-
Notifications
You must be signed in to change notification settings - Fork 29
All issues
Issue creation is restricted in this repository
Issues
is:issue state:open
is:issue state:open
Search results
fix(models): dedicated root-cause debugging of the DeepSeek-V2-Lite greedy nondeterminism (follow-up to #817)
area:modelsModel architectures, weights, loading, metadataModel architectures, weights, loading, metadatapriority:highHigh priorityHigh prioritystatus:investigationFeasibility spike / under investigationFeasibility spike / under investigationtype:bugBug fixes, error corrections, or issue resolutionsBug fixes, error corrections, or issue resolutionsStatus: Open.#824 In lablup/mlxcel;chore(deps): report the MLX CUDA graph-cache lifetime-miss abort upstream and track dropping our local default
area:inferenceGeneration, sampling, decoding (incl. speculative, DRY)Generation, sampling, decoding (incl. speculative, DRY)priority:lowLow priorityLow prioritystatus:backlogIn the backlog, not yet readyIn the backlog, not yet readytype:dependencyDependency updatesDependency updatesStatus: Open.#821 In lablup/mlxcel;fix(server): thinking_mode injection sentinel "enabled" does not match the DeepSeek-V4-flash template's 'thinking' gate
priority:lowLow priorityLow prioritystatus:readyReady to be worked onReady to be worked ontype:bugBug fixes, error corrections, or issue resolutionsBug fixes, error corrections, or issue resolutionsStatus: Open.#819 In lablup/mlxcel;fix(models): DeepSeek-V2-Lite emits nondeterministic garbage on GB10 CUDA (broken since the 2026-05-19 benchmark)
area:modelsModel architectures, weights, loading, metadataModel architectures, weights, loading, metadatapriority:highHigh priorityHigh prioritystatus:readyReady to be worked onReady to be worked ontype:bugBug fixes, error corrections, or issue resolutionsBug fixes, error corrections, or issue resolutionsStatus: Open.#817 In lablup/mlxcel;feat: extend the OpenXLA packed-quant path to bf16 scales/biases
area:inferenceGeneration, sampling, decoding (incl. speculative, DRY)Generation, sampling, decoding (incl. speculative, DRY)priority:lowLow priorityLow prioritystatus:readyReady to be worked onReady to be worked ontype:enhancementNew features, capabilities, or significant additionsNew features, capabilities, or significant additionsStatus: Open.#603 In lablup/mlxcel;track(upstream-iree): fused dequant->matmul for the packed int8 bandwidth win
area:inferenceGeneration, sampling, decoding (incl. speculative, DRY)Generation, sampling, decoding (incl. speculative, DRY)priority:lowLow priorityLow prioritystatus:blockedBlocked by dependencies or other issuesBlocked by dependencies or other issuestype:performancePerformance improvementsPerformance improvementsStatus: Open.#574 In lablup/mlxcel;epic: OpenXLA multimodal / VLM architecture support (Qwen2-VL, Gemma3n, Phi4MM, Molmo, and more)
area:architectureArchitecture and code structure changesArchitecture and code structure changesarea:inferenceGeneration, sampling, decoding (incl. speculative, DRY)Generation, sampling, decoding (incl. speculative, DRY)priority:lowLow priorityLow prioritystatus:readyReady to be worked onReady to be worked ontype:enhancementNew features, capabilities, or significant additionsNew features, capabilities, or significant additionsStatus: Open.#566 In lablup/mlxcel;epic: OpenXLA SSM / hybrid / recurrent architecture support (Mamba2, Nemotron-H, Jamba, and more)
area:architectureArchitecture and code structure changesArchitecture and code structure changesarea:inferenceGeneration, sampling, decoding (incl. speculative, DRY)Generation, sampling, decoding (incl. speculative, DRY)priority:lowLow priorityLow prioritystatus:readyReady to be worked onReady to be worked ontype:enhancementNew features, capabilities, or significant additionsNew features, capabilities, or significant additionsStatus: Open.#564 In lablup/mlxcel;feat(deepseek-v4): clamp shared-expert SwiGLU to swiglu_limit
area:modelsModel architectures, weights, loading, metadataModel architectures, weights, loading, metadatapriority:lowLow priorityLow prioritystatus:blockedBlocked by dependencies or other issuesBlocked by dependencies or other issuestype:enhancementNew features, capabilities, or significant additionsNew features, capabilities, or significant additionsStatus: Open.#550 In lablup/mlxcel;feat(deepseek-v4): implement HiSA hierarchical sparse attention
area:modelsModel architectures, weights, loading, metadataModel architectures, weights, loading, metadatapriority:lowLow priorityLow prioritystatus:blockedBlocked by dependencies or other issuesBlocked by dependencies or other issuestype:enhancementNew features, capabilities, or significant additionsNew features, capabilities, or significant additionsStatus: Open.#549 In lablup/mlxcel;tracking: VLM-reference model families needing architecture investigation before a port
area:modelsModel architectures, weights, loading, metadataModel architectures, weights, loading, metadatapriority:lowLow priorityLow prioritystatus:investigationFeasibility spike / under investigationFeasibility spike / under investigationtype:enhancementNew features, capabilities, or significant additionsNew features, capabilities, or significant additionsStatus: Open.#548 In lablup/mlxcel;tracking: non-generative vision models present in the VLM reference but out of mlxcel's LLM/VLM-generation scope
area:modelsModel architectures, weights, loading, metadataModel architectures, weights, loading, metadatapriority:lowLow priorityLow prioritystatus:investigationFeasibility spike / under investigationFeasibility spike / under investigationtype:enhancementNew features, capabilities, or significant additionsNew features, capabilities, or significant additionsStatus: Open.#524 In lablup/mlxcel;