v0.65.0

What's Changed

• chore(ci): build SDK stuff when Rust / FFI changes by @ndizazzo in #353
• Add per-model parallel config to control the number of parallel request slots per model by @ndizazzo in #357
• Fix startup on nodes with pinned Tegra-based GPU SoCs by @ndizazzo in #360
• Bump iroh to 0.98, move legacy Fly relay to tools/ by @michaelneale in #362
• Corrects duplicated log extension on logfiles by @ndizazzo in #358
• refactor(cli): invert enumerate-host to opt-out with --no-enumerate-host by @ndizazzo in #350
• chore(docs): update docs with directional metrics plan by @ndizazzo in #351
• Fix passing parallel via config.toml by @ndizazzo in #368
• Add support to output to multiple different log files by port for llamacpp servers by @ndizazzo in #367
• Add mesh join age display with backend timestamp support by @ndizazzo in #345
• api: expose read-only model search on the management API by @IvGolovach in #365
• Add delete option to model CLI flags to remove just a single managed model by @ndizazzo in #366
• Add support to explicitly publish mesh via command-line by @ndizazzo in #371
• api: add local explicit-interest submission for canonical model refs by @IvGolovach in #374
• Add persistent OpenCode config writing with --write flag by @ndizazzo in #373
• paginate long model listings in interactive terminals by @i386 in #376
• Show full split GGUF sizes in model search results by @i386 in #377
• Fix: Add full workspace copy to Dockerfiles to resolve missing crates by @intenso in #363
• feature(ui-playground): add a UI playground and components by @ndizazzo in #375
• chore(clippy): address clippy issues and warnings by @ndizazzo in #378
• Fix Windows CUDA builds and dashboard VRAM display by @i386 in #381
• fix(models): handle HF snapshot/ref mismatch in updates flow by @josesantoscorte in #380
• Fix Vulkan device selection across startup and GPU reporting by @i386 in #383
• Build llama.cpp from an upstream patch queue by @i386 in #387
• Fix llama.cpp latest prepare workflow by @i386 in #391
• Add llama upstream pin canary by @i386 in #390
• openai-endpoint plugin: vLLM, TGI, Ollama passthrough by @michaelneale in #395
• feature(console-output): completely revise mesh-llm console output for JSONL output + pretty printing by @ndizazzo in #388
• feature(metrics-collector): add support to centralize storage for met… by @ndizazzo in #386
• Recovered peers can rejoin the mesh after dead_peers TTL expires by @michaelneale in #405
• Fix SIGABRT crash on Mac Studio (iroh 0.98 double-free in Endpoint::online) by @michaelneale in #407
• chore: add JSONL formatter for piped log output by @ndizazzo in #408
• fix(tui): fix TUI interrupting error output for 'mesh-llm model show' by @ndizazzo in #409
• feature(runtime): add model context info to runtime API by @ndizazzo in #410

New Contributors

• @josesantoscorte made their first contribution in #380

Full Changelog: v0.64.0...v0.65.0