All pull requests

Pull requests

• feat(cuda): opt-in greedy MTP inside continuous batching

  #214·win10ogod opened Oct 23, 2026Contributor·

  ·

• test(cuda): make the QSA KV offload parity check restore from SSD

  #213·win10ogod opened Oct 23, 2026Contributor·

  ·

• feat(cuda): KV-budget admission and retraction for continuous batching

  #212·win10ogod opened Oct 23, 2026Contributor·

  ·

• perf(cuda): one parameterized QK norm, RoPE and KV write for FullBlock

  #211·win10ogod opened Oct 23, 2026Contributor·

  ·

• perf(cuda): chunk-parallel GDN prefill for long prompts

  #210·win10ogod opened Oct 23, 2026Contributor·

  ·

• Automate standalone benchmark reports and local history

  #208·xxraincandyxx opened Oct 23, 2026Contributor·

  ·

• perf(cuda): independent-request QSA core for batched decode

  #207·win10ogod opened Oct 23, 2026Contributor·

  ·
  2

• perf(cuda): read MLA sparse-attention query and indices in the tile

  #206·win10ogod opened Oct 23, 2026Contributor·

  ·

• perf(cuda): grouped GEMM for dense routed MoE banks without expert gathers

  #205·win10ogod opened Oct 23, 2026Contributor·

  ·

• perf(cuda): row-bucketed batch decode graphs that survive membership changes

  #204·win10ogod opened Oct 23, 2026Contributor·

  ·

• perf(cuda): batch stochastic sampling for continuous-batching decode

  #203·win10ogod opened Oct 23, 2026Contributor·

  ·
  2

• perf(cuda): expert-major plan for MFE grouped prepare and cache dedup

  #202·win10ogod opened Oct 23, 2026Contributor·

  ·

• perf(cuda): tiled QSA prefill index scores with on-chip head reduction

  #201·win10ogod opened Oct 23, 2026Contributor·

  ·

• fix(cuda): build the Torch extension's bf16 MoE loads

  #200·win10ogod opened Oct 23, 2026Contributor·

  ·

• perf(cuda): incremental QSA pooled-cache updates for chunked prefill

  #199·win10ogod opened Oct 23, 2026Contributor·

  ·

• perf(cuda): parallel block scan for the MoE expert route map

  #198·win10ogod opened Oct 23, 2026Contributor·

  ·

• feat(server): request priority and deadlines through the API, queue and worker

  #197·win10ogod opened Oct 23, 2026Contributor·

  ·

• feat(studio): per-model concurrency settings with live queue updates

  #196·win10ogod opened Oct 23, 2026Contributor·

  ·

• feat(server): Prometheus /metrics for runtime and request telemetry

  #195·win10ogod opened Oct 23, 2026Contributor·

  ·

• perf(studio): stop redundant polling and re-rendering during serving

  #194·win10ogod opened Oct 23, 2026Contributor·

  ·

• feat(server): adaptive multi-request concurrency with queue accounting

  #193·win10ogod opened Oct 23, 2026Contributor·

  ·

• feat: add experimental unified row Trellis codec

  #189·TylogiAdmin opened Oct 22, 2026Contributor·

  ·

• refactor(cuda): reorganize native execution, state ownership, and build boundaries

  #163·DawningKr opened Oct 20, 2026Contributor·

  ·
  1

• fix: resolve Studio and runtime issues and add i18n

  #48·Liz-Lian opened Oct 6, 2026Contributor·

  ·

• perf(cuda): specialize quantized decode hotpaths

  #42·Camelliacloudui opened Oct 3, 2026Contributor·

  ·