Dense GEMM shape census

Kernel selection starts from an observed model workload. A microbenchmark shape alone is not sufficient.

The first census targets the paired Mistral.rs Qwen2.5-1.5B proof of concept. The engine records each resolved unquantized dense-linear call before backend selection. In the current Qwen path, these calls later use Mistral.rs GEMV or Candle CUDA matmul.

The record is a call census, not a kernel trace. host_calls counts successful linear dispatches. It does not count CUDA kernels, Graph replays, retries, or warmup work. The capture must label prefill and decode from engine metadata. It must not infer the phase from M.

Version 1 rejects runs that enable CUDA Graphs. A Graph replay does not return through the Rust linear dispatch hook, so a host-call census would undercount effective execution frequency.

Version 1 also admits one request and one unchunked prefill forward. The decode_forward_steps field counts successful model decode forwards. It does not count generated tokens across a batch.

Request identity

The request digest covers the UTF-8 bytes of canonical_request. The producer uses sorted keys, no insignificant whitespace, and unescaped UTF-8 characters.

The object contains ordered role/content messages and the generation settings that affect the fixed smoke run.

Record contract

The schema requires:

• a clean producer source and immutable source hashes
• exact model, hardware, toolchain, and workload identity
• local tensor shapes and strides after tensor-parallel sharding
• activation and weight dtypes, layouts, transpose flags, and post-ops
• the observed engine path and exact integer call count
• complete, unsampled, untimed capture

Run the local validator and summary tool:

python3 tools/gemm/shape_census.py validate run.jsonl
python3 tools/gemm/shape_census.py summarize run.jsonl > summary.json

The summary ranks shapes by call count, then FLOPs. It reports both call and FLOP coverage because a frequent small-M launch and a large prefill GEMM have different optimization value.

The tool does not choose an algorithm. A reviewer uses the validated profile to freeze the first Oxide admission contract. The first provider candidate must then pass correctness, lifecycle, sanitizer, Graph, matched operator, and real model gates.

For Mistral.rs decode, M <= 8 and even K currently select its custom CUDA GEMV before cuBLASLt. An Oxide small-M algorithm must beat both that engine path and the cuBLASLt baseline. Winning only the vendor microbenchmark is not enough for engine admission.

Do not combine records when source, model, workload, hardware, capture boundary, dtype, layout, transpose, or post-op metadata differs. Do not present a single-request census as a production traffic distribution.

The shape census profiles the engine baseline. It does not compile or execute an Oxide kernel. Record cuda-oxide source, backend, and codegen provenance in the later Oxide provider benchmark, not in this host-dispatch record.