29 Commits
Author SHA1 Message Date
John Smith 2d4c3d2211 feat: add kingjones30/ROCmFPX build workflow and documentation 2026-09-05 15:17:42 -04:00
John Smith 01b82f790c feat(q38rocm): add Vulkan RADV Wave64 detection, tuned MTP K=4, and workload profiles 2026-09-02 23:58:42 -04:00
John Smith 4914630519 fix(yarn): remove unsupported --checkpoint-min-step flag for q38rocm binary 2026-09-02 23:45:52 -04:00
John Smith 235e4aab78 fix(yarn): optimize hybrid recurrent checkpoints to prevent cache drop and forced re-prefill 2026-09-02 23:43:02 -04:00
John Smith 1ffd4e8256 fix(yarn): resolve prompt cache invalidation and optimize prefill in run_yarn_1m 2026-09-02 23:30:09 -04:00
John SmithandClaude dc0ad7d584 fix(scripts): auto-detect the Radeon 8060S as the ROCm target in run_yarn_1m
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]>
2026-08-31 00:39:44 -04:00
John SmithandClaude c291aa74ab fix(scripts): make 1M YaRN context actually work on Qwen 3.8 27B (qwen35)
The server silently capped the slot at the model's native 262K training
context because --override-kv targeted qwen2/qwen2vl/qwen3 arch keys while
the GGUF architecture is qwen35 (context_length=262144). Fix the override
to qwen35.context_length and align the launcher with the upstream
(q38rocm / ROCmFPX, pinned engine build 244) validated recipe:

- --override-kv qwen35.context_length=int:1048576 (removes the cap)
- TurboQuant KV (-ctk q8_0 -ctv turbo4) + matching draft KV
- MTP speculative decode (--spec-type draft-mtp, n_max 6 / p_min 0.6)
- RAM context checkpoints sized for 1M on 128 GB (64 @ 32768 MiB)
- Strix Halo env inlined, -dev ROCm0, HIP_VISIBLE_DEVICES=0
- --no-context-shift, --no-mmap, -fit off, -t 16 -tb 32, --metrics

Co-Authored-By: Claude <[email protected]>
2026-08-31 00:24:57 -04:00
John Smith 6f15cc25e7 fix(yarn): add override-kv and fit off to prevent 262k slot capping on 1M YaRN 2026-08-30 23:25:46 -04:00
John Smith 7a4b4ae87e feat: add generic 1M YaRN launch script and documentation for Qwen3.8-27B 2026-08-30 23:21:34 -04:00
John Smith a304089dd7 fix(workflows): improve tap dispatch token handling and executable permissions 2026-08-30 23:00:24 -04:00
John Smith cc35daccb4 fix(ci): disable WebUI asset download with -DLLAMA_BUILD_WEBUI=OFF 2026-08-30 21:03:18 -04:00
John Smith e4ae52ce29 docs: add Homebrew Tap badge to README 2026-08-30 21:02:02 -04:00
John Smith 9ba23c6bbc fix(ci): point upstream ROCmFPX repo to charlie12345/ROCmFPX and support repo override across workflows 2026-08-30 20:32:50 -04:00
John Smith 125f14bee0 fix(ci): use ubuntu-24.04 runner for glslc and spirv-headers support in q38rocm build 2026-08-30 20:25:15 -04:00
John Smith 9802d92e86 feat(ci): add q38rocm build pipeline for AMD Strix Halo gfx1151 2026-08-30 20:19:01 -04:00
John SmithandClaude ba52442ff6 fix(ci): absolute paths for CK clone and Kairic Edge patch
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]>
2026-08-24 12:27:06 -04:00
John SmithandClaude 5c95236e98 feat(ci): on-demand builds for certified profiles (Kairic Edge, PromptForge)
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]>
2026-08-24 12:14:37 -04:00
John SmithandClaude 7fe1b06230 feat(ci): build multiarch target in every nightly release
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]>
2026-08-24 11:54:39 -04:00
John SmithandClaude e3be2d088e fix(ci): enable dotglob when zipping release archives
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]>
2026-08-23 00:14:18 -04:00
John SmithandClaude 00c502cd10 fix(ci): stage kpacks visibly, restore .kpack geometry at release time
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]>
2026-08-22 23:38:40 -04:00
John SmithandClaude ee35f5af79 fix(ci): explicitly match .kpack in upload-artifact paths
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]>
2026-08-22 23:11:27 -04:00
John SmithandClaude 94d718231b fix(ci): multiarch bundling size and rocm_kpack geometry
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]>
2026-08-22 22:34:24 -04:00
John SmithandClaude 4d10fdd235 fix(ci): filter multiarch from Windows matrix in prepare-matrix
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]>
2026-08-22 21:41:04 -04:00
John SmithandClaude cd9e413e25 feat(ci): add multiarch target for combined gfx1100+gfx1151 builds
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]>
2026-08-22 21:40:08 -04:00
John Smith e9fe662041 fix(ci): avoid pipe termination error in PowerShell binary verification 2026-08-21 15:30:21 -04:00
John Smith 8a8ba95e4e fix(ci): add in-job binary verification, unblock release creation, and add timeouts to self-hosted tests 2026-08-21 12:58:17 -04:00
John Smith fa9a59158c docs: add Homebrew quickstart and tap trigger dispatch 2026-08-20 23:55:43 -04:00
John Smith a062f04e1e docs: add AGENTS.md, CLAUDE.md, GEMINI.md, optimize README for SEO, and polish comments 2026-08-20 19:00:31 -04:00
John Smith ab77426bd6 Initial commit: ROCmFPX automated multi-architecture build pipeline 2026-08-20 18:43:21 -04:00