Issues

docs(reports): realign Korean technical reports 1863, 2127, and 2537 with their English originals

lablup/mlxcel#3038 · opened

on Nov 13, 2026

1

feat(prompt-cache): keep the previous turn's restore point for last-turn edits

lablup/mlxcel#3035 · opened

on Nov 12, 2026

fix(metal): bound the materializing prefill SDPA on Metal, where query-axis chunking is gated CUDA-only and long prompts OOM

lablup/mlxcel#3033 · opened

on Nov 12, 2026

fix(loading): Qwen3.5 VLM route skips the ModelOpt NVFP4 repack, so packed uint8 planes reach the text model and abort at the first forward pass

lablup/mlxcel#3029 · opened

on Nov 12, 2026

fix(quant): load ModelOpt W4A16_NVFP4 checkpoints in text families without a ModelOpt group size

lablup/mlxcel#3024 · opened

on Nov 12, 2026

fix(loading): resolve the Qwen3.5 tokenizer eos_token through tokenizer.json

lablup/mlxcel#3022 · opened

on Nov 12, 2026

fix(server): Unix socket bind removes any existing path, including a live server's socket

lablup/mlxcel#3020 · opened

on Nov 12, 2026

fix(core): reject packed uint8 weights in dense linear loaders instead of aborting

lablup/mlxcel#3018 · opened

on Nov 12, 2026

fix(python): managed server start unlinks any existing socket path, including a live server's socket

lablup/mlxcel#3016 · opened

on Nov 12, 2026

fix(python): AsyncLLM.close() blocks the event loop while the managed server shuts down

lablup/mlxcel#3014 · opened

on Nov 12, 2026

ci(python): Python 3.9 test job fails in setup-python on ubuntu-latest (now Ubuntu 26.04)

lablup/mlxcel#3012 · opened

on Nov 12, 2026

fix(python): a failed or cancelled HTTP close leaves the managed server running

lablup/mlxcel#3010 · opened

on Nov 12, 2026

fix(python): a slow /health response is accepted after startup_timeout has expired

lablup/mlxcel#3008 · opened

on Nov 12, 2026

fix(python): Unix socket path length limit is checked in characters, not bytes

lablup/mlxcel#3006 · opened

on Nov 12, 2026

fix(python): a client that never spawned a server deletes whatever is at its socket path

lablup/mlxcel#3004 · opened

on Nov 12, 2026

fix(python): managed mode builds an invalid URL for an IPv6 host such as ::1

lablup/mlxcel#3002 · opened

on Nov 12, 2026

fix(python): closing a streaming generator early leaves its HTTP response open

lablup/mlxcel#3000 · opened

on Nov 12, 2026

fix(models): run a mixed-bit Gemma MLP projection by projection

lablup/mlxcel#2998 · opened

on Nov 12, 2026

fix(loading): load text-only Gemma 4 and Gemma 3n exports by resolving the weight prefix

lablup/mlxcel#2996 · opened

on Nov 12, 2026

fix(models): accept a Mamba config that names a setting both ways

lablup/mlxcel#2994 · opened

on Nov 12, 2026

fix(vision): apply EXIF orientation to image inputs before preprocessing

lablup/mlxcel#2992 · opened

on Nov 12, 2026

test(cuda): isolate the Windows runtime packager tests from an exported CUDNN_PATH

lablup/mlxcel#2989 · opened

on Nov 12, 2026

test(cuda): make verify-test-cuda fails the cache-default test on every CUDA host since #2876

lablup/mlxcel#2987 · opened

on Nov 12, 2026

fix(xla): Gemma 3 eager-MLX/IREE boundary gate fails on main from a stale reference activation

lablup/mlxcel#2985 · opened

on Nov 12, 2026

Cascade decode: why it loses to the flat launch, and a shared-prefix kernel that beats flat on M5 (measured on a rewrite, not on mlxcel)

lablup/mlxcel#2979 · opened

on Nov 11, 2026

4

Navigation Menu

Repository navigation

Issues

docs(reports): realign Korean technical reports 1863, 2127, and 2537 with their English originals

feat(prompt-cache): keep the previous turn's restore point for last-turn edits

fix(metal): bound the materializing prefill SDPA on Metal, where query-axis chunking is gated CUDA-only and long prompts OOM

fix(loading): Qwen3.5 VLM route skips the ModelOpt NVFP4 repack, so packed uint8 planes reach the text model and abort at the first forward pass

fix(quant): load ModelOpt W4A16_NVFP4 checkpoints in text families without a ModelOpt group size

fix(loading): resolve the Qwen3.5 tokenizer eos_token through tokenizer.json

fix(server): Unix socket bind removes any existing path, including a live server's socket

fix(core): reject packed uint8 weights in dense linear loaders instead of aborting

fix(python): managed server start unlinks any existing socket path, including a live server's socket

fix(python): AsyncLLM.close() blocks the event loop while the managed server shuts down

ci(python): Python 3.9 test job fails in setup-python on ubuntu-latest (now Ubuntu 26.04)

fix(python): a failed or cancelled HTTP close leaves the managed server running

fix(python): a slow /health response is accepted after startup_timeout has expired

fix(python): Unix socket path length limit is checked in characters, not bytes

fix(python): a client that never spawned a server deletes whatever is at its socket path

fix(python): managed mode builds an invalid URL for an IPv6 host such as ::1

fix(python): closing a streaming generator early leaves its HTTP response open

fix(models): run a mixed-bit Gemma MLP projection by projection

fix(loading): load text-only Gemma 4 and Gemma 3n exports by resolving the weight prefix

fix(models): accept a Mamba config that names a setting both ways

fix(vision): apply EXIF orientation to image inputs before preprocessing

test(cuda): isolate the Windows runtime packager tests from an exported CUDNN_PATH

test(cuda): make verify-test-cuda fails the cache-default test on every CUDA host since #2876

fix(xla): Gemma 3 eager-MLX/IREE boundary gate fails on main from a stale reference activation

Cascade decode: why it loses to the flat launch, and a shared-prefix kernel that beats flat on M5 (measured on a rewrite, not on mlxcel)