|
Download compatibility/VALIDATION.md from 4E-AI/Fidel2.0-1B: direct link, hf CLI and curl.
- Browser
- Download file 2.16 kB
-
https://huggingface.co/4E-AI/Fidel2.0-1B/resolve/main/compatibility/VALIDATION.md
- Command line
-
hf download hf://4E-AI/Fidel2.0-1B/compatibility/VALIDATION.md
-
curl -L -o VALIDATION.md https://huggingface.co/4E-AI/Fidel2.0-1B/resolve/main/compatibility/VALIDATION.md
2.16 kB
| # Validation scope | |
| These results apply to the exact wheel identified in `SHA256SUMS`, with all 20 bundled source/assets files matching the source published here. They measure numerical agreement and API/cache behavior, not programming benchmark accuracy or generation speed. | |
| | Device family | Native model frontier/logit checks | Public generation/API checks | Guarded cache checks | Mamba chunk-state cache checks | | |
| | --- | ---: | ---: | ---: | ---: | | |
| | L4 (SM89) | 180/180 exact | 8/8 exact | 14/14 exact | 14/14 exact | | |
| | A10 (SM86) | 180/180 exact | 8/8 exact | 14/14 exact | 14/14 exact | | |
| | A100 (SM80) | 180/180 exact | 8/8 exact | 14/14 exact | 14/14 exact | | |
| | H100 (SM90) | 180/180 exact | 8/8 exact | 14/14 exact | 14/14 exact | | |
| | RTX PRO 6000 (SM120) | 180/180 exact | 8/8 exact | 14/14 exact | 14/14 exact | | |
| The model cases span sequence boundaries 15, 16, 17, 31, 32, 33, 127, 128 and 129, plus six retained text frontiers. Public generation checks cover three-token greedy continuations. Cache checks compare cached results to uncached results at the relevant boundaries and verify request/prompt guards; actual complete Mamba chunks are reused. | |
| Separate operator checks establish native attention reduction profiles through selected columns at all key lengths 1 through 8,192. This is not a complete 8,192-token model validation. CPU checks passed 46 tests and 114 quantization goldens; installed-package six-shard save/reload preserved all 359 original tensor hashes, shapes and mixed data types. CPU checks are not CPU-generation support. | |
| The original tokenizer, weights, numerical recipe, seeds, fixed row/chunk semantics and 0.25 blend are preserved. Blackwell-specific native arithmetic is reproduced by the compatibility helpers on the admitted non-Blackwell families. These results do not establish support for every GPU or software version. | |
| H100 and RTX PRO 6000 were additionally verified in a fresh environment pulled from the published model repository. Both passed all 264 required operator/dense checks, and the chunk-state tests reused 8, 32 and 176 complete Mamba chunks across the three boundary requests. | |