HIP_VISIBLE_DEVICES=0 defaulted the server onto the RX 7900 XTX (gfx1100)
instead of the 8060S iGPU (gfx1151). The launcher's HSA_OVERRIDE_GFX_VERSION
=11.5.1 forces gfx1151 ISA, which fails to launch on the 7900 XTX at the
first compute ("ROCm error: unspecified launch failure" in ggml_cuda_cpy /
hipMemcpyAsync; "current device: -1" because cudaGetDevice itself fails,
ggml-cuda.cu:129). The 24 GB card also can't hold 1M KV + MTP + mmproj.
- Auto-detect the 8060S by name from `--list-devices` (common/arg.cpp:1101)
and set HIP_VISIBLE_DEVICES to it; respect an explicit env pin / --hip-device
- Loud fallback warning to HIP_VISIBLE_DEVICES=0 if detection fails
- Add optional --strict (--spec-mtp-strict-qwen) for greedy==no-spec output
- Banner/help updated; mmproj auto-attach unchanged
Co-Authored-By: Claude <[email protected]>
The prep step ran in ROCmFPX/ but resolved the patch path relative to
the workspace parent; use absolute GITHUB_WORKSPACE paths throughout.
Co-Authored-By: Claude <[email protected]>
Adds build-rocmfpx-profile.yml: workflow_dispatch builds of Ciru's
frozen model-specific release branches with their certified pins —
- profile: kairic-edge (release/kairic-edge-qwen38-27b-v1.2) or
promptforge (qwen3.8-activefpx-promptforge-v2.3)
- gfx_target: any family; default gfx1151 (the only certified target;
others run the branch's guarded fallback routes)
- pinned TheRock 7.15.0a20260718 toolchain for all families (verified
present on the nightlies index)
- Composable Kernel fdf4bb7f + Kairic Edge IU4 patch (sha256-checked)
- certified CMake flag set from Ciru's build recipes
- release notes carry the per-profile sidecar/model contract and the
gfx1151-only qualification warning
Co-Authored-By: Claude <[email protected]>
Adds multiarch to the default gfx_target list so scheduled nightlies and
default dispatches produce the combined gfx1100+gfx1151 artifact. This
also lets the tap's auto-bump track it (matching asset exists in each
new release) instead of relying on manual pins.
Co-Authored-By: Claude <[email protected]>
The kpacks -> .kpack rename succeeded, but the shell glob in
'zip -r ... *' skipped the newly hidden directory. dotglob makes '*'
match it.
Co-Authored-By: Claude <[email protected]>
GitHub expressions cannot return a sequence from an inline conditional,
and upload-artifact excludes dot-paths in both plain and glob paths.
Stage packs as dist/kpacks/, rename to .kpack/ just before zipping.
Co-Authored-By: Claude <[email protected]>
Globs skip dot-paths (dot:false), so ** alone still excluded the kernel
packs; pass dist/bin and dist/.kpack as explicit patterns.
Co-Authored-By: Claude <[email protected]>
The first multiarch run failed release upload: the zip exceeded GitHub's
2 GiB asset limit because the wholesale hipblaslt/rocblas library copies
grabbed every architecture's kernels (~3 GB) from the all-arch dist.
Also, rocm_kpack resolves each host library's embedded
"../.kpack/<name>_@[email protected]" relative to the library itself, so
kernel packs must ship as a sibling of bin/ (TheRock's install geometry),
not inside it — and upload-artifact drops hidden directories from plain
directory paths, so the staged tree is uploaded via a wildcard pattern.
- Skip hipblaslt/rocblas library trees for multiarch; filtered kpacks
(gfx1100 + gfx1151 only) supply BLAS kernels instead
- Skip compiler-only libLLVM/libclang-cpp for multiarch
- Stage dist/{bin,.kpack} after RPATH patching and upload ROCmFPX/dist/**
Co-Authored-By: Claude <[email protected]>
Job-level if cannot reference the matrix context; exclude the Linux-only
target when building the Windows matrix instead.
Co-Authored-By: Claude <[email protected]>
Use TheRock's therock-dist-linux-multiarch dist (all-arch kernel packs in
.kpack/) to produce a single Ubuntu artifact with HIP code objects for
both gfx1100 and gfx1151. Windows job skips the Linux-only target.
Co-Authored-By: Claude <[email protected]>
John Smith