fix release publish, tidy the docker eval stack

Author: anindex

.github/workflows/release.yml

@@ -138,9 +138,12 @@ jobs:
     if: startsWith(github.ref, 'refs/tags/')
     runs-on: ubuntu-24.04
     steps:
+      # Tarballs only. The docker job also leaves a .dockerbuild build record
+      # artifact behind, and pulling that one fails the whole download.
       - uses: actions/download-artifact@v4
-        with: { path: dist, merge-multiple: true }
+        with: { path: dist, pattern: '{linux,macos}-*', merge-multiple: true }
       - uses: softprops/action-gh-release@v2
         with:
           files: dist/*.tar.gz
+          fail_on_unmatched_files: true
           generate_release_notes: true

README.md

@@ -70,6 +70,7 @@ export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
 
 Check [docs/backend](docs/backend) for compiling `vla.cpp` on other platforms.
 WSL2, Apple Silicon, and Intel GPU are all tested.
+To build and run in containers instead, see [docs/DOCKER.md](docs/DOCKER.md).
 
 ---
 

docs/DOCKER.md

@@ -52,11 +52,12 @@ Build args accepted by the server `Dockerfile`:
 | Arg | Default | Notes |
 |-----|---------|-------|
 | `BACKEND` | `cuda` | `cuda` or `cpu` |
-| `CUDA_ARCH` | `120` | Blackwell; `89` for RTX40, `87` for Orin, `86` for RTX30 |
+| `CUDA_ARCH` | `120` in Compose, `89` in the `Dockerfile` | Blackwell; `89` for RTX40, `87` for Orin, `86` for RTX30 |
 | `BASE_IMAGE` | `nvidia/cuda:12.9.1-devel-ubuntu24.04` | Set to `ubuntu:24.04` when building a CPU image |
 | `JOBS` | `nproc` | Lower if nvcc segfaults on flash-attn kernels |
 
-Override via e.g. `docker compose -f eval/docker-compose.yml build --build-arg CUDA_ARCH=89 server`.
+Override the arch from the environment, `CUDA_ARCH=89 docker compose -f eval/docker-compose.yml build server`,
+or per build, `docker compose -f eval/docker-compose.yml build --build-arg CUDA_ARCH=89 server`.
 
 ### 3. Start the server
 

eval/Dockerfile.client

@@ -55,7 +55,6 @@ RUN update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.10 1
 RUN curl -sS https://bootstrap.pypa.io/get-pip.py | python3.10
 
 WORKDIR /workspace/vla.cpp
-COPY . .
 
 # ---------------------------------------------------------------------------
 # Python dependencies — order matters to avoid version conflicts
@@ -96,7 +95,7 @@ RUN pip3 install lerobot==0.4.3 && \
 RUN pip3 install \
     transformers==4.51.3 \
     gymnasium==0.29.1 \
-    zmq \
+    pyzmq \
     msgpack==1.1.0 \
     msgpack-numpy==0.4.8 \
     pillow \
@@ -112,7 +111,7 @@ RUN pip3 install \
     albumentations==1.4.18 \
     pyarrow==12.0.1 \
     diffusers==0.30.1 \
-    huggingface-hub
+    huggingface-hub==0.35.3
 
 # Re-pin numpy to 1.x — later deps (albumentations, tianshou, etc.) pull in 2.x
 RUN pip3 install numpy==1.26.4
@@ -124,4 +123,8 @@ RUN sed -i 's/np_core = np._core if is_numpy_available("2.0.0") else np.core/np_
 # Clean up pip cache to keep image size small
 RUN rm -rf /root/.cache/pip
 
+# The repo goes last: it changes every commit, and everything above it is a
+# half-hour of installs worth caching.
+COPY . .
+
 CMD ["bash"]

eval/docker-compose.yml

@@ -22,15 +22,16 @@ services:
   # vla-server — C++ inference daemon (upstream Dockerfile)
   # Build args:
   #   BACKEND=cuda|cpu   (default: cuda)
-  #   CUDA_ARCH          (default: 120, use 89 for RTX40, 87 for Orin, etc.)
+  #   CUDA_ARCH          (from the environment, default 120; 89 for RTX40, 87 for Orin)
+  # e.g. CUDA_ARCH=89 docker compose -f eval/docker-compose.yml build server
   # --------------------------------------------------------------------------
   server:
     build:
       context: ..
       dockerfile: Dockerfile
       args:
         BACKEND: cuda
-        CUDA_ARCH: "120"
+        CUDA_ARCH: "${CUDA_ARCH:-120}"
     image: vla-cpp-server
     container_name: vla-cpp-server
     devices: