New
yuhai-china/ds4-spark — Markov draft logit computation on the Metal GPU side
The latest commit in ds4-spark moved the output head computation for Markov draft to the Metal GPU side, eliminating the CPU-side Q8_0 dequantization step.
The latest commit in ds4-spark moved the output head computation for Markov draft to the Metal GPU side, eliminating the CPU-side Q8_0 dequantization step.
View on GitHub ↗The latest commit in ds4-spark moved the output head computation for Markov draft to the Metal GPU side, eliminating the CPU-side Q8_0 dequantization step.
No matching stories.