Audit: doc freshness + one-pass-success fixes from the W7900 verification - #16
Merged
Conversation
…tion Documentation (verified against the 2026-08-16 three-ROCm-pass data): - handoff.md still said 'the Radeon dGPU track remains pending' — stale since PR #6 landed; updated with the W7900 track status and evidence. - docs/results/README.md W7900 section described only Study 2 / one pass; now states Study 1+2 and the three labeled passes (7.2.4 submission, 7.2.1 reproduction, full 7.14.0 matrix) with cell-bundle links. - README gfx-vs-APU ratio recomputed from the 7.14 numbers now shown (~6.4-6.5x at c=16, was '~6.6x' carried over from the 7.2.4 tables). - w7900-gfx1100.md preset VRAM note labeled with its ROCm version (7.2.4). - docs/getting-started.md: upfront redirect for non-gfx1151 users to the validated W7900 path and the gfx110X-all requirement. - troubleshooting.md: two new entries distilled from this verification — rocblas-wrong-arch-tarball (gfx1151 tarball core-dumps multi-slot on W7900; c=1 masks it) and orphan-server-contaminates-bench (stale llama-server poisons cells; clean-before/after ritual). One-pass success (scripts): - gguf-quickstart.sh: fail fast when rocminfo shows a non-gfx1151 GPU, instead of building a wrong-arch llama.cpp for ~15 min and failing at server start; empty rocminfo output is left to 00-check-env.sh. - install-rocm-7.14.sh: non-fatal 10 s warning when the host GPU is detectably not gfx1151 (the archive is gfx1151-only); skipped for the ROCM714_MANIFEST test seam. Verified firing on a real W7900 host. - gguf-bench-cell.sh: check 'uv' before launching anything — a missing uv previously surfaced as 'command not found' after the model loaded. - w7900-repro/_repro_driver.sh: provenance pass skips unreadable (truncated-by-interrupt) cell JSONs with a notice instead of crashing, so RESUME always works. - w7900-repro/README.md: resume/cleanliness ritual + gfx110X-all pointer. All gates green: 180 passed / 1 skipped, claim-consistency ok, shellcheck (error-level) and bash -n clean. grep-under-pipefail regression caught by the UX tests and fixed (|| true on the substitution pipelines).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Systematic audit after the W7900 verification passes (PR #6 + 7.2.1/7.14.0 reproductions), in two dimensions:
1. Documentation freshness — every public doc cross-checked against the measured data
handoff.mdclaimed the dGPU track "remains pending"rocblas-wrong-arch-tarballandorphan-server-contaminates-bench2. One-pass success — every user-facing script traced along its documented path
gguf-quickstart.shhardcodesAMDGPU_TARGET=gfx1151: non-gfx1151 users paid a ~15-min wrong-arch build before an opaque HIP failureinstall-rocm-7.14.shsilently installs a stack that core-dumps multi-slot decode on W7900 (rocBLAS has no gfx1100 kernels; c=1 masks it)gguf-bench-cell.shreports a missinguvonly after the model loaded_repro_driver.shprovenance pass crashes on a cell truncated by an interrupted run, breaking RESUMEVerification
check_claim_consistencyok, bash -n + shellcheck (error level) clean.set -eo pipefailsilently aborting the script) — fixed and kept as coverage.