attnres AGENTS.md

AGENTS.md

Project Identity

attnres is a Rust library that implements Attention Residuals for burn-based Transformer experiments, plus examples, benchmarks, and a web demo.

Current State

• Status: alpha as of March 16, 2026.
• Suitable for: research, examples, local experimentation, integration work on trusted inputs.
• Not yet suitable for: production inference services, validated GPU claims, PyTorch checkpoint interchange, or a stable 1.0 API promise.
• Important gap: there is no dedicated spec.md in this checkout. Use ARCHITECTURE.md, README, module docs, and tests as the current source of truth.

Verified Commands

These commands were run successfully during the latest quality pass:

cargo fmt -- --check
cargo clippy -- -D warnings
cargo test --all-features
cargo build --examples
cd web-demo && npm run build

Additional useful commands:

cargo bench
cargo doc --open

Architecture Map

• src/config.rs: AttnResConfig, ConfigError, validation helpers.
• src/attn_res_op.rs: core depth-attention residual operator.
• src/block_state.rs: completed blocks + current partial block.
• src/layer.rs: one Transformer layer with two AttnRes operations.
• src/model.rs: full model, hidden-state forward, two-phase forward.
• src/two_phase.rs: batched inter-block pass + online softmax merge.
• src/attention.rs: multi-head self-attention.
• src/feed_forward.rs: two-layer GELU MLP.
• src/rms_norm.rs: RMSNorm for 3D and 4D tensors.
• src/serialization.rs: burn-record save/load helpers.
• tests/: unit, integration, property, and differential coverage.
• examples/demo_tui.rs: terminal demo with live routing visualization.
• web-demo/: WASM crate plus Vite frontend.

Non-Negotiable Invariants

• Pseudo-query vectors start at zero.
• Depth softmax is over block/layer sources, not tokens.
• Each Transformer layer uses two AttnRes operations.
• Block boundaries are defined in sublayer space.
• BlockState.blocks[0] is the embedding block.
• Internal invariant failures should panic loudly rather than produce silent wrong outputs.