gfx940/gfx1010/gfx1030/gfx1100 cause C++ narrowing errors in ollama's
mma.cuh with hipcc. Since we only have MI50 (gfx906) cards, compile
for gfx906 only. Reduces build time and avoids upstream code bugs.
CPU backends compiled with GCC (fixes AVX intrinsic errors from hipcc).
HIP backend compiled with hipcc (legacy mode skips enable_language(HIP)).
Go binary built with GCC for CGo linking.
This avoids both CMAKE_HIP_COMPILER rejection and CXX=hipcc CPU failures.
The old Dockerfile used the deprecated llama.cpp/ subdirectory approach
which doesn't exist in ollama v0.23.2. Now using the official CMake
presets (ROCm 6 preset) with AMDGPU_TARGETS including gfx906:xnack-.
- Add ollama/Dockerfile that builds ollama from source with AMDGPU_TARGETS=gfx906
- Uses ROCm 6.1 (rocm/dev-ubuntu-22.04:6.1.2-complete) for MI50 support
- Builds llama.cpp runner with HIPBLAS for gfx906 architecture
- Updates compose.yml to build from this Dockerfile instead of pulling ollama/ollama:latest
The migration from debian:stable-slim to nousresearch/hermes-agent:latest
dropped several packages that were previously installed. This restores:
- poppler-utils, imagemagick (PDF/image processing)
- texlive-latex-base, latex-extra, fonts-recommended, xetex, science
- qemu-user-static, binfmt-support (cross-compilation)
- emacs-nox (text editing)
These were added in PRs 3/5, 4/5, 5/5 and earlier commits of the
compose repo. The official image already has git, curl, ffmpeg,
python3, gcc, openssh, ripgrep, tini, docker-cli, etc.
- Patch now matches the current tts_tool.py (newer version ships in
nousresearch/hermes-agent:latest with different Edge fallback text)
- Adds dedicated elif provider == 'piper' block before else:
- Replaces else: fallback to use Piper instead of Edge
- Patches ALL copies (venv site-packages + /opt/hermes/tools/)
- Removes Edge TTS entirely as default/provider
The nousresearch/hermes-agent:latest image creates its venv
as root. Running 'uv pip install' as USER hermes fails with
Permission denied on the site-packages directory.
Fix: keep USER root while modifying the venv, then switch
back to USER hermes for runtime.
The nousresearch/hermes-agent:latest base image already has a
venv with hermes-agent installed at /opt/hermes/.venv/.
Running 'uv venv' on top of it either fails or wipes the
existing install.
Fix: activate the existing venv first, then pip install into it.
Starting from debian:stable-slim required re-installing everything
(Hermes source, Node.js, Playwright, etc.) which was redundant
and fragile. The official nousresearch/hermes-agent image already
has all that.
Now the Dockerfile:
- FROM nousresearch/hermes-agent:latest (has tts_tool.py, Playwright, etc.)
- Install Piper + voice model on top
- Patch tts_tool.py at build time (Edge fallback -> Piper)
- Runtime fallback in fix-permissions.sh for volume resilience
Cleaner, smaller Dockerfile, and the build-time patch can find
tts_tool.py because it's in the base image's venv.
The Dockerfile starts from debian:stable-slim, not from the official
Hermes image. Without installing hermes-agent from pip, there is no
tools/tts_tool.py in the image at build time, so the patch script
crashes with FileNotFoundError.
Adding hermes-agent to uv pip install gives us tts_tool.py in the
venv site-packages, so the COPY+RUN patch step works cleanly.
Also keep the runtime fallback in fix-permissions.sh for cases where
the volume's site-packages differ from the image.
The build-time COPY+RUN of patch_tts_tool.py failed because
the Dockerfile starts from debian:stable-slim and only copies
the ai/ build context — there's no tools/tts_tool.py in the
image at build time (Hermes is on the mounted data volume).
Move patching to fix-permissions.sh which runs at container
startup when the data volume is mounted, so tts_tool.py is
available via the venv site-packages.
Also make patch_tts_tool.py robust: searches multiple paths
for tts_tool.py, accepts path as argument, exits 0 instead
of 1 when file/pattern not found (build must not fail).
Commit 8e9a75f removed the COPY+RUN of patch_tts_tool.py
because the build context was thought to be insufficient.
The build context is ai/ which contains both the Dockerfile
and patch_tts_tool.py, so COPY works fine.
Without this step the tts_tool.py silently falls through
to Edge TTS as its default provider even when
config.yaml says provider: piper, because 'piper' is not
a recognized provider in the unpatched code. This caused
the female Edge TTS voice (AriaNeural) instead of the
configured Ryan High male voice.
