[fix][nvbugs/5399355] Fix Lamport buffer clear issue for MNNVL TwoShot Allreduce and add FP16 support. #6237
Conversation
Signed-off-by: Shiyu Li <[email protected]>
WalkthroughThe changes update buffer clearing and flag management in the MNNVL allreduce CUDA kernels and their Python interface. They introduce a new buffer flag for tracking tokens to clear, adjust buffer clearing logic for correctness, extend support for half-precision data types, and ensure fallback strategies are handled explicitly in the Python layer. The unit tests are enhanced to support multiple sequence lengths simultaneously. No public interfaces are removed. Changes
Estimated code review effort3 (~45 minutes) Suggested reviewers
Poem
✨ Finishing Touches
🪧 TipsChatThere are 3 ways to chat with CodeRabbit:
SupportNeed help? Create a ticket on our support page for assistance with any issues or questions. Note: Be mindful of the bot's finite context window. It's strongly recommended to break down tasks such as reading entire modules into smaller chunks. For a focused discussion, use review comments to chat about specific files and their changes, instead of using the PR comments. CodeRabbit Commands (Invoked using PR comments)
Other keywords and placeholders
Documentation and Community
|
timlee0212
changed the title
timlee0212
changed the title
Signed-off-by: Shiyu Li <[email protected]>
Signed-off-by: Shiyu Li <[email protected]>
|
/bot run |
|
PR_Github #12515 [ run ] triggered by Bot |
|
PR_Github #12515 [ run ] completed with state |
|
/bot run |
|
PR_Github #12586 [ run ] triggered by Bot |
|
PR_Github #12586 [ run ] completed with state |
Signed-off-by: Shiyu Li <[email protected]>
Signed-off-by: Shiyu Li <[email protected]>
timlee0212
changed the title
|
/bot run |
|
PR_Github #12610 [ run ] triggered by Bot |
|
/bot kill |
Signed-off-by: Shiyu Li <[email protected]>
![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
🔭 Outside diff range comments (1)
tests/unittest/_torch/multi_gpu/test_mnnvl_allreduce.py (1)
94-153: Remove unused loop variableOnly the unused
ineeds to be cleaned up—reusing the same AllReduce instance across different sequence lengths is safe, as it doesn’t cache tensor shapes.• In tests/unittest/_torch/multi_gpu/test_mnnvl_allreduce.py, replace the loop:
- for i, (x, residual, reference_output) in enumerate( - zip(x_list, residual_list, reference_output_list)): + for x, residual, reference_output in zip( + x_list, residual_list, reference_output_list):
🧹 Nitpick comments (1)
tests/unittest/_torch/multi_gpu/test_mnnvl_allreduce.py (1)
188-230: Consider improving readability of the complex list comprehension.The test logic correctly handles multiple sequence lengths, but the list comprehension in the MPIPoolExecutor call (lines 211-226) is quite complex and could benefit from improved readability.
Consider extracting the data preparation logic for better readability:
+ # Prepare per-rank data for each sequence length + per_rank_args = [] + for i in range(tensor_parallel_size): + rank_x_list = [x[i, :, :] for x in x_list] + per_rank_args.append(( + tensor_parallel_size, + row_linear_residual_norm_fusion_forward, + rank_x_list, + residual_list, + norm_weight, + eps, + hidden_size, + dtype, + fusion, + reference_output_list, + )) + with MPIPoolExecutor(max_workers=tensor_parallel_size) as executor: results = executor.map( run_single_rank, - *zip(*[ - ( - tensor_parallel_size, - row_linear_residual_norm_fusion_forward, - [ - x[i, :, :] for x in x_list - ], # Extract the i-th rank's data from each sequence length - residual_list, - norm_weight, - eps, - hidden_size, - dtype, - fusion, - reference_output_list, - ) for i in range(tensor_parallel_size) - ]), + *zip(*per_rank_args), )
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
tests/unittest/_torch/multi_gpu/test_mnnvl_allreduce.py(3 hunks)
⏰ 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 (2)
tests/unittest/_torch/multi_gpu/test_mnnvl_allreduce.py (2)
47-78: LGTM! Function signature correctly updated for multi-sequence support.The refactoring from single tensors to lists of tensors is well-implemented and aligns with the goal of supporting multiple sequence lengths simultaneously.
158-179: Excellent parameterization for comprehensive test coverage.The updated parameterization effectively tests:
- Multiple sequence lengths simultaneously (including the multi-length case
[31, 11, 27, 4])- FP16 support as mentioned in the PR objectives
- Various hidden sizes and fusion options
This provides thorough coverage of the enhanced MNNVL allreduce functionality.
|
PR_Github #12617 [ kill ] triggered by Bot |
|
PR_Github #12610 [ run ] completed with state |
|
PR_Github #12617 [ kill ] completed with state |
|
/bot run |
|
PR_Github #12618 [ run ] triggered by Bot |
|
PR_Github #12618 [ run ] completed with state |
|
/bot run |
1 similar comment
|
/bot run |
|
PR_Github #12663 [ run ] triggered by Bot |
|
PR_Github #12663 [ run ] completed with state |
Summary by CodeRabbit
Summary by CodeRabbit
New Features
Bug Fixes
Chores
Description
tl;dr:
The MNNVL twoshot kernel employs three buffers for Lamport synchronization in a circular manner. For each kernel call, it clears the preceding buffer and utilizes the current buffer for communication. In practical scenarios, it is possible that the preceding call had a lower number of tokens compared to the current call, resulting in some elements remaining ambiguous and potentially leading to race conditions for subsequent calls.
This PR addresses this issue by introducing an additional variable in the buffer flags, capturing the number of tokens in the preceding call. Consequently, the kernel will clear the buffer based on this variable rather than the current kernel grid.
Furthermore, this PR resolves an issue that arises when
allreduce_strategyis set toMNNVLbut the allreduce operation necessitates a fallback. The fallback path fails to acknowledgeMNNVLas a valid strategy. Therefore, it is imperative to modify the strategy as well at the fallback level.Test Coverage
pytest tests/unittest/_torch/multi_gpu/test_mnnvl_allreduce.pyGitHub 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.