forked from ccf-ai-infra/TileOPs-Metax
83244ccbc3
Closes #718 ## Summary - Add 12 attention operator entries (MHA/GQA fwd/bwd, decode, paged decode, sliding window, DeepSeek MLA/DSA) to `ops_manifest.yaml` with `status: implemented` and Tier 2 roofline stubs - Add `status` field (`implemented` | `spec-only`) to the manifest schema header documentation - Create `tileops/perf/formulas.py` with roofline stub functions for all attention op families (returns `{flops, bytes}` dicts) - Enhance `tileops/manifest.py` `eval_roofline()` to support `func`-mode dispatch (calls formula functions with workload params) - Add comprehensive tests: manifest loading validation and attention-specific entry/signature/roofline tests ## Test plan - [x] AC-1: All 12 attention ops present in `ops_manifest.yaml` with `status: implemented` - [x] AC-2: `status` field documented in manifest schema header comment - [x] AC-3: `tileops/perf/formulas.py` exists with callable stub functions for all 12 ops returning `{flops, bytes}` - [x] AC-4: `python -c "import yaml; yaml.safe_load(open('tileops/ops_manifest.yaml'))"` passes - [x] AC-5: Existing tests unaffected (1989 passed, 13 skipped) ## Additional context Workloads include both 8B and 70B model configurations with short (2K/4K) and long (8K/32K) sequence lengths. Roofline functions are Tier 2 stubs (formula-based, not benchmarked) following the pattern established by norm ops. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| profiles | ||
| __init__.py | ||
| formulas.py | ||
| profile.py | ||