llama.cpp
ggml-metal: add chunked SSD MMA for Mamba-2 prefill optimization
#26647
Merged
Overview
Commits
15
Changes
View On GitHub
Commits
  • metal: WIP chunked SSD SSM_SCAN kernels for multi-token prefill
    dpantaleoni committed 6 days ago
  • metal: drop scalar SSD path; MMA + sequential tail
    dpantaleoni committed 6 days ago
  • drop WIP ssm scan test noise
    dpantaleoni committed 6 days ago
  • remove state_from_dst and rename CS and NSG constants
    dpantaleoni committed 6 days ago
  • remove unrelated added whitespace padding
    dpantaleoni committed 6 days ago
  • added clarity to mma_tokens calculation
    dpantaleoni committed 6 days ago
  • added clarity to use_mma bool checks
    dpantaleoni committed 6 days ago
  • added comments to metal ssd op constants for clarity
    dpantaleoni committed 6 days ago
  • reserve K tokens for sequential kernel rollback snapshots
    dpantaleoni committed 6 days ago
  • reset concurrency between mma and seq tail
    dpantaleoni committed 6 days ago
  • remove print args no longer used
    dpantaleoni committed 6 days ago
  • fixed comment to no longer point to specific line
    dpantaleoni committed 6 days ago
  • add FC_SSM_SCAN so seq path skips token offlset unless it's mma tail
    dpantaleoni committed 6 days ago
  • added changes to new ssm.metal for rebase after ggml-metal.metal refactor
    dpantaleoni committed 6 days ago
  • specialize ssm_scan tail with a template instead of a function constant
    forforever73 committed 6 days ago
FAQ Terms Privacy Refunds Impressum
Loading