Grand Diomande Research · Full HTML Reader

AGP TurboQuant + ANE Benchmarks

This benchmark package makes the local `cog-rlm` TurboQuant and Apple Neural Engine research measurable inside the AGP research track.

Agents That Account for Themselves experiment experiment writeup candidate score 24 .md

Full Public Reader

AGP TurboQuant + ANE Benchmarks

This benchmark package makes the local `cog-rlm` TurboQuant and Apple Neural Engine research measurable inside the AGP research track.

Source Libraries

  • `Desktop/cog-rlm/scripts/turboquant.py`
  • `Desktop/cog-rlm/scripts/ane_bridge.py`
  • `Desktop/cog-rlm/scripts/ane_whisper_spike.py`
  • `Desktop/cog-rlm/scripts/ane_mlx_train.py`
  • `Desktop/cog-rlm/scripts/ane_trainer.py`
  • `Desktop/Comp-Core/core/retrieval/cc-turboquant-index`

Run

bash
cd Desktop/Comp-Core/benchmarks/agp-turboquant-ane
python3 agp_turboquant_ane_bench.py

Larger run:

bash
python3 agp_turboquant_ane_bench.py \
  --n-vectors 10000 \
  --queries 50 \
  --activation-shape 1,1280,1,512

Rust packed-code sidecar:

bash
cd Desktop/Comp-Core/core/retrieval/cc-turboquant-index
cargo run --release --bin cc-turboquant-index -- benchmark \
  --vectors 4096 \
  --dim 768 \
  --queries 32 \
  --k 10 \
  --bits 4

Persist and inspect a packed snapshot:

bash
mkdir -p Desktop/Comp-Core/benchmarks/agp-turboquant-ane/snapshots

cargo run --release --bin cc-turboquant-index -- benchmark \
  --vectors 4096 \
  --dim 768 \
  --queries 32 \
  --k 10 \
  --bits 4 \
  --snapshot Desktop/Comp-Core/benchmarks/agp-turboquant-ane/snapshots/cc-turboquant-index-4096x768-b4.tqidx

cargo run --release --bin cc-turboquant-index -- inspect \
  Desktop/Comp-Core/benchmarks/agp-turboquant-ane/snapshots/cc-turboquant-index-4096x768-b4.tqidx

Benchmark exported Orbit/RAG++ embeddings:

bash
cd Desktop/Comp-Core/core/retrieval/cc-turboquant-index
cargo run --release --bin cc-turboquant-index -- benchmark-jsonl \
  --input Desktop/orbit-rag-embeddings.jsonl \
  --limit 10000 \
  --queries 64 \
  --k 10 \
  --bits 4 \
  --snapshot Desktop/orbit-rag-embeddings-b4.tqidx

The JSONL importer accepts raw vector arrays or object rows with `embedding`, `vector`, `values`, or `embedding_vector` fields.

What This Measures

  • TurboQuant embedding search quality and memory footprint.
  • Rust packed-code TurboQuant candidate generation over bit-packed corpus rows.
  • Rust packed snapshot persistence and metadata inspection.
  • JSONL ingestion for exported Orbit/RAG++ embedding slices.
  • TurboQuant activation packet compression for AGP-PTP transport.
  • Local ANE private MIL bridge availability.
  • Minimal ANE route-head compile/eval status through `ane_route_head_bench.py`.

What This Does Not Claim

  • It does not prove final retrieval latency; the current Python path pre-dequantizes rotated vectors to fp16.
  • The Rust sidecar v0 is packed-code scalar scan, not the final hand-tuned Apple Silicon SIMD kernel.
  • It does not claim ANE trains the full transformer.
  • It does not claim ANE route-head execution is currently stable; the current local bridge compiles but eval fails with Apple ANE status `0x2/statusType=0x9`.
  • It does not replace MLX/GPU as the stable training path.

Architecture Target

The target path is:

text
AGP hidden/semantic state
  -> CompressionBottleneck
  -> TurboQuant packet
  -> Thunderbolt transfer
  -> receiver-side TransferAdapter
  -> Rust admissibility / Graph Kernel provenance

The ANE target path is:

text
hidden state
  -> route / vitality / semantic projection head
  -> Core ML or private MIL ANE sidecar
  -> routing decision
  -> GPU only if deeper work is justified

Promotion Decision

Attach run IDs, datasets, metrics, and reproduction commands.

Source Anchor

Comp-Core/benchmarks/agp-turboquant-ane/README.md

Detected Structure

Method · Evaluation · Code Anchors · Architecture