ci: linux-x64-rocm (HIP) backend build AMD support#49
Merged
Conversation
Promote dev to master: sched split-input fix for Gemma 4 E-series
M1 of AMD support. Adds a ROCm build job to the dev pipeline modeled on upstream's ubuntu-22-rocm and our linux-x64-cuda job: installs ROCm 7.2.1 from repo.radeon.com, builds with GGML_HIP + rocWMMA flash-attn for RDNA targets only (gfx1030/1100/1101/1102/1151/1200/1201 — GCN/CDNA excluded by scope, those users use Vulkan), bundles libggml-hip.so plus the HIP runtime and rocBLAS/hipBLASLt Tensile data so the archive is self-contained. Build-only: no AMD GPU on the runner, the backend is a dlopen'd plugin (GGML_BACKEND_DL). Also adds the pre-existing linux-x64-cuda-13.3 to the dev-latest EXPECTED list (was missing). Kernels are already RDNA-aware (mmvq-tq.cu). Real-hardware validation (test-backend-ops on gfx1151/gfx1100) is the next milestone.
worthant
changed the title
dev-build.yml alone only puts ROCm in dev-latest; add the same job to release-turboquant.yml (needs: verify-version) and to publish-release's needs list so tagged releases carry llama-turboquant-linux-x64-rocm. Same recipe as the dev-build job — RDNA-only gfx targets, HIP runtime + Tensile data bundled.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
First milestone of AMD/ROCm support: a CI job that builds a self-contained
libggml-hip.soarchive. No AMD GPU needed to compile — the backend is a dlopen'd plugin (GGML_BACKEND_DL).What it does
-DGGML_HIP=ON -DGGML_HIP_ROCWMMA_FATTN=ONfor RDNA targets only:gfx1030(RDNA2),gfx1100/1101/1102(RDNA3),gfx1151(RDNA 3.5 / Strix Halo),gfx1200/1201(RDNA4). GCN/CDNA excluded by scope — those users use the Vulkan backend.dev-latestpublish (and fixes theEXPECTEDlist, which was missinglinux-x64-cuda-13.3).Why this is safe / small
The TurboQuant kernels are already RDNA-aware (
mmvq-tq.cuforks dp4a/sudot4). This PR is pure CI — no source changes.Not in this PR (next milestones)
test-backend-opson gfx1151/gfx1100 (rented HostKey RDNA3/4 or a Strix Halo box). This is where wavefront-width and rocWMMA-FA correctness get proven.The PR build here compiles the job (PRs into dev build everything, publish nothing), so a green check = ROCm backend compiles for all RDNA targets. Full reference:
atomic-forge/docs/ROCM-HIP.md.