Pull down to go back
Has anyone tested speculative decoding in llama.cpp with Gemma 4 31B IT or Qwen 3.5 27B?
opensourcereddit

Has anyone tested speculative decoding in llama.cpp with Gemma 4 31B IT or Qwen 3.5 27B?

有人在 llama.cpp 上測試過 Gemma 4 31B IT 或 Qwen 3.5 27B 的推測解碼嗎？

A developer is asking the community about their experience with speculative decoding—a technique that speeds up AI model inference—using llama.cpp with two popular large language models: Gemma 4 31B IT and Qwen 3.5 27B. They're specifically interested in which smaller draft models work best as companions for faster generation and whether users actually saw real performance improvements. The question suggests some uncertainty about whether Qwen 3.5 even works well with llama.cpp's speculative decoding implementation.

Read original →

Keywords

speculative decodingdraft modelsinference optimizationllama.cppGemma 4Qwen 3.5performance

Related Articles

TRELLIS.2 image-to-3D now runs on Mac (Apple Silicon) - no NVIDIA GPU needed
Open SourceRead

TRELLIS.2 image-to-3D now runs on Mac (Apple Silicon) - no NVIDIA GPU needed

I ported Microsoft's TRELLIS.2 to run on Apple Silicon via PyTorch MPS. The original depends on five CUDA-only compiled extensions (flex_gemm, flash_attn, o_voxel, cumesh, nvdiffrast) that have no Mac equivalent. Wrote replacement backends from scratch: Pure-PyTorch sparse 3D convolution (replacing flex_gemm), Python mesh extraction using spatial hashing (replacing CUDA hashmap ops in o_voxel), SDPA attention for sparse transformers (replacing flash_attn), and GPU-accelerated trilinear interpolation (replacing cumesh and nvdiffrast).

Reddit
Show HN: Run TRELLIS.2 Image-to-3D generation natively on Apple Silicon
Open SourceRead

Show HN: Run TRELLIS.2 Image-to-3D generation natively on Apple Silicon

I ported Microsoft's TRELLIS.2 (4B parameter image-to-3D model) to run on Apple Silicon via PyTorch MPS. The original requires CUDA with flash_attn, nvdiffrast, and custom sparse convolution kernels: none of which work on Mac. I replaced the CUDA-specific ops with pure-PyTorch alternatives: a gather-scatter sparse 3D convolution, SDPA attention for sparse transformers, and a Python-based mesh extraction replacing CUDA hashmap operations. Total changes are a few hundred lines across 9 files.

Hacker News
157
Which Qwen 3.6 Quantization Works Best for M5 Pro with 24GB RAM?
Open SourceRead

Which Qwen 3.6 Quantization Works Best for M5 Pro with 24GB RAM?

A user with an M5 Pro machine featuring 24GB of RAM is trying to figure out the best quantization level for running Qwen 3.6 via Ollama. They're unsure whether to go with Q4 (4-bit quantization) but can't find a solid Q3 (3-bit) alternative. They're looking for recommendations on which version would run smoothly on their setup without compromising too much on quality.

Reddit