gpu-dev 0.5.19__tar.gz → 0.5.21__tar.gz

{gpu_dev-0.5.19 → gpu_dev-0.5.21}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gpu-dev
-Version: 0.5.19
+Version: 0.5.21
 Summary: CLI tool for PyTorch GPU developer server reservations
 Author: PyTorch Team
 Requires-Python: >=3.10

{gpu_dev-0.5.19 → gpu_dev-0.5.21}/cli-tools/gpu-dev-cli/gpu_dev.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gpu-dev
-Version: 0.5.19
+Version: 0.5.21
 Summary: CLI tool for PyTorch GPU developer server reservations
 Author: PyTorch Team
 Requires-Python: >=3.10

{gpu_dev-0.5.19 → gpu_dev-0.5.21}/cli-tools/gpu-dev-cli/gpu_dev_cli/cli.py

@@ -1398,6 +1398,27 @@ def submit(ctx, gpu_type, gpus, hours, disk, no_persistent_disk, dockerfile, doc
         rprint("[red]❌ Provide a command after --, e.g. gpu-dev submit --runtime ./ -- python train.py[/red]")
         sys.exit(2)
 
+    # Catch the common typo where the user drops the leading -- on an option name and
+    # the option's value gets swept into the command (because submit accepts arbitrary
+    # commands via ignore_unknown_options). Without this guard the remote shell happily
+    # runs `gpus 1 bash run.sh` and the user wonders why.
+    _submit_flag_names = {"gpus", "gpu-type", "hours", "disk", "no-persistent-disk",
+                          "dockerfile", "dockerimage", "preserve-entrypoint", "runtime",
+                          "no-pull", "keep-alive", "name", "timeout"}
+    if command[0] in _submit_flag_names:
+        rprint(f"[red]❌ '{command[0]}' looks like a missing '--'. Did you mean '--{command[0]}'? "
+               f"Put your command after '--', e.g. gpu-dev submit --{command[0]} <value> ... -- bash run.sh[/red]")
+        sys.exit(2)
+
+    # rsync is on macOS by default and on virtually every Linux distro; bail early with a
+    # readable message if the user has somehow uninstalled it locally rather than failing
+    # mid-flight after the reservation has already been created.
+    if runtime:
+        import shutil
+        if not shutil.which("rsync"):
+            rprint("[red]❌ rsync not found on PATH locally. Install it (Mac: 'brew install rsync', Debian/Ubuntu: 'sudo apt install rsync') and retry.[/red]")
+            sys.exit(2)
+
     gt = gpu_type.lower()
     # Per-type max GPUs (mirrors gpu_configs in reserve flow)
     max_per_node = {

{gpu_dev-0.5.19 → gpu_dev-0.5.21}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "gpu-dev"
-version = "0.5.19"
+version = "0.5.21"
 description = "CLI tool for PyTorch GPU developer server reservations"
 authors = [{name = "PyTorch Team"}]
 readme = "cli-tools/gpu-dev-cli/README.md"

{gpu_dev-0.5.19 → gpu_dev-0.5.21}/terraform-gpu-devservers/docker/Dockerfile

@@ -6,15 +6,15 @@ FROM pytorch/pytorch:2.11.0-cuda12.8-cudnn9-devel
 ENV DEBIAN_FRONTEND=noninteractive
 ENV TZ=UTC
 
-# Update package lists with retries and install essential packages
+# Update apt and install base packages in a single layer so the package metadata
+# is always fresh relative to the install. Splitting update from install causes
+# Ubuntu to 404 on stale cached versions whenever security updates retire old .debs.
 RUN for attempt in 1 2 3; do \
         echo "Package update attempt $attempt..." && \
         apt-get update -qq && break || \
         ([ $attempt -lt 3 ] && echo "Update failed, waiting 30s..." && sleep 30) \
-    done
-
-# Install system packages in layers for better caching
-RUN apt-get install -y --no-install-recommends \
+    done && \
+    apt-get install -y --no-install-recommends \
         openssh-server \
         sudo \
         curl \
@@ -36,7 +36,8 @@ RUN apt-get install -y --no-install-recommends \
         unzip \
         ccache \
         htop \
-        tree
+        tree \
+        rsync
 # Install Node.js 20 from NodeSource (required for Claude CLI)
 RUN curl -fsSL https://deb.nodesource.com/setup_20.x | bash - && \
     apt-get install -y nodejs

{gpu_dev-0.5.19 → gpu_dev-0.5.21}/terraform-gpu-devservers/kubernetes.tf

@@ -416,6 +416,17 @@ resource "kubernetes_manifest" "image_prepuller_daemonset" {
           app = "image-prepuller"
         }
       }
+      # The prepuller does nothing but pull an image + sit in a pause container.
+      # Default maxUnavailable=1 makes a 32GB image roll out across 26 nodes take
+      # ~2.5h, so user pods landing on un-updated nodes still pull from ECR fresh.
+      # Parallel is safe here — restarting the prepuller doesn't disrupt anything,
+      # and the kubelet's image cache is independent of the prepuller pod's lifecycle.
+      updateStrategy = {
+        type = "RollingUpdate"
+        rollingUpdate = {
+          maxUnavailable = "100%"
+        }
+      }
       template = {
         metadata = {
           labels = {