Issues

• fix(nemotron_h): prefill memory grows roughly quadratically with prompt length and aborts with Metal out-of-memory at ~6K tokens

  lablup/mlxcel#2281 · Filter by author YEONFEEL96 opened on Oct 12, 2026

• perf(rocm): flash SDPA dominates long-context chunked prefill on gfx1151

  area:inferenceplatform:linuxpriority:mediumstatus:readytype:performance
  lablup/mlxcel#2251 · Filter by author inureyes opened on Oct 8, 2026

• fix(rocm): other HIP launchers size an uncapped int grid from the element count

  area:corepriority:lowstatus:readytype:bug
  lablup/mlxcel#2234 · Filter by author inureyes opened on Oct 8, 2026

• perf(rocm): turn the f16 dequant LRU off by default

  area:coreplatform:linuxpriority:lowstatus:in-progresstype:performance
  lablup/mlxcel#2232 · Filter by author inureyes opened on Oct 8, 2026

• fix: small decode-path defects found while mapping CLI vs server decode

  area:cliarea:inferencepriority:lowstatus:readytype:bug
  lablup/mlxcel#2177 · Filter by author inureyes opened on Oct 6, 2026

• track(server): reproducible cold/warm prompt-cache output difference on Qwen3-1.7B / GB10

  status:review
  lablup/mlxcel#2175 · Filter by author Jaeyeong-CHOI opened on Oct 6, 2026
  1

• test(cuda): Llama-3.2-1B fresh-process divergence on GB10 and deterministic-SDPA controls

  status:review
  lablup/mlxcel#2174 · Filter by author Jaeyeong-CHOI opened on Oct 6, 2026
  1

• test: Apple Silicon verification pass for 2026-10-06 CUDA-only changes

  area:inferenceplatform:macospriority:highstatus:readytype:test
  lablup/mlxcel#2158 · Filter by author inureyes opened on Oct 6, 2026
  5

• fix(cuda): iquestloopcoder chunked prefill disagrees with single-pass on GB10

  area:modelsplatform:linuxpriority:mediumstatus:readytype:bug
  lablup/mlxcel#2143 · Filter by author inureyes opened on Oct 6, 2026
  1

• feat: Detect refusal-style VLM image generations in bench_mlxlm.py

  area:benchmarkmodeltype:vlmpriority:mediumstatus:readytype:enhancement
  lablup/mlxcel#2134 · Filter by author inureyes opened on Oct 6, 2026

• fix: investigate Gemma 4 12B long-context MTP parity on Metal

  area:inferenceplatform:macospriority:highstatus:readytype:bug
  lablup/mlxcel#1986 · Filter by author inureyes opened on Sep 26, 2026
  2

• chore(deps): bump the MLX pin and declare a base commit for each overlay

  area:coreplatform:linuxplatform:macospriority:mediumstatus:readytype:chore
  lablup/mlxcel#1932 · Filter by author inureyes opened on Sep 19, 2026

• ci(webui): the Activity performance gate fails on runner noise

  priority:mediumstatus:readytype:bug
  lablup/mlxcel#1925 · Filter by author inureyes opened on Sep 19, 2026
  4

• epic: make the bundled WebUI usable, on the shared ui-common foundation

  area:architecturepriority:highstatus:readytype:enhancement
  lablup/mlxcel#1910 · Filter by author inureyes opened on Sep 18, 2026

• docs(webui): ship release-time user docs and surface --webui in --help

  area:cliarea:docspriority:highstatus:readytype:docs
  lablup/mlxcel#1900 · Filter by author inureyes opened on Sep 16, 2026

• test(cuda): run the deferred CUDA arm for the merged ROCm routing work on a GB10 node

  area:coreplatform:linuxpriority:highstatus:blockedtype:test
  lablup/mlxcel#1873 · Filter by author inureyes opened on Sep 13, 2026
  3

• perf(moe): attribute the MoE decode gap at the M = 1 expert gather

  area:corearea:inferencepriority:mediumstatus:readytype:performance
  lablup/mlxcel#1859 · Filter by author inureyes opened on Sep 12, 2026
  1

• fix(bench): bench_mlxlm.py hardcodes a 128 GB memory budget

  area:benchmarkpriority:mediumstatus:blockedtype:bug
  lablup/mlxcel#1858 · Filter by author inureyes opened on Sep 12, 2026

• refactor(mlx-cpp): organize mlxcel's MLX-side layer as mlxcelverse (backend overlays and mlxcel kernels)

  area:architecturepriority:lowstatus:backlogtype:refactor
  lablup/mlxcel#1816 · Filter by author inureyes opened on Sep 11, 2026
  

• perf(rocm): allocator footprint and ROCm ports of mlxcel fused kernels

  area:coreplatform:linuxpriority:lowstatus:backlogtype:performance
  lablup/mlxcel#1814 · Filter by author inureyes opened on Sep 11, 2026
  2

• ci(rocm): self-hosted gfx1151 runner that builds, links and smoke-tests ROCm changes

  area:coreplatform:linuxpriority:mediumstatus:backlogtype:chore
  lablup/mlxcel#1811 · Filter by author inureyes opened on Sep 11, 2026

• epic: AMD GPU (ROCm) backend on Linux via mlxcelverse

  area:coreplatform:linuxpriority:mediumstatus:in-progresstype:enhancement
  lablup/mlxcel#1801 · Filter by author inureyes opened on Sep 11, 2026
  1

• fix(cache): bound the turbo4 sliding append to visible_len

  area:corearea:inferencepriority:mediumstatus:readytype:bug
  lablup/mlxcel#1793 · Filter by author inureyes opened on Sep 11, 2026

• fix(mlx-cpp): stop casting comparison inputs to bool in compiled kernels

  area:corepriority:mediumstatus:readytype:bug
  lablup/mlxcel#1792 · Filter by author inureyes opened on Sep 11, 2026

• fix(core): make array_evaluated_bytes fallible so GPU faults don't abort

  area:corearea:inferencepriority:mediumstatus:in-progresstype:bug
  lablup/mlxcel#1791 · Filter by author inureyes opened on Sep 11, 2026