[#11086][feat] Optimize Auto Deploy weight loading by preloading weights to CPU #11059
Conversation
Signed-off-by: Taylor Yeonbok Lee <[email protected]>
| # Choose loading method based on environment variable | ||
| # Default behavior: preload checkpoint files to CPU | ||
| # Set AD_DISABLE_PRELOAD=1 to use accelerate's load_checkpoint_in_model (no CPU preload) | ||
| disable_preload = os.environ.get("AD_DISABLE_PRELOAD", "0") == "1" |
There was a problem hiding this comment.
Do you want to keep this configurability or remove it?
Seems to me we don't need to keep it around
There was a problem hiding this comment.
I was not sure, but would it be useful for trying turning off the preloading on host machines w/ small memory?
(Though PT backend is not allowing turning this off though..)
There was a problem hiding this comment.
To continue to the discussion in https://github.com/NVIDIA/TensorRT-LLM/pull/11045/changes#r2744152312
📝 WalkthroughWalkthroughModified checkpoint loading in HuggingFace model factory to support configurable preload behavior. When disabled via AD_DISABLE_PRELOAD flag, uses accelerate's direct device loading; when enabled, preloads checkpoint to CPU memory first. Added support for loading from index.json files by aggregating weights across referenced checkpoints. Changes
Sequence Diagram(s)sequenceDiagram
participant User
participant Factory as AutoModelForCausalLMFactory
participant Accelerate as accelerate
participant CPUMem as CPU Memory
participant Model as nn.Module
participant Checkpoint as Checkpoint Files
User->>Factory: Load HF model with checkpoint
Factory->>Factory: Check AD_DISABLE_PRELOAD flag
alt AD_DISABLE_PRELOAD enabled (default)
Factory->>Checkpoint: Read checkpoint/index.json
Checkpoint-->>Factory: Checkpoint data/file list
Factory->>CPUMem: Load full checkpoint to CPU
CPUMem-->>Factory: Aggregated weights dict
Factory->>Model: Load weights into model
Model-->>Factory: Model loaded
Factory->>CPUMem: Free CPU memory
else AD_DISABLE_PRELOAD disabled
Factory->>Accelerate: Use load_checkpoint_in_model
Accelerate->>Checkpoint: Load directly on device
Checkpoint-->>Accelerate: Weights on device
Accelerate->>Model: Populate model
Model-->>Accelerate: Model loaded
end
Factory-->>User: Model ready
Estimated code review effort🎯 4 (Complex) | ⏱️ ~45 minutes 🚥 Pre-merge checks | ✅ 1 | ❌ 2❌ Failed checks (1 warning, 1 inconclusive)
✅ Passed checks (1 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing touches
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 |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Fix all issues with AI agents
In `@tensorrt_llm/_torch/auto_deploy/models/hf.py`:
- Around line 460-477: The method _load_checkpoint_with_preload currently
ignores the device parameter when placing tensors; wrap the
model.load_state_dict(all_weights, strict=False) call with the same
hf_load_state_dict_with_device(device) context manager used elsewhere (or
otherwise map/move the CPU-stored tensors to the target device before/while
loading) so weights land on the requested device; reference the helper
hf_load_state_dict_with_device and the loader _load_full_checkpoint_to_cpu to
locate where to add the context manager and then delete all_weights after
loading as before.
🧹 Nitpick comments (1)
tensorrt_llm/_torch/auto_deploy/models/hf.py (1)
479-521: Implementation is sound with appropriate format handling.The method correctly handles both index.json multi-file checkpoints and single checkpoint files. Good security practice using
weights_only=Truefortorch.load.Minor style note: Static analysis flagged long exception messages on lines 519 and 521. Consider extracting to a custom exception class if this pattern is used elsewhere, but this is not critical.
♻️ Optional: Extract exception messages to improve readability
+class CheckpointLoadError(ValueError): + """Exception for checkpoint loading failures.""" + pass + def _load_full_checkpoint_to_cpu(self, ckpt_file: str) -> dict: """Load the full checkpoint to CPU memory.""" # ... existing code ... else: - raise ValueError(f"Unsupported checkpoint format: {ckpt_file}") + raise CheckpointLoadError(f"Unsupported checkpoint format: {ckpt_file}") else: - raise ValueError(f"Checkpoint file not found or unsupported: {ckpt_file}") + raise CheckpointLoadError(f"Checkpoint file not found or unsupported: {ckpt_file}")
taylor-yb-lee
changed the title
Signed-off-by: Taylor Yeonbok Lee <[email protected]>
|
/bot run |
|
PR_Github #34146 [ run ] triggered by Bot. Commit: |
|
PR_Github #34146 [ run ] completed with state |
Signed-off-by: Taylor Yeonbok Lee <[email protected]>
3 participants
Improved the AutoDeploy backend loading efficiency by replacing the accelerate module's parameter-wise verification with a file-based preloading strategy. By loading weight files directly to the CPU before transferring them to the device, we bypass the overhead of individual parameter checks. For models exceeding 40K parameters, this change reduced loading latency by 80%.
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
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
/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 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.
Summary by CodeRabbit
Release Notes
✏️ Tip: You can customize this high-level summary in your review settings.