diff --git a/Dockerfile b/Dockerfile
index c0416b7..119236d 100644
--- a/Dockerfile
+++ b/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.
diff --git a/README.md b/README.md
index 9b070a8..07a314c 100644
--- a/README.md
+++ b/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
 
diff --git a/docker-bake.hcl b/docker-bake.hcl
index c1a9ebc..9e7f9d7 100644
--- a/docker-bake.hcl
+++ b/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"]
 }
diff --git a/docker/pytorch-cpu.toml b/docker/pytorch-cpu.toml
new file mode 100644
index 0000000..1ee698a
--- /dev/null
+++ b/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" }]
diff --git a/docs/installation.md b/docs/installation.md
index bf47c85..fd4a09d 100644
--- a/docs/installation.md
+++ b/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
 
diff --git a/docs/models/hub.md b/docs/models/hub.md
index 0e13c7f..9f82539 100644
--- a/docs/models/hub.md
+++ b/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
 
diff --git a/docs/reference/development.md b/docs/reference/development.md
index 4683c2a..3e05dd8 100644
--- a/docs/reference/development.md
+++ b/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
diff --git a/docs/server/docker.md b/docs/server/docker.md
index f4b7812..c2b72aa 100644
--- a/docs/server/docker.md
+++ b/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:
diff --git a/docs/server/index.md b/docs/server/index.md
index 154b9cc..8afd905 100644
--- a/docs/server/index.md
+++ b/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`:
 
diff --git a/docs/server/pip.md b/docs/server/pip.md
index e856b0a..6c003d3 100644
--- a/docs/server/pip.md
+++ b/docs/server/pip.md
@@ -16,8 +16,6 @@ Python >= 3.12. Install TServe from PyPI with [uv](https://docs.astral.sh/uv/) o
     pip install "tserve[server,hub]"
     ```
 
-    The `gpu` extra does not change a pip install. Family extras already install CUDA torch from PyPI (MPS on macOS). To force a CPU wheel, install torch separately first — [CPU-only install](#cpu-only-install).
-
 `server` is enough to serve `naive` (a test baseline). The `hub` extra above covers Chronos Bolt/T5, TTM, and TimesFM 2.x. Do not add `client` on a machine that only serves.
 
 ## Dependencies
@@ -26,25 +24,13 @@ Python >= 3.12. Install TServe from PyPI with [uv](https://docs.astral.sh/uv/) o
 
 Replace `hub` in the install above with another extra from the table. `full` is the union extra. `all-extras` is a pip convenience for `client,server,full` and is not a Docker tag. Each extra's command and models: [catalog](../models/index.md#start-a-server).
 
-`gpu` is not a model family. A CPU wheel: [CPU-only install](#cpu-only-install).
+A CPU build of torch: [CPU-only install](#cpu-only-install).
 
 ## CPU-only install
 
-Family extras pull `torch`, and every install here takes the CUDA wheel from PyPI (MPS on macOS). On a GPU host that is already what you want, so nothing below is needed.
-
-Without a GPU, that wheel is a large download you will never use. Install torch from the CPU index first, then TServe. If a later install replaces that wheel with CUDA, run the torch line again.
-
-```bash
-pip install torch --index-url https://download.pytorch.org/whl/cpu
-```
-
-```bash
-pip install "tserve[server,hub]"
-```
-
-The same order works with `uv pip install`. Swap `hub` for any other family extra from [Dependencies](#dependencies).
+Family extras pull `torch`. Pick the CPU build for your OS on the [PyTorch install page](https://pytorch.org/get-started/locally/), install it, then install TServe as above. If a later install replaces that build, run the PyTorch command again.
 
-The `gpu` extra only selects the torch index in a clone's uv lockfile: [From source](source.md#gpu). Containers pick the wheel through the tag instead: [Docker](docker.md#gpu-images).
+Swap `hub` for any other family extra from [Dependencies](#dependencies). A GPU host can also use a `*-gpu` image: [GPU images](docker.md#gpu-images).
 
 --8<-- "includes/serve.md"
 
diff --git a/docs/server/source.md b/docs/server/source.md
index 0394a02..9c3a9e0 100644
--- a/docs/server/source.md
+++ b/docs/server/source.md
@@ -18,8 +18,6 @@ Python >= 3.12, and a clone over HTTPS. [uv](https://docs.astral.sh/uv/) is the
     pip install -e ".[server,hub]"
     ```
 
-    The `gpu` extra does not work with pip. Family extras already install CUDA torch from PyPI (MPS on macOS). To force a CPU wheel, install torch separately first — [GPU](#gpu).
-
 `server` is enough to serve `naive` (a test baseline). The `hub` extra above covers Chronos Bolt/T5, TTM, and TimesFM 2.x. Do not add `client` on a machine that only serves.
 
 ## Dependencies
@@ -28,28 +26,7 @@ Python >= 3.12, and a clone over HTTPS. [uv](https://docs.astral.sh/uv/) is the
 
 Replace `hub` in the install above with another extra from the table. uv repeats `--extra` (`uv sync --extra server --extra chronos`). pip takes one list (`".[server,chronos]"`). `full` is the union extra. `all-extras` is a pip convenience for `client,server,full` and is not a Docker tag. Each extra's command and models: [catalog](../models/index.md#start-a-server).
 
-`gpu` is not a model family. Torch CPU vs GPU: [GPU](#gpu).
-
-## GPU
-
-Family extras pull `torch`. Which wheel you get depends on the installer.
-
-**uv** defaults to the CPU index. Add `--extra gpu` for the PyPI wheel: CUDA on Linux and Windows, MPS on Apple silicon. Pair it with the family extras you need.
-
-```bash
-uv sync --extra server --extra hub --extra gpu
-```
-
-**pip** does not honor the `gpu` extra. `pip install -e ".[server,hub,gpu]"` is the same as without `gpu`. A normal pip install always takes CUDA torch from PyPI (MPS on macOS).
-
-To **force CPU torch with pip**, install torch from the CPU index first, then TServe. If a later `pip install` replaces that wheel with CUDA, run the torch line again.
-
-```bash
-pip install torch --index-url https://download.pytorch.org/whl/cpu
-pip install -e ".[server,hub]"
-```
-
-Swap `hub` for any other family extra from [Dependencies](#dependencies). Containers use `*-gpu` tags instead: [Docker](docker.md#gpu-images).
+A CPU build of torch: [CPU-only install](pip.md#cpu-only-install).
 
 --8<-- "includes/serve.md"
 
diff --git a/includes/model-dependencies.md b/includes/model-dependencies.md
index 715f42a..5894de2 100644
--- a/includes/model-dependencies.md
+++ b/includes/model-dependencies.md
@@ -1,8 +1,8 @@
 The extra name is the CPU image tag. GPU tags are `{extra}-gpu`. There is no `:base-gpu`.
 
-`kronos` is built on `base`, so it cannot load Chronos Bolt, TTM, or TimesFM. Every extra that pulls `hf` — `hub`, `chronos`, `granite`, `moirai`, `tirex`, `tirex2`, `toto`, `mantis`, `timesfm3`, `t0`, `tafsut`, and `full` — can.
+`kronos` is built on `base`, so it cannot load Chronos Bolt, TTM, or TimesFM. `hub` can, and so can every extra that includes it: `chronos`, `granite`, `moirai`, `tirex`, `tirex2`, `toto`, `mantis`, `timesfm3`, `t0`, `tafsut`, and `full`.
 
-`full` is `chronos`, `kronos`, `granite`, `moirai`, `tirex`, `tirex2`, `toto`, `mantis`, `timesfm3`, `t0`, and `tafsut`. `client`, `http`, `gpu`, `dev`, `docs`, and `all-extras` are not model families. `gpu` selects the torch index for uv on a clone only. Pip ignores `gpu` and installs CUDA torch from PyPI (MPS on macOS). Moirai pins `gluonts`, `lightning`, and `hydra-core` when `python_version < '3.14'`.
+`full` is `chronos`, `kronos`, `granite`, `moirai`, `tirex`, `tirex2`, `toto`, `mantis`, `timesfm3`, `t0`, and `tafsut`. `client`, `http`, `dev`, `docs`, and `all-extras` are not model families. Moirai pins `gluonts`, `lightning`, and `hydra-core` when `python_version < '3.14'`.
 
 **added** counts checkpoints that extra contributes. `full` is the total, including `naive`.
 
diff --git a/pyproject.toml b/pyproject.toml
index 3c3c079..3bcaa0b 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -42,17 +42,16 @@ sktime = [
   "sktime>=1.2.0",
   "skpro",
 ]
-# estimators-general-deps, torch excluded so CPU vs `--extra gpu` can pick an index
-hf = [
+# ChronosForecaster/TinyTimeMixerForecaster/TimesFM2Forecaster. Also shared stack for the families below.
+hub = [
   "tserve[sktime]",
   "transformers",
   "accelerate>=0.26.0",
+  "torch",
 ]
-# ChronosForecaster/TinyTimeMixerForecaster/TimesFM2Forecaster
-hub = ["tserve[hf]", "torch"]
 # Chronos2Forecaster
-chronos = ["tserve[hf]", "chronos-forecasting>=2.0.0", "torch"]
-# KronosForecaster/WindFMForecaster
+chronos = ["tserve[hub]", "chronos-forecasting>=2.0.0"]
+# KronosForecaster/WindFMForecaster. Not built on hub.
 kronos = [
   "tserve[sktime]",
   "einops",
@@ -61,38 +60,34 @@ kronos = [
   "torch",
 ]
 # FlowState/PatchTSMixer
-granite = ["tserve[hf]", "granite-tsfm>=0.3.5", "torch"]
+granite = ["tserve[hub]", "granite-tsfm>=0.3.5"]
 # MOIRAI/Moirai2/LagLlama
 moirai = [
-  "tserve[hf]",
+  "tserve[hub]",
   "gluonts>=0.14.0; python_version < '3.14'",
   "lightning>=2.0; python_version < '3.14'",
   "hydra-core; python_version < '3.14'",
   "einops",
-  "torch",
 ]
 # TiRexForecaster
-tirex = ["tserve[hf]", "tirex-ts", "torch"]
+tirex = ["tserve[hub]", "tirex-ts"]
 # TiRex2Forecaster
-tirex2 = ["tserve[hf]", "tirex-2", "torch"]
+tirex2 = ["tserve[hub]", "tirex-2"]
 # Toto2Forecaster (toto-ts for Toto v1 pins transformers and is omitted)
-toto = ["tserve[hf]", "toto-models", "skpro", "torch"]
+toto = ["tserve[hub]", "toto-models", "skpro"]
 # MantisForecaster
-mantis = ["tserve[hf]", "mantis-tsfm>=1.0.0", "torch"]
+mantis = ["tserve[hub]", "mantis-tsfm>=1.0.0"]
 # TimesFM3Forecaster
-timesfm3 = ["tserve[hf]", "timesfm[torch]>=3.0.0,<4.0.0", "torch"]
+timesfm3 = ["tserve[hub]", "timesfm[torch]>=3.0.0,<4.0.0"]
 # T0Forecaster
-t0 = ["tserve[hf]", "tfc-t0", "torch"]
+t0 = ["tserve[hub]", "tfc-t0"]
 # TafsutForecaster
-tafsut = ["tserve[hf]", "tafsut>=0.1.0", "torch"]
+tafsut = ["tserve[hub]", "tafsut>=0.1.0"]
 full = [
   "tserve[chronos,kronos,granite,moirai,tirex,tirex2,toto,mantis,timesfm3,t0,tafsut]",
-  "torch",
 ]
 # for pip (alternate for `uv sync --all-extras`)
 all-extras = ["tserve[client,server,full]"]
-# pair with any extra above: `uv sync --extra hub --extra gpu`
-gpu = ["torch"]
 # also extras so `pip install -e ".[dev]"` / `".[docs]"` work (no uv, no PEP 735)
 dev = [
   "pytest",
@@ -121,33 +116,6 @@ Changelog = "https://tserve.readthedocs.io/en/latest/reference/changelog/"
 # same packages as the extras, so `uv sync --group dev` stays valid
 dev = ["tserve[dev]"]
 docs = ["tserve[docs]"]
-# empty, only exists so the lock can fork CPU vs GPU torch
-cpu-fork = []
-
-[tool.uv]
-# pytorch-cpu also mirrors a few generic wheels at old versions; consider
-# both indexes so those packages still resolve from PyPI.
-index-strategy = "unsafe-best-match"
-# fork the lock: default extras use the CPU index, `--extra gpu` uses PyPI
-conflicts = [
-  [{ extra = "gpu" }, { group = "cpu-fork" }],
-]
-
-[tool.uv.sources]
-# `--extra gpu` takes PyPI torch (CUDA on Linux/Windows, MPS on macOS).
-# Without it, the extra CPU index below supplies the +cpu wheel.
-torch = [
-  { index = "pypi", extra = "gpu" },
-]
-
-[[tool.uv.index]]
-name = "pypi"
-url = "https://pypi.org/simple"
-default = true
-
-[[tool.uv.index]]
-name = "pytorch-cpu"
-url = "https://download.pytorch.org/whl/cpu"
 
 [tool.hatch.build]
 # repo .gitignore has `*.html`; hatchling would drop the dashboard from