diff --git a/.github/workflows/tests_unit.yml b/.github/workflows/tests_unit.yml
index ba08558c4..b6d1df49f 100644
--- a/.github/workflows/tests_unit.yml
+++ b/.github/workflows/tests_unit.yml
@@ -50,6 +50,24 @@ jobs:
           #
           poetry install --with dev
 
+      - name: Simple Template
+        run: |
+          source $(poetry env info -p)/bin/activate
+          milabench new --name simplebench --template simple
+          cd benchmarks/simplebench
+          make tests
+          cd ..
+          rm -rf simplebench
+        
+      - name: Voir Template
+        run: |
+          source $(poetry env info -p)/bin/activate
+          milabench new --name  voirbench --template voir
+          cd benchmarks/voirbench
+          make tests
+          cd ..
+          rm -rf voirbench
+        
       - name: tests
         run: |
           source $(poetry env info -p)/bin/activate
diff --git a/.gitignore b/.gitignore
index 5c661d462..f4ccd7a37 100644
--- a/.gitignore
+++ b/.gitignore
@@ -47,3 +47,8 @@ dry/
 
 stderr.txt
 stdout.txt
+
+base/
+
+benchmarks/simple
+benchmarks/voir
diff --git a/benchmarks/_templates/simple/Makefile b/benchmarks/_templates/simple/Makefile
new file mode 100644
index 000000000..f295ac5bb
--- /dev/null
+++ b/benchmarks/_templates/simple/Makefile
@@ -0,0 +1,4 @@
+tests:
+	milabench install --config dev.yaml --base base
+	milabench prepare --config dev.yaml --base base
+	milabench run --config dev.yaml --base base
diff --git a/benchmarks/_template/README.md b/benchmarks/_templates/simple/README.md
similarity index 83%
rename from benchmarks/_template/README.md
rename to benchmarks/_templates/simple/README.md
index bb348e387..239a2dfaf 100644
--- a/benchmarks/_template/README.md
+++ b/benchmarks/_templates/simple/README.md
@@ -1,4 +1,4 @@
 
-# {{STEM}}
+# Template
 
 Rewrite this README to explain what the benchmark is!
diff --git a/benchmarks/_template/benchfile.py b/benchmarks/_templates/simple/benchfile.py
similarity index 88%
rename from benchmarks/_template/benchfile.py
rename to benchmarks/_templates/simple/benchfile.py
index 633ad2235..08a51cef0 100644
--- a/benchmarks/_template/benchfile.py
+++ b/benchmarks/_templates/simple/benchfile.py
@@ -1,7 +1,7 @@
 from milabench.pack import Package
 
 
-class TheBenchmark(Package):
+class Template(Package):
     # Requirements file installed by install(). It can be empty or absent.
     base_requirements = "requirements.in"
 
@@ -26,8 +26,6 @@ async def install(self):
     async def prepare(self):
         await super().prepare()  # super() call executes prepare_script
 
-    async def run(self):
-        return await super().run()
 
 
-__pack__ = TheBenchmark
+__pack__ = Template
diff --git a/benchmarks/_template/dev.yaml b/benchmarks/_templates/simple/dev.yaml
similarity index 90%
rename from benchmarks/_template/dev.yaml
rename to benchmarks/_templates/simple/dev.yaml
index 93c661ad5..f25e977b2 100644
--- a/benchmarks/_template/dev.yaml
+++ b/benchmarks/_templates/simple/dev.yaml
@@ -1,5 +1,5 @@
 
-{{STEM}}:
+template:
   inherits: _defaults
   definition: .
   install-variant: unpinned
diff --git a/benchmarks/_template/main.py b/benchmarks/_templates/simple/main.py
similarity index 66%
rename from benchmarks/_template/main.py
rename to benchmarks/_templates/simple/main.py
index 99f0f0adc..b9fe484dd 100644
--- a/benchmarks/_template/main.py
+++ b/benchmarks/_templates/simple/main.py
@@ -5,33 +5,45 @@
 # be deleted.
 
 import time
+import random
 
 import torchcompat.core as accelerator
 from benchmate.observer import BenchObserver
 
 
+def criterion(*args, **kwargs):
+    return random.normalvariate()
+
+
 def main():
     device = accelerator.fetch_device(0) # <= This is your cuda device
-    
-    observer = BenchObserver(batch_size_fn=lambda batch: 1)
+
+    observer = BenchObserver(
+        batch_size_fn=lambda batch: 1
+    )
     # optimizer = observer.optimizer(optimizer)
     # criterion = observer.criterion(criterion)
     
-    dataloader = [1, 2, 3, 4]
+    dataloader = list(range(6000))
     
-    for epoch in range(10):
+    for epoch in range(10000):
         for i in observer.iterate(dataloader):
             # avoid .item()
             # avoid torch.cuda; use accelerator from torchcompat instead
             # avoid torch.cuda.synchronize or accelerator.synchronize
             
             # y = model(i)
-            # loss = criterion(y)
+            loss = criterion()
             # loss.backward()
             # optimizer.step()
+
+            observer.record_loss(loss)
             
             time.sleep(0.1)
 
+    assert epoch < 2, "milabench stopped the train script before the end of training"
+    assert i < 72, "milabench stopped the train script before the end of training"
+
 
 if __name__ == "__main__":
     main()
diff --git a/benchmarks/_template/prepare.py b/benchmarks/_templates/simple/prepare.py
similarity index 100%
rename from benchmarks/_template/prepare.py
rename to benchmarks/_templates/simple/prepare.py
diff --git a/benchmarks/_template/requirements.in b/benchmarks/_templates/simple/requirements.in
similarity index 77%
rename from benchmarks/_template/requirements.in
rename to benchmarks/_templates/simple/requirements.in
index cd5826d4a..94575179d 100644
--- a/benchmarks/_template/requirements.in
+++ b/benchmarks/_templates/simple/requirements.in
@@ -1 +1,2 @@
 voir>=0.2.9,<0.3
+torch
\ No newline at end of file
diff --git a/benchmarks/_template/voirfile.py b/benchmarks/_templates/simple/voirfile.py
similarity index 76%
rename from benchmarks/_template/voirfile.py
rename to benchmarks/_templates/simple/voirfile.py
index 8008df299..d93f886cd 100644
--- a/benchmarks/_template/voirfile.py
+++ b/benchmarks/_templates/simple/voirfile.py
@@ -26,11 +26,6 @@ class Config:
 
 @configurable
 def instrument_main(ov, options: Config):
-    try:
-        import torch
-    except ImportError:
-        torch = None
-
     yield ov.phases.init
 
     if options.dash:
@@ -38,13 +33,6 @@ def instrument_main(ov, options: Config):
 
     ov.require(
         log("value", "progress", "rate", "units", "loss", "gpudata", context="task"),
-        rate(
-            interval=options.interval,
-            skip=options.skip,
-            sync=torch.cuda.synchronize
-            if torch and torch.cuda.is_available()
-            else None,
-        ),
         early_stop(n=options.stop, key="rate", task="train"),
         monitor_monogpu(poll_interval=options.gpu_poll),
     )
diff --git a/benchmarks/_templates/voir/Makefile b/benchmarks/_templates/voir/Makefile
new file mode 100644
index 000000000..f295ac5bb
--- /dev/null
+++ b/benchmarks/_templates/voir/Makefile
@@ -0,0 +1,4 @@
+tests:
+	milabench install --config dev.yaml --base base
+	milabench prepare --config dev.yaml --base base
+	milabench run --config dev.yaml --base base
diff --git a/benchmarks/_templates/voir/README.md b/benchmarks/_templates/voir/README.md
new file mode 100644
index 000000000..239a2dfaf
--- /dev/null
+++ b/benchmarks/_templates/voir/README.md
@@ -0,0 +1,4 @@
+
+# Template
+
+Rewrite this README to explain what the benchmark is!
diff --git a/benchmarks/_templates/voir/benchfile.py b/benchmarks/_templates/voir/benchfile.py
new file mode 100644
index 000000000..4ab9e833f
--- /dev/null
+++ b/benchmarks/_templates/voir/benchfile.py
@@ -0,0 +1,42 @@
+from milabench.pack import Package
+
+
+SOURCE_DIR = "src"
+REPO_URL = "https://github.com/Delaunay/extern_example.git"
+BRANCH = "a524286ab6364bca6729dd6ef4936e175a87c7e4"
+
+
+class Template(Package):
+    # Requirements file installed by install(). It can be empty or absent.
+    base_requirements = "requirements.in"
+
+    # The preparation script called by prepare(). It must be executable,
+    # but it can be any type of script. It can be empty or absent.
+    prepare_script = "prepare.py"
+
+    # The main script called by run(). It must be a Python file. It has to
+    # be present.
+    main_script = f"{SOURCE_DIR}/main.py"
+
+    # You can remove the functions below if you don't need to modify them.
+
+    def make_env(self):
+        # Return a dict of environment variables for prepare_script and
+        # main_script.
+        return super().make_env()
+
+    async def install(self):
+        await super().install()
+
+        source_destination = self.dirs.code / SOURCE_DIR
+        if not source_destination.exists():
+            source_destination.clone_subtree(
+                REPO_URL, BRANCH
+            )
+
+    async def prepare(self):
+        await super().prepare()  # super() call executes prepare_script
+
+
+
+__pack__ = Template
diff --git a/benchmarks/_templates/voir/dev.yaml b/benchmarks/_templates/voir/dev.yaml
new file mode 100644
index 000000000..f25e977b2
--- /dev/null
+++ b/benchmarks/_templates/voir/dev.yaml
@@ -0,0 +1,7 @@
+
+template:
+  inherits: _defaults
+  definition: .
+  install-variant: unpinned
+  plan:
+    method: per_gpu
diff --git a/benchmarks/_templates/voir/prepare.py b/benchmarks/_templates/voir/prepare.py
new file mode 100755
index 000000000..32bd5901d
--- /dev/null
+++ b/benchmarks/_templates/voir/prepare.py
@@ -0,0 +1,16 @@
+#!/usr/bin/env python
+
+import os
+
+if __name__ == "__main__":
+    # If you need the whole configuration:
+    # config = json.loads(os.environ["MILABENCH_CONFIG"])
+
+    data_directory = os.environ["MILABENCH_DIR_DATA"]
+
+    # Download (or generate) the needed dataset(s). You are responsible
+    # to check if it has already been properly downloaded or not, and to
+    # do nothing if it has been.
+    print("Hello I am doing some data stuff!")
+
+    # If there is nothing to download or generate, just delete this file.
diff --git a/benchmarks/_templates/voir/requirements.in b/benchmarks/_templates/voir/requirements.in
new file mode 100644
index 000000000..94575179d
--- /dev/null
+++ b/benchmarks/_templates/voir/requirements.in
@@ -0,0 +1,2 @@
+voir>=0.2.9,<0.3
+torch
\ No newline at end of file
diff --git a/benchmarks/_templates/voir/voirfile.py b/benchmarks/_templates/voir/voirfile.py
new file mode 100644
index 000000000..854ced70b
--- /dev/null
+++ b/benchmarks/_templates/voir/voirfile.py
@@ -0,0 +1,69 @@
+from dataclasses import dataclass
+
+from voir.phase import StopProgram
+from voir import configurable
+from voir.instruments import dash, early_stop, log
+from benchmate.monitor import monitor_monogpu
+from benchmate.observer import BenchObserver
+
+
+@dataclass
+class Config:
+    """voir configuration"""
+
+    # Whether to display the dash or not
+    dash: bool = False
+
+    # How often to log the rates
+    interval: str = "1s"
+
+    # Number of rates to skip before logging
+    skip: int = 5
+
+    # Number of rates to log before stopping
+    stop: int = 20
+
+    # Number of seconds between each gpu poll
+    gpu_poll: int = 3
+
+
+@configurable
+def instrument_main(ov, options: Config):
+    yield ov.phases.init
+
+
+    yield ov.phases.load_script
+
+    if options.dash:
+        ov.require(dash)
+
+    ov.require(
+        log("value", "progress", "rate", "units", "loss", "gpudata", context="task"),
+        early_stop(n=options.stop, key="rate", task="train"),
+        monitor_monogpu(poll_interval=options.gpu_poll),
+    )
+
+    #
+    # Insert milabench tools
+    #
+    observer = BenchObserver(
+        earlystop=options.stop + options.skip,
+        batch_size_fn=lambda x: 1
+    )
+
+    probe = ov.probe("//my_dataloader_creator() as loader", overridable=True)
+    probe['loader'].override(observer.loader)
+
+    probe = ov.probe("//my_criterion_creator() as criterion", overridable=True)
+    probe['criterion'].override(observer.criterion)
+
+    probe = ov.probe("//my_optimizer_creator() as optimizer", overridable=True)
+    probe['optimizer'].override(observer.optimizer)
+    
+    #
+    # Run the benchmark
+    #
+    try:
+        yield ov.phases.run_script
+    except StopProgram:
+        print("early stopped")
\ No newline at end of file
diff --git a/benchmarks/torchvision_ddp/benchfile.py b/benchmarks/torchvision_ddp/benchfile.py
index 3fc2cb054..4b2d56c8e 100644
--- a/benchmarks/torchvision_ddp/benchfile.py
+++ b/benchmarks/torchvision_ddp/benchfile.py
@@ -8,7 +8,6 @@ class TorchvisionBenchmarkDDP(Package):
 
     def build_run_plan(self) -> "Command":
         import milabench.commands as cmd
-
         pack = cmd.PackCommand(self, *self.argv, lazy=True)
         pack = cmd.ActivatorCommand(pack, use_stdout=True)
         return pack
diff --git a/benchmate/benchmate/observer.py b/benchmate/benchmate/observer.py
index 0ada095b4..83a6dfa08 100644
--- a/benchmate/benchmate/observer.py
+++ b/benchmate/benchmate/observer.py
@@ -68,7 +68,7 @@ def override_return_value(self, function, override):
             raise RuntimeError("Not running through voir")
 
     def iterate(self, iterator):
-        return self.loader(loader)
+        return self.loader(iterator)
     
     def loader(self, loader):
         """Wrap a dataloader or an iterable which enable accurate measuring of time spent in the loop's body"""
@@ -94,7 +94,7 @@ def new_backward(*args, **kwargs):
 
                 loss.backward = new_backward
 
-            self.record_loss(loss.detach())
+            self.record_loss(loss)
             return loss
 
         return wrapped
diff --git a/docs/index.rst b/docs/index.rst
index 3ac990fcf..ebbb27383 100644
--- a/docs/index.rst
+++ b/docs/index.rst
@@ -7,12 +7,15 @@ Welcome to milabench's documentation!
    :caption: Contents:
 
    usage.rst
+   recipes.rst
+   new_benchmarks.rst
+
    docker.rst
    dev-usage.rst
-   new_benchmarks.rst
    reference.rst
    sizer.rst
 
+
 Indices and tables
 ==================
 
diff --git a/docs/new_benchmarks.rst b/docs/new_benchmarks.rst
index 058b99c0d..e348a28be 100644
--- a/docs/new_benchmarks.rst
+++ b/docs/new_benchmarks.rst
@@ -2,23 +2,30 @@
 Creating a new benchmark
 ------------------------
 
-To define a new benchmark (let's assume it is called ``ornatebench``), make a copy of ``benchmarks/_template`` using ``cp-template``:
+To define a new benchmark (let's assume it is called ``ornatebench``), 
 
 .. code-block:: bash
 
-    cp-template benchmarks/_template/ benchmarks/ornatebench
+    git clone https://github.com/mila-iqia/milabench.git
+    git checkout -b ornatebench
+
+    pip install -e milabench/
+    milabench new --name ornatebench
+
 
 You should see a directory with the following structure:
 
 .. code-block::
 
-    ornatebench
-    |- README.md          # Document the benchmark here
-    |- benchfile.py       # Benchmark definition file
-    |- main.py            # Executed by milabench run
-    |- prepare.py         # Executed by milabench prepare (EXECUTABLE)
-    |- requirements.in    # Python requirements to install from pip
-    |- voirfile.py        # Probes and extra instruments
+    milabench
+    └── benchmarks
+        └── ornatebench
+            ├── README.md          # Document the benchmark here
+            ├── benchfile.py       # Benchmark definition file
+            ├── main.py            # Executed by milabench run
+            ├── prepare.py         # Executed by milabench prepare (EXECUTABLE)
+            ├── requirements.in    # Python requirements to install from pip
+            └── voirfile.py        # Probes and extra instruments
 
 Some of these files may be unnecessary depending on the benchmark.
 
@@ -26,30 +33,37 @@ First of all, if you want to verify that everything works, you can use the ``dev
 
 .. code-block:: bash
 
-    # You can also use --config
-    export MILABENCH_CONFIG=benchmarks/ornatebench/dev.yaml
+    cd milabench/benchmarks/ornatebench
+
+    milabench install --config dev.yaml --base .
 
-    milabench install
-    milabench prepare
-    milabench run
+    milabench prepare --config dev.yaml --base .
+
+    milabench run     --config dev.yaml --base .
 
 
 Overview
 ~~~~~~~~
 
-
 benchfile.py
 ++++++++++++
 
-``benchfile.py`` defines what to do on ``milabench install/prepare/run``. It is run from the benchmark directory directly, in the *current* virtual environment, but it can create *new processes* in the virtual environment of the benchmark.
+``benchfile.py`` defines what to do on ``milabench install/prepare/run``. 
+It is run from the benchmark directory directly, in the *current* virtual environment, 
+but it can create *new processes* in the virtual environment of the benchmark.
 
-By default it will dispatch to ``requirements.in`` for install requirements, ``prepare.py`` for prep work and downloading datasets, and ``main.py`` for running the actual benchmark. If that is suitable you may not need to change it at all.
+By default it will dispatch to ``requirements.in`` for install requirements, 
+``prepare.py`` for prep work and downloading datasets, and
+``main.py`` for running the actual benchmark. 
+If that is suitable you may not need to change it at all.
 
 
 requirements.in
 +++++++++++++++
 
-Write all of the benchmark's requirements in this file. Use ``milabench install --config benchmarks/ornatebench/dev.yaml`` to install them during development (add ``--force`` if you made changes and want to reinstall.)
+Write all of the benchmark's requirements in this file. 
+Use ``milabench install --config benchmarks/ornatebench/dev.yaml`` 
+to install them during development (add ``--force`` if you made changes and want to reinstall.)
 
 
 prepare.py
@@ -123,7 +137,6 @@ voirfile.py
 
 The voirfile contains instrumentation for the main script. You can usually just leave it as it is. By default, it will:
 
-* Compute the train "rate" (number of samples per second) using events from ``voir.iterate``.
 * Forcefully stop the program after a certain number of rate measurements.
 * Monitor GPU usage.
 
@@ -131,9 +144,11 @@ The voirfile contains instrumentation for the main script. You can usually just
 Development
 ~~~~~~~~~~~
 
-To develop the benchmark, first run ``milabench dev --config benchmarks/BENCHNAME/dev.yaml``. This will activate the benchmark's virtual environment and put you into a shell.
+To develop the benchmark, first run ``milabench dev --config benchmarks/BENCHNAME/dev.yaml``. 
+This will activate the benchmark's virtual environment and put you into a shell.
 
-Then, try and run ``voir --dash main.py``. This should show you a little dashboard and display losses, train rate calculations and one or more progress bars.
+Then, try and run ``voir --dash main.py``. This should show you a little dashboard and display losses, 
+train rate calculations and one or more progress bars.
 
 From there, you can develop as you would any other Python program.
 
@@ -159,9 +174,43 @@ This will create ``requirements.<arch>.txt`` for these two architectures. These
     ``--variant unpinned`` means installing directly from ``requirements.in``. This can be useful during development, but less stable over time since various dependencies may break.
 
 
-.. Adapting existing code
-.. ~~~~~~~~~~~~~~~~~~~~~~
+Adapting existing repository
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+To simplify the creation of benchmarks, milabench can use ptera and voir to override or wrap code from a third party
+without modifying the third party code.
+
+.. code-block:: bash
+
+    git clone https://github.com/mila-iqia/milabench.git
+    git checkout -b ornatebench
+
+    pip install -e milabench/
+    milabench new --name ornatebench --repo-url https://github.com/Delaunay/extern_example.git
+
+
+The instrumentation is inserted inside ``voirfile.py`` using the ``Overseer.probe``, examples can 
+be found in ptera `documentation <https://ptera.readthedocs.io/en/latest/guide.html#probing>`_
+
+
+Wrap a return value
++++++++++++++++++++
+
+
+.. code-block:: python
+
+    class Wrapper:
+        def __init__(self, value):
+            self.value = value
+
+    def wrap(original):
+        return Wrapper(original)
+
+    probe = ov.probe("//my_optimizer_creator() as optimizer", overridable=True)
+    probe['optimizer'].override(wrap)
+
 
-.. Now, let's say you want to adapt code from the repo at ``https://github.com/snakeoilplz/agi``, more specifically the ``train.py`` script.
+* ``//my_optimizer_creator() as optimizer``: get the return value of a function inside the main script
+* ``/module.path.function() as optimizer``: get the return value of a function inside a module
+* ``/module.path.function > loss_fn``: get a variable inside a function inside a module
 
-.. TODO
diff --git a/docs/start.rst b/docs/recipes.rst
similarity index 99%
rename from docs/start.rst
rename to docs/recipes.rst
index ce5c8b8d3..91b3f9620 100644
--- a/docs/start.rst
+++ b/docs/recipes.rst
@@ -1,6 +1,5 @@
-
-Getting Started
-===============
+Running Milabench
+=================
 
 Base Setup
 ----------
@@ -12,17 +11,22 @@ Base Setup
    mkdir milabench
    cd milabench
    git clone https://github.com/mila-iqia/milabench.git
+
    conda activate base
    python --version
    Python 3.11.4
+
    virtualenv ./env
    source ./env/bin/activate
    pip install -e milabench/
+
    export MILABENCH_WORDIR="$(pwd)"
    export MILABENCH_BASE="$MILABENCH_WORDIR/results"
    export MILABENCH_CONFIG="$MILABENCH_WORDIR/milabench/config/standard.yaml"
    export BENCHMARK_VENV="$MILABENCH_WORDIR/results/venv/torch"
+
    module load cuda/12.3.2                                          # <= or set CUDA_HOME to the right spot
+   
    milabench install
    milabench prepare
    milabench run
diff --git a/milabench/cli/__init__.py b/milabench/cli/__init__.py
index e0d57c1e9..205942e47 100644
--- a/milabench/cli/__init__.py
+++ b/milabench/cli/__init__.py
@@ -20,9 +20,14 @@
 from .sql import cli_sqlsetup
 from .summary import cli_summary
 from .resolve import cli_resolve
+from .new import cli_new
 
 
 class Main:
+    def new():
+        """Create a new benchmark from template"""
+        return cli_new()
+
     def run():
         """Run the benchmarks."""
         return cli_run()
diff --git a/milabench/cli/new.py b/milabench/cli/new.py
new file mode 100644
index 000000000..9f80c5629
--- /dev/null
+++ b/milabench/cli/new.py
@@ -0,0 +1,115 @@
+
+from dataclasses import dataclass
+import os
+import pathlib
+import argparse
+
+from coleo import Option, tooled
+
+benchmark = (pathlib.Path(__file__).parent / '..' / '..' / "benchmarks").resolve()
+template =  benchmark / "_templates"
+
+
+multigpu = "\n".join([
+    "method: njobs", 
+    "    n: 1\n",
+])
+
+multinode = "\n".join([
+    "  num_machines: 2",
+    "  requires_capabilities:",
+    "    - \"len(nodes) >= ${num_machines}\"\n",
+])
+
+placeholder_repo = "https://github.com/Delaunay/extern_example.git"
+
+
+# fmt: off
+@dataclass
+class Arguments:
+    name       : str
+    template   : str = "simple"
+    repo_url   : str = None
+    multi_gpu  : bool = False
+    multi_node : bool = False
+# fmt: on
+
+
+@tooled
+def arguments():
+    # Name of the benchmark
+    name: Option & str
+
+    # Number of times to repeat the benchmark
+    template: Option & str = "simple"
+
+    # Repo URL to clone
+    repo_url: Option & str = None
+
+    # is benchmark multi gpu
+    multi_gpu: Option & bool = False
+
+    # is the benchmark is multi node
+    multi_node: Option & bool = False
+
+    return Arguments(name, template, repo_url, multi_gpu, multi_node)
+
+
+@tooled
+def cli_new(args=None):
+    """Create a new benchmark from the template"""
+
+    if args is None:
+        args = arguments()
+
+    if args.repo_url is not None:
+        args.template = "voir"
+
+    package_name =  args.name.capitalize()
+
+    template_dir = template / args.template
+    destination = benchmark / args.name
+    os.makedirs(destination, exist_ok=True)
+
+    for file in os.listdir(template_dir):
+        if file in ("base",):
+            continue
+
+        source = template_dir / file
+        dest = destination / file
+
+        with open(source, "r") as fp:
+            content = fp.read()
+
+        placeholders = [
+            ("Template", package_name),
+            ("template", args.name),
+            (placeholder_repo, args.repo_url),
+        ]
+
+        if args.multi_gpu:
+            placeholders.append(("method: per_gpu\n", multigpu))
+
+        if args.repo_url:
+            placeholders.append(("method: per_gpu\n",  args.repo_url))
+
+        if args.multi_node:
+            placeholders((None, multinode))
+
+        for placeholder, value in placeholders:
+            if value is not None:
+                if placeholder is not None:
+                    content = content.replace(placeholder, value)
+                else:
+                    content += value
+        
+        with open(dest, "w") as fp:
+            fp.write(content)
+
+        st = os.stat(source)
+        os.chown(dest, st.st_uid, st.st_gid)
+        os.chmod(dest, st.st_mode)
+
+
+if __name__ == "__main__":
+    cli_new()
diff --git a/milabench/commands/__init__.py b/milabench/commands/__init__.py
index f6f19a87a..76bec08b9 100644
--- a/milabench/commands/__init__.py
+++ b/milabench/commands/__init__.py
@@ -612,7 +612,7 @@ class PerGPU(ListCommand):
     """
 
     def __init__(self, executor: Command, gpus: list = None, **kwargs) -> None:
-        if gpus is None:
+        if gpus is None or len(gpus) == 0:
             gpus = [{"device": 0, "selection_variable": "CPU_VISIBLE_DEVICE"}]
 
         self.devices = gpus
diff --git a/milabench/pack.py b/milabench/pack.py
index 5b30d05b8..5ad45ccff 100644
--- a/milabench/pack.py
+++ b/milabench/pack.py
@@ -374,7 +374,7 @@ async def install(self):
         """
         assert self.phase == "install"
 
-        for reqs in self.requirements_files(self.config.get("install_variant", None)):
+        for reqs in self.requirements_files(self.config.get("install-variant", None)):
             if reqs.exists():
                 await self.pip_install("-r", reqs)
             else: