mw00/project-maya v1.1.12

Project Maya v1.1.12

on GitHub

EXL3 on one GPU with the CPU's help, batched answers across a layer split (opt-in), the VRAM tier following the conversation while DFlash2 drafts, and AMD RDNA 2 cards as experimental.

Update: the dashboard's About > Update, or git pull, then ./setup.sh (Windows: START-MAYA.bat). The engine recompiles; nothing is downloaded again.

What's new

• EXL3 on one GPU: the CPU takes more of the experts, faster (#188 by @needmorevram). On one GPU most of an EXL3 model's experts run on the CPU. The CPU now reads each expert in memory order (split-K pieces) and runs a route's experts as one batch of jobs, and CPUs without AVX-512 (Intel desktop chips, Zen 2 and 3) get an AVX2 tier of their own. Prompts give their least-used experts to the CPU too, as GGUF prompts already did. On an RTX 3090 alone: EXL3-3.05bpw decode (writing the answer) 14.3-16.8 -> 20.9-22.1 tokens/s, 22.4 -> 29.2 with DFlash2; EXL3-4.05bpw 14.4-15.1 -> 18.9-19.6; an 8K prompt's prefill (reading the prompt) 384-396 -> 422-489 tokens/s. The answers are the same tokens as before.
• Batched answers across a layer split, opt-in (#204 by @zetlaw). With STRATA_GLM_SPLIT_ROWS=k (1-8) and several requests at once, each GPU of a split takes up to k requests in one step, reading the shared weights once for all of them. On four RTX 4000 Ada GPUs with Maya-M and eight requests, k=2 gave 94.5 tokens/s in all against 74.4 (+27%); larger groups were slower there, so it stays off by default. See several requests at once.
• The VRAM tier keeps following the conversation while DFlash2 drafts (#198 by @merbanan). With the drafter's verify window on two CPU lanes, the background swaps that move the experts you use into VRAM had stopped. They run again: on a V100 with a Ryzen 9 7950X, the share of experts found in VRAM went from 49-62% to 54-67%.
• AMD RDNA 2, experimental (#205). The setup accepts the RX 6800 / 6900 series and the Radeon PRO V620 / W6800 (gfx1030), and compiles for them only when one is chosen. Maya has not been validated on such a card yet: please tell us how it goes in an issue.

NVIDIA GPUs give v1.1.11's tokens.

Checked

• Two Tesla V100s: the same tokens as v1.1.11; the engine's tests; the EXL3 CPU kernels exact on an AVX2 CPU; several requests at once on the split the same tokens and speed as v1.1.11, and with STRATA_GLM_SPLIT_ROWS the same tokens; two requests on one GPU the same as v1.1.11; the server end to end.
• One Tesla V100 with a Core i5 (AVX2): EXL3-3.05bpw the same tokens as v1.1.11, with the CPU now taking part of the experts.
• Radeon 8065S (Gorgon Halo, Windows 11): the same tokens with MTP drafting on and off; two requests on one GPU the same tokens as one alone; the server end to end.
• The engine builds on Linux, Windows and Windows + AMD; the GitHub checks.