Clef could not run under SGLang on V100 (needs CUDA 13, which dropped Volta kernels). It runs under plain transformers in fp16 on two cards, serving the model own systemone() endpoint.
Clef could not run under SGLang on V100 (needs CUDA 13, which dropped Volta kernels). It runs under plain transformers in fp16 on two cards, serving the model own systemone() endpoint.
Restores the route removed when the model could not run. It runs now, on a different
runtime: SGLang needs v0.5.22+ which upstream publishes only on CUDA 13, and CUDA 13
dropped Volta kernels. clef-serve loads the same weights through plain transformers in
fp16, sharded over two V100s, and serves the model's own systemone() on /v1/systemone.
Weights measured at 17.75 GiB in BF16; fp16 costs the same bytes per parameter.
rock
merged commit 1f906e95d8 into main2026-10-30 06:06:49 +00:00
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Clef could not run under SGLang on V100 (needs CUDA 13, which dropped Volta kernels). It runs under plain transformers in fp16 on two cards, serving the model own systemone() endpoint.