Issues

• all pi05 implementation should share the same pi05_bench

  infinigence/ApxInf#166 · Filter by author zealoct opened on Oct 23, 2026

• 关于amd 平台的支持有排期吗?

  infinigence/ApxInf#156 · Filter by author WangFengtu1996 opened on Oct 22, 2026
  1

• [Perf] Qwen3.8 vision tower: CUDA Graph capture for fixed profiles and device-side positional interpolation

  enhancement
  infinigence/ApxInf#151 · Filter by author zealoct opened on Oct 20, 2026

• [Perf] cuda-new segmented_attention only has a scalar L0 fallback (173–348x slower than dense FA2)

  enhancement
  infinigence/ApxInf#150 · Filter by author zealoct opened on Oct 20, 2026

• cuda-new: eager autoregressive decode leaks prepared executions (pointer-keyed cache, no eviction) -> device OOM

  infinigence/ApxInf#123 · Filter by author ChenxiHe231 opened on Oct 13, 2026

• cuda-new: eliminate hand-mirrored FFI ABI declarations (bindgen + single-source version)

  infinigence/ApxInf#96 · Filter by author ChenxiHe231 opened on Sep 26, 2026

• Feature request: FP8 GEMM kernels for sm_89 (Ada hardware supports FP8 tensor cores)

  infinigence/ApxInf#83 · Filter by author lixcwq opened on Sep 18, 2026
  1

• Bug: split-KV FA2 kernel produces NaN on sm_89 (RTX 4090M, 76 SMs)

  infinigence/ApxInf#82 · Filter by author lixcwq opened on Sep 18, 2026
  3

• tuning: shared GemmPlanCache vs thread_local cuBLASLt plan state — threading contract undocumented

  infinigence/ApxInf#37 · Filter by author zealoct opened on Sep 1, 2026

• tuning: merge_record_atomic destructively prunes version-incompatible records from the hardware DB

  infinigence/ApxInf#36 · Filter by author zealoct opened on Sep 1, 2026