-
Notifications
You must be signed in to change notification settings - Fork 27
Milestones
List view
Native Windows + CUDA builds and packaging hardening carried over from 0.4: the x86_64-pc-windows-msvc build (build.rs Windows branches, static-link delay-load), the Windows release artifact with bundled CUDA/cuDNN/OpenBLAS DLLs, the code-signing decision, and a developer Windows build guide. Tracked by #25 (umbrella) and #58 (feasibility spike + porting plan).
No due date•0/2 issues closedPlanned work: harden the OpenAI-compatible audio endpoints (TTS + STT, MLX FFI and panic safety), new models (Mellum 2) and Gemma 4 video input, a ComputeBackend abstraction seam, MoE and dense decode performance gaps vs mlx-lm, disaggregated-serving expansion (/v1/completions, multi-node routing and failover), and native Windows + CUDA builds with packaging hardening.
Overdue by 19 day(s)•Due by June 26, 2026•18/22 issues closedNew model families (BitNet 1.58-bit, PLaMo 2, Apertus, Seed-OSS, dots.llm1, Falcon-H1, LFM2/LFM2-MoE, IBM Granite) and audio I/O (TTS speech + Whisper ASR endpoints). Fused decode-MoE kernel wired across MoE families, Turbo KV cache modes with a per-family advisor, CUDA (GB10) warmup/JIT fixes, bf16 and mixed-bit quant correctness, and decode-performance work for M1 through M4.
Overdue by 1 month(s)•Due by June 15, 2026•131/131 issues closedFirst public release-related issues / PRs
Overdue by 1 month(s)•Due by May 28, 2026•146/146 issues closedUnified paged KV cache (epic): radix-tree-indexed paged attention with shared, refcounted copy-on-write blocks and a fused Metal paged-attention kernel. Server-side prefix-cache reuse and disaggregated serving (chunked-prefill handoff, per-frame token streaming), batched MTP speculative-decoding hardening, tool-calling and chat-template fidelity, and DiffusionGemma block-diffusion generation.
Overdue by 1 month(s)•Due by June 12, 2026•81/81 issues closed