Updated 23 October 2026.
- Main Swift package:
/Users/lakshyaprajapati/Repositories/Particle. - Package name: Particle. Sarvam modules:
ParticleGGUF,ParticleCore, andParticleCLI. Upstream Gemma module names remain intact. - Xcode:
/Volumes/Particle/Xcode.app(27.1, build 27A9275). Compiler: Swift 6.4. - Metal Toolchain: 27A266a, installed as an Apple-managed component. The exported installer is retained under
/Volumes/Particle/Caches/Particle/Downloads/. - Model:
/Volumes/Particle/Models/sarvam-30b/sarvam-30b-q4_k_m.gguf(19,582,216,032 bytes). Particle reads it directly; no full copy or conversion is required. - Official tokenizer sidecars:
/Volumes/Particle/Models/sarvam-30b/tokenizer/, pinned to Hugging Face revision071ae95e933605ca1104a6b4524a36a98488efa4with checksums inScripts/fetch_sarvam_tokenizer.py. - Build and Derived Data:
/Volumes/Particle/Build/Particle/. - Dependency/compiler caches:
/Volumes/Particle/Caches/Particle/. Scripts/particleandScripts/particle-env.shselect Xcode locally without changing system-widexcode-select.Particle.code-workspaceconfigures the editor, terminal, and ParticleCLI build task for the same tools.- Pinned dependency resolution and release compilation succeeded with the SD-card Xcode.
- The Metal shaders compile and pass independent CPU comparisons on this M2. See the CLI guide for validation commands.
- The root package can be opened with
Scripts/particle xcode. - Current CLI defaults: 128-token batched prefill, bounded embedding rows, 12 RAM expert slots, and the approved exclusive SSD expert cache capped at 4096 MiB in
~/Library/Caches/Particle/Experts/. Use--no-ssd-cachefor original-volume-only expert reads. Fifteen focused tests pass; see current performance.
The earlier Xcode launch failure applied to the prior installation. Historical setup notes and the original Derived Data preference backup are retained under setup/; they do not describe the active toolchain.
See the implementation plan for the Sarvam integration and optimization milestones.