Conversation
New windows-x64-rocm job in dev-build and release: HIP backend built with the clang from the ROCm 10.0 wheels (TheRock), consumer gfx targets only (RDNA2-RDNA4, Ryzen AI 300, Strix Halo), --offload-compress. The publish job glues the CPU archive with ggml-hip.dll, the HIP runtime and rocBLAS/hipBLAS plus gfx-filtered Tensile kernels, so the zip loads with only an Adrenalin driver (upstream stopped bundling rocBLAS and its zip needs a ROCm SDK, ggml-org#26996). ROCm CI hygiene: drop the dead GGML_HIP_ROCWMMA_FATTN flag (removed upstream), add gfx950/gfx1150/gfx1103 to the Linux targets, fix the stale CDNA comment, update the archive tables. docs/amd: status of HIP in the fork, ROCm vs Vulkan bench matrix for TurboQuant, roadmap. scripts/bench-amd.sh + bench-amd-report.py run the matrix per backend build and render one table. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…iagnostics Temporary workflow_dispatch-only copy of the windows-x64-rocm job: prints where the ROCm 10 wheels keep the rocBLAS kernels, checks that --offload-compress reaches clang, silences -Wignored-attributes, and makes the runtime bundling non-fatal. To be deleted before merge. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…l stores Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The ROCm 10 Windows wheels ship no rocBLAS kernel library; rocblas.dll imports libhipblaslt.dll and kernels come from amd_comgr at run time. Walk the DLL imports with dumpbin instead of looking for a library dir. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
- --offload-compress goes through CMAKE_CXX_FLAGS: on Windows ggml-hip compiles the .cu files as CXX via hip::device, so CMAKE_HIP_FLAGS is ignored. ggml-hip.dll drops from 545 MB to 47 MiB for 9 targets. - Bundle the import closure of ggml-hip.dll (amdhip64_7, rocm_kpack, amd_comgr, hipblas, rocblas, libhipblaslt, rocsolver, origami; 230 MiB, 92 MiB zipped) instead of a rocBLAS kernel library: the ROCm 10 wheels ship none, kernels come from amd_comgr at run time. - -Wno-ignored-attributes: 850k GGML_API dllimport warnings made the job log unfetchable. - Drop the scratch workflow used for iteration. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
ci: windows-x64-rocm archive + AMD ROCm bench plan
Sign hip-out after bundling, the same rule as the CUDA archives: the code-sign action skips DLLs that already carry a valid signature, so AMD-signed DLLs stay AMD's and unsigned ones get ours. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
github-actions
Bot
added
documentation
ci(windows-rocm): sign the whole HIP payload + 1.6.0 changelog
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.
Promote dev to master for b10269-1.6.0. What lands:
llama-turboquant-windows-x64-rocm.zip(self-contained, signed, beta until run on a Radeon)Same commits already built on dev; dev-latest 2026-09-09 carries the new archive.
🤖 Generated with Claude Code