fix(docker): build the ROCm image on ROCm 7.2.4

[davide221]

What

The :rocm Docker image builds on ROCm 7.2.4 instead of 6.4.1, and one DS4 indexer call moves inside its rocWMMA guard.

Why

• The image was the only build left on ROCm 6.4.1. The self-hosted AMD CI runner compiles with /opt/rocm-7.2.4, the Lucebox boxes run 7.2.4, and the README lists ROCm 7.2 for both supported AMD GPUs (R9700 and Strix Halo).
• Every main push since 10-07 failed the ROCm image build. The last green one was 48f49bb1. Two changes compile on 7.2 but not on 6.4.1:
  • mmid.cu:109 calls __syncwarp() (feat(qwen4exp): Qwen3.8-Flash-Next backend — parity, prefill/decode perf #774). ROCm 6 declares the warp-sync builtins only on opt-in; ROCm 7 declares them by default.
  • ds4-indexer.cu:986, the tokenwise path (feat(qwen4exp): split Qwen3.8-Flash-Next across an R9700 and a Strix Halo #827), launched ds4_indexer_score_decode_wmma_kernel outside #if DS4_INDEXER_WMMA_AVAILABLE. ROCm 6.4 ships rocWMMA 1.x, so the kernel isn't compiled there.
• PR checks never saw it. The Docker workflow runs on PRs only when Docker files change, and the AMD runner already builds with 7.2.4.
• A ROCm 7.x host already needed 7.x userspace. The Dockerfile notes that a 6.4.x userspace segfaults at model load against a 7.x host driver.

Changes

• Dockerfile.rocm, docker-bake.hcl: ROCM_VERSION 6.4.1 → 7.2.4, on Ubuntu 24.04 (base rocm/dev-ubuntu-24.04:7.2.4). ROCm 7.2's CMake package files use block(), which needs CMake 3.25; Ubuntu 22.04 ships 3.22, and the 22.04 image stopped at configure. 24.04 has a Python 3.12 of its own, so UV_PYTHON_PREFERENCE=only-managed keeps uv's managed build under /opt/uv, as in the CUDA image (without it the final chmod of /opt/uv failed). The notes say 7.2.4 is the tested stack and that the engine no longer builds on 6.x.
• ds4-indexer.cu: the tokenwise call is inside the existing rocWMMA 2.x guard, so a source build without rocWMMA 2.x compiles again. Every score then takes the scalar kernel, one token each, as the rest of that build already does.
• hip_compat/cuda_runtime.h: a comment said the CI runner was on ROCm 6.4.x.

Tested

• Full main-config image: LUCE_HIP_ARCHES="gfx1151;gfx1100;gfx1200;gfx1201" docker buildx bake rocm-local on a Strix Halo + R9700 box. Configure, all 424 build steps and the runtime stage pass; the image is 26.6 GB.
• Image smoke test on that box (host ROCm 7.2 driver):
  • /opt/rocm/.info/version in the image is 7.2.4;
  • test_server_unit: 658 passed, 2 skipped;
  • luce_server --list-devices sees the R9700 (gfx1201) and the Strix Halo iGPU (gfx1151);
  • Qwen3-0.6B (BF16), served on each GPU, answers a chat request: "Three prime numbers are 2, 3, and 5."
• CI: this PR touches Dockerfile.rocm, so the Docker workflow builds the ROCm image (gfx1201) here, and the AMD GPU jobs build the indexer change.

Not tested

• A host still on a ROCm 6.x kernel driver running the 7.2.4 image.
• gfx1100 and gfx1200 cards: their code objects build, but there's no such card to run them on.

🤖 Generated with Claude Code

[davide221]

Code review

No issues found. Checked for bugs and logic errors in the changed code.

_{Reviewed ac5477ca · 0 candidate issues checked, kept at ≥70% confidence · add the code-review label to review again}