Search results

perf(models): re-bench fused QKV+RMSNorm+RoPE on CUDA and decide the MLXCEL_FUSED_QK_NORM default

area:coremlxcel-core: MLX FFI, primitives, KV cache, layersmlxcel-core: MLX FFI, primitives, KV cache, layers

area:modelsModel architectures, weights, loading, metadataModel architectures, weights, loading, metadata

priority:mediumMedium priorityMedium priority

status:readyReady to be worked onReady to be worked on

type:performancePerformance improvementsPerformance improvements

Status: Open.

#355 In lablup/mlxcel;

· inureyes opened on Jun 18, 2026

perf(core): record 4k/16k/32k KV-cache sweeps to back the Turbo KV advisor recommendations

area:coremlxcel-core: MLX FFI, primitives, KV cache, layersmlxcel-core: MLX FFI, primitives, KV cache, layers

area:inferenceGeneration, sampling, decoding (incl. speculative, DRY)Generation, sampling, decoding (incl. speculative, DRY)

priority:backlogFuture considerationsFuture considerations

status:readyReady to be worked onReady to be worked on

type:performancePerformance improvementsPerformance improvements

Status: Open.

#354 In lablup/mlxcel;

· inureyes opened on Jun 18, 2026

perf(core): hardware-gated MLX_MAX_OPS_PER_BUFFER default (Gemma3n +11-13% on M1 Ultra, regresses on M5)

area:coremlxcel-core: MLX FFI, primitives, KV cache, layersmlxcel-core: MLX FFI, primitives, KV cache, layers

area:inferenceGeneration, sampling, decoding (incl. speculative, DRY)Generation, sampling, decoding (incl. speculative, DRY)

priority:mediumMedium priorityMedium priority

status:readyReady to be worked onReady to be worked on

type:performancePerformance improvementsPerformance improvements

Status: Open.

#353 In lablup/mlxcel;

· inureyes opened on Jun 18, 2026

fix(server): gemma-4-12b-it defaults to enable_thinking, yielding verbose thinking and truncated/empty answers at normal max_tokens

area:coremlxcel-core: MLX FFI, primitives, KV cache, layersmlxcel-core: MLX FFI, primitives, KV cache, layers

area:inferenceGeneration, sampling, decoding (incl. speculative, DRY)Generation, sampling, decoding (incl. speculative, DRY)

priority:mediumMedium priorityMedium priority

status:reviewUnder reviewUnder review

type:bugBug fixes, error corrections, or issue resolutionsBug fixes, error corrections, or issue resolutions

Status: Open.

#352 In lablup/mlxcel;

· inureyes opened on Jun 18, 2026

fix(server): /v1/messages returns HTTP 422 for Claude Code >= 2.1.156 (system-role turns inside the messages array)

area:coremlxcel-core: MLX FFI, primitives, KV cache, layersmlxcel-core: MLX FFI, primitives, KV cache, layers

area:inferenceGeneration, sampling, decoding (incl. speculative, DRY)Generation, sampling, decoding (incl. speculative, DRY)

priority:highHigh priorityHigh priority

status:readyReady to be worked onReady to be worked on

type:bugBug fixes, error corrections, or issue resolutionsBug fixes, error corrections, or issue resolutions

Status: Open.

#349 In lablup/mlxcel;

· w200kr opened on Jun 18, 2026

feat: introduce a ComputeBackend seam to abstract the forward-execution engine

area:architectureArchitecture and code structure changesArchitecture and code structure changes

priority:mediumMedium priorityMedium priority

status:readyReady to be worked onReady to be worked on

type:enhancementNew features, capabilities, or significant additionsNew features, capabilities, or significant additions

Status: Open.

#338 In lablup/mlxcel;

· inureyes opened on Jun 17, 2026

perf(core): adaptive selector for the native paged-attention decode kernel

area:coremlxcel-core: MLX FFI, primitives, KV cache, layersmlxcel-core: MLX FFI, primitives, KV cache, layers

area:inferenceGeneration, sampling, decoding (incl. speculative, DRY)Generation, sampling, decoding (incl. speculative, DRY)

priority:mediumMedium priorityMedium priority

type:performancePerformance improvementsPerformance improvements

Status: Open.

#331 In lablup/mlxcel;

· inureyes opened on Jun 17, 2026

perf(moe): backend-aware fused-MoE Dff cap (CUDA crossover) and dispatch heuristic

area:coremlxcel-core: MLX FFI, primitives, KV cache, layersmlxcel-core: MLX FFI, primitives, KV cache, layers

area:modelsModel architectures, weights, loading, metadataModel architectures, weights, loading, metadata

priority:mediumMedium priorityMedium priority

type:performancePerformance improvementsPerformance improvements

Status: Open.

#330 In lablup/mlxcel;

· inureyes opened on Jun 17, 2026

fix(olmoe): OLMoE-1B-7B produces corrupted/degenerate output beyond the q_norm fix (pre-existing, not MoE-related)

area:inferenceGeneration, sampling, decoding (incl. speculative, DRY)Generation, sampling, decoding (incl. speculative, DRY)

area:modelsModel architectures, weights, loading, metadataModel architectures, weights, loading, metadata

platform:macosmacOS (Apple Silicon) specificmacOS (Apple Silicon) specific

priority:highHigh priorityHigh priority

status:readyReady to be worked onReady to be worked on

type:bugBug fixes, error corrections, or issue resolutionsBug fixes, error corrections, or issue resolutions

Status: Open.

#318 In lablup/mlxcel;

· inureyes opened on Jun 16, 2026

perf(moe): wire minimax to the fused decode-MoE kernel

area:inferenceGeneration, sampling, decoding (incl. speculative, DRY)Generation, sampling, decoding (incl. speculative, DRY)

area:modelsModel architectures, weights, loading, metadataModel architectures, weights, loading, metadata

platform:macosmacOS (Apple Silicon) specificmacOS (Apple Silicon) specific

priority:lowLow priorityLow priority

status:blockedBlocked by dependencies or other issuesBlocked by dependencies or other issues

type:performancePerformance improvementsPerformance improvements

Status: Open.

#304 In lablup/mlxcel;

· inureyes opened on Jun 16, 2026

perf(nemotron-h): decode gap is MoE-block op-density (routed + shared expert), not SSM/attention

area:inferenceGeneration, sampling, decoding (incl. speculative, DRY)Generation, sampling, decoding (incl. speculative, DRY)

area:modelsModel architectures, weights, loading, metadataModel architectures, weights, loading, metadata

platform:macosmacOS (Apple Silicon) specificmacOS (Apple Silicon) specific

priority:mediumMedium priorityMedium priority

type:performancePerformance improvementsPerformance improvements

Status: Open.

#284 In lablup/mlxcel;

· inureyes opened on Jun 14, 2026

chore: Bump mlxcel-core to Rust edition 2024 to match the root crate

priority:lowLow priorityLow priority

status:readyReady to be worked onReady to be worked on

type:choreMaintenance tasks (build, CI, etc.)Maintenance tasks (build, CI, etc.)

Status: Open.

#272 In lablup/mlxcel;

· inureyes opened on Jun 13, 2026