Starstruk Logo MarkStarstruk

v0.44.0

Mesh-LLM/mesh-llmv0.44.0Mar 24, 2026by michaelneale

AI Summary

Rebases the llama.cpp fork on the latest upstream, incorporating security patches and new capabilities while retaining mesh-specific patches.

Key Highlights

  • Rebased llama.cpp fork on latest upstream (419 commits)
  • Upstream RPC remote code execution security patch
  • Metal CONV_3D support and bf16 native flash attention (CUDA)
  • Mesh-specific patches including Zero-transfer tensor loading and B2B transfers

New Features

  • Upstream security patch for RPC
  • Metal CONV_3D support
  • bf16 native flash attention (CUDA)
  • Zero-transfer tensor loading (SET_TENSOR_GGUF)
  • B2B direct server-to-server transfers
  • MoE expert routing tools

Full Release Notes

Rebased llama.cpp fork on latest upstream (419 commits).

**Upstream highlights picked up:**
- RPC remote code execution security patch
- Metal CONV_3D support
- bf16 native flash attention (CUDA)
- Server Host header fix, httplib dynamic threads
- Grammar parsing fix (stack overflow prevention)
- Memory fix for recurrent models

**Our patches (all rebased clean):**
- Zero-transfer tensor loading (SET_TENSOR_GGUF)
- RPC probing skip
- get_alloc_size cache
- B2B direct server-to-server transfers
- MoE expert routing + analysis/split tools

Also: removed stale PLAN.md