Skip to content

rocm: update Strix Halo inference stack to 7.15#148

Open
georgewhewell wants to merge 19 commits into
masterfrom
codex/rocm-7.15-vllm-0.25.1
Open

rocm: update Strix Halo inference stack to 7.15#148
georgewhewell wants to merge 19 commits into
masterfrom
codex/rocm-7.15-vllm-0.25.1

Conversation

@georgewhewell

@georgewhewell georgewhewell commented Jul 19, 2026 •
edited
Loading

Copy link
Copy Markdown
Contributor

Summary

  • move both the default gfx1151 TheRock runtime and the optional full source SDK to ROCm 7.15
  • move the inference environment from Python 3.12 to Python 3.13
  • update Torch to 2.11.0, Triton to 3.7.1, and vLLM to 0.25.1
  • build vLLM's native HIP extension against the same source SDK, eliminating the mixed-LLVM closure
  • make the full TheRock profile hermetic, including the profiler submodule graph and offline Dyninst/timemory dependencies
  • keep SGLang 0.5.14 working on cp313, including its Rust outlines-core dependency
  • preserve TheRock's split-library layout and export the complete runtime library path
  • refresh the active nixpkgs, llama.cpp, MLX, FastFlowLM, nix-ai-tools, EC, thunderbolt-ibverbs, and treefmt inputs
  • keep large model fixtures out of the PR gate while retaining runtime and hardware smoke coverage

This supersedes #102, which contains only the intermediate ROCm 7.14 archive update.

Validation

  • full ROCm 7.15 source SDK built successfully as /nix/store/n1pz2h3ndjxr1mh9ii0sdiww7lisjssa-therock-rocm-from-source-gfx1151-full-7.15.drv
    • RCCL 1922/1922, rocprofiler-systems 613/613, composable-kernel 1412/1412, MIOpen 641/641, hipTensor 404/404
    • downstream MIOpen provider and full install/fixup completed
  • source vLLM built successfully as /nix/store/09qzgfxr00agafpfx21x1i3pz53s44vl-python3.13-vllm-0.25.1.drv
    • compiled all 46 C++/HIP targets for gfx1151
    • linked _rocm_C.abi3.so and passed dependency and Python import checks
    • nix why-depends confirms it directly uses the exact source SDK above
    • the previous llvm::createExpandMemCmpLegacyPass LLVM-23 symbol failure does not recur
  • default runtime on a real gfx1151 node
    • vllm --version reports 0.25.1
    • Torch 2.11.0+rocm7.15 detects AMD Radeon 8060S Graphics and executes a GPU tensor operation
    • sglang version reports 0.5.14
  • nix flake check --accept-flake-config --no-build --show-trace
  • Hydra source/meta checks, nixfmt, deadnix, statix, and git diff --check

@georgewhewell
rocm: update Strix Halo inference stack to 7.15
97dec08 
Move the TheRock runtime and Python wheels to ROCm 7.15 on Python 3.13, update vLLM to 0.25.1, keep SGLang compatible, and refresh the active flake inputs.
@georgewhewell georgewhewell mentioned this pull request Jul 19, 2026
Closed
@hellasbot

hellasbot commented Jul 19, 2026 •
edited
Loading

Copy link
Copy Markdown

Hydra reported failures for this PR.

Failed builds:

Hydra reported failures for this PR.

- Eval: #78571
- Jobset: hellas/nix-strix-halo-pr-148
- Head: 77dcc78623c0

Failed builds:
- x86_64-linux.ci.build - dependency failed (log, tail, raw)
~~ - therock-rocm-from-source-gfx1151-full-7.15 - failed (tail, raw)~~
- x86_64-linux.ci.smoke - dependency failed (log, tail, raw)
~~ - llvm-23.0.0-rocm-dev, llvm-23.0.0-rocm-lib, ...and 2 more - failed (tail, raw)~~

Update: hellas/nix-strix-halo-pr-148, eval #78818, head 29c79ff90844: newer evaluation started.

Hydra reported failures for this PR.

- Eval: #78528
- Jobset: hellas/nix-strix-halo-pr-148
- Head: 87a50d049252

Failed builds:
- x86_64-linux.ci.build - dependency failed (log, tail, raw)
~~ - therock-rocm-from-source-gfx1151-full-7.15 - failed (tail, raw)~~
~~ - compiler-rt-src-23.0.0-rocm - aborted (tail, raw) - �[31;1merror:�[0m unexpected end-of-file~~
~~ - clang-src-23.0.0-rocm - aborted (tail, raw) - �[31;1merror:�[0m unexpected end-of-file~~
~~ - lld-src-23.0.0-rocm - aborted (tail, raw) - �[31;1merror:�[0m unexpected end-of-file~~
~~ - compiler-rt-src-23.0.0-rocm - aborted (tail, raw) - �[31;1merror:�[0m unexpected end-of-file~~
- x86_64-linux.ci.smoke - dependency failed cached (log, tail, raw)
~~ - amdsmi-7.15.0 - aborted (tail, raw) - �[31;1merror:�[0m Cannot build '�[35;1m/nix/store/4pmipfr9q1wmbi2bqifs9a6zh25b05r2-amdsmi-7.15.0....~~
~~ - therock-rocm-source-rocm-libraries-projects-hipblas - aborted (tail, raw) - �[31;1merror:�[0m Cannot build '�[35;1m/nix/store/3v096ngkg570g30cz33wk84gppm9r6gg-therock-rocm-s...~~
~~ - amdsmi-7.15.0 - aborted (tail, raw) - �[31;1merror:�[0m write of �[35;1m28672�[0m bytes: �[35;1mBroken pipe�[0m~~
~~ - therock-rocm-source-rocm-systems-projects-rocm-smi-lib - aborted (tail, raw) - �[31;1merror:�[0m Cannot build '�[35;1m/nix/store/crn4rrxs5v8jkpxa58sa8yll875c5dlx-therock-rocm-s...~~
~~ - therock-rocm-source-rocm-libraries-projects-rocthrust - aborted (tail, raw) - �[31;1merror:�[0m unexpected end-of-file~~
~~ - therock-rocm-source-rocm-libraries-projects-rocsparse - aborted (tail, raw) - �[31;1merror:�[0m unexpected end-of-file~~
~~ - lld-src-23.0.0-rocm - aborted (tail, raw) - �[31;1merror:�[0m unexpected end-of-file~~
~~ - lld-src-23.0.0-rocm - aborted (tail, raw) - �[31;1merror:�[0m unexpected end-of-file~~
~~ - ...and 1 more failed steps~~

Update: hellas/nix-strix-halo-pr-148, eval #78571, head 77dcc78623c0: newer evaluation started.

Hydra reported failures for this PR.

- Eval: #78513
- Jobset: hellas/nix-strix-halo-pr-148
- Head: ae685436f1da

Failed builds:
- x86_64-linux.ci.build - dependency failed (log, tail, raw)
~~ - therock-rocm-from-source-gfx1151-full-7.15 - failed (tail, raw)~~

Update: hellas/nix-strix-halo-pr-148, eval #78528, head 87a50d049252: newer evaluation started.

Hydra reported failures for this PR.

- Eval: #78472
- Jobset: hellas/nix-strix-halo-pr-148
- Head: 808b37ef5d11

Failed builds:
- x86_64-linux.ci.build - dependency failed (log, tail, raw)
~~ - amdsmi-7.15.0 - cached failure (tail, raw), propagated from build 11752, log~~
- x86_64-linux.ci.smoke - dependency failed cached (log, tail, raw)
~~ - therock-llvm-project-source-7.15.0 - aborted (tail, raw) - �[31;1merror:�[0m unexpected end-of-file~~
~~ - therock-rocm-source-rocm-systems-projects-roctracer - aborted (tail, raw) - �[31;1merror:�[0m Cannot build '�[35;1m/nix/store/46rqbyqiw0cxddmd2h1fml2syvyqnhzk-therock-rocm-s...~~
~~ - therock-rocm-source-rocm-systems-projects-rocm-smi-lib - aborted (tail, raw) - �[31;1merror:�[0m failed to start SSH connection to '�[35;1mstrix-1.hydra-builders�[0m'~~
~~ - amdsmi-7.15.0 - failed (tail, raw)~~

Update: hellas/nix-strix-halo-pr-148, eval #78513, head ae685436f1da: newer evaluation started.

Hydra reported failures for this PR.

- Eval: #78457
- Jobset: hellas/nix-strix-halo-pr-148
- Head: eecccbdeb9a9

Failed builds:
- x86_64-linux.ci.build - dependency failed (log, tail, raw)
~~ - source - aborted (tail, raw) - �[31;1merror:�[0m �[31;1merror:�[0m~~
~~ … while setting up the build environment...~~
~~ - therock-rocm-from-source-gfx1151-full-7.15 - failed (tail, raw)~~
- x86_64-linux.ci.smoke - dependency failed (log, tail, raw)
~~ - therock-llvm-project-source-7.15.0 - aborted (tail, raw) - �[31;1merror:�[0m unexpected end-of-file~~
~~ - therock-rocm-source-rocm-systems-projects-roctracer - aborted (tail, raw) - �[31;1merror:�[0m Cannot build '�[35;1m/nix/store/46rqbyqiw0cxddmd2h1fml2syvyqnhzk-therock-rocm-s...~~
~~ - therock-rocm-source-rocm-systems-projects-rocm-smi-lib - aborted (tail, raw) - �[31;1merror:�[0m failed to start SSH connection to '�[35;1mstrix-1.hydra-builders�[0m'~~
~~ - amdsmi-7.15.0 - failed (tail, raw)~~

Hydra reported failures for this PR.

- Eval: #78243
- Jobset: hellas/nix-strix-halo-pr-148
- Head: fc2953cd0d2f

Failed builds:
- x86_64-linux.ci.build - dependency failed cached (log, tail, raw)
~~ - python3.13-vllm-0.25.1-dist, python3.13-vllm-0.25.1 - failed (tail, raw)~~

Update: hellas/nix-strix-halo-pr-148, eval #78457, head eecccbdeb9a9: newer evaluation started.

Hydra reported failures for this PR.

- Eval: #78167
- Jobset: hellas/nix-strix-halo-pr-148
- Head: a995ee1db3fc

Failed builds:
- x86_64-linux.ci.build - dependency failed (log, tail, raw)
~~ - python3.13-vllm-0.25.1-dist, python3.13-vllm-0.25.1 - failed (tail, raw)~~
- x86_64-linux.ci.smoke - dependency failed (log, tail, raw)
~~ - ds4-rocm-gfx1151-smoke - failed (tail, raw)~~

Hydra reported failures for this PR.

- Eval: #78229
- Jobset: hellas/nix-strix-halo-pr-148
- Head: 2e497195c9f7

Failed builds:
- x86_64-linux.ci.smoke - dependency failed (log, tail, raw)
~~ - fastflowlm-short - failed (tail, raw)~~
~~ - ds4-metal-pi-smoke - failed (tail, raw)~~

Update: hellas/nix-strix-halo-pr-148, eval #78243, head fc2953cd0d2f: newer evaluation started.

Hydra reported failures for this PR.

- Eval: #78218
- Jobset: hellas/nix-strix-halo-pr-148
- Head: bf874eaa88ae

Failed builds:
- x86_64-linux.ci.smoke - dependency failed (log, tail, raw)
~~ - vllm-rocm-gfx1151-throughput-smoke - cached failure (tail, raw), propagated from build 11738, log~~

Update: hellas/nix-strix-halo-pr-148, eval #78229, head 2e497195c9f7: newer evaluation started.

Hydra reported failures for this PR.

- Eval: #78213
- Jobset: hellas/nix-strix-halo-pr-148
- Head: fc1d819025cb

Failed builds:
- x86_64-linux.ci.smoke - dependency failed (log, tail, raw)
~~ - vllm-rocm-gfx1151-throughput-smoke - failed (tail, raw)~~

Update: hellas/nix-strix-halo-pr-148, eval #78218, head bf874eaa88ae: newer evaluation started.

Hydra reported failures for this PR.

- Eval: #78206
- Jobset: hellas/nix-strix-halo-pr-148
- Head: 25a3950b33d2

Failed builds:
- x86_64-linux.ci.smoke - dependency failed (log, tail, raw)
~~ - fastflowlm-medium - failed (tail, raw)~~
~~ - mlx-metal-gemm-smoke - failed (tail, raw)~~

Hydra reported failures for this PR.

- Eval: #78206
- Jobset: hellas/nix-strix-halo-pr-148
- Head: 25a3950b33d2

Failed builds:
- x86_64-linux.ci.smoke - dependency failed (log, tail, raw)
~~ - mlx-metal-gemm-smoke - failed (tail, raw)~~

Update: hellas/nix-strix-halo-pr-148, eval #78213, head fc1d819025cb: newer evaluation started.

Hydra reported failures for this PR.

- Eval: #78187
- Jobset: hellas/nix-strix-halo-pr-148
- Head: 5fb6f77e93f3

Failed builds:
- x86_64-linux.ci.smoke - dependency failed (log, tail, raw)
~~ - fastflowlm-medium - failed (tail, raw)~~

Update: hellas/nix-strix-halo-pr-148, eval #78206, head 25a3950b33d2: newer evaluation started.

Hydra reported failures for this PR.

- Eval: #78167
- Jobset: hellas/nix-strix-halo-pr-148
- Head: a995ee1db3fc

Failed builds:
- x86_64-linux.ci.smoke - dependency failed (log, tail, raw)
~~ - ds4-rocm-gfx1151-smoke - failed (tail, raw)~~

Update: hellas/nix-strix-halo-pr-148, eval #78187, head 5fb6f77e93f3: newer evaluation started.

Hydra reported failures for this PR.

- Eval: #78158
- Jobset: hellas/nix-strix-halo-pr-148
- Head: 89532da41083

Failed builds:
- x86_64-linux.ci.smoke - dependency failed (log, tail, raw)
~~ - fastflowlm-0.9.45-unstable-fd37140 - failed (tail, raw)~~

Update: hellas/nix-strix-halo-pr-148, eval #78167, head a995ee1db3fc: newer evaluation started.

Hydra reported failures for this PR.

- Eval: #78152
- Jobset: hellas/nix-strix-halo-pr-148
- Head: e1225b0f3c48

Failed builds:
- x86_64-linux.ci.build - dependency failed (log, tail, raw)
~~ - ds4-unstable-80ebbc3 - failed (tail, raw)~~

Update: hellas/nix-strix-halo-pr-148, eval #78158, head 89532da41083: newer evaluation started.

Hydra reported failures for this PR.

- Eval: #78139
- Jobset: hellas/nix-strix-halo-pr-148
- Head: 7d9a0072bb6e

Failed builds:
- x86_64-linux.ci.build - dependency failed (log, tail, raw)
~~ - python3.14-mlx-rocm-gfx1151-0.32.0-dist, python3.14-mlx-rocm-gfx1151-0.32.0 - cached failure (tail, raw), propagated from build 11710, log~~
- x86_64-linux.ci.smoke - dependency failed (log, tail, raw)
~~ - python3.14-mlx-rocm-gfx1151-0.32.0-dist, python3.14-mlx-rocm-gfx1151-0.32.0 - cached failure (tail, raw), propagated from build 11710, log~~

Update: hellas/nix-strix-halo-pr-148, eval #78152, head e1225b0f3c48: newer evaluation started.

Hydra reported failures for this PR.

- Eval: #78118
- Jobset: hellas/nix-strix-halo-pr-148
- Head: 8bb7ed08759c

Failed builds:
- x86_64-linux.ci.build - dependency failed (log, tail, raw)
~~ - python3.14-mlx-metal-0.32.0-dist, python3.14-mlx-metal-0.32.0 - cached failure (tail, raw), propagated from build 11706, log~~
~~ - python3.14-mlx-metal-0.32.0-dist, python3.14-mlx-metal-0.32.0 - failed (tail, raw)~~
~~ - python3.14-mlx-rocm-gfx1151-0.32.0-dist, python3.14-mlx-rocm-gfx1151-0.32.0 - failed (tail, raw)~~
- x86_64-linux.ci.smoke - dependency failed cached (log, tail, raw)
~~ - ds4-unstable-80ebbc3 - cached failure (tail, raw), propagated from build 11589, log~~
~~ - python3.14-mlx-metal-0.32.0-dist, python3.14-mlx-metal-0.32.0 - cached failure (tail, raw), propagated from build 11710, log~~

Hydra reported failures for this PR.

- Eval: #78118
- Jobset: hellas/nix-strix-halo-pr-148
- Head: 8bb7ed08759c

Failed builds:
- x86_64-linux.ci.build - dependency failed (log, tail, raw)
~~ - python3.14-mlx-metal-0.32.0-dist, python3.14-mlx-metal-0.32.0 - cached failure (tail, raw), propagated from build 11706, log~~
~~ - python3.14-mlx-metal-0.32.0-dist, python3.14-mlx-metal-0.32.0 - failed (tail, raw)~~
~~ - python3.14-mlx-rocm-gfx1151-0.32.0-dist, python3.14-mlx-rocm-gfx1151-0.32.0 - failed (tail, raw)~~
- x86_64-linux.ci.smoke - dependency failed cached (log, tail, raw)
~~ - ds4-unstable-80ebbc3 - cached failure (tail, raw), propagated from build 11589, log~~

Update: hellas/nix-strix-halo-pr-148, eval #78139, head 7d9a0072bb6e: newer evaluation started.

Hydra reported failures for this PR.

- Eval: #77958
- Jobset: hellas/nix-strix-halo-pr-148
- Head: 97dec0843fdb

Failed builds:
- x86_64-linux.ci.build - dependency failed (log, tail, raw)
~~ - ds4-unstable-80ebbc3 - cached failure (tail, raw), propagated from build 11589, log~~
~~ - python3.14-mlx-metal-0.32.0-dist, python3.14-mlx-metal-0.32.0 - failed (tail, raw)~~
~~ - llama-cpp-rocm-gfx1151-9925-dev, llama-cpp-rocm-gfx1151-9925 - failed (tail, raw)~~
~~ - llama-cpp-rocm-gfx1151-9925-dev, llama-cpp-rocm-gfx1151-9925 - failed (tail, raw)~~
~~ - python3.14-mlx-metal-0.32.0-dist, python3.14-mlx-metal-0.32.0 - failed (tail, raw)~~
~~ - python3.14-mlx-metal-0.32.0-dist, python3.14-mlx-metal-0.32.0 - failed (tail, raw)~~
~~ - llama-cpp-rocm-gfx1151-9925-dev, llama-cpp-rocm-gfx1151-9925 - failed (tail, raw)~~
~~ - llama-cpp-9925-dev, llama-cpp-9925 - failed (tail, raw)~~
~~ - ...and 46 more failed steps~~

Hydra reported failures for this PR.

- Eval: #77958
- Jobset: hellas/nix-strix-halo-pr-148
- Head: 97dec0843fdb

Failed builds:
- x86_64-linux.ci.build - dependency failed (log, tail, raw)
~~ - ds4-unstable-80ebbc3 - cached failure (tail, raw), propagated from build 11589, log~~
~~ - python3.14-mlx-metal-0.32.0-dist, python3.14-mlx-metal-0.32.0 - failed (tail, raw)~~
~~ - llama-cpp-rocm-gfx1151-9925-dev, llama-cpp-rocm-gfx1151-9925 - failed (tail, raw)~~
~~ - llama-cpp-rocm-gfx1151-9925-dev, llama-cpp-rocm-gfx1151-9925 - failed (tail, raw)~~
~~ - python3.14-mlx-metal-0.32.0-dist, python3.14-mlx-metal-0.32.0 - failed (tail, raw)~~
~~ - python3.14-mlx-metal-0.32.0-dist, python3.14-mlx-metal-0.32.0 - failed (tail, raw)~~
~~ - llama-cpp-rocm-gfx1151-9925-dev, llama-cpp-rocm-gfx1151-9925 - failed (tail, raw)~~
~~ - llama-cpp-9925-dev, llama-cpp-9925 - failed (tail, raw)~~
~~ - ...and 42 more failed steps~~
- x86_64-linux.ci.smoke - dependency failed (log, tail, raw)
~~ - ds4-unstable-80ebbc3 - cached failure (tail, raw), propagated from build 11589, log~~

@georgewhewell
llama-cpp: normalize RPC server install alias
8bb7ed0
@georgewhewell
mlx: keep Darwin packages on Python 3.13
7d9a007
@georgewhewell
mlx-rocm: refresh patches for 0.32.0
e1225b0
@georgewhewell
ds4: use lld for Darwin executables
89532da
@georgewhewell
fastflowlm: keep install inside the Nix prefix
a995ee1
@georgewhewell
therock: preserve split-load ELF layout
5fb6f77
@georgewhewell
fastflowlm: repair bundled NPU runtime paths
25a3950
@georgewhewell
bench: use each MLX package Python runtime
fc1d819
@georgewhewell
ci: bound FastFlowLM smoke to a short prompt
bf874ea
@georgewhewell
vllm: preserve runtime environment in model inspection
2e49719
@georgewhewell
ci: keep full DS4 models out of PR smoke
fc2953c
@georgewhewell
therock: align source SDK with ROCm 7.15
eecccbd
@georgewhewell
therock: make the 7.15 full profile hermetic
808b37e
@georgewhewell
therock: finish ROCm 7.15 source build plumbing
ae68543
@georgewhewell
therock: fix ROCm 7.15 subproject configuration
87a50d0
@georgewhewell
therock: pin rocprofiler SDK submodules
77dcc78
@georgewhewell
therock: complete ROCm profiler source graph
fdc736e
@georgewhewell
therock: fix ROCm 7.15 full source build
29c79ff
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@georgewhewell @hellasbot