[None][feat] Enable EPLB for trtllm-gen and cutlass backend #8886
Conversation
|
/bot run --disable-fail-fast |
|
PR_Github #23395 [ run ] triggered by Bot. Commit: |
|
/bot run --disable-fail-fast |
|
PR_Github #23402 [ run ] triggered by Bot. Commit: |
|
PR_Github #23402 [ run ] completed with state |
|
/bot run --disable-fail-fast |
|
PR_Github #23429 [ run ] triggered by Bot. Commit: |
dongxuy04
force-pushed
the
user/dongxuy/trtllmgen_eplb
branch
from
dfdc029 to
abe56ca
Compare
|
/bot run --disable-fail-fast |
|
PR_Github #23450 [ run ] triggered by Bot. Commit: |
|
PR_Github #23429 [ run ] completed with state |
|
PR_Github #23450 [ run ] completed with state |
|
/bot run --disable-fail-fast |
dongxuy04
force-pushed
the
user/dongxuy/trtllmgen_eplb
branch
from
60d5d1e to
495e260
Compare
|
/bot run --disable-fail-fast |
|
PR_Github #23708 [ run ] triggered by Bot. Commit: |
|
/bot run --disable-fail-fast |
|
PR_Github #23723 [ run ] triggered by Bot. Commit: |
|
PR_Github #23708 [ run ] completed with state |
📝 WalkthroughWalkthroughIntegrates load-balancer support across MoE backends with slot-based token routing, GPU/CPU stage synchronization, and expert statistics tracking. Refactors weight loading for shared-weight scenarios and expands test coverage to exercise multiple MoE backends with online expert-parallel load balancing. Changes
Sequence Diagram(s)sequenceDiagram
participant App as Application
participant MoE as MoE Module
participant LB as Load Balancer
participant Expert as Expert Slot
participant GPU as GPU Stage
participant CPU as CPU Stage
App->>MoE: forward_chunk(x, router_logits, repeating_info=(is_first, is_last))
rect rgb(200, 220, 240)
Note over MoE,GPU: GPU Wait Stage (is_first_call)
MoE->>LB: _load_balancer_start_wait_gpu_stage(is_first)
LB->>GPU: wait for prior stage completion
GPU-->>LB: ready
end
rect rgb(220, 240, 220)
Note over MoE,LB: Routing & Statistics
MoE->>LB: _load_balancer_update_statistic(token_selected_experts, is_first, is_last)
MoE->>LB: _load_balancer_route(token_selected_experts)
LB-->>MoE: token_selected_slots
end
rect rgb(240, 220, 220)
Note over MoE,Expert: Expert Computation
MoE->>Expert: fused_moe(x, token_selected_slots, weights/biases)
Expert-->>MoE: output
end
rect rgb(240, 240, 200)
Note over MoE,GPU: GPU Done Stage
MoE->>LB: _load_balancer_done_wait_gpu_stage(is_first)
LB->>GPU: signal completion
end
rect rgb(220, 240, 240)
Note over MoE,CPU: CPU Set Stage (is_last_call)
MoE->>LB: _load_balancer_start_set_cpu_stage(is_last)
LB->>CPU: update statistics, prepare next layer
CPU-->>LB: ready
MoE->>LB: _load_balancer_done_set_cpu_stage(is_last)
end
MoE-->>App: output
Estimated code review effort🎯 4 (Complex) | ⏱️ ~50 minutes Areas requiring extra attention:
Pre-merge checks and finishing touches❌ Failed checks (2 warnings)
✅ Passed checks (1 passed)
✨ Finishing touches
🧪 Generate unit tests (beta)
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: 2
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (7)
tensorrt_llm/_torch/models/modeling_gpt_oss.py (1)
143-175: Guardself.num_slots // moe_ep_sizeagainst non-divisible slot counts.
self.num_slotsnow comes from the load balancer config and may not be divisible byconfig.mapping.moe_ep_size. When that's the case, the integer floor division silently drops remainder slots and produces the wrong tensor length forswiglu_*; whennum_slots < moe_ep_sizeit also returns 0, leading to empty tensors and downstream crashes. Please either validate divisibility (e.g., assertnum_slots % moe_ep_size == 0) or adjust the construction to reflect the actual per-rank slot distribution.tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py (2)
432-486: Requeue statistic update for non-last MNNVL chunks.In the MNNVL path, you call
_load_balancer_get_local_statistic_tensor()only whenis_last_callis true, so chunked executions skip exporting intermediate stats. However,_load_balancer_update_statisticstill runs for every chunk withignore_allreducetrue. If chunking happens, we never push the accumulated local tensor to the balancer, causing outdated global stats. Please either gather stats on every chunk (when tensor is non-None) or deferignore_allreduce=Trueto the final chunk only.
649-692: Ensure CPU-stage sync wrapsalltoall_combine.
_load_balancer_start_set_cpu_stageis invoked beforealltoall_combine, but for chunked MNNVL you may return early (when not using alltoall) without calling_load_balancer_done_set_cpu_stage. Please move both start/done calls into atry/finallyor otherwise guarantee they execute around every return path so the balancer doesn’t get stuck waiting.tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py (4)
81-186: Restore SM==120 escape hatch.Prior code short-circuited on
sm_version >= 120before load-balancer routing. After extending aux_stream_dict and load balancer support, we still abort execution on new hardware. If TRTLLM-Gen remains unsupported on SM120+, that’s fine, but we now reject the load balancer at creation time without fallback. Please document or downgrade this to a runtime feature flag so other backends (e.g., Cutlass) can still progress under EPLB.
333-380: Missing load-balancer routing in non-post-comm branch.When
post_quant_commis false (no alltoall/allgather needed), we skip token routing entirely and never call_load_balancer_route, so the balancer’s decisions are ignored. Please invoke the load balancer prior to_quantize_for_post_quant_commor ensure routing always happens before data leaves the rank.
398-417: Guard gathered statistic reshape by tensor size.
gathered_loadbalancer_local_statistic_info.view((self.mapping.moe_ep_size, self.num_experts))assumes a dense tensor. When some ranks return None (e.g., static routing), the gather op hands back an empty tuple, and.viewraises. Add a size check before reshaping or ensure the op always returns a filled tensor.
800-820: Keep repeat index increment outside load-balancer guard.
self.repeat_idxis only advanced whenself.layer_load_balanceris set. Without the balancer,repeat_idxremains 0, but upstream code (e.g., chunked execution guards) still relies on it. Move the increment out of the guard so the original behavior is preserved.
🧹 Nitpick comments (2)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_cute_dsl.py (1)
143-151: Remove unused load-balancer parameters.
forward_chunksignature now acceptsrepeating_info,all_rank_num_tokens, anduse_dp_paddingbut doesn’t use them or interact with the load balancer. That makes the interface inconsistent and silences actual load-balancer wiring for CuteDSL. Either plumb the parameters through (start/done GPU waits, route slots, statistics update) or drop them to avoid misleading call sites.tests/integration/defs/accuracy/test_llm_api_pytorch.py (1)
3926-3961: Consider parameterizing overmoe_backendfor broader online EPLB coverage.The test hardcodes
backend="TRTLLM"(line 3947), unlike the DeepSeekV3Lite online EPLB tests above which are parameterized over multiple backends. Given the PR enables EPLB for both TRTLLM and CUTLASS, consider adding parameterization to test both backends:+ @pytest.mark.parametrize("moe_backend", ["CUTLASS", "TRTLLM"], + ids=["cutlass", "trtllm"]) def test_w4_4gpus_online_eplb(self, kv_cache_dtype, mocker): + def test_w4_4gpus_online_eplb(self, kv_cache_dtype, moe_backend, mocker): ... pytorch_config = dict(disable_overlap_scheduler=False, cuda_graph_config=CudaGraphConfig(), - moe_config=MoeConfig(backend="TRTLLM", + moe_config=MoeConfig(backend=moe_backend, load_balancer=eplb_config))If TRTLLM-only testing is intentional, please document the rationale in the test docstring.
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (17)
tensorrt_llm/_torch/models/modeling_gpt_oss.py(2 hunks)tensorrt_llm/_torch/modules/fused_moe/create_moe.py(2 hunks)tensorrt_llm/_torch/modules/fused_moe/fused_moe_cute_dsl.py(1 hunks)tensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py(16 hunks)tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py(11 hunks)tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py(9 hunks)tensorrt_llm/_torch/modules/fused_moe/interface.py(4 hunks)tensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py(7 hunks)tensorrt_llm/_torch/modules/fused_moe/quantization.py(27 hunks)tests/integration/defs/.test_durations(2 hunks)tests/integration/defs/accuracy/test_llm_api_pytorch.py(3 hunks)tests/integration/test_lists/qa/llm_function_core.txt(1 hunks)tests/integration/test_lists/qa/llm_function_core_sanity.txt(1 hunks)tests/integration/test_lists/qa/llm_function_nim.txt(1 hunks)tests/integration/test_lists/test-db/l0_gb200_multi_gpus.yml(1 hunks)tests/integration/test_lists/test-db/l0_gb300_multi_gpus.yml(1 hunks)tests/integration/test_lists/test-db/l0_rtx_pro_6000.yml(1 hunks)
🧰 Additional context used
📓 Path-based instructions (3)
**/*.{h,hpp,hh,hxx,cpp,cxx,cc,cu,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Use only spaces, no tabs; indent with 4 spaces.
Files:
tensorrt_llm/_torch/models/modeling_gpt_oss.pytensorrt_llm/_torch/modules/fused_moe/create_moe.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_cute_dsl.pytensorrt_llm/_torch/modules/fused_moe/quantization.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.pytests/integration/defs/accuracy/test_llm_api_pytorch.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.pytensorrt_llm/_torch/modules/fused_moe/interface.pytensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py
**/*.py
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.py: Python code must target Python 3.8+.
Indent Python code with 4 spaces; do not use tabs.
Maintain module namespace when importing; prefer 'from package.subpackage import foo' then 'foo.SomeClass()' instead of importing the class directly.
Python filenames should be snake_case (e.g., some_file.py).
Python classes use PascalCase names.
Functions and methods use snake_case names.
Local variables use snake_case; prefix 'k' for variables that start with a number (e.g., k_99th_percentile).
Global variables use upper SNAKE_CASE prefixed with 'G' (e.g., G_MY_GLOBAL).
Constants use upper SNAKE_CASE (e.g., MY_CONSTANT).
Avoid shadowing variables from an outer scope.
Initialize all externally visible members of a class in the constructor.
Prefer docstrings for interfaces that may be used outside a file; comments for in-function or file-local interfaces.
Use Google-style docstrings for classes and functions (Sphinx-parsable).
Document attributes and variables inline so they render under the class/function docstring.
Avoid reflection when a simpler, explicit approach suffices (e.g., avoid dict(**locals()) patterns).
In try/except, catch the most specific exceptions possible.
For duck-typing try/except, keep the try body minimal and use else for the main logic.
Files:
tensorrt_llm/_torch/models/modeling_gpt_oss.pytensorrt_llm/_torch/modules/fused_moe/create_moe.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_cute_dsl.pytensorrt_llm/_torch/modules/fused_moe/quantization.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.pytests/integration/defs/accuracy/test_llm_api_pytorch.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.pytensorrt_llm/_torch/modules/fused_moe/interface.pytensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py
**/*.{cpp,cxx,cc,h,hpp,hh,hxx,cu,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Prepend the NVIDIA Apache-2.0 copyright header with current year to the top of all source files (e.g., .cpp, .h, .cu, .py).
Files:
tensorrt_llm/_torch/models/modeling_gpt_oss.pytensorrt_llm/_torch/modules/fused_moe/create_moe.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_cute_dsl.pytensorrt_llm/_torch/modules/fused_moe/quantization.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.pytests/integration/defs/accuracy/test_llm_api_pytorch.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.pytensorrt_llm/_torch/modules/fused_moe/interface.pytensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py
🧠 Learnings (16)
📓 Common learnings
Learnt from: djns99
Repo: NVIDIA/TensorRT-LLM PR: 6915
File: cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_kernels.cu:4010-4012
Timestamp: 2025-08-14T23:23:27.449Z
Learning: For MOE (Mixture of Experts) code reviews in TensorRT-LLM, avoid repeatedly suggesting finalize fusion validation checks and safety assertions. The user djns99 has indicated these suggestions are repetitive and unwanted across multiple MOE-related changes.
📚 Learning: 2025-09-09T09:40:45.658Z
Learnt from: fredricz-20070104
Repo: NVIDIA/TensorRT-LLM PR: 7645
File: tests/integration/test_lists/qa/llm_function_core.txt:648-648
Timestamp: 2025-09-09T09:40:45.658Z
Learning: In TensorRT-LLM test lists, it's common and intentional for the same test to appear in multiple test list files when they serve different purposes (e.g., llm_function_core.txt for comprehensive core functionality testing and llm_function_core_sanity.txt for quick sanity checks). This duplication allows tests to be run in different testing contexts.
Applied to files:
tests/integration/test_lists/qa/llm_function_nim.txttests/integration/test_lists/test-db/l0_gb300_multi_gpus.ymltests/integration/test_lists/test-db/l0_rtx_pro_6000.ymltests/integration/test_lists/qa/llm_function_core_sanity.txttests/integration/test_lists/qa/llm_function_core.txttests/integration/test_lists/test-db/l0_gb200_multi_gpus.yml
📚 Learning: 2025-08-21T02:39:12.009Z
Learnt from: djns99
Repo: NVIDIA/TensorRT-LLM PR: 7104
File: cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_kernels.cu:1475-1480
Timestamp: 2025-08-21T02:39:12.009Z
Learning: The min latency mode functionality in TensorRT-LLM MOE kernels (cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_kernels.cu) is deprecated and no longer being maintained/updated, as confirmed by djns99. Bug reports and optimization suggestions for the computeStridesTmaWarpSpecializedLowLatencyKernel and related min latency code paths should be deprioritized.
Applied to files:
tests/integration/test_lists/qa/llm_function_nim.txttests/integration/test_lists/qa/llm_function_core_sanity.txttests/integration/test_lists/qa/llm_function_core.txttensorrt_llm/_torch/modules/fused_moe/create_moe.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.pytensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py
📚 Learning: 2025-09-17T02:48:52.732Z
Learnt from: tongyuantongyu
Repo: NVIDIA/TensorRT-LLM PR: 7781
File: tests/integration/test_lists/waives.txt:313-313
Timestamp: 2025-09-17T02:48:52.732Z
Learning: In TensorRT-LLM, `tests/integration/test_lists/waives.txt` is specifically for waiving/skipping tests, while other test list files like those in `test-db/` and `qa/` directories are for different test execution contexts (pre-merge, post-merge, QA tests). The same test appearing in both waives.txt and execution list files is intentional - the test is part of test suites but will be skipped due to the waiver.
Applied to files:
tests/integration/test_lists/qa/llm_function_nim.txttests/integration/test_lists/test-db/l0_gb300_multi_gpus.ymltests/integration/test_lists/test-db/l0_rtx_pro_6000.ymltests/integration/test_lists/qa/llm_function_core_sanity.txttests/integration/test_lists/qa/llm_function_core.txttests/integration/test_lists/test-db/l0_gb200_multi_gpus.yml
📚 Learning: 2025-08-14T23:23:27.449Z
Learnt from: djns99
Repo: NVIDIA/TensorRT-LLM PR: 6915
File: cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_kernels.cu:4010-4012
Timestamp: 2025-08-14T23:23:27.449Z
Learning: For MOE (Mixture of Experts) code reviews in TensorRT-LLM, avoid repeatedly suggesting finalize fusion validation checks and safety assertions. The user djns99 has indicated these suggestions are repetitive and unwanted across multiple MOE-related changes.
Applied to files:
tests/integration/test_lists/qa/llm_function_nim.txttests/integration/test_lists/test-db/l0_rtx_pro_6000.ymltensorrt_llm/_torch/models/modeling_gpt_oss.pytests/integration/test_lists/qa/llm_function_core_sanity.txttests/integration/test_lists/qa/llm_function_core.txttensorrt_llm/_torch/modules/fused_moe/create_moe.pytensorrt_llm/_torch/modules/fused_moe/quantization.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.pytensorrt_llm/_torch/modules/fused_moe/interface.pytensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py
📚 Learning: 2025-08-19T03:35:20.866Z
Learnt from: djns99
Repo: NVIDIA/TensorRT-LLM PR: 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:
tests/integration/test_lists/qa/llm_function_nim.txttests/integration/test_lists/qa/llm_function_core.txttensorrt_llm/_torch/modules/fused_moe/create_moe.pytensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py
📚 Learning: 2025-07-28T17:06:08.621Z
Learnt from: moraxu
Repo: NVIDIA/TensorRT-LLM PR: 6303
File: tests/integration/test_lists/qa/examples_test_list.txt:494-494
Timestamp: 2025-07-28T17:06:08.621Z
Learning: In TensorRT-LLM testing, it's common to have both CLI flow tests (test_cli_flow.py) and PyTorch API tests (test_llm_api_pytorch.py) for the same model. These serve different purposes: CLI flow tests validate the traditional command-line workflow, while PyTorch API tests validate the newer LLM API backend. Both are legitimate and should coexist.
Applied to files:
tests/integration/test_lists/qa/llm_function_nim.txttests/integration/test_lists/test-db/l0_gb300_multi_gpus.ymltests/integration/test_lists/test-db/l0_rtx_pro_6000.ymltests/integration/test_lists/qa/llm_function_core_sanity.txttests/integration/test_lists/qa/llm_function_core.txttests/integration/test_lists/test-db/l0_gb200_multi_gpus.ymltests/integration/defs/accuracy/test_llm_api_pytorch.py
📚 Learning: 2025-08-26T09:49:04.956Z
Learnt from: pengbowang-nv
Repo: NVIDIA/TensorRT-LLM PR: 7192
File: tests/integration/test_lists/test-db/l0_dgx_b200.yml:56-72
Timestamp: 2025-08-26T09:49:04.956Z
Learning: In TensorRT-LLM test configuration files, the test scheduling system handles wildcard matching with special rules that prevent duplicate test execution even when the same tests appear in multiple yaml files with overlapping GPU wildcards (e.g., "*b200*" and "*gb200*").
Applied to files:
tests/integration/test_lists/test-db/l0_gb300_multi_gpus.ymltests/integration/test_lists/test-db/l0_rtx_pro_6000.ymltests/integration/test_lists/qa/llm_function_core_sanity.txttests/integration/test_lists/qa/llm_function_core.txttests/integration/test_lists/test-db/l0_gb200_multi_gpus.ymltests/integration/defs/accuracy/test_llm_api_pytorch.py
📚 Learning: 2025-08-14T06:36:40.701Z
Learnt from: timlee0212
Repo: NVIDIA/TensorRT-LLM PR: 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:
tests/integration/test_lists/test-db/l0_rtx_pro_6000.ymltensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py
📚 Learning: 2025-08-20T07:43:36.447Z
Learnt from: ChristinaZ
Repo: NVIDIA/TensorRT-LLM PR: 7068
File: cpp/tensorrt_llm/kernels/moeTopKFuncs.cuh:169-172
Timestamp: 2025-08-20T07:43:36.447Z
Learning: In TensorRT-LLM MOE kernels, when processing up to 128 experts across 32 threads, each thread handles at most 4 experts (N < 5 constraint), where N represents candidates per thread rather than total system capacity.
Applied to files:
tensorrt_llm/_torch/models/modeling_gpt_oss.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py
📚 Learning: 2025-08-06T13:58:07.506Z
Learnt from: galagam
Repo: NVIDIA/TensorRT-LLM PR: 6487
File: tests/unittest/_torch/auto_deploy/unit/singlegpu/test_ad_trtllm_bench.py:1-12
Timestamp: 2025-08-06T13:58:07.506Z
Learning: In TensorRT-LLM, test files (files under tests/ directories) do not require NVIDIA copyright headers, unlike production source code files. Test files typically start directly with imports, docstrings, or code.
Applied to files:
tests/integration/test_lists/qa/llm_function_core_sanity.txttests/integration/test_lists/qa/llm_function_core.txt
📚 Learning: 2025-08-08T22:03:40.707Z
Learnt from: sklevtsov-nvidia
Repo: NVIDIA/TensorRT-LLM PR: 3294
File: cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_kernels.cu:1198-1209
Timestamp: 2025-08-08T22:03:40.707Z
Learning: In the CUTLASS MoE kernels (cpp/tensorrt_llm/cutlass_extensions), when `layout_info.fusion` is set to `TmaWarpSpecializedGroupedGemmInput::EpilogueFusion::FINALIZE`, the `router_scales` parameter must be non-null by design. The fused finalize kernel epilogue does not perform nullptr checks and requires valid router scales to function correctly. This is an implicit contract that callers must satisfy when enabling the FINALIZE fusion mode.
Applied to files:
tensorrt_llm/_torch/modules/fused_moe/create_moe.pytensorrt_llm/_torch/modules/fused_moe/quantization.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py
📚 Learning: 2025-08-09T20:57:04.084Z
Learnt from: sklevtsov-nvidia
Repo: NVIDIA/TensorRT-LLM PR: 3294
File: cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_gemm_tma_warp_specialized_input.cu:118-127
Timestamp: 2025-08-09T20:57:04.084Z
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/create_moe.pytensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py
📚 Learning: 2025-08-19T12:45:11.997Z
Learnt from: amitz-nv
Repo: NVIDIA/TensorRT-LLM PR: 7033
File: tensorrt_llm/_torch/pyexecutor/model_engine.py:0-0
Timestamp: 2025-08-19T12:45:11.997Z
Learning: In tensorrt_llm/_torch/pyexecutor/model_engine.py, DoRA (Delta Orthogonal Rank Adaptation) functionality was removed from the PyTorch flow to eliminate issues with inverted DoRA detection logic. The original is_dora condition was checking if scaling_vec_pointer == 0, which was potentially incorrect.
Applied to files:
tensorrt_llm/_torch/modules/fused_moe/quantization.py
📚 Learning: 2025-09-19T21:28:13.751Z
Learnt from: jhaotingc
Repo: NVIDIA/TensorRT-LLM PR: 7856
File: cpp/tensorrt_llm/thop/fp8BlockScaleMoe.cpp:159-166
Timestamp: 2025-09-19T21:28:13.751Z
Learning: In TensorRT-LLM blockScaleMoe routing (cpp/tensorrt_llm/kernels/trtllmGenKernels/blockScaleMoe/runner.cu), the DeepSeek routing method performs reinterpret_cast<float*>(routingLogits) at line 89, which could cause issues if routing_logits are BF16. However, Qwen3-FP8 models use RenormalizeNaive routing method and are not affected by this dtype casting issue.
Applied to files:
tensorrt_llm/_torch/modules/fused_moe/quantization.py
📚 Learning: 2025-08-21T21:48:35.135Z
Learnt from: djns99
Repo: NVIDIA/TensorRT-LLM PR: 7104
File: cpp/tensorrt_llm/cutlass_extensions/include/cutlass_extensions/epilogue/fusion/sm90_visitor_scatter.hpp:399-417
Timestamp: 2025-08-21T21:48:35.135Z
Learning: CUTLASS extensions in TensorRT-LLM (located under cpp/tensorrt_llm/cutlass_extensions/) are designed to integrate with and extend functionality in the external CUTLASS repository. When analyzing these extensions, their consumers and functionality wiring may exist in the CUTLASS codebase rather than within TensorRT-LLM itself.
Applied to files:
tensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py
🧬 Code graph analysis (9)
tensorrt_llm/_torch/modules/fused_moe/create_moe.py (3)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py (1)
WideEPMoE(26-945)tensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py (1)
CutlassFusedMoE(28-782)tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py (1)
TRTLLMGenFusedMoE(29-853)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_cute_dsl.py (1)
tensorrt_llm/_torch/utils.py (1)
Fp4QuantizedTensor(110-117)
tensorrt_llm/_torch/modules/fused_moe/quantization.py (4)
tensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py (2)
need_load_shared_weights(462-463)get_load_expert_ids(452-454)tensorrt_llm/logger.py (1)
warning(132-133)tensorrt_llm/_torch/modules/fused_moe/interface.py (2)
_add_raw_shared_weights_for_unmap(295-299)register_all_parameter_slot_and_to_fix_weight_fns(424-456)tensorrt_llm/_torch/modules/linear.py (2)
load_weight_shard(65-118)TensorParallelMode(50-62)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py (5)
tensorrt_llm/_torch/utils.py (1)
Fp4QuantizedTensor(110-117)tensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py (3)
_supports_load_balancer(254-256)enable_alltoall(243-246)moe_alltoall_backend(249-252)tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py (2)
_supports_load_balancer(933-935)enable_alltoall(243-246)tensorrt_llm/_torch/modules/fused_moe/interface.py (11)
_supports_load_balancer(301-306)_load_balancer_start_wait_gpu_stage(317-320)_load_balancer_done_wait_gpu_stage(322-325)enable_alltoall(610-613)AlltoallMethodType(26-34)_load_balancer_update_statistic(327-352)_load_balancer_route(354-361)_load_balancer_get_local_statistic_tensor(373-378)_load_balancer_update_statistic_with_gathered_statistic(380-385)_load_balancer_start_set_cpu_stage(363-366)_load_balancer_done_set_cpu_stage(368-371)tensorrt_llm/_mnnvl_utils.py (1)
mnnvl_moe_alltoallv_prepare_without_allgather(402-446)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py (3)
tensorrt_llm/_torch/distributed/ops.py (2)
allgather(265-298)reducescatter(363-428)tensorrt_llm/_torch/modules/fused_moe/interface.py (11)
_load_balancer_start_wait_gpu_stage(317-320)_load_balancer_done_wait_gpu_stage(322-325)enable_alltoall(610-613)AlltoallMethodType(26-34)_load_balancer_update_statistic(327-352)_load_balancer_route(354-361)_load_balancer_get_local_statistic_tensor(373-378)_load_balancer_update_statistic_with_gathered_statistic(380-385)_load_balancer_start_set_cpu_stage(363-366)_load_balancer_done_set_cpu_stage(368-371)_supports_load_balancer(301-306)tensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py (3)
enable_alltoall(243-246)moe_alltoall_backend(249-252)_supports_load_balancer(254-256)
tests/integration/defs/accuracy/test_llm_api_pytorch.py (4)
tests/integration/defs/conftest.py (1)
parametrize_with_ids(1818-1844)tensorrt_llm/llmapi/llm_args.py (2)
KvCacheConfig(1265-1409)MoeConfig(268-302)tensorrt_llm/_torch/model_config.py (1)
MoeLoadBalancerConfig(30-64)tests/integration/defs/accuracy/accuracy_core.py (3)
GSM8K(334-349)evaluate(184-247)evaluate(766-776)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py (4)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py (1)
_supports_load_balancer(183-185)tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py (2)
_supports_load_balancer(933-935)forward_chunk(382-694)tensorrt_llm/_torch/modules/fused_moe/interface.py (9)
_supports_load_balancer(301-306)_load_balancer_start_wait_gpu_stage(317-320)_load_balancer_done_wait_gpu_stage(322-325)_load_balancer_update_statistic(327-352)_load_balancer_route(354-361)_load_balancer_get_local_statistic_tensor(373-378)_load_balancer_update_statistic_with_gathered_statistic(380-385)_load_balancer_start_set_cpu_stage(363-366)_load_balancer_done_set_cpu_stage(368-371)tensorrt_llm/_mnnvl_utils.py (3)
MnnvlMoe(352-624)mnnvl_moe_alltoallv_prepare_without_allgather(402-446)mnnvl_moe_alltoallv(531-592)
tensorrt_llm/_torch/modules/fused_moe/interface.py (6)
tensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py (23)
get_moe_load_balancer(1146-1156)add_layer(907-943)get_repeat_count(456-457)set_initial_weight_assignments(534-543)is_static_routing(459-460)is_static_routing(880-882)_add_raw_host_weight_for_unmap(502-511)start_wait_gpu_stage(588-604)done_wait_gpu_stage(606-615)update_local_statistic(648-681)update_statistic_with_local_ids(729-763)route(797-814)start_set_cpu_stage(617-634)done_set_cpu_stage(636-646)get_local_statistic_tensor(683-697)update_statistic_with_gathered_statistic(699-727)register_weight_slot(489-500)make_tensor_host_accessible(561-563)add_register_weight_fn(551-559)add_to_migrate_weight_fn(545-549)get_load_expert_ids(452-454)share_host_tensor_with_shape(212-237)pre_register_host_tensor_with_shape(187-210)tensorrt_llm/_torch/model_config.py (4)
num_local_slots(46-47)get_layer_initial_global_assignments(57-64)slot_start(50-51)slot_end(54-55)tensorrt_llm/_torch/modules/fused_moe/fused_moe_cutlass.py (1)
_supports_load_balancer(254-256)tensorrt_llm/_torch/modules/fused_moe/fused_moe_trtllm_gen.py (1)
_supports_load_balancer(183-185)tensorrt_llm/_torch/modules/fused_moe/fused_moe_wide_ep.py (1)
_supports_load_balancer(933-935)tensorrt_llm/_torch/distributed/ops.py (1)
AllReduce(554-710)
tensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py (1)
tensorrt_llm/_utils.py (2)
data_ptr(974-975)numel(1002-1003)
🪛 Ruff (0.14.3)
tensorrt_llm/_torch/modules/fused_moe/fused_moe_cute_dsl.py
148-148: Unused method argument: all_rank_num_tokens
(ARG002)
149-149: Unused method argument: use_dp_padding
(ARG002)
150-150: Unused method argument: repeating_info
(ARG002)
tensorrt_llm/_torch/modules/fused_moe/quantization.py
2134-2135: Consider iterable unpacking instead of concatenation
Replace with iterable unpacking
(RUF005)
2139-2140: Consider iterable unpacking instead of concatenation
Replace with iterable unpacking
(RUF005)
tensorrt_llm/_torch/modules/fused_moe/interface.py
431-431: Loop control variable expert_id not used within loop body
(B007)
tensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py
59-59: Avoid specifying long messages outside the exception class
(TRY003)
89-89: Avoid specifying long messages outside the exception class
(TRY003)
🔇 Additional comments (11)
tests/integration/test_lists/test-db/l0_rtx_pro_6000.yml (1)
96-99: Inconsistency between AI summary and code state.The AI-generated summary claims these tests are "Added" but the current code shows them commented out with a GDRCopy infrastructure dependency note. Additionally, the summary states that
test_guided_decoding[xgrammar-mtp_nextn=0]is deleted, yet line 102 shows it remains in the file.Please clarify: are these test entries intentionally disabled pending infrastructure setup, or should they be enabled as part of this PR? The presence of moe_backend parameters (WIDEEP, TRTLLM, CUTLASS) suggests these are new test variants designed to validate the PR's EPLB support across multiple MoE backends.
tests/integration/test_lists/test-db/l0_gb200_multi_gpus.yml (3)
45-45: No issues found. Test method and fp8 parameter both exist and are correctly supported.Verification confirms that
test_w4_4gpus_online_eplbexists inTestGPTOSSclass at line 3931 oftests/integration/defs/accuracy/test_llm_api_pytorch.pywith the parameterization["auto", pytest.param("fp8", marks=skip_pre_blackwell)]. Thefp8variant is explicitly supported and properly marked for Blackwell devices. The YAML entry is valid.
37-40: All test methods and moe_backend parameters are valid and supported.The test methods
test_nvfp4_4gpus_online_eplbandtest_bfloat16_4gpus_online_eplbexist in the test suite and are correctly parameterized with moe_backend values. The YAML entries use only supported backend combinations:
test_nvfp4_4gpus_online_eplbsupports moe_backend values: WIDEEP, TRTLLMtest_bfloat16_4gpus_online_eplbsupports moe_backend values: WIDEEP, CUTLASS, TRTLLMAll parameter combinations in lines 37-40 match the test definitions.
37-45: Test matrix alignment verified—all entries are valid.All test methods and parameterized variants referenced in the yml file exist in
tests/integration/defs/accuracy/test_llm_api_pytorch.pyand correctly match the test definitions:
test_bfloat16_4gpus_online_eplbsupports moe_backend variants (WIDEEP, CUTLASS, TRTLLM) and mtp_nextn=[0, 2]test_nvfp4_4gpus_online_eplbsupports moe_backend variants (WIDEEP, TRTLLM) and fp8kv=[True, False]test_w4_4gpuscorrectly parameterized with backend and tp/dp variantstest_w4_4gpus_online_eplbsupports kv_cache_dtype=[auto, fp8]test_fp8_tp4andtest_nvfp4_tp4both existThe yml file lists a selected subset of generated test variants, which is standard practice for targeted test execution across different environments.
tests/integration/test_lists/qa/llm_function_core.txt (1)
467-470: Quiet add/remove; no issues spotted.The test list entries simply switch backend parameters to WIDEEP and remain well-formed. Nothing to change.
tests/integration/defs/.test_durations (1)
228-229: LGTM! Test duration metadata updated consistently.The test identifiers now include the
moe_backend=WIDEEPparameter, which aligns with the broader parameterization of online EPLB tests across MOE backends introduced in this PR.Also applies to: 255-256
tests/integration/test_lists/test-db/l0_gb300_multi_gpus.yml (1)
42-45: Backend coverage expanded appropriately for GB300 hardware.The changes expand online EPLB test coverage across multiple MOE backends (WIDEEP, TRTLLM, CUTLASS), which is appropriate for comprehensive multi-GPU testing on GB300 hardware. The test variants cover key combinations:
- nvfp4 with both WIDEEP and TRTLLM backends (fp8kv=True)
- bfloat16 with both WIDEEP and CUTLASS backends (mtp_nextn=2)
Verify whether any of these online EPLB tests require the
ISOLATIONmarker (like Line 41) to prevent resource conflicts or ensure deterministic behavior, especially given the multi-backend and multi-GPU nature of these tests.tests/integration/test_lists/qa/llm_function_nim.txt (1)
280-283: LGTM! Online EPLB tests now explicitly specify WIDEEP backend.The test identifiers are updated consistently to include
moe_backend=WIDEEP, covering the key parameter combinations for online EPLB scenarios (bfloat16 with mtp_nextn variations and nvfp4 with fp8kv variations).tests/integration/test_lists/qa/llm_function_core_sanity.txt (1)
55-56: LGTM! Sanity test list updated with explicit backend parameterization.The online EPLB test entries now consistently include
moe_backend=WIDEEP, maintaining appropriate coverage for sanity testing with the key parameter variations (mtp_nextn and fp8kv).Also applies to: 60-61
tests/integration/defs/accuracy/test_llm_api_pytorch.py (2)
1620-1641: LGTM! MoE backend parameterization enables comprehensive online EPLB testing.The parameterization correctly extends test coverage to verify online expert parallel load balancing across WIDEEP, CUTLASS, and TRTLLM backends. The implementation properly integrates the
moe_backendparameter into the test logic and follows established patterns.
1645-1667: Verify CUTLASS exclusion from NVFP4 online EPLB testing.The parameterization for this test includes only
["WIDEEP", "TRTLLM"], excluding CUTLASS which was present in the bfloat16 variant above (line 1620). Please confirm this is intentional—likely due to NVFP4 quantization not supporting CUTLASS with online EPLB, or a known issue on specific SM versions.Run the following to check if there are any related comments or constraints in the MoE backend implementations:
Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>
Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>
Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>
Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>
Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>
Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>
Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>
Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>
Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>
Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>
Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>
Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>
Signed-off-by: Dongxu Yang <78518666+dongxuy04@users.noreply.github.com>
dongxuy04
force-pushed
the
user/dongxuy/trtllmgen_eplb
branch
from
aaaf4bb to
b94a262
Compare
|
/bot run --disable-fail-fast |
|
/bot kill |
dongxuy04
force-pushed
the
user/dongxuy/trtllmgen_eplb
branch
from
b94a262 to
8378276
Compare
|
/bot run --disable-fail-fast |
|
PR_Github #24269 [ run ] triggered by Bot. Commit: |
|
PR_Github #24226 [ run ] completed with state |
|
PR_Github #24272 [ kill ] triggered by Bot. Commit: |
|
PR_Github #24269 [ run ] completed with state |
|
PR_Github #24272 [ kill ] completed with state |
|
PR_Github #24273 [ run ] triggered by Bot. Commit: |
|
PR_Github #24273 [ run ] completed with state |
7 participants
Summary by CodeRabbit
New Features
Tests
Description
Enable EPLB for trtllm-gen and cutlass backend:
Known issues:
For now, there might be some issues when using EPLB backend with TRTLLM-Gen backend and number of expert per group is larger than 32. The reason is that our current TRTLLM-Gen fused routing kernel doesn't support that as no models using that. But when using EPLB with redundant experts like EPLB288 for DeepSeek-R1, it seems that we have more than that. In thant case, ideally, we should not run the router inside TRTLLM-Gen, but since there are some issues with AutoTuner when using topk input in that case, it is not enabled now.
Test with GPT-OSS 120B on 4 GB200 nodes (16 GPU) with EP=16, BS=512, trtllm-bench IFB mode.
Without EPLB:
0: ===========================================================
0: = PERFORMANCE OVERVIEW
0: ===========================================================
0: Request Throughput (req/sec): 214.6129
0: Total Output Throughput (tokens/sec): 211175.4890
0: Total Token Throughput (tokens/sec): 441672.5608
0: Total Latency (ms): 38171.0446
0: Average request latency (ms): 35473.2092
0: Per User Output Throughput [w/ ctx] (tps/user): 27.7429
0: Per GPU Output Throughput (tps/gpu): 13198.4681
With 144 slot EPLB (16 redundant experts):
0: ===========================================================
0: = PERFORMANCE OVERVIEW
0: ===========================================================
0: Request Throughput (req/sec): 224.4187
0: Total Output Throughput (tokens/sec): 220824.1976
0: Total Token Throughput (tokens/sec): 461852.7904
0: Total Latency (ms): 36503.1961
0: Average request latency (ms): 33789.7656
0: Per User Output Throughput [w/ ctx] (tps/user): 29.1216
0: Per GPU Output Throughput (tps/gpu): 13801.5124
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.
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.