Issues - sleepy/sleepy-llm - Sleepy Git (Forgejo)

sleepy/sleepy-llm

Search syntax

`assignee:<username>`	Filter by assigned user
`author:<username>`	Filter by the author
`is:all`	All
`is:closed`	Closed
`is:open`	Open
`mentions:<username>`	Filter by mentioned user
`modified:[>\|<]<date>`	Filter by last modified date
`review:<username>`	Filter by user who reviewed
`sort:<by>:[asc\|desc]`	Sort by: `created\|comments\|updated\|deadline`

New issue

bug: forward pass produces incorrect logits — greedy decode diverges from MLX

#63 opened 2026-05-22 17:26:51 +02:00 by sleepy

bug: Tokenizer test hardcoded to Qwen3.5-4B path which may not exist

#56 opened 2026-05-20 23:58:14 +02:00 by sleepy

perf: Reach llama.cpp/MLX decode parity (37 tok/s target)

#53 opened 2026-05-15 21:47:43 +02:00 by sleepy

feat: Multi-token prediction (MTP) speculative decoding

#48 opened 2026-05-15 19:09:57 +02:00 by sleepy

perf: Pipeline decode — overlap GPU commit with next token embedding

#47 opened 2026-05-15 19:02:26 +02:00 by sleepy

feat: GPU-accelerated prefill path

#45 opened 2026-05-15 19:02:26 +02:00 by sleepy

perf: MPP matmul2d for Apple Silicon M3+

#42 opened 2026-05-15 19:02:26 +02:00 by sleepy

perf: Quantized weight loading (4-bit) for 4x bandwidth reduction

#39 opened 2026-05-15 19:02:26 +02:00 by sleepy

Single command buffer decode — full GPU pipeline for all 32 layers

#36 opened 2026-05-15 12:21:00 +02:00 by sleepy

Metal kernels for GPU linear attention (conv1d, delta rule, norm/gate)

#35 opened 2026-05-15 12:20:34 +02:00 by sleepy

Reach 37+ tok/s decode target (match llama.cpp)

#34 opened 2026-05-15 00:50:48 +02:00 by sleepy