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

test(models): validate Inkling omni on real checkpoints Open

arch:moe area:inference area:models modelsize:xlarge modeltype:omni platform:macos priority:medium status:backlog type:test

lablup/mlxcel#1549 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(cuda/quant): quantized prefill is not bitwise reproducible on sm_70, and its determinism test never runs Open

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

lablup/mlxcel#1558 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(inference): fused_sample_probs differs by 1 ULP at temperature 1.0 on sm_70 Open

area:inference platform:linux priority:medium status:investigation type:bug

lablup/mlxcel#1563 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
perf: first-token latency is dominated by lazy weight materialization Open

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

lablup/mlxcel#1564 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(lora): fail the load when staged runtime adapter terms are never claimed Open

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

lablup/mlxcel#1577 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[nightly-verify] main is red Open

priority:high status:ready type:bug

lablup/mlxcel#1599 · 7 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(youtu_vl): processor ignores preprocessor max_num_patches Open

area:models modeltype:vlm priority:medium status:ready type:bug

lablup/mlxcel#1611 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
perf(bench): Confirm or clear three small-model prefill drops at 0.6.0 Open

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

lablup/mlxcel#1614 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
chore: bench_decode.sh all measures duplicate checkpoints twice Open

area:benchmark priority:low status:review type:chore

lablup/mlxcel#1615 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(youtu_vl): multi-window images still described wrong after the patch-order fix Open

area:models modeltype:vlm priority:high status:ready type:bug

lablup/mlxcel#1618 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(test): Inkling-VL mixed-prefill scatter test fails intermittently under the full suite and passes in isolation Open

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

lablup/mlxcel#1622 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
chore(release): refresh Linux aarch64 CUDA artifact for specialized serving Open

platform:linux priority:high status:backlog type:chore

lablup/mlxcel#1653 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
perf(cuda/quant): Turing (sm_75) needs its own tensor-core arm Open

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

lablup/mlxcel#1654 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs(cli): unify subcommand summary punctuation and add missing Examples blocks Open

priority:low status:ready type:docs

lablup/mlxcel#1657 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs: source comments reference docs/model_tests.md and docs/testing.md, which do not exist Open

priority:low status:ready type:docs

lablup/mlxcel#1658 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs(bench): CONTINUOUS_BATCHING.md cites a benchmark record that was never committed Open

priority:low status:ready type:docs

lablup/mlxcel#1659 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
chore(scripts): harden deploy_webpage.sh with set -euo pipefail Open

priority:low status:ready type:chore

lablup/mlxcel#1660 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
chore(scripts): honor MLXCEL_BIN and MODEL overrides in measure_sparse_v_skip_rate.sh Open

priority:low status:ready type:chore

lablup/mlxcel#1661 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
test(server): add unit coverage for the store_budget LRU key ordering and JSON length helpers Open

priority:low status:ready type:test

lablup/mlxcel#1662 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
chore(scripts): mark the shebang-carrying Python scripts executable Open

priority:low status:ready type:chore

lablup/mlxcel#1663 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(cli): reject whitespace-only prompts in embed to match rerank Open

priority:low status:ready type:bug

lablup/mlxcel#1664 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
refactor(server): hoist the duplicated INITIAL_HASH_CAPACITY_LIMIT into store_budget Open

priority:low status:ready type:refactor

lablup/mlxcel#1665 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(bench): run_bench.sh hardcodes the run date and uses GNU-only grep -oP, failing on macOS Open

priority:medium status:ready type:bug

lablup/mlxcel#1666 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(bench): speculative_bench panics instead of erroring when --prompt fails to tokenize Open

priority:low status:ready type:bug

lablup/mlxcel#1667 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
test(audio): pin the AudioConfig derived attention geometry Open

priority:low status:ready type:test

lablup/mlxcel#1668 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
chore(bench): fix the --help defects in the bench_speculative/bench_block_width script pair Open

priority:low status:ready type:chore

lablup/mlxcel#1669 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
chore(ci): extend Dependabot to the python/ and webpage/site dependency roots Open

priority:low status:ready type:chore

lablup/mlxcel#1670 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs(server): document that --api-prefix removes the health endpoints from the public set Open

priority:low status:ready type:docs

lablup/mlxcel#1671 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs(cli): sync the serve --tp-size family list with generate and mlxcel-server Open

priority:low status:ready type:docs

lablup/mlxcel#1672 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs(reports): Korean report 1059 is missing section 3.2 from the English original Open

priority:low status:ready type:docs

lablup/mlxcel#1673 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
test(python): cover the _common.py URL normalization helpers Open

priority:low status:ready type:test

lablup/mlxcel#1674 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
chore(deps): add the shipped x86_64 Linux CUDA target to deny.toml graph targets Open

priority:low status:ready type:chore

lablup/mlxcel#1675 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs: update cascade-attention prose from "benchmark missing" to "measured and rejected" Open

priority:low status:ready type:docs

lablup/mlxcel#1676 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs(recipes): add a recipes/README explaining the registry snapshot lifecycle Open

priority:low status:ready type:docs

lablup/mlxcel#1677 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(granite_vision): descriptive prompts draw a refusal, colour questions work Open

area:models modelsize:small modeltype:vlm priority:high status:ready type:bug

lablup/mlxcel#1683 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
perf(qwen_vl): text decode falls to half of mlx-lm as context grows Open

area:models priority:medium status:ready type:performance

lablup/mlxcel#1686 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
refactor(server): add ErrorResponse::internal_server_error and collapse the eight hand-rolled 500s Open

priority:low status:ready type:refactor

lablup/mlxcel#1690 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs(bench): speculative_bench module docs contradict its own code in three places Open

priority:low status:ready type:docs

lablup/mlxcel#1691 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
refactor(server): extract the duplicated image fetch/decode helper from the embeddings and rerank routes Open

priority:low status:ready type:refactor

lablup/mlxcel#1692 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
test(surgery): cover the SurgeryPipeline::apply error path Open

priority:low status:ready type:test

lablup/mlxcel#1693 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs(surgery): mlxcel-surgery comments still describe shipped work as future Open

priority:low status:ready type:docs

lablup/mlxcel#1694 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
fix(server): audio routes answer 400 for server-side failures where siblings return 500 Open

priority:medium status:ready type:bug

lablup/mlxcel#1695 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
chore(examples): xla_oracle_check defaults --model to a developer home directory Open

priority:low status:ready type:chore

lablup/mlxcel#1696 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
test(surgery): cover ModelDims::from_config and validate_ids directly Open

priority:low status:ready type:test

lablup/mlxcel#1697 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
test: add reason strings to the bare #[ignore] attributes Open

priority:low status:ready type:test

lablup/mlxcel#1698 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
refactor(cli): dedupe the --estimate-memory preflight block shared by serve and generate Open

priority:low status:ready type:refactor

lablup/mlxcel#1699 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs(embeddings): the detection table omits model_type aliases the loader accepts Open

priority:low status:ready type:docs

lablup/mlxcel#1700 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs(webpage): add a webpage/site README covering the pnpm/uv/zensical build and deploy contract Open

priority:low status:ready type:docs

lablup/mlxcel#1701 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs: CONTRIBUTING says clippy is not gated at PR time, but ci.yml runs -D warnings clippy on PRs Open

priority:medium status:ready type:docs

lablup/mlxcel#1702 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
docs(python): the model-plus-connect-target rule contradicts the recommended socket= recipe Open

priority:low status:ready type:docs

lablup/mlxcel#1703 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026

lablup/mlxcel

Get new issues in your inbox