fix(docker): build native binaries for every arch and un-pin the base image

[maple52046]

Why

Built as-is, the image is tied to a single GPU architecture and to one specific torch base. Three of those are correctness problems that are independent of which base you use.

The three generic fixes

hip-stream was built with no offload arch at all. vendor/babelstream-hip/HIP.make never passes --offload-arch, so the target is left to hipcc's detection — and there is no GPU to detect during docker build. It does append EXTRA_FLAGS, so the archs now go in as one repeated flag each.

p2pBandwidthLatencyTest was deliberately limited to the first arch. On a host running any other arch from PYTORCH_ROCM_ARCH the binary just does not run. The previous comment attributed this to the Makefile interpolating GPU_TARGETS into a single --offload-arch, which is true, but the actual failure with a ;-separated value is that the shell splits the command in two:

hipcc ... --offload-arch=gfx942;gfx950 -I/opt/rocm/include ... -o p2pBandwidthLatencyTest p2p...cpp
ld.lld: error: undefined symbol: main

Everything after ; becomes a second command, so hipcc gets one arch and no source file — hence the confusing undefined symbol: main. The Makefile appends CPPFLAGS verbatim, so passing the first arch through GPU_TARGETS and the remaining ones as extra --offload-arch flags yields a genuine fat binary with no Makefile or source change.

ENV PYTHONPATH="/opt/aiter:${PYTHONPATH}" leaves a trailing colon when the base image does not set PYTHONPATH. An empty PYTHONPATH entry means "current directory", so any directory containing e.g. ./cmake shadows the installed module. Now only appended when there is something to append: /opt/aiter on a base without it, /opt/aiter:<inherited> on a base with it (both verified).

Net effect: all three native binaries cover every arch in PYTORCH_ROCM_ARCH, so one image runs on gfx942 and gfx950 and nothing is compiled at run time.

Base image becomes an ARG

BASE_IMAGE now defaults to the combination this was verified against: ROCm 7.2.4 / Python 3.13 / PyTorch 2.12.0 (rocm/ali-private:ubuntu22.04_rocm7.2.4_cp313_torch2.12.0_20260726). Labels and ROCM_VERSION were updated to match that default.

The two base-specific steps are made self-checking rather than deleted, so the file still works on the previous base:

• apt remove -y amdgpu-install-internal now runs only when that package is actually installed. It is legitimate for bases that carry the internal repo, but apt remove on a missing package exits non-zero and fails the build on bases that do not.
• cmake is reinstalled only when the base image's cmake is broken. Some torch bases install the pip cmake package under one interpreter while /usr/local/bin/cmake runs another, so from cmake import cmake resolves to the /opt/rocm/lib/cmake namespace directory and raises cannot import name 'cmake' from 'cmake'. TransferBench and rccl-tests both need cmake.

hipBLASLt client packages

The GEMM benchmarks need hipblaslt-bench, and the public radeon repo does not publish hipblaslt-benchmarks — it only ever existed in the offline bundle. The .debs are still bind-mounted from the build context, but now from a pkgs/ subdirectory so that COPY . /opt/benchtestkit does not drag them into the image. Defaults point at the 1.2.2 client packages built from the matching rocm-7.2.4 tag; both are ARGs.

vLLM

Unchanged — still intentionally absent, per this repo's existing rationale. paged_attention uses the aiter backend; the commented-out opt-in block is untouched. This is now also stated in the image description label.

Verification

Built and run on an MI355X (gfx950) host from the new default base.

benchtestkit check-env in the resulting image:

[ok] mpirun: /opt/mpich/bin/mpirun
[ok] python3: /usr/bin/python3
[ok] hip-stream: /usr/local/bin/hip-stream
[ok] p2pBandwidthLatencyTest: /usr/local/bin/p2pBandwidthLatencyTest
[ok] TransferBench: /usr/local/bin/TransferBench
[ok] hipblaslt-bench: /opt/rocm-7.2.4/bin/hipblaslt-bench

roc-obj-ls on each native binary reports gfx942 gfx950 — all three are fat binaries.

Real workloads on the gfx950 host:

• hipblaslt-bench -m 1024 -n 1024 -k 1024 → 55924 Gflops (f16)
• hip-stream → Triad 3031719 MB/s (~3.0 TB/s)
• import aiter succeeds and JIT-compiles module_aiter_core in 9.0s; is_flydsl_available() is True

The framework stack is untouched by this image: torch 2.12.0+rocm7.2.4.gitcf5ea6e, triton 3.7.1, numpy 2.5.1, flash_attn 2.8.4 — identical to the base image. The only Python package added is pydantic 2.13.4 (benchtestkit's own dependency) plus its transitive deps; nothing pulls or downgrades torch or numpy.

Note for reviewers

BuildKit emits an UndefinedVar: '$PYTHONPATH' lint warning on bases that do not set the variable. That is exactly the case ${var:+...} exists to handle. Do not silence it by declaring ARG PYTHONPATH — an ARG shadows the inherited ENV and the append is lost (verified).