Skip to content

[Perf][gfx1250] Add combined benchmark driver - #5454

Draft
JiaoliangYu wants to merge 7 commits into
ROCm:mainfrom
JiaoliangYu:gfx1250/bench-combo
Draft

JiaoliangYu wants to merge 7 commits into
ROCm:mainfrom
JiaoliangYu:gfx1250/bench-combo

Conversation

@JiaoliangYu

@JiaoliangYu JiaoliangYu commented Sep 11, 2026

Copy link
Copy Markdown
Contributor

The gfx1250 microbench coverage is spread across several operator test entry points, which makes full performance and DeepSeek-V4 sweeps cumbersome to run and compare. This adds a single combo driver with curated operator sets, shared shape and initialization controls, optional SMI telemetry, kernel summaries, and keep-going failure reporting.

The driver uses the original structured transport: child UTs emit record-oriented JSON, the combo validates and forwards those records, and captured compiler, profiler, NUMA, and logger output is not treated as benchmark data.

This draft depends on the test layout, JSON summary helper, and aiter.smi_monitor support from #5391. It is intentionally kept as one file in one commit so the driver can be cherry-picked directly where needed.

Validation:

  • ruff check --no-cache op_tests/bench_gfx1250_combo.py
  • black --check op_tests/bench_gfx1250_combo.py
  • python3 -m py_compile op_tests/bench_gfx1250_combo.py

@github-actions

Copy link
Copy Markdown
Contributor

🏷️ CI Guide

Runs automatically on every PR:

  • ✅ Pre-checks (submodule verification, code formatting)
  • ✅ Aiter op tests (gfx942 + gfx950)
  • ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label Tests
ci:gfx1250-ffm-triton Run the five-shard gfx1250 FFM Triton test suite
ci:triton-300x Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
multigpu Aiter multi-GPU tests on the 8-GPU runner
ci:sglang SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 5454 --add-label <label>

PR title tags & labels:
Component tags ([Triton/Gluon], [HIP], [CK], [ASM], ...) are added to the PR title and as PR labels automatically from the changed files and re-synced on every push — change-type tags like [fix]/[Perf], op tags like [MLA], and human labels (ci:*) are left untouched. Add the no-auto-title label to opt this PR out.

@JiaoliangYu
JiaoliangYu force-pushed the gfx1250/bench-combo branch 6 times, most recently from 0192e0a to 13c8b41 Compare September 12, 2026 02:34
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant