All pull requests

Pull requests

• fix(mlx): interleave prefill with active decode

  #2312·STRML opened Sep 15, 2026·

  ·
  1

• feat(profiling): report negotiated and supported link speed per interface

  #2308·ovurrsl opened Sep 11, 2026·

  ·

• feat(models): add Qwen3-4B-4bit model card

  #2307·ovurrsl opened Sep 11, 2026·

  ·

• fix(models): reject model ids that escape the models directory

  #2306·ovurrsl opened Sep 11, 2026·

  ·

• fix(api): reject image URLs pointing at private or loopback addresses (SSRF)

  #2305·ovurrsl opened Sep 11, 2026·

  ·

• docs: running the tests needs the workspace packages installed

  #2304·ovurrsl opened Sep 11, 2026·

  ·

• docs: EXO_LIBP2P_NAMESPACE was renamed to EXO_ZENOH_NAMESPACE

  #2303·ovurrsl opened Sep 11, 2026·

  ·

• Support Qwen3.8-27B Q8 and BF16 processor loading and model discovery

  #2301·marksmeltzer opened Sep 10, 2026·

  ·

• fix(mlx): set memory/wired/cache limits on CUDA backend to prevent OOM during layer loading

  #2300·cdvankammen opened Sep 9, 2026·

  ·

• feat: add Windows mlx-cpu runtime support

  #2298·hvkeyn opened Sep 8, 2026·

  ·

• feat: place pipeline shards by measured bandwidth and latency

  #2297·abendrothj opened Sep 8, 2026·

  ·
  1

  1

• fix(utils): stop retrying SIGKILL forever on an unkillable child process

  #2294·zhast opened Sep 2, 2026·

  ·
  1

• fix(mlx): reject image input when the runner has no vision processor

  #2293·zhast opened Sep 2, 2026·

  ·
  2

• fix: auto-detect CUDA_HOME from bundled nvidia headers on Linux

  #2291·abendrothj opened Sep 2, 2026·

  ·

• fix: disable loguru variable inspection so runner exceptions surface

  #2290·abendrothj opened Sep 2, 2026·

  ·

• fix: repair the test suite on main

  #2289·abendrothj opened Sep 2, 2026·

  ·

• fix(mlx): let sharded models re-derive layer indices after pipeline split

  #2287·zhast opened Sep 1, 2026·

  ·

• fix(discovery): multicast discovery never forms a cluster on macOS

  #2286·zhast opened Sep 1, 2026·

  ·
  3

• fix: file-log rotation never fires again after process start

  #2284·raajkumars opened Sep 1, 2026·

  ·

• feat(mlx): support Qwen3.8 Flash Next (qwen4_exp)

  #2283·tharalds opened Aug 31, 2026·

  ·

• Accept sanitised Kimi tool-call ids and return the canonical form

  #2282·Bdwg222 opened Aug 31, 2026·

  ·
  1

• feat(download): pull oci:// models via llmman serve

  #2280·ericcurtin opened Aug 30, 2026·

  ·

• fix(download): recover stale partial files after HTTP 416

  #2278·RezaGolriz opened Aug 26, 2026·

  ·
  1

• fix(mlx): validate vision prefix cache fail closed

  #2277·RezaGolriz opened Aug 26, 2026·

  ·
  1

• fix(mlx): resolve tokenizer behavior from model metadata

  #2276·RezaGolriz opened Aug 26, 2026·

  ·
  2