Issues

• perf(rocm): flash SDPA dominates long-context chunked prefill on gfx1151

  area:inferenceplatform:linuxpriority:mediumstatus:readytype:performance
  lablup/mlxcel#2251 · Filter by author inureyes opened on Oct 8, 2026

• fix(core): accumulate CPU quantized matmul in f32 for bf16 activations

  area:corepriority:mediumstatus:readytype:bug
  lablup/mlxcel#2248 · Filter by author inureyes opened on Oct 8, 2026

• perf(rocm): give the dense mxfp4 qmv kernels the warp-shared word path

  area:coreplatform:linuxpriority:lowstatus:readytype:performance
  lablup/mlxcel#2246 · Filter by author inureyes opened on Oct 8, 2026

• perf(moe): route granite-4.0-h single-token MoE through the fused kernel

  area:modelspriority:lowstatus:readytype:performance
  lablup/mlxcel#2245 · Filter by author inureyes opened on Oct 8, 2026

• chore(rocm): remove dead overlay env knobs and range-check the dequant byte cap

  area:coreplatform:linuxpriority:lowstatus:in-progresstype:chore
  lablup/mlxcel#2242 · Filter by author inureyes opened on Oct 8, 2026

• fix(moe): sorted gather_qmm prefill lost right_sorted since #2137, MoE prefill 1.7x to 4.8x slower

  area:inferencepriority:highstatus:in-progresstype:bug
  lablup/mlxcel#2241 · Filter by author inureyes opened on Oct 8, 2026

• perf(rocm): pipelined B=1 decode on Gemma 3 is 1.3 percent below the pre-epic loop at 2048 tokens on gfx1151

  area:inferenceplatform:linuxpriority:mediumstatus:readytype:performance
  lablup/mlxcel#2239 · Filter by author inureyes opened on Oct 8, 2026

• fix(rocm): other HIP launchers size an uncapped int grid from the element count

  area:corepriority:lowstatus:readytype:bug
  lablup/mlxcel#2234 · Filter by author inureyes opened on Oct 8, 2026

• perf(rocm): turn the f16 dequant LRU off by default

  area:coreplatform:linuxpriority:lowstatus:readytype:performance
  lablup/mlxcel#2232 · Filter by author inureyes opened on Oct 8, 2026

• chore: post-epic cleanup of compat paged kernels, the model-owned paged flag, and the CUDA SSM bf16 parity failures

  area:corearea:modelspriority:lowstatus:readytype:chore
  lablup/mlxcel#2230 · Filter by author inureyes opened on Oct 8, 2026

• perf: pipeline penalty and DRY requests in the engine's B=1 client, and offer prompt lookup on the server

  area:corearea:inferencepriority:mediumstatus:readytype:performance
  lablup/mlxcel#2229 · Filter by author inureyes opened on Oct 8, 2026

• perf: choose the prefill chunk by whether other sequences are decoding

  area:corearea:inferencepriority:mediumstatus:readytype:performance
  lablup/mlxcel#2228 · Filter by author inureyes opened on Oct 8, 2026

• perf: bring mlxcel run and chat B=1 decode within the 1.0 percent threshold

  area:cliarea:inferencepriority:highstatus:readytype:performance
  lablup/mlxcel#2227 · Filter by author inureyes opened on Oct 8, 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

• epic: one batch-native engine for CLI and server decode

  area:architecturearea:inferencepriority:highstatus:in-progresstype:refactor
  lablup/mlxcel#2166 · Filter by author inureyes opened on Oct 6, 2026
  2

• 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
  4

• 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

• fix(cuda): ssm_update_kernel bf16 parity tests fail on GB10

  area:coreplatform:linuxpriority:highstatus:readytype:bug
  lablup/mlxcel#2142 · Filter by author inureyes opened on Oct 6, 2026

• fix(scripts): A/B gate scripts fail on GNU mktemp and then write to /

  area:benchmarkplatform:linuxpriority:mediumstatus:readytype:bug
  lablup/mlxcel#2141 · Filter by author inureyes opened on Oct 6, 2026

• fix: granite-vision text-only requests drop the user message

  area:cliarea:inferencearea:modelsmodeltype:vlmpriority:highstatus:readytype:bug
  lablup/mlxcel#2135 · Filter by author inureyes opened on Oct 6, 2026

• 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:readytype:bug
  lablup/mlxcel#1986 · Filter by author inureyes opened on Sep 26, 2026
  1

• chore(deps): bump the MLX pin and declare a base commit for each overlay

  area:coreplatform:linuxplatform:macospriority:mediumstatus:readytype:chore
  lablup/mlxcel#1932 · Filter by author inureyes opened on Sep 19, 2026