forked from ccf-ai-infra/TileOPs-Metax
d5b3279a16
## Summary - bump TileLang references from `0.1.8` to `0.1.9` - add a centralized temporary TileLang 0.1.9 skip set in `tests/conftest.py` - update existing temporary skip messages from the old `5f70374c` tracker to #1039 - keep the suite green while the skipped buckets are fixed and restored under #1041 The skip set currently covers 10 file-level skips, 2 nodeid-prefix skips, and 71 exact nodeid skips. In response to review, `test_mha.py` and `test_mha_decode.py` are no longer skipped at file level, so MHA forward and the short-context MHA decode case continue to run. `test_gqa.py` remains file-level skipped because narrowing it exposed a smoke-only hang in `test_gqa_fwd[1-1024-8-4-64-False-dtype0-False]` under TileLang 0.1.9. Closes #1042 Refs #1039 Refs #1041 ## Validation - `git diff --check` - `python3 -m ruff check tests/conftest.py tests/test_ci_venv_hash.py pyproject.toml` inside `tileops-runner:nightly-tl019-fullstack-no-tileops-ldfix` - TileLang 0.1.9 full validation with this skip set: `2160 passed, 232 skipped, 196 warnings in 40.21s` |
||
|---|---|---|
| .. | ||
| README.md | ||
| elementwise-evidence.md | ||
| elementwise.md | ||
README.md
Performance Guides
Empirical performance lessons per op category. Each category has:
- Checklist — heuristic rules for audit (lightweight, always load)
- Evidence — measured data and reasoning (load on demand)
Test Environment
All conclusions are scoped to this configuration. Re-validate when any component changes.
| Component | Value |
|---|---|
| GPU | NVIDIA H200 (HBM3e, 4.8 TB/s peak, SM_90a) |
| Driver / CUDA | 575.57.08 / 12.8 |
| PyTorch | 2.9.1+cu128 |
| TileLang | 0.1.9 |
| Profiler | CUPTI (primary); CUDA event+median fallback when CUPTI singleton unavailable |
Index
| Category | Checklist | Evidence |
|---|---|---|
| Elementwise | elementwise.md | elementwise-evidence.md |