Fetching from the wire…
01
02
03
04
05
06
07
08
Models2026-09-26 · source-backed
PR #57934 (+1,933 lines) ports LiLiCorrDraftModel from SGLang onto vLLM's DFlash path. Its head reranks global candidate lattices using target embeddings, full-vocabulary log-probs and draft hidden states, reuses DFlash2 convolution layers, supports mixed BF16/NVFP4 checkpoints, and matched SGLang's acceptance length exactly on 4x GB200.
Each link below shares sources, entities, or timing with this story.
PR 27342 adds grouped dynamic depthwise convolution and a candidate selector for DFlash2 draft models, reporting 1.81x on Apple M5 Pro running Qwen3.8-27B Q4_K_M at about 5.03 token acceptance, 1.77x to 1.85x across BF16 and Q8_0, and one Nvidia user seeing a consistent 2x dec...
1. Set Up Cursor Automations (intermediate) — Event-driven agents from PagerDuty/GitHub/Slack triggers with isolated sandboxes. Cursor Blog 2. Apply Context Engineering to Cut Agent Costs 60-80% (advanced) — Hierarchical token budgets, dynamic tool filtering (max 15), automati...
Lily is Apache 2.0 inside pplx-garden, a small Metal inference server built for Qwen3.6-35B-A3B converted to MLX affine 4-bit, exposing a minimal OpenAI-compatible chat API with greedy decoding. The README explicitly rules out dense and smaller Qwen checkpoints, BF16, GGUF, AW...
One file handles tokenizer, transformer, KV cache, sampling and CPU kernels with no external library doing the interesting parts, producing about a 5.0 GB model file (GitHub). Weights are int8 with FP16 scales, linear-layer inputs dynamically quantized to int8, other activatio...
Intel's release on September 16 covers Kokoro-82M, Qwen3-VL-4B with EAGLE-3 speculative decoding, Qwen3-ASR, Muse Glimmer 30B, Hy-MT2-1.8B and Granite 4.0 H Micro, with Gemma-3n on CPU and FLUX.2-Klein 4B plus Kokoro-82M on NPUs. Early-release CPU/GPU enablement adds Qwen-Imag...
anuj0456/OpenArch has single-file implementations for GPT-2 XL, Llama 2/3/4 Maverick, OLMo 2, DeepSeek R1, Gemma 3, Mistral 3, Qwen 3, Kimi K2, GLM 4.5, GPT-OSS and PaliGemma, each making attention type, normalization and positional encoding explicit. Stated goal is clarity ov...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.