[BANADDA]

Part of #152, after #153.

GGUF for mt-1 and mt-2

• artifact/gguf.py reads GGUF v2 and v3 headers: metadata, tensor shapes and types, parameter count (a tied output head that llama.cpp stores twice counts once), file type, and the KV cache size at a context.
• An SLM repository may now be one *.gguf with README.md and LICENSE.
• check_slm_repo refuses any tensor below 4 bit (Q2_K, Q3_K, IQ1/2/3, TQ) and a file type outside the arena's precision list.
• Parameter bounds come from the arena's budgets instead of a fixed 1B to 8B. The contract floor is now 100M so mt-1 bases fit.

CPU runner

• When the definition's envelope is on CPU, the SLM runs in the llama.cpp server image with --cpus, --cpuset-cpus 0..n-1 and --memory set from the group's reference: 4 threads and 2 GB for mt-1, 8 threads and 4 GB for mt-2. Context is max input plus max output.
• Peak memory is the container's cgroup memory.peak, with the GGUF footprint as fallback.
• The SLM phase runs one submission at a time on CPU groups so measurements do not share cores.
• The epoch refuses to run when the host architecture differs from the group's (arm64 for mt-1, x86_64 for mt-2).
• Token counts fall back to counting streamed pieces when a server sends no usage.

mt envelope check --arena <id> --slm <dir> [--items dev.jsonl]

• Reads the arena's resolved budgets from the server.
• Starts the same container validators use, vLLM on one GPU for GPU groups or llama.cpp at the group's caps, published on 127.0.0.1.
• Times prompts at the arena's full input length: dev items rendered with the arena prompt, or filler.
• Prints parameters, memory, first token p95, and per token p95 or tokens per second against the limits. It exits 1 on any fail.

Provenance

• The fingerprint dequantises AWQ (AutoAWQ GEMM nibble order) and GPTQ (with g_idx and the zero offset) attention weights.
• It reads GGUF through the gguf package, mapping blk.N.attn_q/k/v/output to the HF names. Row permutation in GGUF conversion leaves singular values unchanged.
• Weights load through torch and convert to float32, so bf16 bases still work.
• gguf joins the validator extra.

Testing. Local tests, not committed:

• AWQ and GPTQ round trips reproduce the spectrum (cosine above 0.999, and both agree element-wise).
• A synthetic GGUF header counts a tied head once and flags Q3_K.
• The CPU container pins threads and memory.

I also parsed a real header (Qwen2.5 0.5B Q4_K_M): version 3, file type Q4_K_M, no tensors below 4 bit. Full local suite, newest ruff and mypy are clean, apart from the Windows only resource findings already on main.

Not in this PR: lower group (edge) tests, which no arena can use yet, and pinning the llama.cpp image tag after the benchmark.

🤖 Generated with Claude Code