Termux-LlamaCpp

Advanced Parameters & Security Tuning

Production guidelines for memory sizing, thread scheduling, loopback network security, and cryptographic supply-chain verification.

1. Mobile Context Allocation & Slots (-np 1)

By default on multi-slot servers, context window memory is divided equally across slots (e.g., 256 context / 4 slots = 64 tokens per slot). On single-user mobile devices, this leads to rapid context truncation. Termux-LlamaCpp enforces -np 1 and allocates the entire 2048 token buffer directly to the active session.

llama-server \
  -m model.gguf \
  --host 127.0.0.1 \
  --port 18080 \
  -c 2048 \
  -np 1 \
  -t 4 \
  --no-mmap

2. Android Bionic Direct Memory Loading (--no-mmap)

Standard Linux memory-mapping (mmap) on Android flash storage (/data/data/com.termux/files/...) frequently triggers kernel I/O locks, memory thrashing, and process timeouts. The --no-mmap flag instructs the llama.cpp engine to sequentially load model tensors directly into RAM, guaranteeing deterministic cold-start times (~1.8s for 1.92GB models).

3. Ed25519 Cryptographic Manifest Verification

Termux-LlamaCpp implements strict supply-chain security protocols:

• Fail-Closed Anti-Downgrade: If a signed release manifest is present, the runtime will strictly reject untrusted local build receipts.
• Symlink Traversal Prevention: Model weight paths, trust root public keys, and revocation files are verified with is_symlink() to block TOCTOU injection.
• Loopback Reverse Proxy Isolation: The native backend binds strictly to loopback 127.0.0.1:18080, protected by CORS origin filtering.