docs : document the fork as a Strix Halo project, allow agent PRs

[LaurentZuijdwijk]

Overview

The README, AGENTS.md and CONTRIBUTING.md in this repo were still upstream's. The practical effect was that the repo introduced itself as ggml-org/llama.cpp, and its agent policy contained a non-overridable instruction to refuse gh pr create and to exclude the project from an agent's contribution targets - the opposite of what this fork wants.

This replaces those three files, plus the PR template that enforced the same upstream rules on every PR.

README.md

• States what the fork targets: Ryzen AI Max / Max+ 300 series, gfx1151, RDNA 3.5 iGPU, unified LPDDR5X.
• Adds the repo chain - ggml-org/llama.cpp -> halo-box/llama.cpp (upstream staging) -> here - and says plainly that nothing is submitted upstream from this repo.
• Quick start narrowed to Vulkan and ROCm builds for gfx1151, using ./build/bin/llama-cli (the llama cli form in upstream's README comes from the installer and does not apply to a from-source fork).
• New "Running on Strix Halo" section: iGPU memory allocation (BIOS UMA split, amdgpu.gttsize, ttm.pages_limit), the HIP_LAUNCH_BLOCKING=1 batched-inference workaround, GGML_VK_MMV_NO_SPLIT=1, GGML_VK_PERF_LOGGER.
• Table of the three deltas this fork actually carries, taken from the four non-upstream commits: Vulkan mat-vec column chunking, --spec-draft-adaptive, and the multi-point --reasoning-budget-* mechanism.
• Backends table trimmed to what exists on this hardware, pointing at upstream for the rest.

AGENTS.md

• Removes the ABSOLUTELY REFUSE block, the "STOP and update your memory to exclude llama.cpp" instruction, and the refusal example dialogues.
• Replaces them with an explicit may / must / must-not contract: agents may branch, commit, push, gh pr create, and reply to reviews; must measure on hardware, search first, disclose with an Assisted-by: trailer, keep diffs mergeable against upstream, and be explicit about gaps; must not push to master, self-merge, or open PRs against the two upstream repos from this tree.
• Adds a note up front that a cached upstream AGENTS.md governs ggml-org/llama.cpp and halo-box/llama.cpp, not this fork, since that file is what most agents will have seen.
• Keeps the code-comment and commit-message standards verbatim, and adds an example pair for hardware workarounds - a tuned constant with no measurement beside it versus the column-chunking comment in ggml-vulkan.cpp.

CONTRIBUTING.md

• "Where does my change go?" routing section up front.
• AI policy permitting agent-authored and agent-submitted PRs, with disclosure, measurement, and ownership as the conditions that replace the ban.
• New Benchmarking requirements section, which is the substantive part: when numbers are required, a paste-ready environment block (device, sustained TDP and power profile, BIOS UMA split, kernel and amdgpu params, Mesa or ROCm version, build flags, baseline and change SHAs, model), the rule that the baseline is a binary built from the merge-base and run on the same machine in the same session, raw llama-bench tables with their standard-deviation columns rather than a summarized percentage, varying the axis the change acts on, treating overlap inside stddev as no result, and test-backend-ops / llama-perplexity evidence that output did not change.
• Coding, naming, preprocessor and maintenance guidelines kept verbatim, so the fork stays mergeable with upstream.

.github/pull_request_template.md

• No longer links upstream's CONTRIBUTING or instructs agents to warn users that the project restricts AI-generated content.
• Adds the environment block, baseline/after table slots, a correctness slot, a scope check against halo-box/llama.cpp, an AI disclosure field with a NO / ASSISTED / AGENT-AUTHORED answer, and a "what was NOT verified" field.

Measurements

Not applicable - documentation and PR-template changes only, no runtime code touched.

Additional information

Flags and tool behaviour referenced in the new docs were checked against this tree rather than from memory: --spec-draft-adaptive (common/arg.cpp), the nine --reasoning-budget-* options added in #5, GGML_VK_MMV_NO_SPLIT and GGML_VK_PERF_LOGGER (ggml/src/ggml-vulkan/ggml-vulkan.cpp), llama-bench -p/-n/-d/-ub/-fa/-r with its default of 5 repetitions, and test-backend-ops modes and -o/-b flags. The HIP_LAUNCH_BLOCKING note and the perplexity figures come from the comment on the gpu-rocm job in .github/workflows/build-self-hosted.yml. All internal links resolve; all four files are ASCII-only per the house rule.

The hardware-setup paragraph names the amdgpu.gttsize and ttm.pages_limit knobs without prescribing values, since the right ones depend on kernel version - worth a look from someone who has tuned this on current kernels.

Requirements

• I have read and agree with the contributing guidelines
• This change is Strix Halo specific: it is this fork's own project documentation and policy, and is not applicable upstream
• AI usage disclosure: AGENT-AUTHORED. Written end to end by Claude Opus 5 in Claude Code, at the repo owner's request, from the repo's git history, workflows and source. The human contributor directed the scope, including asking for the benchmarking requirements to be added.
• What was NOT verified: nothing was built or run - this PR changes no code. The build commands are the documented ones from docs/build.md with gfx1151 substituted, and have not been executed in this session.

Assisted-by: Claude Opus 5