Issues

• feat(embeddings): support EmbeddingGemma 2 (embedding_gemma2) multimodal embedder

  arch:densearea:docsarea:inferencearea:modelsmodelsize:smallmodeltype:omnipriority:mediumstatus:readytype:enhancement
  lablup/mlxcel#2314 · Filter by author inureyes opened on Oct 14, 2026

• [nightly-verify] fix/issue-1599-metal-gather-qmm-rhs-rows is red (fix/issue-1599-metal-gather-qmm-rhs-rows)

  priority:highstatus:readytype:bug
  lablup/mlxcel#2302 · Filter by author github-actions opened on Oct 13, 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:donetype:bug
  lablup/mlxcel#1986 · Filter by author inureyes opened on Sep 26, 2026
  3

• 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

• 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

• 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(drafter): validate quantization against SUPPORTED_AFFINE_BITS

  area:corepriority:lowstatus:readytype:bug
  lablup/mlxcel#1790 · Filter by author inureyes opened on Sep 11, 2026

• fix(cli): render video content items without image support

  area:cliarea:inferencepriority:lowstatus:readytype:bug
  lablup/mlxcel#1789 · Filter by author inureyes opened on Sep 11, 2026

• fix(server): count expanded media tokens in input_tokens

  area:inferencepriority:mediumstatus:readytype:bug
  lablup/mlxcel#1788 · Filter by author inureyes opened on Sep 11, 2026

• chore(core): correct the sink attention Used by comment

  area:corepriority:lowstatus:readytype:chore
  lablup/mlxcel#1786 · Filter by author inureyes opened on Sep 11, 2026

• fix(examples): anchor logit_trace chunk 0 with the BOS token

  area:benchmarkpriority:lowstatus:readytype:bug
  lablup/mlxcel#1785 · Filter by author inureyes opened on Sep 11, 2026

• fix(tokenizer): non-special added tokens do not encode as single ids on the SentencePiece path

  area:corepriority:mediumstatus:readytype:bug
  lablup/mlxcel#1758 · Filter by author inureyes opened on Sep 10, 2026

• fix(server): gate media parts on /v1/messages/count_tokens like the other non-generating routes

  area:inferencepriority:mediumstatus:readytype:bug
  lablup/mlxcel#1748 · Filter by author inureyes opened on Sep 10, 2026