tokenspeed-mla python library

tokenspeed-mla

Subscription: Developer
Last updated a month ago
Active
Subscription: Developer

Speed-of-light TokenSpeed MLA kernels for Blackwell SM100 and SM103. Includes CuTe DSL JIT backend for ragged varlen FMHA, optional AOT binary backend for FP8 E4M3 prefill, BF16 output, LSE support, causal/non-causal modes, PDL support.

pip install tokenspeed_mla==0.1.2
  • 0.1.2
    Linux
    Last updated a month ago