[https://nvbugs/5481434][feat] Reuse pytorch memory segments occupied by cudagraph pool #7390
Conversation
📝 WalkthroughWalkthroughIntroduces a buffer reuse cache for workspace allocations in DeepGemmFusedMoE.get_workspace, adding a class-level recorder and a helper to retrieve or allocate CUDA tensors. Replaces direct allocations with cache lookups, imports math for element product, and keeps the external API unchanged. Changes
Sequence Diagram(s)sequenceDiagram
autonumber
actor Caller
participant DeepGemmFusedMoE
participant Cache as allocated_buffer_recorder
Caller->>DeepGemmFusedMoE: get_workspace(args)
activate DeepGemmFusedMoE
loop For each required buffer (workspace_0, workspace_1, workspace_sf)
DeepGemmFusedMoE->>Cache: get_empty(cache_name, shape, dtype, device)
alt cached buffer fits
Cache-->>DeepGemmFusedMoE: return view of existing tensor
else allocate new
DeepGemmFusedMoE->>DeepGemmFusedMoE: torch.empty(..., device='cuda')
DeepGemmFusedMoE->>Cache: store buffer under cache_name
Cache-->>DeepGemmFusedMoE: return new tensor
end
end
DeepGemmFusedMoE-->>Caller: workspace tensors
deactivate DeepGemmFusedMoE
note over DeepGemmFusedMoE,Cache: Buffers reused across calls/graph capture
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~20 minutes Possibly related PRs
Suggested reviewers
Tip 🔌 Remote MCP (Model Context Protocol) integration is now available!Pro plan users can now connect to remote MCP servers from the Integrations page. Connect with popular remote MCPs such as Notion and Linear to add more context to your reviews and chats. ✨ 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
|
|
/bot run |
|
PR_Github #17058 [ 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: 3
🧹 Nitpick comments (3)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py (3)
395-402: Validate that reused buffers don’t require zero-init; otherwise enable it per call.If any consumer reads partially-written regions (e.g., padding, masked lanes), stale data in
workspace_0/1can leak into results. With the proposedzero_on_reusedefaulting to True, behavior matches priortorch.zerossemantics; consider settingzero_on_reuse=Falsehere only after confirming full overwrites.
407-411: Same concern for scale-factor workspace; ensure overwrite or zero on reuse.
workspace_sfis later viewed/returned; if not fully overwritten each step, keep zeroing on reuse to preserve correctness.
317-388: Key cache by device/dtype to avoid cross-context reuse.Even with dtype/device checks, consider folding them into the cache key to reduce linear scans and accidental cross-context sharing:
key = (cache_name, dtype, torch.cuda.current_device()).Also applies to: 395-411
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (1)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py(3 hunks)
🧰 Additional context used
📓 Path-based instructions (2)
**/*.{cpp,cc,cxx,cu,h,hpp,hh,hxx,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.{cpp,cc,cxx,cu,h,hpp,hh,hxx,cuh,py}: Use spaces only; no tabs; indent with 4 spaces
Prepend NVIDIA copyright header (current year) to all source files (.cpp, .h, .cu, .py, etc.)
Files:
tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py
**/*.py
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.py: Python code must target Python 3.8+
Indent Python with 4 spaces; no tabs
Preserve module namespaces when importing: from package.subpackage import foo; then call foo.SomeClass() instead of importing the class directly
Python naming: files snake_case; classes PascalCase; functions/methods snake_case; locals snake_case (prefix k_ when starting with a number); globals UPPER_SNAKE_CASE with G_ prefix; constants UPPER_SNAKE_CASE
Avoid shadowing outer-scope variables; initialize all externally visible members in init
Prefer docstrings for interfaces used outside a file; limit comments to function-internal or file-local interfaces
Use Google-style docstrings for classes and functions; document attributes/variables inline so Sphinx can render them
Avoid reflection when simpler alternatives exist; prefer explicit parameters and return dicts over locals()/dynamic tricks
In try/except, catch the narrowest exceptions possible; keep try bodies minimal and use else for the main logic when doing duck-typing checks
Files:
tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py
🧠 Learnings (2)
📓 Common learnings
Learnt from: djns99
PR: NVIDIA/TensorRT-LLM#6915
File: cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_kernels.cu:4616-4626
Timestamp: 2025-08-19T03:35:20.866Z
Learning: In the MOE profiler TMA workspace preparation (cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_kernels.cu), the overlapping of TMA WS regions for NONE and FINALIZE variants is deliberate design to save memory space, as confirmed by djns99. The comment "reuse the same pointers to save space" reflects this intentional behavior.
📚 Learning: 2025-08-19T03:35:20.866Z
Learnt from: djns99
PR: NVIDIA/TensorRT-LLM#6915
File: cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_kernels.cu:4616-4626
Timestamp: 2025-08-19T03:35:20.866Z
Learning: In the MOE profiler TMA workspace preparation (cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_kernels.cu), the overlapping of TMA WS regions for NONE and FINALIZE variants is deliberate design to save memory space, as confirmed by djns99. The comment "reuse the same pointers to save space" reflects this intentional behavior.
Applied to files:
tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py
🧬 Code graph analysis (1)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py (1)
tensorrt_llm/quantization/utils/fp8_utils.py (2)
align(24-25)ceil_div(10-21)
⏰ 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
| def get_empty(tensor_shape: list[int], dtype: torch.dtype, | ||
| cache_name: str) -> torch.Tensor: | ||
| if DeepGemmFusedMoE.allocated_buffer_recorder is not None: | ||
| # Safely get the list of candidates. Defaults to an empty list if key is missing. | ||
| candidate_buffers = DeepGemmFusedMoE.allocated_buffer_recorder.get( | ||
| cache_name, []) | ||
| numel_like = math.prod(tensor_shape) | ||
|
|
||
| for buffer in candidate_buffers: | ||
| numel_buffer = buffer.numel() | ||
|
|
||
| # buffer just needs to be large enough. | ||
| if numel_buffer >= numel_like: | ||
| return buffer[0:numel_like].view( | ||
| tensor_shape) # Found a fit, return immediately. | ||
|
|
||
| # If we get here, no suitable buffer was found in the cache. Create a new one. | ||
| new_buffer = torch.zeros(tensor_shape, device='cuda', dtype=dtype) | ||
| if DeepGemmFusedMoE.allocated_buffer_recorder is not None: | ||
| DeepGemmFusedMoE.allocated_buffer_recorder.setdefault( | ||
| cache_name, []).append(new_buffer) | ||
| return new_buffer | ||
|
|
There was a problem hiding this comment.
🛠️ Refactor suggestion
Make buffer reuse safe across dtype/device and preserve zero-initialization semantics.
Issues:
- No dtype/device check before reuse; mixing devices or dtypes is possible.
- Previous behavior allocated zeros; reused views may contain stale data. Unless every element is overwritten by kernels, this can cause correctness bugs.
- The cache list can grow unbounded per key.
Proposed fix keeps behavior safe-by-default and 3.8-compatible; you can later disable zeroing per-call if proven unnecessary.
Apply this diff:
- def get_empty(tensor_shape: list[int], dtype: torch.dtype,
- cache_name: str) -> torch.Tensor:
+ def get_empty(tensor_shape: List[int],
+ dtype: torch.dtype,
+ cache_name: str,
+ zero_on_reuse: bool = True) -> torch.Tensor:
- if DeepGemmFusedMoE.allocated_buffer_recorder is not None:
- # Safely get the list of candidates. Defaults to an empty list if key is missing.
- candidate_buffers = DeepGemmFusedMoE.allocated_buffer_recorder.get(
- cache_name, [])
- numel_like = math.prod(tensor_shape)
-
- for buffer in candidate_buffers:
- numel_buffer = buffer.numel()
-
- # buffer just needs to be large enough.
- if numel_buffer >= numel_like:
- return buffer[0:numel_like].view(
- tensor_shape) # Found a fit, return immediately.
-
- # If we get here, no suitable buffer was found in the cache. Create a new one.
- new_buffer = torch.zeros(tensor_shape, device='cuda', dtype=dtype)
- if DeepGemmFusedMoE.allocated_buffer_recorder is not None:
- DeepGemmFusedMoE.allocated_buffer_recorder.setdefault(
- cache_name, []).append(new_buffer)
- return new_buffer
+ if DeepGemmFusedMoE.allocated_buffer_recorder is not None:
+ candidates = DeepGemmFusedMoE.allocated_buffer_recorder.get(cache_name, [])
+ numel_like = math.prod(tensor_shape)
+ device = torch.device('cuda', torch.cuda.current_device())
+ for buf in candidates:
+ # Reuse only if dtype/device match and capacity is sufficient.
+ if buf.dtype is dtype and buf.device == device and buf.numel() >= numel_like:
+ view = buf[:numel_like].view(tensor_shape)
+ if zero_on_reuse:
+ view.zero_()
+ return view
+ # No suitable buffer found; allocate fresh and record.
+ device = torch.device('cuda', torch.cuda.current_device())
+ new_buffer = torch.zeros(tensor_shape, device=device, dtype=dtype)
+ if DeepGemmFusedMoE.allocated_buffer_recorder is not None:
+ DeepGemmFusedMoE.allocated_buffer_recorder.setdefault(cache_name, []).append(new_buffer)
+ return new_bufferOptional follow-ups:
- Keep only the largest tensor per (cache_name, device, dtype) to cap memory growth.
Committable suggestion skipped: line range outside the PR's diff.
🤖 Prompt for AI Agents
tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py around lines
366-388: the current get_empty may reuse cached buffers across differing
dtypes/devices and returns views that may contain stale (non-zero) data, and the
per-key cache can grow unbounded; fix by partitioning the cache key by device
and dtype (e.g., use (cache_name, device, dtype) or maintain per-device/dtype
sublists), when scanning candidates only consider buffers matching the requested
device and dtype, and before returning a reused buffer ensure
zero-initialization (either allocate torch.zeros for new buffers or call
.zero_() on the selected buffer slice) so semantics match the original zeroed
allocation; also cap growth by keeping only the largest buffer per (cache_name,
device, dtype) or replace smaller entries when appending so the per-key list
does not grow unbounded (implement in a 3.8-compatible way without using newer
language features).
|
PR_Github #17058 [ run ] completed with state |
HuiGao-NV
changed the title
| new_buffer = torch.zeros(tensor_shape, device='cuda', dtype=dtype) | ||
| if DeepGemmFusedMoE.allocated_buffer_recorder is not None: | ||
| DeepGemmFusedMoE.allocated_buffer_recorder.setdefault( | ||
| cache_name, []).append(new_buffer) |
There was a problem hiding this comment.
Will this cause more memory consumption since we are allocating and appending new buffers if not fit with existing ones?
There was a problem hiding this comment.
Thx for the SOL change/fix BTW
There was a problem hiding this comment.
I won't cause more memory consumption. KV cache manager has the same memory size as before.
|
/bot run --add-multi-gpu-test --disable-fail-fast |
|
Rerunning to include the test cases of 8xB200 R1 added by #7192 |
|
PR_Github #17084 [ run ] triggered by Bot |
|
PR_Github #17084 [ run ] completed with state |
HuiGao-NV
force-pushed
the
reuse_buffer_in_deepgemm_moe
branch
from
756313f to
2bc58c1
Compare
|
/bot run --add-multi-gpu-test --disable-fail-fast |
|
PR_Github #17164 [ run ] triggered by Bot |
| """ | ||
|
|
||
| # To reuse pytorch memory segments allocated during graph capture. | ||
| allocated_buffer_recorder: dict[str, list[torch.Tensor]] = {} |
There was a problem hiding this comment.
Can we use simpler name like buffer_cache?
|
PR_Github #17164 [ run ] completed with state |
HuiGao-NV
force-pushed
the
reuse_buffer_in_deepgemm_moe
branch
from
2bc58c1 to
e326632
Compare
|
/bot run --add-multi-gpu-test --disable-fail-fast |
|
PR_Github #17223 [ ] completed with state |
|
/bot run --add-multi-gpu-test --disable-fail-fast |
|
PR_Github #17256 [ ] completed with state |
|
Closed by mistake. reopen it. |
5 participants
cudagraph
Summary by CodeRabbit
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 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.