Commit History
19 days ago
-
Warm FP8/linear-attention kernels once at startup Running
Verified signatured34150c -
FP8 vs bf16 dose-response check: FP8 reproduces the effect at every dose
Verified signature159f917 -
Persist Triton autotune; size GPU durations from measured speed; one padded batch per validation dose
Verified signaturee75de23
20 days ago
-
Rat-Qwen on ZeroGPU: Qwen3.5-27B-FP8, L51 f26585
Verified signature0838b38 -
initial commit
Verified signature7441271