_dynamic_spec_decode.py
🧬 Code graph analysis (1)
tests/unittest/_torch/speculative/test_dynamic_spec_decode.py (2)
cpp/include/tensorrt_llm/batch_manager/llmRequest.h (1)
LlmRequestState(45-206)tensorrt_llm/_torch/pyexecutor/llm_request.py (2)
get_draft_token_length(574-585)get(102-111)
🪛 Ruff (0.12.2)
tests/unittest/_torch/speculative/test_dynamic_spec_decode.py
64-64: Unused function argument: max_batch_size
(ARG001)
65-65: Unused function argument: max_num_tokens
(ARG001)
65-65: Unused function argument: max_draft_len
(ARG001)
⏰ 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 (4)
tests/unittest/_torch/speculative/test_dynamic_spec_decode.py (2)
11-12: LGTM: imports for LlmRequestState and get_draft_token_length.These are needed for the new mock logic.
100-105: LGTM: env var restoration in finally.Correct cleanup for out‑of‑band env mutation (if not adopting monkeypatch).
tensorrt_llm/_torch/pyexecutor/cuda_graph_runner.py (1)
77-79: enable_spec_decode exists on the engine but does not supersede is_spec_decode
- model_engine.py still defines self.is_spec_decode and then sets self.enable_spec_decode = self.is_spec_decode (model_engine.py:298–299) — engine.enable_spec_decode is present, so cuda_graph_runner's access is safe.
- Many code paths (model_engine.py, py_executor.py:969, tests) still reference is_spec_decode, so the claim that enable_spec_decode "supersedes" is_spec_decode is incorrect; remove or reword that assertion.
Likely an incorrect or invalid review comment.
tensorrt_llm/_torch/pyexecutor/sampler.py (1)
581-583: Rewind preallocated draft pages on zero-draft pathWhen get_draft_token_length(req) == 0 the code sets py_rewind_len = 0 which can skip releasing preallocated draft pages — rewind the allocated pages instead.
- else: - req.py_num_accepted_draft_tokens = 0 - req.py_rewind_len = 0 + else: + req.py_num_accepted_draft_tokens = 0 + # Rewind any preallocated pages for zero-draft requests + req.py_rewind_len = getattr(req, "py_draft_pages_allocated", 0)Confirm whether py_draft_pages_allocated can be non‑zero when get_draft_token_length(req) == 0; if yes apply the change, otherwise replace with an assertion documenting the invariant.
|
PR_Github #18618 [ run ] triggered by Bot |
|
PR_Github #18618 [ run ] completed with state |
ziyixiong-nv
force-pushed
the
dev-fxiong-bug5517404
branch
from
a9189c8 to
a1fea25
Compare
|
/bot run |
|
PR_Github #18638 [ run ] triggered by Bot |
|
PR_Github #18638 [ run ] completed with state |
ziyixiong-nv
force-pushed
the
dev-fxiong-bug5517404
branch
from
a1fea25 to
33ef542
Compare
|
/bot run |
|
PR_Github #18674 [ run ] triggered by Bot |
|
PR_Github #18674 [ run ] completed with state |
|
/bot run |
|
PR_Github #18728 [ run ] triggered by Bot |
|
PR_Github #18728 [ run ] completed with state |
ziyixiong-nv
force-pushed
the
dev-fxiong-bug5517404
branch
from
75d09ca to
68de057
Compare
|
/bot run |
|
PR_Github #19131 [ run ] triggered by Bot |
|
PR_Github #19131 [ run ] completed with state |
ziyixiong-nv
force-pushed
the
dev-fxiong-bug5517404
branch
from
68de057 to
3a57ce7
Compare
|
/bot run |
|
PR_Github #19162 [ run ] triggered by Bot |
|
/bot run |
|
PR_Github #19225 [ run ] triggered by Bot |
|
PR_Github #19225 [ run ] completed with state |
|
/bot run |
|
PR_Github #19259 [ run ] triggered by Bot |
ziyixiong-nv
force-pushed
the
dev-fxiong-bug5517404
branch
from
3a57ce7 to
78712e1
Compare
|
PR_Github #19259 [ run ] completed with state |
…spec dec Signed-off-by: ziyixiong-nv <[email protected]>
Signed-off-by: ziyixiong-nv <[email protected]>
Signed-off-by: ziyixiong-nv <[email protected]>
ziyixiong-nv
force-pushed
the
dev-fxiong-bug5517404
branch
from
78712e1 to
f021f5e
Compare
Signed-off-by: ziyixiong-nv <[email protected]>
ziyixiong-nv
force-pushed
the
dev-fxiong-bug5517404
branch
from
f021f5e to
0b770ab
Compare
|
/bot run |
|
PR_Github #19328 [ run ] triggered by Bot |
|
PR_Github #19328 [ run ] completed with state |
|
/bot run |
|
PR_Github #19352 [ run ] triggered by Bot |
Signed-off-by: ziyixiong-nv <[email protected]>
|
/bot run |
|
PR_Github #19390 [ run ] triggered by Bot |
|
PR_Github #19390 [ run ] completed with state |
…spec dec (NVIDIA#7728) Signed-off-by: ziyixiong-nv <[email protected]>
…spec dec (NVIDIA#7728) Signed-off-by: ziyixiong-nv <[email protected]>
3 participants
…spec dec
Summary by CodeRabbit
Bug Fixes
Refactor
Tests
Description
The spec dec could be turned on/off dynamically, so we need to check engine's
enable_spec_decodefor the value.With this value, we can figure out the cuda graph that should be used.
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 thestage-listparameter 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.mdand the
scripts/test_to_stage_mapping.pyhelper.kill
killKill all running builds associated with pull request.
skip
skip --comment COMMENTSkip 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-pipelineReuse 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.