[https://nvbugs/5498478][fix] Fix eagle3 fp8 kv target model + bf16 draft model + chunked prefill #8910
Conversation
Signed-off-by: Dylan Chen <[email protected]>
📝 WalkthroughWalkthroughThe changes introduce output dtype support for e4m3 data type handling. An optional output_dtype parameter is added to the cubin header selection logic, with conditional disabling when combining e4m3 with bf16/fp16 outputs. XQA kernel selection is adjusted to allow generation with these configurations. A comprehensive end-to-end test for Qwen3-Eagle3 speculative decoding is added. Changes
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~20 minutes
Pre-merge checks and finishing touches✅ Passed checks (2 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: 1
🧹 Nitpick comments (2)
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplJIT/decoderXQAImplJIT.cpp (1)
127-134: Select JIT XQA when FMHA lacks support (correct); fix log wording.The gating to force JIT XQA for {E4M3 KV cache + FP16/BF16 inputs + non-FP8 outputs} despite no perf gain is correct and matches the Eagle3 BF16 draft requirement. Minor nit: the log says “MMHA” but the project consistently uses “FMHA”.
Apply:
- TLLM_LOG_DEBUG( - "JIT XQA is selected in the generation phase for fp16/bf16 input and e4m3 kv cache because MMHA " - "does not support this combination."); + TLLM_LOG_DEBUG( + "JIT XQA is selected in the generation phase for fp16/bf16 input and E4M3 KV cache because FMHA " + "does not support this combination.");cpp/kernels/fmha_v2/setup.py (1)
3234-3236: Fix late‑binding of loop variables in nested helper (Ruff B023).get_lname_from_kname captures sm/head_size/prec/output_prec from the surrounding loop; late binding risks wrong values if reused. Bind them via defaults.
Apply:
- def get_lname_from_kname(kname: str) -> str: - if use_cubin_header(int(sm), int(head_size), prec.lower(), - output_prec.lower()): + def get_lname_from_kname( + kname: str, + sm=sm, + head_size=head_size, + prec=prec, + output_prec=output_prec) -> str: + if use_cubin_header(int(sm), int(head_size), + str(prec).lower(), + str(output_prec).lower()): return 'nullptr'This preserves current behavior and silences B023. Based on static analysis hints.
Also applies to: 3254-3256
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (3)
cpp/kernels/fmha_v2/setup.py(5 hunks)cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplJIT/decoderXQAImplJIT.cpp(1 hunks)tests/unittest/_torch/speculative/test_eagle3.py(1 hunks)
🧰 Additional context used
📓 Path-based instructions (6)
**/*.{h,hpp,hh,hxx,cpp,cxx,cc,cu,cuh}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.{h,hpp,hh,hxx,cpp,cxx,cc,cu,cuh}: Namespace closing braces must include a trailing comment with the namespace name (e.g., '} // namespace foo').
Prefer const or constexpr variables over #define for constants.
Declare variables that are not modified after initialization as const.
Avoid magic literals in code; except for 0, nullptr, true, false. Use named constants for comparisons and logic.
Use Allman brace style for formatting.
Place the semicolon of an empty for/while loop on a new line.
Bodies of switch/while/do-while/for must be compound statements (brace-delimited), and if/else must always be followed by brace-delimited statements.
Type names (e.g., classes) must be CamelCase starting with an uppercase letter (e.g., FooBar).
Local variables, methods, and namespaces use lowerCamelCase (e.g., localFooBar).
Non-magic-number global variables that are non-static and not in an anonymous namespace must be lowerCamelCase prefixed with 'g' (e.g., gDontUseGlobalFoos).
Non-magic-number globals that are static or in an anonymous namespace use lowerCamelCase prefixed with 's' (e.g., sMutableStaticGlobal).
Locally visible static variables use lowerCamelCase with 's' prefix (e.g., static std::once_flag sFlag).
Private/protected member variables use 'm' prefix with CamelCase (e.g., mNbFooValues). Public members may omit, but 'm' is encouraged for clarity.
Constants (enums, global constants, static constants, and function-scope magic/literal constants) use uppercase SNAKE_CASE with 'k' prefix (e.g., kDIGIT_NUM).
Function-scope constants that are not magic numbers or literals are named like non-constant variables (e.g., bool const pass = a && b).
If macros are necessary, name them in UPPER_SNAKE_CASE (e.g., FOO_VERSION) and prefer constants over #define.
Use LLVM clang-format; wrap lines at a maximum of 120 columns; use '// clang-format off/on' sparingly with justification.
Use smart pointers for heap allocations; prefer unique_ptr for sole ownership, shared_ptr for shared...
Files:
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplJIT/decoderXQAImplJIT.cpp
**/*.{cpp,cxx,cc,cu,h,hpp,hh,hxx,cuh}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
C++ filenames should be lowerCamelCase (first letter lowercase) and must be case-insensitive unique within a compilation target.
Files:
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplJIT/decoderXQAImplJIT.cpp
**/*.{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:
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplJIT/decoderXQAImplJIT.cpptests/unittest/_torch/speculative/test_eagle3.pycpp/kernels/fmha_v2/setup.py
**/*.{h,hpp,hh,hxx,cpp,cxx,cc}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.{h,hpp,hh,hxx,cpp,cxx,cc}: Prefer anonymous namespaces over 'static' for internal linkage of functions.
All templates (class/function/member/static) must be instantiated at least once; non-POD classes should have private data members.
Files:
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplJIT/decoderXQAImplJIT.cpp
**/*.{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:
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplJIT/decoderXQAImplJIT.cpptests/unittest/_torch/speculative/test_eagle3.pycpp/kernels/fmha_v2/setup.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:
tests/unittest/_torch/speculative/test_eagle3.pycpp/kernels/fmha_v2/setup.py
🧠 Learnings (4)
📓 Common learnings
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.
📚 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:
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplJIT/decoderXQAImplJIT.cpp
📚 Learning: 2025-08-14T21:04:50.248Z
Learnt from: thorjohnsen
Repo: NVIDIA/TensorRT-LLM PR: 6910
File: cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp:0-0
Timestamp: 2025-08-14T21:04:50.248Z
Learning: In KV cache onboarding logic during prefill in cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp, when calculating which blocks fall within the attention window, use getTokensPerBlock() to advance token indices rather than block->getUniqueTokens().size(), because the calculation needs to consider the post-prefill state where blocks will be filled to capacity, not their current token count.
Applied to files:
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplJIT/decoderXQAImplJIT.cpp
📚 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/unittest/_torch/speculative/test_eagle3.py
🧬 Code graph analysis (2)
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplJIT/decoderXQAImplJIT.cpp (1)
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplPrecompiled.cpp (4)
xqaParams(106-135)xqaParams(106-106)xqaParams(137-154)xqaParams(137-137)
tests/unittest/_torch/speculative/test_eagle3.py (4)
tests/scripts/perf-sanity/run_benchmark_serve.py (1)
llm_models_root(174-175)tensorrt_llm/llmapi/llm_args.py (4)
KvCacheConfig(1265-1409)CudaGraphConfig(106-163)EagleDecodingConfig(575-692)speculative_model_dir(1764-1765)tensorrt_llm/_torch/models/modeling_llama.py (2)
dtype(1092-1093)tokenizer(1080-1081)tests/unittest/llmapi/test_llm.py (1)
encode(316-317)
🪛 Ruff (0.14.3)
tests/unittest/_torch/speculative/test_eagle3.py
630-630: zip() without an explicit strict= parameter
Add explicit value for parameter strict=
(B905)
cpp/kernels/fmha_v2/setup.py
3234-3234: Function definition does not bind loop variable sm
(B023)
3234-3234: Function definition does not bind loop variable head_size
(B023)
3234-3234: Function definition does not bind loop variable prec
(B023)
3235-3235: Function definition does not bind loop variable output_prec
(B023)
⏰ 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 (3)
cpp/kernels/fmha_v2/setup.py (3)
3078-3080: Call site updated to pass output_dtype (good).Plumbs kspec.output_dtype into use_cubin_header; consistent with the new signature.
3788-3801: Confirm skip policy for FP8 WS + return_softmax.Setting
skip_combination = return_softmaxskips generation whenever softmax stats are requested. Intentional? It disables FP8 WS kernels with softmax return for all input layouts, not just non‑MLA. If that’s desired, add a brief comment to prevent future regressions.Would you confirm test coverage exercises a “return_softmax=True” FP8 WS path (expected to skip) and a “return_softmax=False” path (expected to generate)?
3066-3069: All call sites correctly use the new 4-argument signature with theoutput_dtypeparameter. No stale calls with the old signature remain. The early-return logic for E4M3×(BF16|FP16) is properly implemented. Verification passed.
| eagle_model_dir = "/ziqingc_large/03_Data/models/Zhi-Create-Qwen3-32B-Eagle3" # temp | ||
| target_model_dir = f"{models_path}/Qwen3/Qwen3-32B" | ||
| if fp8_target: | ||
| target_model_dir = f"{models_path}/Qwen3/Qwen3-32B-FP8/" | ||
|
|
There was a problem hiding this comment.
Remove hard-coded absolute draft-model path
eagle_model_dir points to /ziqingc_large/..., which only exists on the author’s workstation. In CI (or any other dev machine) this path resolves to nothing, so the LLM load will throw before the test even starts. Please source the draft checkpoint via llm_models_root() (or a fixture that stages the model) so the test remains runnable everywhere.
Apply this diff as a starting point:
- eagle_model_dir = "/ziqingc_large/03_Data/models/Zhi-Create-Qwen3-32B-Eagle3" # temp
+ eagle_model_dir = os.path.join(
+ models_path, "Zhi-Create-Qwen3-32B-Eagle3")(Adjust the relative directory if the checkpoint sits elsewhere under LLM_MODELS_ROOT.)
🤖 Prompt for AI Agents
In tests/unittest/_torch/speculative/test_eagle3.py around lines 543 to 547,
remove the hard-coded absolute path assigned to eagle_model_dir
(/ziqingc_large/...), and instead construct the path relative to the test's
model root (use llm_models_root() or a provided fixture that stages models).
Replace the literal with a join of llm_models_root() and the relative checkpoint
directory (or use the fixture value) so the test resolves the draft checkpoint
on CI and other machines; adjust the relative subpath as needed to match where
the checkpoint lives under LLM_MODELS_ROOT.
1 participant
Summary by CodeRabbit
Release Notes
New Features
Improvements
Description
Fix eagle3 FP8 kv target model + BF16 draft model + chunked prefill by supporting FP8 FMHA with BF16 out in draft layers, because the attention of the 2nd chunk context phase needs to load FP8 KV cache and the output of attention should be converted to BF16 in BF16 draft layers.
Test Coverage
A test of eagle3 + FP8 KV target model + bf16 draft model + chunked prefill is added:
tests/unittest/_torch/speculative/test_eagle3.py::test_qwen3_eagle3[True-True-True-True]
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.