llama.cpp
ggml : update CUDA/Metal radix Top-K to make tied selections deterministic
#31084
Merged

Commits
  • ggml : fix for the radix TOP_K thread timing bug
    sanmai committed 7 days ago
  • make the key an ordered 64-bit pair
    sanmai committed 7 days ago
  • a rewrite without shift-skip logic
    sanmai committed 7 days ago
  • fix casts
    sanmai committed 7 days ago
  • warp-contiguous layout
    sanmai committed 7 days ago
  • reduce the scope removing vulkan+metal changes
    sanmai committed 7 days ago
  • turn it off for Metal/ROCm, note about cub::DeviceTopK
    sanmai committed 7 days ago
  • fix the hip build
    sanmai committed 7 days ago
  • skip Vulkan as well
    sanmai committed 7 days ago
  • popc lambda from fattn
    sanmai committed 6 days ago
  • do block_cols math in int64_t
    sanmai committed 6 days ago
  • make it allowlist bc SYCL looks non deterministic too
    sanmai committed 6 days ago
  • ggml : make Metal radix Top-K tied selections deterministic
    ggerganov committed 5 days ago
  • tests : enable deterministic top-k check on Metal, add radix-path tie cases
    ggerganov committed 5 days ago
  • ggml : use coalesced accesses in the Metal radix Top-K compaction
    ggerganov committed 5 days ago
  • tests : remove the extra top-k tie and perf cases
    ggerganov committed 5 days ago
  • metal: scan the radix bins and warp totals with a simdgroup
    ServeurpersoCom committed 5 days ago
  • tests: enable the deterministic top-k check on ROCm
    ServeurpersoCom committed 5 days ago
  • cuda: remove the redundant radix top-k counter reset
    ServeurpersoCom committed 5 days ago
Loading