Support for non-saturating mode for fp8
Open
- Dominant language
- C++
- Stars
- 361
- Forks
- 60
- Avg merge
- 18h 41m
- Merged PRs (30d)
- 3
Description
OFP8 (https://www.opencompute.org/documents/ocp-8-bit-floating-point-specification-ofp8-revision-1-0-2023-12-01-pdf-1) defines two saturation modes - saturating and non-saturating mode.
It looks like the current code runs in saturating mode by default.
```
>>> math.isnan(np.float32(449).astype(float8_e4m3fn))
False
```
Wondering if there's any plan to support both.
Thank you.
Contributor guide
Research direction
No implementation files, tests, or entry points are identified in the issue. Start by reading the OFP8 specification and tracing the current float8_e4m3fn conversion behavior; done would require a decided API for selecting saturation modes and coverage demonstrating both modes.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp, numpy
- Domain
- machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 30/100