Summary

Extend .apr format to support web-scale models (10B+ parameters) while maintaining WASM-first design.

Current State

APR v1 is excellent for small-medium models but lacks:

1. Tensor alignment - No 32/64-byte alignment for zero-copy mmap
2. Streaming decompression - LZ4 mentioned in spec but not implemented
3. Sharding - Bundle system exists but not integrated with APR reader

Proposed APR v2 Features

Tier 1: Alignment (mmap-ready)

• 32-byte tensor alignment (matches GGUF)
• Padding bytes between tensors
• Zero-copy slice access

Tier 2: Compression

• LZ4 block compression (64KB blocks)
• Per-tensor compression flag
• Streaming decompression for WASM

Tier 3: Sharding (web-scale)

• Manifest file for multi-part models
• Tensor-level sharding
• Progressive loading support

Backward Compatibility

• Magic: APR2 (vs APR1)
• APR1 reader can skip unknown sections
• Converter tool: apr1-to-apr2

Non-Goals

• Not replacing GGUF (interop via export)
• Not supporting training (inference only)

References

• Current spec: docs/specifications/model-format-spec.md
• GGUF reference: format/gguf.rs
• Bundle system: bundle/