Pull requests / #387

#387 AMD: ship the gfx1200 hipBLASLt table, with the with/without measurements

closed · @Efeisot · 0 comentarios · En GitHub

AMD / HIPModels & quantsDocumentation

Descripción

The table #256 left out (b0dda4b): tools/hip/gfx1200-hipblaslt-100202.txt, 67 rows calibrated on the RX 9060 XT with tools/hip/tune_hipblaslt (system ROCm 7.2, hipBLASLt 1.2.2). Same binary, prompt speed with the table vs plain hipBLAS, two runs each, Coder IQ1_M greedy MTP, STRATA_PREFILL_MMQ=1:

| prompt | table | plain hipBLAS | gain |
| 2,374  | 597-599 | 396-398 | 1.50x |
| 65,045 | 754-757 | 380     | 1.99x |
| 130,091| 719     | 363-370 | 1.98x |

docs/AMD_HIP.md: the community-validated section carries the numbers; ctest 37/37 (without ple_parity and platform_memory_test).

Continues #256.

🤖 Generated with Codebuff

En el sitio

Enlaces a install, modelos, releases.