- Notifications
You must be signed in to change notification settings - Fork1.9k
[https://nvbugs/5440241][fix] Fix 70B GSM8K Accuracy drop#7075
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to ourterms of service andprivacy statement. We’ll occasionally send you account related emails.
Already on GitHub?Sign in to your account
[https://nvbugs/5440241][fix] Fix 70B GSM8K Accuracy drop#7075
Uh oh!
There was an error while loading.Please reload this page.
Conversation
Signed-off-by: Chenfei Zhang <chenfeiz@nvidia.com>
coderabbitaibot commentedAug 20, 2025 • edited
Loading Uh oh!
There was an error while loading.Please reload this page.
edited
Uh oh!
There was an error while loading.Please reload this page.
📝 WalkthroughWalkthroughUpdates integration accuracy references for Llama-3.3-70B-Instruct on GSM8K and MMLU and adjusts two PyTorch accuracy tests to new FP8/FP4 model paths, introduce KvCacheConfig (free_gpu_memory_fraction=0.5), set max_tokens=256, add max_batch_size=32 in FP4 test, and change GPQADiamond evaluator args. Changes
Sequence Diagram(s)sequenceDiagram autonumber actor Tester participant Test as PyTorch Test participant LLM as LLM(init with KvCacheConfig) participant Eval as Task/Evaluator Tester->>Test: run test_fp8_tp4 / test_nvfp4_tp4 Test->>LLM: init(model_path, kv_cache_config[, max_batch_size]) Note right of LLM: KV cache reserves GPU memory per free_gpu_memory_fraction Test->>Eval: evaluate(llm, sampling_params(max_tokens=256)[, extra_evaluator_kwargs]) Eval->>LLM: generate() LLM-->>Eval: outputs Eval-->>Test: metrics Test-->>Tester: assert accuracy vs referencesEstimated code review effort🎯 2 (Simple) | ⏱️ ~10 minutes Possibly related PRs
Suggested labels
Suggested reviewers
Tip 🔌 Remote MCP (Model Context Protocol) integration is now available!Pro plan users can now connect to remote MCP servers from theIntegrations 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 withCodeRabbit:
SupportNeed help? Create a ticket on oursupport page for assistance with any issues or questions. CodeRabbit Commands (Invoked using PR/Issue comments)Type Other keywords and placeholders
Status, Documentation and Community
|
chenfeiz0326 commentedAug 20, 2025
/bot run |
tensorrt-cicd commentedAug 20, 2025
PR_Github #15881 [ run ] triggered by Bot |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others.Learn more.
Actionable comments posted: 0
🧹 Nitpick comments (2)
tests/integration/defs/accuracy/test_llm_api_pytorch.py (2)
435-444:NVFP4: memory safety knobs look right; consider aligning max_seq_len with the FP8 test
- model_path updated to the FP4 artifact, kv_cache_config set to 0.5 fraction, max_batch_size=32, and max_tokens=256 — all make sense to address the illegal memory access and stabilize accuracy.
For parity and reproducibility across devices, you may also set max_seq_len=8192 here (as you did in the FP8 test). Proposed local change:
- with LLM(model_path,- tensor_parallel_size=4,- max_batch_size=32,- kv_cache_config=kv_cache_config) as llm:+ with LLM(model_path,+ tensor_parallel_size=4,+ max_seq_len=8192,+ max_batch_size=32,+ kv_cache_config=kv_cache_config) as llm:
411-421:Residual reference to old FP8 path in Eagle3 test
- tests/integration/defs/accuracy/test_llm_api_pytorch.py:390 still points to
modelopt-hf-model-hub/Llama-3.3-70B-Instruct-fp8.
Update it to the new layout for consistency:- model_path = f"{llm_models_root()}/modelopt-hf-model-hub/Llama-3.3-70B-Instruct-fp8"+ model_path = f"{llm_models_root()}/llama-3.3-models/Llama-3.3-70B-Instruct-FP8"
- Verify that
llama-3.3-models/Llama-3.3-70B-Instruct-FP8exists in the CI model store.
📜 Review details
Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled
- Linear integration is disabled
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (3)
tests/integration/defs/accuracy/references/gsm8k.yaml(1 hunks)tests/integration/defs/accuracy/references/mmlu.yaml(1 hunks)tests/integration/defs/accuracy/test_llm_api_pytorch.py(3 hunks)
🧰 Additional context used
📓 Path-based instructions (2)
**/*.py
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
**/*.py: Python code must target Python 3.8+
Python indentation: 4 spaces, no tabs
Maintain module namespace in imports (from package.subpackage import foo; then use foo.SomeClass())
Python file names use snake_case
Python class names use PascalCase
Python functions/methods and local variables use snake_case; variables starting with a number get k_ prefix (e.g., k_99th_percentile)
Global variables use G_ prefixed UPPER_SNAKE_CASE (e.g., G_MY_GLOBAL)
Constants use UPPER_SNAKE_CASE in Python
Avoid shadowing variables from outer scopes in Python
Initialize all externally visible members of a Python class ininit
Prefer docstrings for interfaces used outside a file; comments for local code
Use Google-style docstrings for classes and functions (Sphinx-parsable)
Document attributes/variables inline with short docstrings
Avoid reflection when simple alternatives exist (e.g., prefer explicit parameters over dict(**locals()))
In try/except, catch the narrowest exceptions possible
For duck-typing with try/except, keep try body minimal and put logic in else
Files:
tests/integration/defs/accuracy/test_llm_api_pytorch.py
**/*.{cpp,cxx,cc,cu,h,hpp,hxx,hh,cuh,py}
📄 CodeRabbit Inference Engine (CODING_GUIDELINES.md)
Prepend NVIDIA copyright header (current year) to all source files
Files:
tests/integration/defs/accuracy/test_llm_api_pytorch.py
🧠 Learnings (1)
📚 Learning: 2025-07-28T17:06:08.621Z
Learnt from: moraxuPR: NVIDIA/TensorRT-LLM#6303File: tests/integration/test_lists/qa/examples_test_list.txt:494-494Timestamp: 2025-07-28T17:06:08.621ZLearning: 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/defs/accuracy/test_llm_api_pytorch.py
🧬 Code Graph Analysis (1)
tests/integration/defs/accuracy/test_llm_api_pytorch.py (4)
tests/integration/defs/conftest.py (1)
llm_models_root(77-83)tensorrt_llm/llmapi/llm_args.py (1)
KvCacheConfig(923-1002)tensorrt_llm/quantization/mode.py (1)
QuantAlgo(23-44)tensorrt_llm/sampling_params.py (1)
SamplingParams(125-483)
🔇 Additional comments (3)
tests/integration/defs/accuracy/references/gsm8k.yaml (1)
16-21:Confirm disambiguation for FP8 entries in GSM8K referencesI ran the verification script but it failed sorting entries missing both
quant_algoandkv_cache_quant_algo. Please manually verify that the harness selection logic:
- Matches on both
quant_algoandkv_cache_quant_algowhen present, so the specific FP8+KV entry wins over the generic FP8 entry.- Never falls back to the generic
quant_algo: FP8row when the model is actually using FP8 KV–cache.File under review (lines 16–21):
tests/integration/defs/accuracy/references/gsm8k.yaml-quant_algo:NVFP4kv_cache_quant_algo:FP8accuracy:87.33-quant_algo:FP8kv_cache_quant_algo:FP8accuracy:90.30-quant_algo:FP8accuracy:90.30tests/integration/defs/accuracy/references/mmlu.yaml (1)
67-72:MMLU entries include dual FP8 variants; ensure resolver specificity
- File: tests/integration/defs/accuracy/references/mmlu.yaml
Section: meta-llama/Llama-3.3-70B-Instruct- Confirmed entries:
• NVFP4 + kv_cache_FP8 → 78.78
• FP8 + kv_cache_FP8 → 80.40
• FP8 (plain) → 80.40- Verify these values match the source benchmarks and that your matcher will prefer the “FP8 + kv_cache_FP8” entry over the plain “FP8” when both criteria apply.
tests/integration/defs/accuracy/test_llm_api_pytorch.py (1)
429-431:Switching GPQA to extra_evaluator_kwargs is consistent with recent harness changesMoving GPQADiamond to use extra_evaluator_kwargs=dict(apply_chat_template=True) (without sampling_params) matches the pattern used elsewhere in this file.
Also applies to: 452-454
tensorrt-cicd commentedAug 20, 2025
PR_Github #15881 [ run ] completed with state |
5acf213 intoNVIDIA:release/1.0Uh oh!
There was an error while loading.Please reload this page.
Signed-off-by: Chenfei Zhang <chenfeiz@nvidia.com>
Signed-off-by: Chenfei Zhang <chenfeiz@nvidia.com>Signed-off-by: Wangshanshan <30051912+dominicshanshan@users.noreply.github.com>
Signed-off-by: Chenfei Zhang <chenfeiz@nvidia.com>Signed-off-by: Wangshanshan <30051912+dominicshanshan@users.noreply.github.com>
Signed-off-by: Chenfei Zhang <chenfeiz@nvidia.com>Signed-off-by: Wangshanshan <30051912+dominicshanshan@users.noreply.github.com>
Signed-off-by: Chenfei Zhang <chenfeiz@nvidia.com>Signed-off-by: Wangshanshan <30051912+dominicshanshan@users.noreply.github.com>
Signed-off-by: Chenfei Zhang <chenfeiz@nvidia.com>Signed-off-by: Wangshanshan <30051912+dominicshanshan@users.noreply.github.com>
Uh oh!
There was an error while loading.Please reload this page.
Description
This PR fixes Llama3.3 70B fp8/fp4 gsm8k's accuracy drop by adding
max_tokens=256.This PR fixes Llama3.3 70B fp4's illegal memory access by adding
free_gpu_mem_fraction=0.5, max_batch_size=32.FP8 MMLU on H200:
MMLU weighted average accuracy: 80.51 (4104)
FP8 GSM8K on H200:
FP8 GPQA_Diamond on H200:
FP8 MMLU on B200:
MMLU weighted average accuracy: 80.48 (4104)
FP8 GSM8K on B200:
FP8 GPQA_Diamond on B200:
FP4 MMLU on B200:
MMLU weighted average accuracy: 78.78 (4104)
FP4 GSM8K on B200:
FP4 GPQA_Diamond on B200:
Test Coverage
GitHub Bot Help
/bot [-h] ['run', 'kill', 'skip', 'reuse-pipeline'] ...Provide a user friendly way for developers to interact with a Jenkins server.
Run
/bot [-h|--help]to print this help message.See details below for each supported subcommand.
run [--reuse-test (optional)pipeline-id --disable-fail-fast --skip-test --stage-list "A10-PyTorch-1, xxx" --gpu-type "A30, H100_PCIe" --test-backend "pytorch, cpp" --add-multi-gpu-test --only-multi-gpu-test --disable-multi-gpu-test --post-merge --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" --detailed-log --debug(experimental)]Launch build/test pipelines. All previously running jobs will be killed.
--reuse-test (optional)pipeline-id(OPTIONAL) : Allow the new pipeline to reuse build artifacts and skip successful test stages from a specified pipeline or the last pipeline if no pipeline-id is indicated. If the Git commit ID has changed, this option will be always ignored. The DEFAULT behavior of the bot is to reuse build artifacts and successful test results from the last pipeline.--disable-reuse-test(OPTIONAL) : Explicitly prevent the pipeline from reusing build artifacts and skipping successful test stages from a previous pipeline. Ensure that all builds and tests are run regardless of previous successes.--disable-fail-fast(OPTIONAL) : Disable fail fast on build/tests/infra failures.--skip-test(OPTIONAL) : Skip all test stages, but still run build stages, package stages and sanity check stages. Note: DoesNOT update GitHub check status.--stage-list "A10-PyTorch-1, xxx"(OPTIONAL) : Only run the specified test stages. Examples: "A10-PyTorch-1, xxx". Note: DoesNOT 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: DoesNOT 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: DoesNOT update GitHub pipeline status.--only-multi-gpu-test(OPTIONAL) : Only run the multi-GPU tests. Note: DoesNOT update GitHub check status.--disable-multi-gpu-test(OPTIONAL) : Disable the multi-GPU tests. Note: DoesNOT 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: DoesNOT 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.
Summary by CodeRabbit
Tests
Chores