Repository navigation

perf(glm): reuse inference metadata and streamline FP8 prefill - #1685

Merged

shihaobai merged 11 commits into

mainfrom

codex/glm-parallel-nsa-indices

Nov 10, 2026

Merged

perf(glm): reuse inference metadata and streamline FP8 prefill#1685
shihaobai merged 11 commits into
mainfrom
codex/glm-parallel-nsa-indices

Commits on Nov 3, 2026
- perf(nsa): parallelize long ragged index copies
  shihaobai
  committed
  37a9d4c
Commits on Nov 10, 2026
- perf(glm): eliminate redundant ragged index preparation
  shihaobai
  committed
  fb559a8
- refactor(glm): remove legacy ragged prefill compatibility
  shihaobai
  committed
  4106af4
- refactor(glm): align NSA names with inference state
  shihaobai
  committed
  c8466d6
- perf(glm): simplify FP8 prefill workspaces and dispatch
  shihaobai
  committed
  d0c1db8
- refactor(glm): unify FP8 prefill on FlashAttention
  shihaobai
  committed
  914e949
- Merge origin/main and adapt GLM request-table regressions
  shihaobai
  committed
  c41b123
- perf(glm): streamline FP8 prefill KV gathering
  shihaobai
  committed
  e40c932
- refactor(glm): move FP8 prefill into attention backend
  shihaobai
  committed
  4e62741
- refactor(glm): allocate prefill workspaces through the backend
  shihaobai
  committed
  8cd6129
- perf(glm): reserve reusable prefill workspaces before KV sizing
  shihaobai
  committed
  645114d