README.md

act-feather

FEATHER specific patches for ACT-generated compiler backend.

Docker

./docker/build.sh builds the image.

./docker/run.sh runs a ephemeral container with the current repository mounted inside.

Run Q-sliced Attention

The XLA-HLO IR for Q-sliced attention is at ./attention_q_slice.hlo.

To run the compiler,

cd act-backend
cargo build --release
./target/release/act-feather --input ../attention_q_sliced.hlo --log ./log/

Understanding the log

You should see a structure like:

log/
	metadata.json
	pii_0/
		0.pii
		variant_stats.csv
		layout_variants/
			pii_0_v0.json
			...
		layout_mapper_variants/
			mapped_pii_0_v0.json
			...
		minisa_traces/
			final_assembly_v0.txt
			...
			final_assembly.txt

File meanings:

• metadata.json: HLO-level metadata dumped by the frontend/parser stage.
• pii_0/0.pii: extracted PII graph for this candidate.
• pii_0/layout_variants/*.json: layout variants generated before mapper execution.
• pii_0/layout_mapper_variants/mapped_*.json: mapper output JSON per variant (with tile mapping and latency).
• pii_0/minisa_traces/final_assembly_v*.txt: emitted instruction traces per variant.
• pii_0/minisa_traces/final_assembly.txt: best variant trace (minimum total latency, earliest-index tie-break).
• pii_0/variant_stats.csv: per-layer CSV rows written after each variant mapper run.

variant_stats.csv columns:

• variant: variant index.
• layer: layer name (for example L1, L3).
• latency: per-layer latency from mapper stdout.
• execute_tiles: number of ExecuteTile instructions for the layer.
• M,K,N: GEMM dimensions parsed from mapper stdout.
• overall_utilization,compute_utilization: utilization percentages parsed from mapper stdout.