fix: remove gpu extra and update docs and docker files

Dockerfile

@@ -17,24 +17,40 @@ ENV UV_LINK_MODE=copy \
 RUN apt-get update && apt-get install -y --no-install-recommends libgomp1 \
     && rm -rf /var/lib/apt/lists/*
 
-# `server` and `sktime` are always in, so the process can load naive. Heavier extras
-# come from the build arg, one word each, e.g.
+# `server` is always in, and it pulls `sktime`, so the process can load naive. Heavier
+# extras come from the build arg, one word each, e.g.
 #   docker build --build-arg TSERVE_EXTRAS=hub .
-#   docker build --build-arg TSERVE_EXTRAS="chronos gpu" .
 ARG TSERVE_EXTRAS=""
+# Empty keeps PyPI torch (CUDA on Linux). Any value appends the CPU index, e.g.
+#   docker build --build-arg TSERVE_CPU=1 .
+ARG TSERVE_CPU=""
+
+# Optional `--extra` for every sync. Empty when TSERVE_EXTRAS is unset.
+ENV SYNC_EXTRAS="${TSERVE_EXTRAS:+--extra ${TSERVE_EXTRAS}}"
 
 # Resolve and install third-party deps from pyproject.toml only. Source changes
-# then do not rebuild this layer. uv.lock is not tracked, so this is not
-# `--frozen` / `--locked`. printf repeats `--extra` once per remaining word.
+# then do not rebuild this layer.
+COPY docker/pytorch-cpu.toml /pytorch-cpu.toml
 COPY pyproject.toml ./
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv sync --no-dev --no-install-project --no-editable \
-        $(printf -- '--extra %s ' server sktime $TSERVE_EXTRAS)
+    if [ -n "$TSERVE_CPU" ]; then cat /pytorch-cpu.toml >> pyproject.toml; fi \
+    && uv sync \
+        --no-dev \
+        --no-install-project \
+        --no-editable \
+        --extra server \
+        $SYNC_EXTRAS
 
+# Install tserve now that the source tree is present. Third-party deps stay in
+# the layer above, so a source edit rebuilds only this step.
 COPY . .
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv sync --no-dev --no-editable \
-        $(printf -- '--extra %s ' server sktime $TSERVE_EXTRAS)
+    if [ -n "$TSERVE_CPU" ]; then cat /pytorch-cpu.toml >> pyproject.toml; fi \
+    && uv sync \
+        --no-dev \
+        --no-editable \
+        --extra server \
+        $SYNC_EXTRAS
 
 # Runtime image: no uv, no source tree. `--no-editable` baked tserve
 # into the venv, so only `.venv` is copied. Python path must match the builder.

README.md

@@ -122,9 +122,9 @@ Multivariate series, covariates, and quantiles differ by family: [Capabilities](
 
 Docker needs no local Python. uv and pip need Python 3.12 or newer. Install the extra, or pull the tag, for the family in the table above.
 
-- **Docker.** [Pull an image](https://tserve.readthedocs.io/en/latest/server/docker/#pull-an-image), then [run the server](https://tserve.readthedocs.io/en/latest/server/docker/#run-the-server). CPU and GPU are separate tags.
-- **uv or pip.** [UV / Pip](https://tserve.readthedocs.io/en/latest/installation/#uv-pip). A PyPI install takes CUDA torch (MPS on macOS). A CPU wheel: [CPU-only install](https://tserve.readthedocs.io/en/latest/server/pip/#cpu-only-install).
-- **A clone.** [From source](https://tserve.readthedocs.io/en/latest/installation/#from-source). On a clone, uv selects the torch index with the [`gpu` extra](https://tserve.readthedocs.io/en/latest/server/source/#gpu).
+- **Docker.** [Pull an image](https://tserve.readthedocs.io/en/latest/server/docker/#pull-an-image), then [run the server](https://tserve.readthedocs.io/en/latest/server/docker/#run-the-server). Each family has a CPU tag and a `-gpu` tag.
+- **uv or pip.** [UV / Pip](https://tserve.readthedocs.io/en/latest/installation/#uv-pip). A CPU-only build for your OS on [CPU-only install](https://tserve.readthedocs.io/en/latest/server/pip/#cpu-only-install).
+- **A clone.** [From source](https://tserve.readthedocs.io/en/latest/server/source/). An editable install for development and unreleased changes.
 
 ## Load a model
 

docker-bake.hcl

@@ -1,6 +1,7 @@
 # TServe images: same Dockerfile, different TSERVE_EXTRAS.
 # Every tag is linux/amd64 + linux/arm64 (Linux, Mac, Windows Docker Desktop).
-# `--extra gpu` is a torch index choice (PyPI / CUDA / MPS), not an arch pin.
+# Empty TSERVE_CPU keeps PyPI torch (CUDA on Linux). CPU tags set TSERVE_CPU
+# so torch comes from the CPU index. :base has no torch.
 #
 # First time on a new machine:
 #   docker run --privileged --rm tonistiigi/binfmt --install all
@@ -55,156 +56,156 @@ target "base" {
 
 target "hub" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "hub" }
+  args     = { TSERVE_EXTRAS = "hub", TSERVE_CPU = "1" }
   tags     = ["${TSERVE_IMAGE}:hub"]
 }
 
 target "chronos" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "chronos" }
+  args     = { TSERVE_EXTRAS = "chronos", TSERVE_CPU = "1" }
   tags     = ["${TSERVE_IMAGE}:chronos"]
 }
 
 target "kronos" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "kronos" }
+  args     = { TSERVE_EXTRAS = "kronos", TSERVE_CPU = "1" }
   tags     = ["${TSERVE_IMAGE}:kronos"]
 }
 
 target "granite" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "granite" }
+  args     = { TSERVE_EXTRAS = "granite", TSERVE_CPU = "1" }
   tags     = ["${TSERVE_IMAGE}:granite"]
 }
 
 target "moirai" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "moirai" }
+  args     = { TSERVE_EXTRAS = "moirai", TSERVE_CPU = "1" }
   tags     = ["${TSERVE_IMAGE}:moirai"]
 }
 
 target "tirex" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "tirex" }
+  args     = { TSERVE_EXTRAS = "tirex", TSERVE_CPU = "1" }
   tags     = ["${TSERVE_IMAGE}:tirex"]
 }
 
 target "tirex2" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "tirex2" }
+  args     = { TSERVE_EXTRAS = "tirex2", TSERVE_CPU = "1" }
   tags     = ["${TSERVE_IMAGE}:tirex2"]
 }
 
 target "toto" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "toto" }
+  args     = { TSERVE_EXTRAS = "toto", TSERVE_CPU = "1" }
   tags     = ["${TSERVE_IMAGE}:toto"]
 }
 
 target "mantis" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "mantis" }
+  args     = { TSERVE_EXTRAS = "mantis", TSERVE_CPU = "1" }
   tags     = ["${TSERVE_IMAGE}:mantis"]
 }
 
 target "timesfm3" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "timesfm3" }
+  args     = { TSERVE_EXTRAS = "timesfm3", TSERVE_CPU = "1" }
   tags     = ["${TSERVE_IMAGE}:timesfm3"]
 }
 
 target "t0" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "t0" }
+  args     = { TSERVE_EXTRAS = "t0", TSERVE_CPU = "1" }
   tags     = ["${TSERVE_IMAGE}:t0"]
 }
 
 target "tafsut" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "tafsut" }
+  args     = { TSERVE_EXTRAS = "tafsut", TSERVE_CPU = "1" }
   tags     = ["${TSERVE_IMAGE}:tafsut"]
 }
 
 target "full" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "full" }
+  args     = { TSERVE_EXTRAS = "full", TSERVE_CPU = "1" }
   tags     = ["${TSERVE_IMAGE}:full"]
 }
 
 target "hub-gpu" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "hub gpu" }
+  args     = { TSERVE_EXTRAS = "hub" }
   tags     = ["${TSERVE_IMAGE}:hub-gpu"]
 }
 
 target "chronos-gpu" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "chronos gpu" }
+  args     = { TSERVE_EXTRAS = "chronos" }
   tags     = ["${TSERVE_IMAGE}:chronos-gpu"]
 }
 
 target "kronos-gpu" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "kronos gpu" }
+  args     = { TSERVE_EXTRAS = "kronos" }
   tags     = ["${TSERVE_IMAGE}:kronos-gpu"]
 }
 
 target "granite-gpu" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "granite gpu" }
+  args     = { TSERVE_EXTRAS = "granite" }
   tags     = ["${TSERVE_IMAGE}:granite-gpu"]
 }
 
 target "moirai-gpu" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "moirai gpu" }
+  args     = { TSERVE_EXTRAS = "moirai" }
   tags     = ["${TSERVE_IMAGE}:moirai-gpu"]
 }
 
 target "tirex-gpu" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "tirex gpu" }
+  args     = { TSERVE_EXTRAS = "tirex" }
   tags     = ["${TSERVE_IMAGE}:tirex-gpu"]
 }
 
 target "tirex2-gpu" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "tirex2 gpu" }
+  args     = { TSERVE_EXTRAS = "tirex2" }
   tags     = ["${TSERVE_IMAGE}:tirex2-gpu"]
 }
 
 target "toto-gpu" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "toto gpu" }
+  args     = { TSERVE_EXTRAS = "toto" }
   tags     = ["${TSERVE_IMAGE}:toto-gpu"]
 }
 
 target "mantis-gpu" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "mantis gpu" }
+  args     = { TSERVE_EXTRAS = "mantis" }
   tags     = ["${TSERVE_IMAGE}:mantis-gpu"]
 }
 
 target "timesfm3-gpu" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "timesfm3 gpu" }
+  args     = { TSERVE_EXTRAS = "timesfm3" }
   tags     = ["${TSERVE_IMAGE}:timesfm3-gpu"]
 }
 
 target "t0-gpu" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "t0 gpu" }
+  args     = { TSERVE_EXTRAS = "t0" }
   tags     = ["${TSERVE_IMAGE}:t0-gpu"]
 }
 
 target "tafsut-gpu" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "tafsut gpu" }
+  args     = { TSERVE_EXTRAS = "tafsut" }
   tags     = ["${TSERVE_IMAGE}:tafsut-gpu"]
 }
 
 target "full-gpu" {
   inherits = ["_common"]
-  args     = { TSERVE_EXTRAS = "full gpu" }
+  args     = { TSERVE_EXTRAS = "full" }
   tags     = ["${TSERVE_IMAGE}:full-gpu"]
 }

docker/pytorch-cpu.toml

@@ -0,0 +1,10 @@
+# Appended to pyproject.toml when the image is built with TSERVE_CPU set.
+# `explicit` keeps every package except torch on PyPI.
+
+[[tool.uv.index]]
+name = "pytorch-cpu"
+url = "https://download.pytorch.org/whl/cpu"
+explicit = true
+
+[tool.uv.sources]
+torch = [{ index = "pytorch-cpu" }]

docs/installation.md

@@ -26,7 +26,7 @@ Other model families use different image tags. Choose the model first, then use
 
 ## UV / Pip
 
-TServe requires Python 3.12 or newer. Install the `server` extra and the extra for the model family you need. The examples below install the `hub` family. A plain install pulls the CUDA build of torch (MPS on macOS); the CPU tabs skip that download on a machine without a GPU.
+TServe requires Python 3.12 or newer. Install the `server` extra and the extra for the model family you need. The examples below install the `hub` family. A CPU build of torch: [CPU-only install](server/pip.md#cpu-only-install).
 
 === "uv"
 
@@ -36,23 +36,9 @@ TServe requires Python 3.12 or newer. Install the `server` extra and the extra f
     uv venv
     ```
 
-    Then install TServe:
-
-    === "GPU (default)"
-
-        ```bash
-        uv pip install "tserve[server,hub]"
-        ```
-
-    === "CPU only"
-
-        ```bash
-        uv pip install torch --index-url https://download.pytorch.org/whl/cpu
-        ```
-
-        ```bash
-        uv pip install "tserve[server,hub]"
-        ```
+    ```bash
+    uv pip install "tserve[server,hub]"
+    ```
 
 === "pip"
 
@@ -72,29 +58,15 @@ TServe requires Python 3.12 or newer. Install the `server` extra and the extra f
         .venv\Scripts\Activate.ps1
         ```
 
-    Then install TServe:
-
-    === "GPU (default)"
-
-        ```bash
-        python -m pip install "tserve[server,hub]"
-        ```
-
-    === "CPU only"
-
-        ```bash
-        python -m pip install torch --index-url https://download.pytorch.org/whl/cpu
-        ```
-
-        ```bash
-        python -m pip install "tserve[server,hub]"
-        ```
+    ```bash
+    python -m pip install "tserve[server,hub]"
+    ```
 
-The `server` extra alone supports the `naive` test baseline. Replace `hub` with another [family extra](models/index.md#dependencies), or use `full` for every family. The same CPU-first order is documented as [CPU-only install](server/pip.md#cpu-only-install).
+The `server` extra alone supports the `naive` test baseline. Replace `hub` with another [family extra](models/index.md#dependencies), or use `full` for every family.
 
 ## From source
 
-Use a source install when developing TServe or testing unreleased changes. It is also the only path where the `gpu` extra selects the torch index, because that choice lives in the repository's uv lockfile. The [From source](server/source.md) guide covers cloning the repository, editable installs, dependency extras, and GPU setup.
+Use a source install when developing TServe or testing unreleased changes. The [From source](server/source.md) guide covers cloning the repository, editable installs, and dependency extras.
 
 ## Next
 

docs/models/hub.md

@@ -6,7 +6,7 @@ Four Hugging Face families, 81 of the catalog's 117 models. The usual starting p
 | --- | --- | --- | --- | --- | --- |
 | `hub` | [`:hub`](https://hub.docker.com/r/sktime/tserve/tags?name=hub) | [`:hub-gpu`](https://hub.docker.com/r/sktime/tserve/tags?name=hub-gpu) | Chronos Bolt, Chronos T5, TTM, TimesFM 2.x | 81 | `chronos_bolt` |
 
-Builds on [`base`](base.md), so `naive` is available here too. Every extra that pulls `hf` builds on `hub`, so those pages can load these models as well.
+Builds on [`base`](base.md), so `naive` is available here too. Every extra that includes `hub` can load these models as well.
 
 ## Start a server
 

docs/reference/development.md

@@ -57,7 +57,7 @@ make docs-serve
 
 ## Docker images
 
-`Dockerfile` always installs `--extra server --extra sktime` and adds whatever `TSERVE_EXTRAS` names, which is how one file produces every tag. `docker-bake.hcl` holds the published matrix: one target per tag, plus `cpu` and `gpu` groups, with `TSERVE_IMAGE` defaulting to `sktime/tserve`.
+`Dockerfile` always installs `--extra server` and adds whatever `TSERVE_EXTRAS` names. `TSERVE_CPU` selects the CPU torch index; leaving it empty keeps the PyPI wheel. That is how one file produces every tag. `docker-bake.hcl` holds the published matrix: one target per tag, plus `cpu` and `gpu` groups, with `TSERVE_IMAGE` defaulting to `sktime/tserve`.
 
 ```bash
 TSERVE_IMAGE=sktime/tserve docker buildx bake --push hub

docs/server/docker.md

@@ -90,7 +90,7 @@ The `*-gpu` tags install torch from PyPI instead of the CPU wheel index. They ne
 docker run --rm --gpus all -p 8000:8000 sktime/tserve:hub-gpu chronos_bolt ttm_r3
 ```
 
-That covers Linux and Windows through WSL2. Docker on macOS has no GPU passthrough, so Apple silicon acceleration means a [local UV / Pip install](pip.md#install), which takes the MPS build of torch.
+That covers Linux and Windows through WSL2. Docker on macOS has no GPU passthrough, so Apple silicon acceleration means a [local UV / Pip install](pip.md#install).
 
 ## Models from a directory
 
@@ -113,16 +113,22 @@ cd tserve
 
 ### One image with `docker build`
 
-The [Dockerfile](https://github.com/sktime/tserve/blob/main/Dockerfile) always installs `--extra server --extra sktime`; `TSERVE_EXTRAS` adds the heavier ones:
+The [Dockerfile](https://github.com/sktime/tserve/blob/main/Dockerfile) always installs `--extra server`. `TSERVE_EXTRAS` adds the heavier ones, one word each. Leave `TSERVE_CPU` empty for the PyPI torch wheel. Any value installs torch from the CPU index, which is what the published CPU tags do:
 
 ```bash
-docker build --build-arg TSERVE_EXTRAS=hub -t tserve:hub .
+docker build --build-arg TSERVE_EXTRAS=hub --build-arg TSERVE_CPU=1 -t tserve:hub .
+```
+
+A GPU image is the same build with `TSERVE_CPU` unset:
+
+```bash
+docker build --build-arg TSERVE_EXTRAS=chronos -t tserve:chronos-gpu .
 ```
 
 Several extras go in one quoted argument:
 
 ```bash
-docker build --build-arg TSERVE_EXTRAS="chronos gpu" -t tserve:chronos-gpu .
+docker build --build-arg TSERVE_EXTRAS="chronos moirai" --build-arg TSERVE_CPU=1 -t tserve:custom .
 ```
 
 This builds for the architecture of the machine you are on and leaves the image in the local store, which is all you need to run it locally:

docs/server/index.md

@@ -28,7 +28,7 @@ A bare `tserve` loads `naive` only. Name models to load them too. `GET /models`
     tserve chronos_bolt ttm_r3
     ```
 
-    This installs CUDA torch (MPS on macOS). A CPU wheel: [CPU-only install](pip.md#cpu-only-install).
+    This installs CUDA torch (MPS on macOS). A CPU build: [CPU-only install](pip.md#cpu-only-install).
 
 Another family is the same command with that row's tag and model. `moirai_2`: