Issues

• perf(xla): remove the per-step LM-head transpose and f32-to-f16 conversion in the single-sequence decode graph

  area:inferencepriority:mediumstatus:readytype:performance
  lablup/mlxcel#2863 · Filter by author inureyes opened on Nov 5, 2026

• perf(metal): fuse the tanh-GELU f32 activation chain into one Metal kernel (counterpart to #2780)

  priority:lowstatus:readytype:performance
  lablup/mlxcel#2857 · Filter by author hephaex opened on Nov 5, 2026

• test(models): the nemotron NVFP4 sorted-expert parity test is order-dependent inside the lib test binary

  area:corearea:modelspriority:mediumstatus:readytype:test
  lablup/mlxcel#2852 · Filter by author inureyes opened on Nov 4, 2026
  2

• Benchmark cuda-oxide RMSNorm and record the adoption decision

  area:benchmarkplatform:linuxpriority:mediumstatus:blockedtype:test
  lablup/mlxcel#2834 · Filter by author inureyes opened on Nov 2, 2026

• Integrate an optional cuda-oxide RMSNorm path with MLX execution

  area:coreplatform:linuxpriority:mediumstatus:blockedtype:enhancement
  lablup/mlxcel#2833 · Filter by author inureyes opened on Nov 2, 2026

• Add a standalone cuda-oxide fused residual RMSNorm kernel

  area:coreplatform:linuxpriority:mediumstatus:readytype:enhancement
  lablup/mlxcel#2832 · Filter by author inureyes opened on Nov 2, 2026

• test(core): fused_probe_matches_the_pinned_dispatch_rules failed once in nightly and did not reproduce

  priority:mediumstatus:readytype:test
  lablup/mlxcel#2824 · Filter by author inureyes opened on Nov 1, 2026

• chore(deps): follow up the dependencies held back in the v0.8.0 update

  priority:lowstatus:readytype:dependency
  lablup/mlxcel#2823 · Filter by author inureyes opened on Nov 1, 2026

• ci: WebUI installed artifact fails when the GB10 host does not become quiet

  priority:mediumstatus:readytype:bug
  lablup/mlxcel#2822 · Filter by author inureyes opened on Nov 1, 2026

• perf: residual batched decode gaps left by the Metal acceptance of #2566

  platform:macospriority:mediumstatus:readytype:performance
  lablup/mlxcel#2817 · Filter by author inureyes opened on Nov 1, 2026

• [nightly-verify] main is red

  priority:highstatus:readytype:bug
  lablup/mlxcel#2806 · Filter by author github-actions opened on Oct 31, 2026
  4

• perf(cuda): MoE batched decode reads top_k * B expert matrices per gather

  area:corepriority:mediumstatus:in-progresstype:performance
  lablup/mlxcel#2784 · Filter by author inureyes opened on Oct 31, 2026
  1

• perf(cuda): single-kernel tanh GELU so tanh-switched families match erf prefill cost

  area:coreplatform:linuxpriority:lowstatus:readytype:performance
  lablup/mlxcel#2780 · Filter by author inureyes opened on Oct 31, 2026

• Explain the temperature 0 flip between plain and block-built KV history

  area:corepriority:mediumstatus:readytype:bug
  lablup/mlxcel#2778 · Filter by author inureyes opened on Oct 30, 2026
  1

• test(rocm): validate NVFP4 kernels on a wave64 (CDNA) GPU

  platform:linuxpriority:lowstatus:blockedtype:test
  lablup/mlxcel#2699 · Filter by author inureyes opened on Oct 28, 2026
  1

• perf(cuda): re-measure the fused-MoE Dff bound and GB10 fused-MoE defaults without the JIT compile in the window

  arch:moearea:benchmarkplatform:linuxpriority:lowstatus:readytype:performance
  lablup/mlxcel#2540 · Filter by author inureyes opened on Oct 22, 2026

• [nightly-verify] the run did not finish

  priority:highstatus:readytype:bug
  lablup/mlxcel#2529 · Filter by author github-actions opened on Oct 21, 2026
  8

• feat: Extend conversation compaction to Chat Completions, compaction_trigger and multi-pass summaries

  area:inferencepriority:lowstatus:readytype:enhancement
  lablup/mlxcel#2523 · Filter by author adrysn opened on Oct 21, 2026

• perf: Add the packed Dk=128 gated-delta Metal kernel for prefill

  arch:hybridarea:coreplatform:macospriority:lowstatus:readytype:performance
  lablup/mlxcel#2407 · Filter by author inureyes opened on Oct 16, 2026

• feat: Add tensor-parallel support for Qwen3 MoE

  arch:moearea:modelsmodeltype:textpriority:lowstatus:readytype:enhancement
  lablup/mlxcel#2405 · Filter by author inureyes opened on Oct 16, 2026

• feat: Add pipeline stage executors for Qwen3 MoE, Qwen2, DeepSeek-V3.2

  area:modelsmodeltype:textpriority:lowstatus:readytype:enhancement
  lablup/mlxcel#2404 · Filter by author inureyes opened on Oct 16, 2026

• feat(vlm): add GLM-5 Next (glm5_next) support

  arch:hybridarea:modelsmodelsize:xlargemodeltype:vlmpriority:mediumstatus:blockedtype:enhancement
  lablup/mlxcel#2403 · Filter by author inureyes opened on Oct 16, 2026

• feat: Render per-group video timestamps for the Qwen3-VL family

  area:modelsmodeltype:vlmpriority:mediumstatus:readytype:enhancement
  lablup/mlxcel#2402 · Filter by author inureyes opened on Oct 16, 2026

• feat(vlm): add Mistral Large 3 (mistral_large3) support

  arch:moearea:modelsmodelsize:xlargemodeltype:vlmpriority:mediumstatus:blockedtype:enhancement
  lablup/mlxcel#2401 · Filter by author inureyes opened on Oct 16, 2026

• feat: Load bundled Qwen 3.5 MTP drafters and detect native MTP heads

  area:inferencemodeltype:textpriority:mediumstatus:readytype:enhancement
  lablup/mlxcel#2400 · Filter by author inureyes opened on Oct 16, 2026