2026-11-08 · America/Los_Angeles · 社区动态 · #20

DeepSeek V4.1-Flash running on a single Arc A770 16GB: 45 tok/s prompt, 5 tok/s generation (community SYCL port of Strata)

I got DeepSeek V4.1-Flash running on my Arc A770 through Strata, which splits a big MoE model across VRAM, RAM and the NVMe. Strata's DeepSeek engine only supports CUDA, HIP and Metal, so this is a SYCL port for Arc. It's unofficial and still a work in progress, but it works. EDIT, since a few people asked: - This run uses all 6 of DeepSeek's experts per token (--experts 6). Strata's default on a 16 GB card is 5 ("faster" mode); I turned that off. - The real quality cost is the quantization (ISTA-DASLab 2.25 bpw). Strata's docs measure it against the original model at perplexity +7.3% to…

热度 44.2 / 100；排名与评分保留该期记录。