[None][feat] wide_ep support block-wise FP8 on blackwell #7376
Conversation
Signed-off-by: xxi <[email protected]> modified: tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py new file: tensorrt_llm/_torch/modules/fused_moe/moe_backend.py modified: tests/unittest/_torch/modules/test_fused_moe.py
|
/bot run |
📝 WalkthroughWalkthroughAdds a pluggable MoE backend layer and integrates it into WideEPMoE. The module now lazily selects and initializes a backend (Cutlass or DeepGemm) and routes MoE execution through backend.run_moe. Introduces backend selection logic (SM100 + FP8 block scales -> DeepGemm), per-layer unpadded size, and new FP8 tests; deduplicated flags and call parameters. Changes
Sequence Diagram(s)sequenceDiagram
autonumber
participant W as WideEPMoE
participant S as MoEBackendSelection
participant C as MoECutlassBackend
participant D as MoEDeepGemmBackend
participant B as Backend (selected)
participant K as Kernels/Libs
Note over W: Forward called (inputs, weights, quant scales)
W->>W: Ensure weights created
W->>S: select_backend(self)
alt SM100 && FP8 block scales
S-->>W: DeepGemm backend
W->>D: lazy init (once)
activate D
W->>D: run_moe(...)
D->>K: permute, grouped FP8 GEMMs
D->>K: finalize scale (fused or separate)
K-->>D: output
D-->>W: output
deactivate D
else
S-->>W: Cutlass backend
W->>C: lazy init (once)
activate C
W->>C: run_moe(...)
C->>K: autotune/finalize tactic (cached)
C->>K: fused_moe compute
K-->>C: output
C-->>W: output
deactivate C
end
W-->>W: Return output tensor
Estimated code review effort🎯 4 (Complex) | ⏱️ ~60 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
|
![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 (6)
tensorrt_llm/_torch/modules/fused_moe/moe_backend.py (2)
1-4: Add NVIDIA copyright header.Per repo guidelines, prepend the current-year NVIDIA copyright header at the top of all Python sources.
+# Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# +""" -MoE Backend abstraction for supporting different MoE computation implementations. -This module provides a unified interface for different MoE backends (Cutlass, DeepGemm, etc.) -""" +MoE Backend abstraction for supporting different MoE computation implementations. +This module provides a unified interface for different MoE backends (Cutlass, DeepGemm, etc.) +"""
348-348: Wrap long line (E501).Split the conditional selection of run_moe to satisfy 120-char limit.
- run_moe = self.moe_runner.fused_moe_runner.run_moe_min_latency if min_latency_mode else self.moe_runner.fused_moe_runner.run_moe + run_moe = (self.moe_runner.fused_moe_runner.run_moe_min_latency + if min_latency_mode + else self.moe_runner.fused_moe_runner.run_moe)tests/unittest/_torch/modules/test_fused_moe.py (1)
641-791: Assert backend selection for Blackwell + FP8 block-scales.Add an assertion to ensure WideEPMoE actually picks DeepGemm backend on SM100 when FP8 block scales are used. This guards the selection heuristic.
with mock.patch.object(WideEPMoE, "select_alltoall_method_type", return_value=alltoall_method_type): alltoall_model = WideEPMoE( @@ ) alltoall_model.to("cuda") alltoall_model.load_weights([weights]) + # Ensure DeepGemm backend is selected on Blackwell with FP8 block scales + from tensorrt_llm._torch.modules.fused_moe.moe_backend import MoEDeepGemmBackend + assert isinstance(alltoall_model.moe_backend_impl, MoEDeepGemmBackend)tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py (3)
1-4: Add NVIDIA copyright header.This file lacks the required header. Please prepend it.
+# Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved. + import os from enum import IntEnum from typing import Dict, List, Optional, Tuple, Union
322-331: Backend-specific quant method duplication.You now select DeepGemm's FP8 block-scales path here and again in MoEBackendSelection. Duplication risks drift. Consider centralizing the SM100+block-FP8 decision in one place (e.g., keep selection here for quant weights and have backend selection consult the same predicate).
569-571: Intentional no-op for FP8 block scales pre-quantization.Add a short comment to clarify that input quantization is handled inside the selected backend (DeepGemm path), to avoid future regressions.
- elif self.has_deepseek_fp8_block_scales: - pass + elif self.has_deepseek_fp8_block_scales: + # FP8 block scales: input quantization is performed inside the backend (e.g., DeepGemm path). + pass
📜 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 (3)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py(9 hunks)tensorrt_llm/_torch/modules/fused_moe/moe_backend.py(1 hunks)tests/unittest/_torch/modules/test_fused_moe.py(1 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:
tests/unittest/_torch/modules/test_fused_moe.pytensorrt_llm/_torch/modules/fused_moe/moe_backend.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.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:
tests/unittest/_torch/modules/test_fused_moe.pytensorrt_llm/_torch/modules/fused_moe/moe_backend.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py
🧠 Learnings (1)
📚 Learning: 2025-08-14T06:36:40.701Z
Learnt from: timlee0212
PR: NVIDIA/TensorRT-LLM#6886
File: tensorrt_llm/_torch/models/modeling_deepseekv3.py:0-0
Timestamp: 2025-08-14T06:36:40.701Z
Learning: In DeepSeek V3 model (tensorrt_llm/_torch/models/modeling_deepseekv3.py), the disagreement between AllReduce.__init__ guard and _compute_mlp_tp_size logic for MNNVL usage is expected by design. The AllReduce component and MLP TP-size computation intentionally use different criteria for MNNVL availability decisions.
Applied to files:
tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py
🧬 Code graph analysis (3)
tests/unittest/_torch/modules/test_fused_moe.py (5)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py (1)
AlltoallMethodType(30-38)tensorrt_llm/mapping.py (1)
Mapping(32-513)tensorrt_llm/quantization/utils/fp8_utils.py (1)
per_block_cast_to_fp8_e8m0(54-79)tests/unittest/_torch/helpers.py (1)
per_block_cast_to_fp8_e8m0(55-68)tensorrt_llm/models/modeling_utils.py (2)
QuantConfig(128-268)quant_algo(547-548)
tensorrt_llm/_torch/modules/fused_moe/moe_backend.py (4)
tensorrt_llm/_torch/modules/fused_moe/interface.py (2)
MoE(22-181)has_deepseek_fp8_block_scales(127-130)tensorrt_llm/_torch/autotuner.py (2)
AutoTuner(271-752)choose_one(329-436)tensorrt_llm/_torch/custom_ops/torch_custom_ops.py (1)
MoERunner(27-121)tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py (5)
deepgemm_fp8_group_blockwise_gemm(298-336)masked_index_copy_group_quant_fp8(88-159)preprocess_after_permute(259-294)set_strides(339-345)triton_masked_index_gather(194-215)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py (2)
tensorrt_llm/_torch/modules/fused_moe/moe_backend.py (5)
MoEBackend(17-177)MoEBackendSelection(750-793)select_backend(762-793)run_moe(109-177)run_moe(379-461)tensorrt_llm/_torch/modules/fused_moe/quantization.py (2)
DeepSeekFP8BlockScalesFusedMoEMethod(604-737)DeepSeekFP8BlockScalesFusedMoEMethodDeepGemm(740-781)
🪛 Ruff (0.12.2)
tensorrt_llm/_torch/modules/fused_moe/moe_backend.py
322-322: Local variable use_deepseek_fp8_block_scale is assigned to but never used
Remove assignment to unused variable use_deepseek_fp8_block_scale
(F841)
326-326: Local variable use_int8_woq_per_channel is assigned to but never used
Remove assignment to unused variable use_int8_woq_per_channel
(F841)
348-348: Line too long (136 > 120)
(E501)
⏰ 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/moe_backend.py (1)
217-241: weight_dtype correctly reflects storage dtype
MoERunner intentionally uses the original tensor’s dtype for weight_dtype (storage layout) and only applies weight_view_dtype when profiling via AutoTuner; the FusedMoeRunner flags (use_w4_group_scaling, use_int8_woq_per_channel, etc.) cover all quant modes.tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py (1)
661-678: Plumbing looks correct; good switch to backend.run_moe.Arguments forwarded (weights, scales, input_sf, tuner_* and fused finalize) align with the new backend contract.
| self.finalize_tactic(module, input, output_dtype, min_latency_mode, | ||
| tuner_top_k) |
There was a problem hiding this comment.
Bug: finalize_tactic receives tuner_top_k as use_fused_finalize due to positional args.
Base run_moe passes the 5th positional argument into the 5th parameter of MoECutlassBackend.finalize_tactic, which is use_fused_finalize, not tuner_top_k. Use keywords to avoid misbinding.
- self.finalize_tactic(module, input, output_dtype, min_latency_mode,
- tuner_top_k)
+ self.finalize_tactic(
+ module,
+ input,
+ output_dtype,
+ min_latency_mode=min_latency_mode,
+ tuner_top_k=tuner_top_k,
+ )📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| self.finalize_tactic(module, input, output_dtype, min_latency_mode, | |
| tuner_top_k) | |
| self.finalize_tactic( | |
| module, | |
| input, | |
| output_dtype, | |
| min_latency_mode=min_latency_mode, | |
| tuner_top_k=tuner_top_k, | |
| ) |
🤖 Prompt for AI Agents
In tensorrt_llm/_torch/modules/fused_moe/moe_backend.py around lines 157-158,
the call self.finalize_tactic(module, input, output_dtype, min_latency_mode,
tuner_top_k) incorrectly binds tuner_top_k into the use_fused_finalize parameter
via positional args; change the call to use explicit keyword arguments for
clarity and correctness (e.g., pass module=..., input=..., output_dtype=...,
min_latency_mode=..., tuner_top_k=tuner_top_k and set use_fused_finalize
explicitly if needed) so tuner_top_k is passed into the correct parameter.
| use_deepseek_fp8_block_scale = getattr(module, | ||
| 'has_deepseek_fp8_block_scales', | ||
| False) | ||
| use_w4_group_scaling = getattr(module, 'has_w4afp8', False) | ||
| use_int8_woq_per_channel = getattr(module, 'has_int8_woq_per_channel', | ||
| False) | ||
|
|
There was a problem hiding this comment.
🛠️ Refactor suggestion
Remove unused locals flagged by Ruff F841.
use_deepseek_fp8_block_scale and use_int8_woq_per_channel are computed but never used.
- use_deepseek_fp8_block_scale = getattr(module,
- 'has_deepseek_fp8_block_scales',
- False)
- use_w4_group_scaling = getattr(module, 'has_w4afp8', False)
- use_int8_woq_per_channel = getattr(module, 'has_int8_woq_per_channel',
- False)
+ use_w4_group_scaling = getattr(module, 'has_w4afp8', False)📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| use_deepseek_fp8_block_scale = getattr(module, | |
| 'has_deepseek_fp8_block_scales', | |
| False) | |
| use_w4_group_scaling = getattr(module, 'has_w4afp8', False) | |
| use_int8_woq_per_channel = getattr(module, 'has_int8_woq_per_channel', | |
| False) | |
| use_w4_group_scaling = getattr(module, 'has_w4afp8', False) |
🧰 Tools
🪛 Ruff (0.12.2)
322-322: Local variable use_deepseek_fp8_block_scale is assigned to but never used
Remove assignment to unused variable use_deepseek_fp8_block_scale
(F841)
326-326: Local variable use_int8_woq_per_channel is assigned to but never used
Remove assignment to unused variable use_int8_woq_per_channel
(F841)
🤖 Prompt for AI Agents
In tensorrt_llm/_torch/modules/fused_moe/moe_backend.py around lines 322 to 328,
two locals (use_deepseek_fp8_block_scale and use_int8_woq_per_channel) are
computed but never used; remove the unused getattr assignments (or if they were
intended for future checks, replace them with direct getattr calls at the point
of use) so that only necessary variables remain—delete the two unused lines or
consolidate any needed flags into actual conditional logic where they are used.
| self.finalize_tactic(module, tuner_input, output_dtype, | ||
| min_latency_mode, tuner_top_k) | ||
|
|
There was a problem hiding this comment.
Bug: Same finalize_tactic misbinding in Cutlass.run_moe.
tuner_top_k is being passed positionally into use_fused_finalize. Pass by keyword (and wire use_fused_finalize).
- self.finalize_tactic(module, tuner_input, output_dtype,
- min_latency_mode, tuner_top_k)
+ self.finalize_tactic(
+ module,
+ tuner_input,
+ output_dtype,
+ min_latency_mode=min_latency_mode,
+ use_fused_finalize=use_fused_finalize,
+ tuner_top_k=tuner_top_k,
+ )📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
| self.finalize_tactic(module, tuner_input, output_dtype, | |
| min_latency_mode, tuner_top_k) | |
| self.finalize_tactic( | |
| module, | |
| tuner_input, | |
| output_dtype, | |
| min_latency_mode=min_latency_mode, | |
| use_fused_finalize=use_fused_finalize, | |
| tuner_top_k=tuner_top_k, | |
| ) |
🤖 Prompt for AI Agents
In tensorrt_llm/_torch/modules/fused_moe/moe_backend.py around lines 441 to 443,
the call to finalize_tactic passes tuner_top_k positionally which misbinds it
into the next parameter (use_fused_finalize); change the call to pass
tuner_top_k as a keyword (tuner_top_k=tuner_top_k) and also explicitly pass
use_fused_finalize by keyword (use_fused_finalize=use_fused_finalize) so the
values are correctly wired into finalize_tactic and onward to
use_fused_finalize.
Signed-off-by: xxi [email protected]
Summary by CodeRabbit
New Features
Performance
Tests
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.