Repository navigation
docs: mark ROCm as supported and refresh the GB10 benchmark record - #2860
Merged
Merged
Conversation
AMD ROCm has been an officially supported backend since 0.8.0 on Linux and on Windows x86_64, but the README, installation matrix, architecture overview, environment-variable reference and contributing guide still called it experimental. Update the support status and rename the installation headings, so links move from the `-experimental` anchors to #linux-with-amd-rocm and #windows-with-amd-rocm. The validated configurations, the source-build requirement and the open gaps are unchanged. Add an Unreleased changelog entry for the status change. Historical release notes and dated benchmark reports keep the wording they shipped with. Refresh the benchmark surface for the 0.8.1 GB10 sweep: the README performance section now records the version-over-version decode ratios, the two fixes that unblocked BitNet and DiffusionGemma, and the first embedding, rerank and image generation ladders on that host. The methodology page points at the GB10 records, says GB10 is current instead of carrying the 0.4.0-rc.1 record, and documents the bench_imagegen.py command. Both pages keep the prefill caveat: the harness changed shape, so only decode is comparable with the older sweep. Verification: every ROCm anchor resolves to a renamed heading, and no current-facing document still describes ROCm as experimental. The GB10 figures come from the committed 2026-11-03 and 2026-11-04 CSVs.
Windows ROCm is an officially supported backend since 0.8 alongside Linux. The platform matrix, feature table, installation section and overlay README already say so, but the README build-from-source section, the docs index, the architecture overview and the contributing prerequisites still described the platform as Linux-only or left the Windows build undocumented. Add the Windows ROCm build command and link to the README, list the platform in the docs index, describe the delay-loaded runtime directory in the architecture overview, and add the Windows CUDA and ROCm prerequisites to the contributing guide. Windows NVIDIA CUDA remains experimental, and no code or validation scope changes.
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
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
AMD ROCm has been an officially supported backend since 0.8.0, but the user-facing docs still described it as experimental. This updates the support status across the README, installation matrix, architecture overview, environment-variable reference, contributing guide and the ROCm overlay README, and refreshes the benchmark surface for the 0.8.1 GB10 sweep.
Changes
## Linux with AMD ROCmand## Windows with AMD ROCm, and update every anchor link.docs/benchmarks.md: point at the GB10 records, correct the cross-host version paragraph, and documentscripts/bench_imagegen.py.Testing
cuda_gb10_*_2026-11-03and2026-11-04CSVs.Notes
Prefill is not compared across the two GB10 sweeps because the harness moved to the 512-token synthetic prompt. Only decode is compared.