[None][fix] Pre-allocate workspaces for DeepGEMM MoE to avoid frequent cudaFree/cudaMalloc #6811
Conversation
📝 WalkthroughWalkthroughReplaces world_size/use_dp gating with a dp_size-derived MoE token cap (Mapping.dp_size). DeepGemm fused MoE was reworked into a workspace-driven, chunked, DP-aware backend (new get_workspace, forward, workspace-backed forward_chunk, modified GEMM/quant signatures). FP8 post-quant now writes into caller-provided buffers and returns only scales. Changes
Sequence Diagram(s)sequenceDiagram
participant Client
participant DeepGemmMoE as DeepGemmFusedMoE
participant Stream0
participant Stream1
participant Collectives as DP_Collectives
Client->>DeepGemmMoE: forward(x, router_logits, ...)
DeepGemmMoE->>DeepGemmMoE: compute token_count, moe_max_num_tokens
alt token_count > threshold
loop per chunk
DeepGemmMoE->>DeepGemmMoE: get_workspace(...)
DeepGemmMoE->>Stream0: forward_chunk(chunk0, workspace0)
Stream0-->>DeepGemmMoE: partial outputs/scales
DeepGemmMoE->>Stream1: forward_chunk(chunk1, workspace1) (overlap)
end
DeepGemmMoE->>Collectives: reducescatter/all_reduce (if DP)
Collectives-->>DeepGemmMoE: reduced pieces
DeepGemmMoE->>DeepGemmMoE: assemble final output
else
DeepGemmMoE->>Stream0: forward_chunk(all, workspace)
Stream0-->>DeepGemmMoE: output
end
DeepGemmMoE-->>Client: final output
Estimated code review effort🎯 4 (Complex) | ⏱️ ~45 minutes Possibly related PRs
Suggested labels
Suggested reviewers
✨ Finishing Touches
🧪 Generate unit tests
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. 🪧 TipsChatThere are 3 ways to chat with CodeRabbit:
SupportNeed help? Create a ticket on our support page for assistance with any issues or questions. CodeRabbit Commands (Invoked using PR/Issue comments)Type Other keywords and placeholders
Status, Documentation and Community
|
lfr-0531
force-pushed
the
user/fanrongl/add_workspace_for_dg_moe
branch
from
4a71d56 to
2f94ddf
Compare
|
/bot run |
|
PR_Github #14910 [ run ] triggered by Bot |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 4
🔭 Outside diff range comments (1)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py (1)
89-160: Critical: Non-contiguous output buffer after transpose operation.The function modifies the
output_sparameter in-place with a transpose operation at line 159, which makes it non-contiguous. This violates the expected contract that the output buffer should remain contiguous for subsequent operations.Apply this diff to fix the issue:
- output_s = output_s.transpose(1, 2)[:, :col_size, :] - return output_s + output_s_transposed = output_s.transpose(1, 2)[:, :col_size, :].contiguous() + return output_s_transposed
🧹 Nitpick comments (3)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py (3)
492-510: Extract magic numbers into named constants.The code uses several magic numbers (128, 4) for alignment and padding without explanation. These should be extracted into named constants for better maintainability.
Add constants at the class level:
class DeepGemmFusedMoE(CutlassFusedMoE): # FP8 alignment requirements FP8_TOKEN_ALIGNMENT = 128 FP8_SCALE_ALIGNMENT = 4 FP8_GROUP_SIZE = 128Then update the code:
- m_max = fp8_utils.align(x.shape[0], 128) + m_max = fp8_utils.align(x.shape[0], self.FP8_TOKEN_ALIGNMENT) act_input_fp8 = set_strides(workspace["workspace_0"], self.expert_size_per_partition, m_max, self.hidden_size) - m_padded = fp8_utils.align(m_max, 4) - scale_k = fp8_utils.ceil_div(self.hidden_size, 128) - scale_k_padded = fp8_utils.align(scale_k, 4) + m_padded = fp8_utils.align(m_max, self.FP8_SCALE_ALIGNMENT) + scale_k = fp8_utils.ceil_div(self.hidden_size, self.FP8_GROUP_SIZE) + scale_k_padded = fp8_utils.align(scale_k, self.FP8_SCALE_ALIGNMENT)
608-610: Improve conditional logic for clarity.The nested condition for determining
num_chunkscould be simplified for better readability.- num_chunks = 1 - if num_rows > self.moe_max_num_tokens * 2: - num_chunks = (num_rows + self.moe_max_num_tokens - - 1) // self.moe_max_num_tokens + # Calculate number of chunks needed (minimum 1) + num_chunks = max(1, (num_rows + self.moe_max_num_tokens - 1) // self.moe_max_num_tokens) + # Only split if we need more than 2 chunks (to use dual-stream overlap) + if num_chunks <= 2: + num_chunks = 1
663-712: Consider refactoring the chunked execution logic.The chunked execution logic with stream overlapping is complex and contains repeated patterns. Consider extracting this into a separate method for better maintainability and testability.
Extract the stream-based chunked execution into a helper method:
def _execute_chunked_with_overlap(self, x_list, router_logits_list, all_rank_num_tokens_list, use_dp_padding, workspace_0, workspace_1): """Execute chunked MoE with stream overlap for latency hiding.""" outputs_list = [] for idx_chunk, (x, router_logits) in enumerate(zip(x_list, router_logits_list)): workspace = workspace_0 if idx_chunk % 2 == 0 else workspace_1 stream = self.aux_stream if idx_chunk % 2 == 0 else None # Forward pass (potentially on aux stream) if stream: with torch.cuda.stream(stream): outputs = self._forward_chunk_wrapper( x, router_logits, idx_chunk, all_rank_num_tokens_list, use_dp_padding, workspace) else: outputs = self._forward_chunk_wrapper( x, router_logits, idx_chunk, all_rank_num_tokens_list, use_dp_padding, workspace) # Reduce previous chunk (overlap with current forward) if idx_chunk > 0: self._reduce_previous_chunk(outputs_list, idx_chunk - 1, all_rank_num_tokens_list, use_dp_padding, stream is None) outputs_list.append(outputs) # Final reduction self._reduce_final_chunk(outputs_list, num_chunks, all_rank_num_tokens_list, use_dp_padding) return outputs_list
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (6)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py(1 hunks)tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py(10 hunks)tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.py(1 hunks)tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py(1 hunks)tensorrt_llm/mapping.py(1 hunks)tensorrt_llm/quantization/utils/fp8_utils.py(2 hunks)
🚧 Files skipped from review as they are similar to previous changes (3)
- tensorrt_llm/mapping.py
- tensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py
- tensorrt_llm/quantization/utils/fp8_utils.py
🧰 Additional context used
📓 Path-based instructions (2)
**/*.py
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.py: Python code should conform to Python 3.8+.
Indent Python code with 4 spaces. Do not use tabs.
Always maintain the namespace when importing in Python, even if only one class or function from a module is used.
Python filenames should use snake_case (e.g., some_file.py).
Python classes should use PascalCase (e.g., class SomeClass).
Python functions and methods should use snake_case (e.g., def my_awesome_function():).
Python local variables should use snake_case. Prefix k for variable names that start with a number (e.g., k_99th_percentile).
Python global variables should use upper snake_case and prefix G (e.g., G_MY_GLOBAL).
Python constants should use upper snake_case (e.g., MY_CONSTANT).
Avoid shadowing variables declared in an outer scope in Python.
Initialize all externally visible members of a Python class in the constructor.
For interfaces that may be used outside a Python file, prefer docstrings over comments.
Comments in Python should be reserved for code within a function, or interfaces that are local to a file.
Use Google style docstrings for Python classes and functions, which can be parsed by Sphinx.
Attributes and variables in Python can be documented inline; attribute docstrings will be rendered under the class docstring.
Avoid using reflection in Python when functionality can be easily achieved without it.
When using try-except blocks in Python, limit the except to the smallest set of errors possible.
When using try-except blocks to handle multiple possible variable types in Python, keep the body of the try as small as possible, using the else block to implement the logic.
Files:
tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py
**/*.{cpp,h,hpp,cc,cxx,cu,py}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
All TensorRT-LLM Open Source Software code should contain an NVIDIA copyright header that includes the current year. This includes .cpp, .h, .cu, .py, and any other source files which are compiled or interpreted.
Files:
tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py
🧠 Learnings (1)
📚 Learning: 2025-08-09T20:57:04.067Z
Learnt from: sklevtsov-nvidia
PR: NVIDIA/TensorRT-LLM#3294
File: cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_gemm_tma_warp_specialized_input.cu:118-127
Timestamp: 2025-08-09T20:57:04.067Z
Learning: In the CUTLASS MoE finalize fusion implementation (cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_gemm_tma_warp_specialized_input.cu), when setting `fused_finalize_epilogue.stride_final_output` with shape `(hidden_size, num_output_tokens, 1)`, the `num_rows_in_final_output` should be set to `num_output_tokens` (not `hidden_size`) because of a swap+transpose operation that maps rows of the output tensor to `hidden_size` and columns to `num_output_tokens`.
Applied to files:
tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py
⏰ 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)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.py (1)
85-86: LGTM! Clean implementation of DP-aware token scaling.The unconditional DP-size-based token scaling is correctly implemented, removing the previous conditional logic based on
world_sizeanduse_dp. The fallback pattern using theoroperator is clean and aligns with the same pattern used in other MoE implementations across the PR.tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py (1)
154-157: LGTM! Consistent DP-size-based token limit calculation.The changes correctly implement DP-size-based token scaling, matching the pattern used in other MoE modules. The auxiliary stream initialization is properly gated on the comparison between the configured and calculated token limits.
|
PR_Github #14910 [ run ] completed with state |
Signed-off-by: Fanrong Li <[email protected]>
Signed-off-by: Fanrong Li <[email protected]>
Signed-off-by: Fanrong Li <[email protected]>
Signed-off-by: Yuxian Qiu <[email protected]>
…max_num_tokens * 2. Signed-off-by: Fanrong Li <[email protected]>
lfr-0531
force-pushed
the
user/fanrongl/add_workspace_for_dg_moe
branch
from
2f94ddf to
d6150c5
Compare
|
/bot run |
|
PR_Github #14953 [ run ] triggered by Bot |
There was a problem hiding this comment.
Actionable comments posted: 1
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (6)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py(1 hunks)tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py(10 hunks)tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.py(1 hunks)tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py(1 hunks)tensorrt_llm/mapping.py(1 hunks)tensorrt_llm/quantization/utils/fp8_utils.py(2 hunks)
🚧 Files skipped from review as they are similar to previous changes (4)
- tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py
- tensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py
- tensorrt_llm/quantization/utils/fp8_utils.py
- tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py
🧰 Additional context used
📓 Path-based instructions (2)
**/*.py
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.py: Python code should conform to Python 3.8+.
Indent Python code with 4 spaces. Do not use tabs.
Always maintain the namespace when importing in Python, even if only one class or function from a module is used.
Python filenames should use snake_case (e.g., some_file.py).
Python classes should use PascalCase (e.g., class SomeClass).
Python functions and methods should use snake_case (e.g., def my_awesome_function():).
Python local variables should use snake_case. Prefix k for variable names that start with a number (e.g., k_99th_percentile).
Python global variables should use upper snake_case and prefix G (e.g., G_MY_GLOBAL).
Python constants should use upper snake_case (e.g., MY_CONSTANT).
Avoid shadowing variables declared in an outer scope in Python.
Initialize all externally visible members of a Python class in the constructor.
For interfaces that may be used outside a Python file, prefer docstrings over comments.
Comments in Python should be reserved for code within a function, or interfaces that are local to a file.
Use Google style docstrings for Python classes and functions, which can be parsed by Sphinx.
Attributes and variables in Python can be documented inline; attribute docstrings will be rendered under the class docstring.
Avoid using reflection in Python when functionality can be easily achieved without it.
When using try-except blocks in Python, limit the except to the smallest set of errors possible.
When using try-except blocks to handle multiple possible variable types in Python, keep the body of the try as small as possible, using the else block to implement the logic.
Files:
tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.pytensorrt_llm/mapping.py
**/*.{cpp,h,hpp,cc,cxx,cu,py}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
All TensorRT-LLM Open Source Software code should contain an NVIDIA copyright header that includes the current year. This includes .cpp, .h, .cu, .py, and any other source files which are compiled or interpreted.
Files:
tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.pytensorrt_llm/mapping.py
🧠 Learnings (1)
📚 Learning: 2025-08-09T20:57:04.067Z
Learnt from: sklevtsov-nvidia
PR: NVIDIA/TensorRT-LLM#3294
File: cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_gemm_tma_warp_specialized_input.cu:118-127
Timestamp: 2025-08-09T20:57:04.067Z
Learning: In the CUTLASS MoE finalize fusion implementation (cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_gemm_tma_warp_specialized_input.cu), when setting `fused_finalize_epilogue.stride_final_output` with shape `(hidden_size, num_output_tokens, 1)`, the `num_rows_in_final_output` should be set to `num_output_tokens` (not `hidden_size`) because of a swap+transpose operation that maps rows of the output tensor to `hidden_size` and columns to `num_output_tokens`.
Applied to files:
tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.py
⏰ 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 (1)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_vanilla.py (1)
85-86: Ignore incorrect NoneType/orconcerns formoe_max_num_tokensModelConfig.max_num_tokens is defined as an
int(default 8192) in the_torchdataclass, so it cannot beNoneat initialization. Usingorto fall back whenmoe_max_num_tokensis unset (i.e.None) is valid, and an explicit0value formoe_max_num_tokenswould likewise be a non‐standard configuration that our APIs do not expect.No changes required here; please proceed.
Likely an incorrect or invalid review comment.
|
PR_Github #14953 [ run ] completed with state |
Summary by CodeRabbit
New Features
Performance Improvements
API Changes
Description
Test Coverage
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.