lablup/mlxcel

View on GitHub

High-performance LLM, VLM, embedding, reranking, and audio inference for Apple Silicon and NVIDIA CUDA systems.

Stars: 467
Forks: 54
Open beginner issues: 0
Indexed issues: 152
Avg merge: 4h 25m
Merged PRs (30d): 310

Dominant language: Rust
License: Apache-2.0
Last GitHub push: Sep 15, 2026
Latest indexed: Sep 15, 2026
Contributing guide: Contributing guide
Code of conduct: Code of conduct
Beginner labels: help wanted

Open All

149 open issues indexed

docs: README claims Homebrew installs on Linux, but the formula is macOS-only Open

priority:medium status:ready type:docs

lablup/mlxcel#1704 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs(install): the Cargo feature table omits three declared features Open

priority:low status:ready type:docs

lablup/mlxcel#1705 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs: the code-guidelines HashMap-order section cites line numbers and counts that no longer exist Open

priority:low status:ready type:docs

lablup/mlxcel#1706 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs(reports): add an index README for the 308-file TECHNICAL_REPORTS directory Open

priority:low status:ready type:docs

lablup/mlxcel#1708 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
test(turbo): delegated steel envelope fails intermittently with an identical RMS Open

area:core priority:low status:ready type:test

lablup/mlxcel#1712 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
perf(core): host graph build at decode is serial with device execution and scales with op count Open

arch:hybrid arch:moe area:benchmark area:core area:inference area:models platform:macos priority:high status:ready type:performance

lablup/mlxcel#1713 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
perf(qwen_vl): multimodal RoPE leaks f32 into the decode residual Open

area:models modeltype:vlm priority:medium status:review type:performance

lablup/mlxcel#1714 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
feat(models): add Arcee AFM-4.5B (arcee, dense ArceeForCausalLM) text model support Open

arch:dense area:models modelsize:small modeltype:text priority:medium status:blocked type:enhancement

lablup/mlxcel#1719 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
feat(distributed): Kimi K3 pipeline stage executor and full-model multi-node validation Open

arch:hybrid area:inference area:models modelsize:xlarge modeltype:text priority:low status:blocked type:enhancement

lablup/mlxcel#1734 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(server): gate media parts on /v1/messages/count_tokens like the other non-generating routes Open

area:inference priority:medium status:ready type:bug

lablup/mlxcel#1748 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(kimi_linear): chunked prefill discards the recurrent state Open

area:inference area:models priority:medium status:ready type:bug

lablup/mlxcel#1749 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(prompt-cache): completion-origin snapshots claim one more token than their caches hold Open

area:core area:inference priority:high status:ready type:bug

lablup/mlxcel#1754 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
feat(scheduler): sequence-aware cache trim hook so model-owned families can use padded prefill Open

area:core area:inference priority:medium status:ready type:enhancement

lablup/mlxcel#1755 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(tokenizer): non-special added tokens do not encode as single ids on the SentencePiece path Open

area:core priority:medium status:ready type:bug

lablup/mlxcel#1758 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(prompt-cache): a whole-prompt cache hit re-runs the last token on top of a cache that already holds it Open

area:inference priority:medium status:ready type:bug

lablup/mlxcel#1760 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(prompt-cache): snapshot bucket sizing is not model-aware for the attention-cache families, so large-context reuse silently never happens Open

area:inference priority:high status:ready type:bug

lablup/mlxcel#1761 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(split-mtp): write the drafter atomically, refuse symlinked outputs Open

area:cli area:surgery priority:low status:ready type:security

lablup/mlxcel#1784 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(examples): anchor logit_trace chunk 0 with the BOS token Open

area:benchmark priority:low status:ready type:bug

lablup/mlxcel#1785 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
chore(core): correct the sink attention Used by comment Open

area:core priority:low status:ready type:chore

lablup/mlxcel#1786 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(server): harden native-video temp files and resolve allowlist once Open

area:inference priority:medium status:ready type:security

lablup/mlxcel#1787 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(server): count expanded media tokens in input_tokens Open

area:inference priority:medium status:ready type:bug

lablup/mlxcel#1788 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(cli): render video content items without image support Open

area:cli area:inference priority:low status:ready type:bug

lablup/mlxcel#1789 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(drafter): validate quantization against SUPPORTED_AFFINE_BITS Open

area:core priority:low status:ready type:bug

lablup/mlxcel#1790 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(core): make array_evaluated_bytes fallible so GPU faults don't abort Open

area:core area:inference priority:medium status:ready type:bug

lablup/mlxcel#1791 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(mlx-cpp): stop casting comparison inputs to bool in compiled kernels Open

area:core priority:medium status:ready type:bug

lablup/mlxcel#1792 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(cache): bound the turbo4 sliding append to visible_len Open

area:core area:inference priority:medium status:ready type:bug

lablup/mlxcel#1793 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(models): stop model loaders from printing progress to stdout Open

area:benchmark area:models priority:low status:ready type:bug

lablup/mlxcel#1794 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
perf(speculative): the default draft block width of 16 loses on GB10 Open

area:core area:inference platform:linux priority:medium status:ready type:performance

lablup/mlxcel#1797 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
epic: AMD GPU (ROCm) backend on Linux via mlxcelverse Open

area:core platform:linux priority:medium status:in-progress type:enhancement

lablup/mlxcel#1801 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(rocm): surface ROCm GPU failures through Event::error instead of NaN or hangs Open

area:core platform:linux priority:medium status:backlog type:bug

lablup/mlxcel#1804 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
feat(quant): backend quantization capability table and load-time convert-or-reject policy Open

area:models platform:linux priority:medium status:backlog type:enhancement

lablup/mlxcel#1806 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
feat(rocm/quant): mxfp8 end-to-end on ROCm, including FP8 block checkpoints and the MoE gather path Open

area:models platform:linux priority:medium status:backlog type:enhancement

lablup/mlxcel#1807 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(rocm/quant): mxfp4 qmm hang and GPU quantize launch failure, with an affine 4-bit fallback Open

arch:moe area:models platform:linux priority:medium status:backlog type:bug

lablup/mlxcel#1808 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
test(rocm): correctness matrix against a Metal baseline and a verify-test-rocm gate Open

area:inference platform:linux priority:medium status:in-progress type:test

lablup/mlxcel#1809 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
perf(bench): ROCm support in the benchmark harness and a published gfx1151 baseline Open

area:benchmark platform:linux priority:medium status:backlog type:performance

lablup/mlxcel#1810 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
ci(rocm): self-hosted gfx1151 runner that builds, links and smoke-tests ROCm changes Open

area:core platform:linux priority:medium status:backlog type:chore

lablup/mlxcel#1811 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
chore(rocm): mlxcelverse ROCm fork sync script, MLX pin-bump procedure, and upstreaming local fixes Open

area:core platform:linux priority:medium status:backlog type:chore

lablup/mlxcel#1813 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
perf(rocm): allocator footprint and ROCm ports of mlxcel fused kernels Open

area:core platform:linux priority:low status:backlog type:performance

lablup/mlxcel#1814 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
refactor(mlx-cpp): organize mlxcel's MLX-side layer as mlxcelverse (backend overlays and mlxcel kernels) Open

area:architecture priority:low status:backlog type:refactor

lablup/mlxcel#1816 · 0 comments · 0 reactions · 1 assignee · Indexed Sep 15, 2026
perf(cuda): bucket the cuDNN SDPA plan-cache key across verify rounds Open

area:core area:inference platform:linux priority:medium status:in-progress type:performance

lablup/mlxcel#1820 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(bench): bench_mlxlm.py hardcodes a 128 GB memory budget Open

area:benchmark priority:medium status:blocked type:bug

lablup/mlxcel#1858 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
perf(moe): attribute the MoE decode gap at the M = 1 expert gather Open

area:core area:inference priority:medium status:ready type:performance

lablup/mlxcel#1859 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
test(cuda): run the deferred CUDA arm for the merged ROCm routing work on a GB10 node Open

area:core platform:linux priority:high status:ready type:test

lablup/mlxcel#1873 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(rocm): get_launch_args clamps the grid without a grid-stride contract Open

area:core platform:linux priority:medium status:ready type:bug

lablup/mlxcel#1874 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
chore(ci): kernel dtype-key checker cannot see launches that move into headers Open

area:core priority:low status:ready type:enhancement

lablup/mlxcel#1875 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(rocm): hipFFT blocks when too many plans are alive, root cause unknown Open

area:core platform:linux priority:medium status:ready type:bug

lablup/mlxcel#1876 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(rocm): the Gumbel-max and rejection sampler entry points have no ROCm arm and abort on metal_kernel Open

status:backlog type:bug

lablup/mlxcel#1885 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(rocm): the WebUI catalog reports every architecture unsupported on ROCm Open

status:backlog type:bug

lablup/mlxcel#1886 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(rocm): CUDA and ROCm hosts share one MTP policy cache key Open

status:backlog type:bug

lablup/mlxcel#1887 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026

lablup/mlxcel

Get new issues in your inbox