[OMNIML-6177] [Draft, not for merge] GPU CI speed-up series (runtime measurement) by shengliangxu · Pull Request #2730 · NVIDIA/Model-Optimizer

Run the unit tests in parallel with pytest-xdist

shengliangxu

committed

6bf4ca4

Write only coverage.xml in CI

shengliangxu

committed

30d9973

Use the sysmon coverage core on the unit, GPU and regression lanes

shengliangxu

committed

832b5b5

Cache JIT extensions per lane, save them on failure and cache Triton kernels

shengliangxu

committed

7f8854a

Give the Mamba search-space test and the Megatron worker pool more time

shengliangxu

committed

a355d34

Notice a dead pool worker within a second and add worker-state helpers

shengliangxu

committed

7faaa68

Share one Megatron worker pool across the gpu_megatron modules

shengliangxu

committed

2daa767

Cache the CUDA wheels the gpu lane compiles

shengliangxu

committed

e71b166

Run the tests marked release on the nightly and on-demand gpu runs

shengliangxu

committed

15c7255

Run the real-weights Megatron MMLU test only with --run-release

shengliangxu

committed

985713a

Size the vLLM MoE engines' KV pool down

shengliangxu

committed

533945e

Cut the DFlash regression to 500 training steps and a 20-step resume

shengliangxu

committed

576c8f5

Skip the Megatron state reset for megatron_bridge tests that use no GPU

shengliangxu

committed

c4872a8

Use the tiny local tokenizer in six preprocess_data tests

shengliangxu

committed

cb1fc12

Run the CPU-only ONNX quantization tests in the unit lane

shengliangxu

committed

9536e27

Calibrate test_concat_elim on CPU with a small model

shengliangxu

committed

c5af12e

Benchmark a small model in the autotune benchmark tests

shengliangxu

committed

1662f96

Skip trtexec's 3-second measurement soak in the autotune workflow test

shengliangxu

committed

aea663a

Shrink the puzzletron test data and fold a duplicate NAS test into another

shengliangxu

committed

7aa6a21

Calibrate the unified HF export test on local text, with no previews or batch probing

shengliangxu

committed

b3bdaf0

Calibrate the GPTQ end-to-end test on local text

shengliangxu

committed

72c16fe

Evaluate one MMLU subtask in test_lm_eval_hf

shengliangxu

committed

35b7eae

Use one MMLU prompt per subject in test_qwen3_eval_fp8

shengliangxu

committed

43eb82e

Quantize once per module in the llm_qat QAT and QAD tests

shengliangxu

committed

987ad4c

Skip Triton GEMM autotune in the 1-GPU EAGLE3 compile smoke test

shengliangxu

committed

3e60ef1

Drop the duplicate tiny-llama offline EAGLE3 training test

shengliangxu

committed

e53b909

Run the gated Nemotron v3 dataset smoke test only on the nightly

shengliangxu

committed

faa41c0

Drop DataLoader workers from the torch_onnx calibration loaders

shengliangxu

committed

bdfa096

Skip swinv2 int8/mxfp8/auto on PRs and run trtexec for one iteration

shengliangxu

committed

2815b82

Add a worker runner so example steps share one python launch

shengliangxu

committed

8a1b465

Run the torch_onnx lane's torch_quant_to_onnx.py steps in one worker

shengliangxu

committed

d8b9cf3

Make test_example_runner.py collectable on Windows

shengliangxu

committed

846c0e4

Restore ConstraintsFunc._get_true_latency after test_search_constraints

shengliangxu

committed

1e2d732

Give the unit tests 120 s per test on every platform

shengliangxu

committed

954ac75

Navigation Menu

Repository navigation

[OMNIML-6177] [Draft, not for merge] GPU CI speed-up series (runtime measurement) - #2730

[OMNIML-6177] [Draft, not for merge] GPU CI speed-up series (runtime measurement)#2730
shengliangxu wants to merge 34 commits into
mainfrom
shengliangx/ci-speed

Commits on Oct 17, 2026

Run the unit tests in parallel with pytest-xdist

Write only coverage.xml in CI

Use the sysmon coverage core on the unit, GPU and regression lanes

Cache JIT extensions per lane, save them on failure and cache Triton kernels

Give the Mamba search-space test and the Megatron worker pool more time

Notice a dead pool worker within a second and add worker-state helpers

Share one Megatron worker pool across the gpu_megatron modules

Cache the CUDA wheels the gpu lane compiles

Run the tests marked release on the nightly and on-demand gpu runs

Run the real-weights Megatron MMLU test only with --run-release

Size the vLLM MoE engines' KV pool down

Cut the DFlash regression to 500 training steps and a 20-step resume

Skip the Megatron state reset for megatron_bridge tests that use no GPU

Use the tiny local tokenizer in six preprocess_data tests

Run the CPU-only ONNX quantization tests in the unit lane

Calibrate test_concat_elim on CPU with a small model

Benchmark a small model in the autotune benchmark tests

Skip trtexec's 3-second measurement soak in the autotune workflow test

Shrink the puzzletron test data and fold a duplicate NAS test into another

Calibrate the unified HF export test on local text, with no previews or batch probing

Calibrate the GPTQ end-to-end test on local text

Evaluate one MMLU subtask in test_lm_eval_hf

Use one MMLU prompt per subject in test_qwen3_eval_fp8

Quantize once per module in the llm_qat QAT and QAD tests

Skip Triton GEMM autotune in the 1-GPU EAGLE3 compile smoke test

Drop the duplicate tiny-llama offline EAGLE3 training test

Run the gated Nemotron v3 dataset smoke test only on the nightly

Drop DataLoader workers from the torch_onnx calibration loaders

Skip swinv2 int8/mxfp8/auto on PRs and run trtexec for one iteration

Add a worker runner so example steps share one python launch

Run the torch_onnx lane's torch_quant_to_onnx.py steps in one worker

Make test_example_runner.py collectable on Windows

Restore ConstraintsFunc._get_true_latency after test_search_constraints

Give the unit tests 120 s per test on every platform