| GQA | ||
| MLA | ||
| DSA | ||
| GDN / KDA | ||
Ring
Ulysses
| W [N,K] | K | N | — | ||
| X [M,K] | K | — | M | ||
| dY [M,N] | — | N | M |
| TP | ||
| DP | ||
| EP | ||
| CP / DCP | ||
| PP | ||
| FSDP | ||
| TP4 × 16 | TP8 × 8 | DEP32 × 2 | DEP64 | |
|---|---|---|---|---|
| 8 | 763 / 10.5 | 927 / 8.6 | 998 / 8.0 | 1,100 / 7.3 |
| 32 | 1,857 / 17.2 | 2,645 / 12.1 | 3,428 / 9.3 | 3,723 / 8.6 |
| 128 | 5,153 / 24.8 | 5,713 / 22.4 | 8,567 / 14.9 | 9,007 / 14.2 |
| MTP | ||||
| EAGLE-3 | ||||
| DFlash / DFlash2 | ||||
| DSpark |
| — | Miles tools/convert_hf_to_{mxfp8,nvfp4}.py | |||
Miles --fp8-recipe mxfp8 / --fp4-recipe nvfp4 + NVTE_* | ||||
| — | ||||
| WxAy | ||
| MX / NV | ||
| RHT | ||
| SR | ||
| PTQ / QAT | ||