For Flash there are some excellent Q2/Q4 hybrids. I know that model was QAT so it handles Q4 better but the meta on quantization seems to be shifting a little bit to be more intelligent about what exactly gets quantized.
replyHow fast are 2x or 4x DGX?
replyI only have one and am wondering what the benefits are of getting another. I feel I will be disappointed…
For DS4 Flash, with 2x Sparks, I am getting 35-85 TPS in single stream, fresh context after quite a bit of RoCe config and the DSpark MTP, on vLLM with Ray and tensor parallel = 2. For multi-stream, it tops out all stream at well north of 100-120. This all degrades with context, but I rarely fill context that much, and if I do it's coding where it's non-real-time.
replyFor something like GLM, it's larger, has a larger number of active experts, and doesn't support tensor parallel. This means performance doesn't really scale with more Sparks. You can layer split, but then you are still seeing each layer in series and so if anything performance gets slightly worse. I would not expect more than 10-20 TPS on GLM with 2-4 Sparks.
If you can afford it, another DGX spark is worth it imo. Especially since, owning just one, you have a $1000 ConnectX7 card that's unused. You can find speeds here: https://spark-arena.com/leaderboard
replyi run flash v4 at 2bit, its pretty great and on my tests against full model It didn't lose any capabilities. It just was thinking more. So you don't have the same efficiency.
reply