gpu-dev 0.4.1__tar.gz → 0.5.0__tar.gz

{gpu_dev-0.4.1 → gpu_dev-0.5.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gpu-dev
-Version: 0.4.1
+Version: 0.5.0
 Summary: CLI tool for PyTorch GPU developer server reservations
 Author: PyTorch Team
 Requires-Python: >=3.10

{gpu_dev-0.4.1 → gpu_dev-0.5.0}/cli-tools/gpu-dev-cli/gpu_dev.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gpu-dev
-Version: 0.4.1
+Version: 0.5.0
 Summary: CLI tool for PyTorch GPU developer server reservations
 Author: PyTorch Team
 Requires-Python: >=3.10

{gpu_dev-0.4.1 → gpu_dev-0.5.0}/cli-tools/gpu-dev-cli/gpu_dev_cli/cli.py

@@ -498,9 +498,9 @@ def main(ctx: click.Context) -> None:
     "--gpu-type",
     "-t",
     type=click.Choice(
-        ["b200", "h200", "h100", "a100", "a10g", "t4", "l4", "t4-small", "cpu-arm", "cpu-x86"], case_sensitive=False
+        ["b200", "h200", "h100", "a100", "g7e", "a10g", "t4", "l4", "t4-small", "cpu-arm", "cpu-x86"], case_sensitive=False
     ),
-    help="GPU type to reserve (b200/h200/h100/a100/a10g/t4/l4/t4-small/cpu-arm/cpu-x86)",
+    help="GPU type to reserve (b200/h200/h100/a100/g7e/a10g/t4/l4/t4-small/cpu-arm/cpu-x86)",
 )
 @click.option(
     "--hours",
@@ -652,6 +652,7 @@ def reserve(
             "t4": {"max_gpus": 4, "instance_type": "g4dn.12xlarge"},
             "l4": {"max_gpus": 4, "instance_type": "g6.12xlarge"},
             "a10g": {"max_gpus": 4, "instance_type": "g5.12xlarge"},
+            "g7e": {"max_gpus": 4, "instance_type": "g7e.24xlarge"},
             "t4-small": {"max_gpus": 1, "instance_type": "g4dn.xlarge"},
             "a100": {"max_gpus": 8, "instance_type": "p4d.24xlarge"},
             "h100": {"max_gpus": 8, "instance_type": "p5.48xlarge"},
@@ -2397,6 +2398,7 @@ def _show_availability() -> None:
                 "a100": "Ampere (sm80)",
                 "a10g": "Ampere (sm80)",
                 "l4": "Ada Lovelace (sm89)",
+                "g7e": "Blackwell (sm120)",
                 "t4": "Turing (sm75)",
                 "cpu-x86": "CPU (x86_64)",
                 "cpu-arm": "CPU (arm64)",
@@ -2405,6 +2407,7 @@ def _show_availability() -> None:
             # Sort order: newest GPU architectures first, then CPUs at the bottom
             arch_priority = {
                 "Blackwell (sm100)": 0,
+                "Blackwell (sm120)": 0,
                 "Hopper (sm90)": 1,
                 "Ada Lovelace (sm89)": 2,
                 "Ampere (sm80)": 3,
@@ -2544,6 +2547,7 @@ def _show_availability_watch(interval: int) -> None:
                             "a100": "Ampere (sm80)",
                             "a10g": "Ampere (sm80)",
                             "l4": "Ada Lovelace (sm89)",
+                            "g7e": "Blackwell (sm120)",
                             "t4": "Turing (sm75)",
                             "cpu-x86": "CPU (x86_64)",
                             "cpu-arm": "CPU (arm64)",
@@ -2552,6 +2556,7 @@ def _show_availability_watch(interval: int) -> None:
                         # Sort order: newest GPU architectures first, then CPUs at the bottom
                         arch_priority = {
                             "Blackwell (sm100)": 0,
+                            "Blackwell (sm120)": 0,
                             "Hopper (sm90)": 1,
                             "Ada Lovelace (sm89)": 2,
                             "Ampere (sm80)": 3,

{gpu_dev-0.4.1 → gpu_dev-0.5.0}/cli-tools/gpu-dev-cli/gpu_dev_cli/interactive.py

@@ -153,7 +153,7 @@ def select_gpu_count_interactive(gpu_type: str, max_gpus: int) -> Optional[int]:
         # CPU instances don't have GPUs, but we still need a "count" for nodes
         valid_counts = [0]  # 0 GPUs for CPU-only instances
         multinode_counts = []  # No multinode for CPU instances
-    elif gpu_type in ["t4", "l4", "a10g"]:
+    elif gpu_type in ["t4", "l4", "a10g", "g7e"]:
         valid_counts = [1, 2, 4]
         # Add multinode options
         multinode_counts = [8, 12, 16, 20, 24]  # multiples of 4

{gpu_dev-0.4.1 → gpu_dev-0.5.0}/cli-tools/gpu-dev-cli/gpu_dev_cli/reservations.py

@@ -535,6 +535,7 @@ class ReservationManager:
                 "t4": {"max_gpus": 4},
                 "l4": {"max_gpus": 4},
                 "a10g": {"max_gpus": 4},
+                "g7e": {"max_gpus": 4},
                 "t4-small": {"max_gpus": 1},
                 "g5g": {"max_gpus": 2},
                 "a100": {"max_gpus": 8},

{gpu_dev-0.4.1 → gpu_dev-0.5.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "gpu-dev"
-version = "0.4.1"
+version = "0.5.0"
 description = "CLI tool for PyTorch GPU developer server reservations"
 authors = [{name = "PyTorch Team"}]
 readme = "cli-tools/gpu-dev-cli/README.md"

{gpu_dev-0.4.1 → gpu_dev-0.5.0}/terraform-gpu-devservers/docker/Dockerfile

@@ -1,6 +1,6 @@
 # Custom PyTorch GPU Development Server Image
-# Based on pytorch/pytorch:2.9.1-cuda12.8-cudnn9-devel
-FROM pytorch/pytorch:2.9.1-cuda12.8-cudnn9-devel
+# Based on pytorch/pytorch:2.11.0-cuda12.8-cudnn9-devel
+FROM pytorch/pytorch:2.11.0-cuda12.8-cudnn9-devel
 
 # Set environment variables for non-interactive installation
 ENV DEBIAN_FRONTEND=noninteractive
@@ -41,23 +41,22 @@ RUN apt-get install -y --no-install-recommends \
 RUN curl -fsSL https://deb.nodesource.com/setup_20.x | bash - && \
     apt-get install -y nodejs
 
-# Install CUDA 13.0 alongside existing CUDA 12.9
+# Install CUDA 12.9, 13.0, 13.1, 13.2 alongside base CUDA 12.8
+# Base image already has NVIDIA repo configured, no need for cuda-keyring
 RUN apt-get update && apt-get install -y --no-install-recommends \
-        software-properties-common \
-    && wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/cuda-keyring_1.0-1_all.deb \
-    && dpkg -i cuda-keyring_1.0-1_all.deb \
-    && apt-get update \
-    && apt-get install -y --no-install-recommends \
+        cuda-toolkit-12-9 \
         cuda-toolkit-13-0 \
-    && rm cuda-keyring_1.0-1_all.deb \
+        cuda-toolkit-13-1 \
+        cuda-toolkit-13-2 \
     && apt-get clean \
     && rm -rf /var/lib/apt/lists/*
 
-# Set CUDA paths for both versions - 12.8 as default for PyTorch compatibility
-ENV CUDA_12_PATH=/usr/local/cuda-12.8
-ENV CUDA_13_PATH=/usr/local/cuda-13.0
-ENV PATH=/usr/local/cuda-12.8/bin:/usr/local/cuda-13.0/bin:${PATH}
-ENV LD_LIBRARY_PATH=/usr/local/cuda-12.8/lib64:/usr/local/cuda-13.0/lib64:${LD_LIBRARY_PATH}
+# CUDA 12.8 is the default (PyTorch compiled against it)
+# All versions available at /usr/local/cuda-{12.8,12.9,13.0,13.1,13.2}/
+# Switch with: export CUDA_HOME=/usr/local/cuda-13.2
+ENV CUDA_HOME=/usr/local/cuda-12.8
+ENV PATH=/usr/local/cuda-12.8/bin:${PATH}
+ENV LD_LIBRARY_PATH=/usr/local/cuda-12.8/lib64:${LD_LIBRARY_PATH}
 
 # Install EFA stack (prebuilt libfabric + OpenMPI + aws-ofi-nccl with GPU/RDMA support)
 # Uses AWS EFA installer which bundles tested, compatible versions of all components
@@ -80,10 +79,12 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     && apt-get clean && rm -rf /var/lib/apt/lists/*
 
 # Clone and build NCCL tests with MPI support for multi-node benchmarking
+RUN apt-get update && apt-get install -y --no-install-recommends libnccl-dev \
+    && apt-get clean && rm -rf /var/lib/apt/lists/*
 RUN cd /opt && \
     git clone https://github.com/NVIDIA/nccl-tests.git && \
     cd nccl-tests && \
-    make MPI=1 MPI_HOME=/opt/amazon/openmpi CUDA_HOME=/usr/local/cuda NCCL_HOME=/usr/lib/x86_64-linux-gnu -j$(nproc)
+    make MPI=1 MPI_HOME=/opt/amazon/openmpi CUDA_HOME=/usr/local/cuda NCCL_HOME=/usr -j$(nproc)
 
 # Set environment variables for EFA and NCCL
 ENV PATH=/opt/amazon/openmpi/bin:/opt/amazon/efa/bin:${PATH}
@@ -101,7 +102,7 @@ ENV NCCL_ASYNC_ERROR_HANDLING=1
 ENV SUPPORTS_EFA=true
 
 # Install Python packages (Jupyter and common ML packages)
-RUN pip install --no-cache-dir \
+RUN pip install --no-cache-dir --break-system-packages \
         jupyterlab \
         ipywidgets \
         matplotlib \

{gpu_dev-0.4.1 → gpu_dev-0.5.0}/terraform-gpu-devservers/docker/shell_env

@@ -1,10 +1,11 @@
 # Clean PATH setup (no duplicates)
 export PATH="/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin"
 
-# CUDA environment
-export CUDA_HOME=/usr/local/cuda
-export PATH="/usr/local/cuda/bin:$PATH"
-export LD_LIBRARY_PATH="/usr/local/cuda/lib64:${LD_LIBRARY_PATH:-}"
+# CUDA environment (12.8 default, also available: 12.9, 13.0, 13.1, 13.2)
+# Switch with: export CUDA_HOME=/usr/local/cuda-13.2 && export PATH="$CUDA_HOME/bin:$PATH"
+export CUDA_HOME=/usr/local/cuda-12.8
+export PATH="$CUDA_HOME/bin:$PATH"
+export LD_LIBRARY_PATH="$CUDA_HOME/lib64:${LD_LIBRARY_PATH:-}"
 
 # EFA and OpenMPI environment for multi-node GPU communication
 export PATH="/opt/amazon/openmpi/bin:/opt/amazon/efa/bin:$PATH"
@@ -31,5 +32,4 @@ export CCACHE_DIR="/ccache_shared"
 export CCACHE_MAXSIZE="10G"
 
 # Claude Code configuration for Bedrock
-export CLAUDE_CODE_USE_BEDROCK=1
-export ANTHROPIC_MODEL="us.anthropic.claude-sonnet-4-20250514-v1:0"
+export CLAUDE_CODE_USE_BEDROCK=1

{gpu_dev-0.4.1 → gpu_dev-0.5.0}/terraform-gpu-devservers/eks.tf

@@ -83,13 +83,13 @@ resource "aws_iam_role_policy" "eks_node_bedrock_policy" {
         Effect = "Allow"
         Action = [
           "bedrock:InvokeModel",
-          "bedrock:InvokeModelWithResponseStream"
-        ]
-        Resource = [
-          "arn:aws:bedrock:*:*:foundation-model/anthropic.claude-*",
-          "arn:aws:bedrock:*:*:inference-profile/us.anthropic.claude-*",
-          "arn:aws:bedrock:*:*:inference-profile/global.anthropic.claude-*"
+          "bedrock:InvokeModelWithResponseStream",
+          "bedrock:ListInferenceProfiles",
+          "bedrock:GetInferenceProfile",
+          "bedrock:ListFoundationModels",
+          "bedrock-mantle:*"
         ]
+        Resource = "*"
       },
       {
         Effect = "Allow"
@@ -185,6 +185,7 @@ locals {
     "t4-az2"   = "t4" # Both t4 and t4-az2 should be labeled as "t4" in Kubernetes
     "l4"       = "l4"
     "a10g"     = "a10g"
+    "g7e"      = "g7e"
     "h100"     = "h100"
     "h200"     = "h200"
     "b200"     = "b200"

{gpu_dev-0.4.1 → gpu_dev-0.5.0}/terraform-gpu-devservers/git-cache.tf

@@ -228,29 +228,29 @@ NGINXCONF
                   fi
                 done
 
-                # Create bare .git tarball (much faster - no checkout needed!)
-                echo "[CACHE] Creating pytorch .git tarball..."
+                # Create tarballs for main repo + ALL submodules
+                # Naming convention: org_repo-git.tar.gz (matches git-clone-cached client)
+                echo "[CACHE] Creating tarballs..."
                 cd /git-cache
-                rm -f pytorch-git.tar.gz.tmp
 
-                # Just tar up the bare repo (pack files only, no working tree)
-                # Client will do git checkout after download (unavoidable anyway)
-                tar -czf pytorch-git.tar.gz.tmp -C /git-cache pytorch.git
-                mv pytorch-git.tar.gz.tmp pytorch-git.tar.gz
+                # Main pytorch repo — name must match org_repo convention
+                echo "[CACHE]   Creating pytorch_pytorch-git.tar.gz..."
+                rm -f pytorch_pytorch-git.tar.gz.tmp
+                tar -czf pytorch_pytorch-git.tar.gz.tmp -C /git-cache pytorch.git
+                mv pytorch_pytorch-git.tar.gz.tmp pytorch_pytorch-git.tar.gz
+                SIZE=$(du -sh pytorch_pytorch-git.tar.gz | awk '{print $1}')
+                echo "[CACHE]   pytorch_pytorch: $SIZE"
 
-                SIZE=$(du -sh pytorch-git.tar.gz | awk '{print $1}')
-                echo "[CACHE] Bare .git tarball created: $SIZE"
-
-                # Create tarballs for largest submodules (top 10 by size)
-                echo "[CACHE] Creating submodule tarballs..."
-                for repo in $(du -s /git-cache/*.git 2>/dev/null | sort -rn | head -11 | tail -10 | awk '{print $2}'); do
+                # All submodule repos (already named org_repo.git by init container)
+                for repo in /git-cache/*.git; do
                   name=$(basename "$repo")
+                  [ "$name" = "pytorch.git" ] && continue
                   tarball="$${name%.git}-git.tar.gz"
                   echo "[CACHE]   Creating $tarball..."
                   rm -f "$tarball.tmp" 2>/dev/null
                   tar -czf "$tarball.tmp" -C /git-cache "$name" 2>/dev/null && mv "$tarball.tmp" "$tarball" || echo "[CACHE]   WARNING: Failed to create $tarball"
                 done
-                echo "[CACHE] Submodule tarballs created"
+                echo "[CACHE] All tarballs created"
               fi
 
               echo "[CACHE] Refresh complete at $(date). Next in 3600s (1 hour)..."

{gpu_dev-0.4.1 → gpu_dev-0.5.0}/terraform-gpu-devservers/lambda/availability_updater/index.py

@@ -150,7 +150,7 @@ def update_gpu_availability(gpu_type: str, k8s_client=None) -> None:
             else:
                 available_gpus = total_gpus
         else:
-            # GPU nodes - use existing logic
+            # GPU nodes - use K8s schedulable node count for total if available
             total_gpus = running_instances * gpus_per_instance
             logger.info(
                 f"ASG calculation: {running_instances} instances * {gpus_per_instance} GPUs = {total_gpus} total GPUs")
@@ -181,6 +181,7 @@ def update_gpu_availability(gpu_type: str, k8s_client=None) -> None:
                 nodes = v1.list_node(label_selector=f"GpuType={gpu_type}")
 
                 single_node_max = 0  # Max available on any single node
+                schedulable_total_gpus = 0  # Total GPUs on schedulable (non-cordoned) nodes
                 for node in nodes.items:
                     if is_node_ready_and_schedulable(node):
                         available_on_node = get_available_gpus_on_node(v1, node)
@@ -192,6 +193,8 @@ def update_gpu_availability(gpu_type: str, k8s_client=None) -> None:
                             except (ValueError, TypeError):
                                 pass
 
+                        schedulable_total_gpus += total_on_node
+
                         # Track max available on any single node
                         single_node_max = max(single_node_max, available_on_node)
 
@@ -199,6 +202,8 @@ def update_gpu_availability(gpu_type: str, k8s_client=None) -> None:
                         if total_on_node > 0 and available_on_node == total_on_node:
                             full_nodes_available += 1
 
+                total_gpus = schedulable_total_gpus
+
                 # Calculate max reservable considering multinode scenarios
                 # Only high-end GPU types support multinode (up to 4 nodes = 32 GPUs)
                 multinode_gpu_types = ['h100', 'h200', 'b200', 'a100']

{gpu_dev-0.4.1 → gpu_dev-0.5.0}/terraform-gpu-devservers/lambda/reservation_processor/index.py

@@ -49,7 +49,7 @@ DEFAULT_TIMEOUT_HOURS = int(os.environ["DEFAULT_TIMEOUT_HOURS"])
 QUEUE_URL = os.environ["QUEUE_URL"]
 PRIMARY_AVAILABILITY_ZONE = os.environ["PRIMARY_AVAILABILITY_ZONE"]
 GPU_DEV_CONTAINER_IMAGE = os.environ.get(
-    "GPU_DEV_CONTAINER_IMAGE", "pytorch/pytorch:2.8.0-cuda12.9-cudnn9-devel")
+    "GPU_DEV_CONTAINER_IMAGE", "pytorch/pytorch:2.11.0-cuda12.8-cudnn9-devel")
 EFS_SECURITY_GROUP_ID = os.environ.get("EFS_SECURITY_GROUP_ID")
 EFS_SUBNET_IDS = os.environ.get("EFS_SUBNET_IDS", "").split(
     ",") if os.environ.get("EFS_SUBNET_IDS") else []
@@ -66,6 +66,7 @@ GPU_CONFIG = {
     "t4": {"instance_type": "g4dn.12xlarge", "max_gpus": 4, "cpus": 48, "memory_gb": 192, "efa_count": 0},
     "l4": {"instance_type": "g6.12xlarge", "max_gpus": 4, "cpus": 48, "memory_gb": 192, "efa_count": 1},
     "a10g": {"instance_type": "g5.12xlarge", "max_gpus": 4, "cpus": 48, "memory_gb": 192, "efa_count": 1},
+    "g7e": {"instance_type": "g7e.24xlarge", "max_gpus": 4, "cpus": 96, "memory_gb": 1024, "efa_count": 2},
     "t4-small": {"instance_type": "g4dn.2xlarge", "max_gpus": 1, "cpus": 8, "memory_gb": 32, "efa_count": 0},
     "g5g": {"instance_type": "g5g.2xlarge", "max_gpus": 2, "cpus": 8, "memory_gb": 32, "efa_count": 0},
     "a100": {"instance_type": "p4d.24xlarge", "max_gpus": 8, "cpus": 96, "memory_gb": 1152, "efa_count": 4},
@@ -2150,7 +2151,7 @@ def validate_reservation_request(request: dict[str, Any]) -> tuple[bool, str]:
     gpu_type = request.get("gpu_type", "")
 
     # Validate GPU type
-    valid_gpu_types = ["t4", "l4", "a10g", "t4-small", "a100",
+    valid_gpu_types = ["t4", "l4", "a10g", "g7e", "t4-small", "a100",
                        "h100", "h200", "b200", "cpu-arm", "cpu-x86"]
     if gpu_type not in valid_gpu_types:
         error_msg = f"Invalid GPU type: {gpu_type}. Must be one of: {', '.join(valid_gpu_types)}"
@@ -2381,6 +2382,7 @@ def update_gpu_availability_table(
             "t4": {"gpus_per_instance": 4},
             "l4": {"gpus_per_instance": 4},
             "a10g": {"gpus_per_instance": 4},
+            "g7e": {"gpus_per_instance": 4},
             "a100": {"gpus_per_instance": 8},
             "h100": {"gpus_per_instance": 8},
             "h200": {"gpus_per_instance": 8},
@@ -4570,7 +4572,7 @@ EOFREADME
 
                         cat > /usr/local/bin/git-clone-cached << 'GITCACHESCRIPT'
 #!/bin/bash
-# Clones from bare .git tarball if available in cache (10x faster than git protocol)
+# Clones repo + submodules from in-cluster cache (much faster than GitHub)
 CACHE_URL="http://git-cache.management.svc.cluster.local:8080"
 GIT="/usr/bin/git"
 GITHUB_URL="${{1}}"
@@ -4582,59 +4584,95 @@ if [ -z "$GITHUB_URL" ]; then
     DEST="${{DEST:-pytorch}}"
 fi
 
+# Handle short names: "pytorch" -> "https://github.com/pytorch/pytorch.git"
+if [[ ! "$GITHUB_URL" =~ ^https?:// ]] && [[ ! "$GITHUB_URL" =~ ^git@ ]]; then
+    GITHUB_URL="https://github.com/pytorch/$GITHUB_URL.git"
+    DEST="${{DEST:-${{1}}}}"
+fi
+
 # Extract org/repo from GitHub URL and create cache tarball name
-# https://github.com/pytorch/pytorch.git -> pytorch_pytorch-git.tar.gz
-# https://github.com/ROCm/aiter.git -> ROCm_aiter-git.tar.gz
 if [[ "$GITHUB_URL" =~ github\.com[/:]([^/]+)/([^/\.]+) ]]; then
     ORG="${{BASH_REMATCH[1]}}"
     REPO="${{BASH_REMATCH[2]}}"
     TARBALL="${{ORG}}_${{REPO}}-git.tar.gz"
 else
-    # Not a GitHub URL, fall back to direct clone
-    exec "$GIT" clone "$GITHUB_URL" "${{DEST:+"$DEST"}}"
-fi
-
-# Default destination to repo name if not specified
-if [ -z "$DEST" ]; then
-    DEST="$REPO"
+    exec "$GIT" clone --recurse-submodules --jobs 8 "$GITHUB_URL" "${{DEST:+"$DEST"}}"
 fi
 
-if [ -d "$DEST" ]; then
-    echo "Error: $DEST already exists"
-    exit 1
-fi
+if [ -z "$DEST" ]; then DEST="$REPO"; fi
+if [ -d "$DEST" ]; then echo "Error: $DEST already exists"; exit 1; fi
 
-# Try to download from cache
-echo "[git-cache] Checking cache for $ORG/$REPO..."
+echo "[git-cache] Cloning $ORG/$REPO..."
 TOTAL_START=$(date +%s)
 
+# --- Main repo ---
 mkdir -p "$DEST/.git"
 START=$(date +%s)
 if curl -sf "$CACHE_URL/$TARBALL" | tar -xz -C "$DEST/.git" --strip-components=1 2>/dev/null; then
     END=$(date +%s)
-    echo "[git-cache] Downloaded .git in $((END - START))s"
+    echo "[git-cache] Main repo .git: $((END - START))s"
 
-    # Configure as non-bare repository and set origin
     cd "$DEST"
     "$GIT" config --file .git/config core.bare false
     "$GIT" config --file .git/config remote.origin.url "$GITHUB_URL"
     "$GIT" config --file .git/config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*"
 
-    echo "[git-cache] Checking out working tree..."
     START=$(date +%s)
     "$GIT" checkout -f HEAD 2>/dev/null
     END=$(date +%s)
-    echo "[git-cache] Checkout took $((END - START))s"
+    echo "[git-cache] Checkout: $((END - START))s"
+
+    # --- Submodules from cache ---
+    if [ -f .gitmodules ]; then
+        echo "[git-cache] Setting up submodules..."
+        SUB_START=$(date +%s)
+
+        "$GIT" submodule init
+        ABS_ROOT="$(pwd)"
+
+        "$GIT" config --file .gitmodules --get-regexp 'submodule\..*\.url' | while read key url; do
+            name=$(echo "$key" | sed 's/^submodule\.//;s/\.url$//')
+            path=$("$GIT" config --file .gitmodules "submodule.$name.path")
+            [ -z "$path" ] && continue
+
+            COMMIT=$("$GIT" ls-tree HEAD "$path" 2>/dev/null | awk '{{print $3}}')
+            [ -z "$COMMIT" ] && continue
+
+            if [[ "$url" =~ github\.com[/:]([^/]+)/([^/.]+) ]]; then
+                SUB_TARBALL="${{BASH_REMATCH[1]}}_${{BASH_REMATCH[2]}}-git.tar.gz"
+                MODULES_DIR="$ABS_ROOT/.git/modules/$name"
+
+                mkdir -p "$MODULES_DIR"
+                if curl -sf "$CACHE_URL/$SUB_TARBALL" | tar -xz -C "$MODULES_DIR" --strip-components=1 2>/dev/null; then
+                    "$GIT" -C "$MODULES_DIR" config core.bare false
+                    "$GIT" -C "$MODULES_DIR" config core.worktree "$ABS_ROOT/$path"
+                    "$GIT" -C "$MODULES_DIR" config remote.origin.url "$url"
+                    "$GIT" -C "$MODULES_DIR" config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*"
+                    mkdir -p "$ABS_ROOT/$path"
+                    echo "gitdir: $MODULES_DIR" > "$ABS_ROOT/$path/.git"
+                    "$GIT" -C "$ABS_ROOT/$path" checkout -f "$COMMIT" 2>/dev/null
+                else
+                    rm -rf "$MODULES_DIR"
+                fi
+            fi
+        done
+
+        # Fetch remaining/recursive submodules from GitHub
+        "$GIT" -c protocol.file.allow=always submodule update --init --recursive --jobs 8 2>/dev/null
+
+        SUB_END=$(date +%s)
+        echo "[git-cache] Submodules: $((SUB_END - SUB_START))s"
+    fi
 
     TOTAL_END=$(date +%s)
-    echo "[git-cache] Total: $((TOTAL_END - TOTAL_START))s (from cache)"
+    echo "[git-cache] Total: $((TOTAL_END - TOTAL_START))s"
     exit 0
 fi
 
 # Fallback: clone from GitHub
 echo "[git-cache] Cache miss, cloning from GitHub..."
 rm -rf "$DEST"
-"$GIT" clone "$GITHUB_URL" "${{DEST:+"$DEST"}}"
+"$GIT" clone --recurse-submodules --jobs 8 "$GITHUB_URL" "${{DEST:+"$DEST"}}"
 GITCACHESCRIPT
                         chmod +x /usr/local/bin/git-clone-cached
                         echo "[STARTUP] ✓ git-clone-cached available (opt-in: use 'git-clone-cached pytorch' for cache)"

{gpu_dev-0.4.1 → gpu_dev-0.5.0}/terraform-gpu-devservers/lambda.tf

@@ -180,8 +180,8 @@ resource "aws_lambda_function" "reservation_processor" {
       HOSTED_ZONE_ID                     = local.effective_domain_name != "" ? local.hosted_zone_id : ""
       SSH_DOMAIN_MAPPINGS_TABLE          = local.effective_domain_name != "" ? aws_dynamodb_table.ssh_domain_mappings.name : ""
       SSL_CERTIFICATE_ARN                = local.effective_domain_name != "" ? aws_acm_certificate.wildcard[0].arn : ""
-      LAMBDA_VERSION                     = "0.4.1"
-      MIN_CLI_VERSION                    = "0.4.0"
+      LAMBDA_VERSION                     = "0.5.0"
+      MIN_CLI_VERSION                    = "0.5.0"
       DISK_CONTENTS_BUCKET               = aws_s3_bucket.disk_contents.bucket
       OPERATIONS_TABLE                   = aws_dynamodb_table.operations.name
     }, local.alb_env_vars)

{gpu_dev-0.4.1 → gpu_dev-0.5.0}/terraform-gpu-devservers/main.tf

@@ -79,7 +79,7 @@ locals {
         "cpu-arm" = {
           instance_type       = "c7g.4xlarge"
           instance_types      = null
-          instance_count      = 3
+          instance_count      = 1
           gpus_per_instance   = 0
           use_placement_group = false
           architecture        = "arm64"
@@ -88,7 +88,7 @@ locals {
         "cpu-x86" = {
           instance_type       = "c7i.4xlarge"
           instance_types      = null
-          instance_count      = 3
+          instance_count      = 1
           gpus_per_instance   = 0
           use_placement_group = false
           architecture        = "x86_64"
@@ -97,7 +97,7 @@ locals {
         "t4" = {
           instance_type       = "g4dn.12xlarge"
           instance_types      = null
-          instance_count      = 2 # 2 instances in primary AZ
+          instance_count      = 1
           gpus_per_instance   = 4
           use_placement_group = true
           architecture        = "x86_64"
@@ -106,7 +106,7 @@ locals {
         "t4-az2" = {
           instance_type       = "g4dn.12xlarge"
           instance_types      = null
-          instance_count      = 2 # 2 instances in secondary AZ
+          instance_count      = 0 # Disabled - use primary AZ only for testing
           gpus_per_instance   = 4
           use_placement_group = true
           architecture        = "x86_64"
@@ -115,7 +115,7 @@ locals {
         "h100" = {
           instance_type       = "p5.48xlarge"
           instance_types      = null
-          instance_count      = 2 # Fallback default (not used when capacity_reservations defined)
+          instance_count      = 0 # Disabled - only use via CR when needed
           gpus_per_instance   = 8
           use_placement_group = false
           architecture        = "x86_64"
@@ -124,7 +124,7 @@ locals {
         "t4-small" = {
           instance_type       = "g4dn.2xlarge"
           instance_types      = null
-          instance_count      = 1
+          instance_count      = 0 # Disabled
           gpus_per_instance   = 1
           use_placement_group = false
           architecture        = "x86_64"
@@ -183,7 +183,7 @@ locals {
         "t4" = {
           instance_type       = "g4dn.12xlarge"
           instance_types      = null
-          instance_count      = 5 # Fallback default (not used when capacity_reservations defined)
+          instance_count      = 2
           gpus_per_instance   = 4
           use_placement_group = true
           architecture        = "x86_64"
@@ -192,7 +192,7 @@ locals {
         "l4" = {
           instance_type       = "g6.12xlarge"
           instance_types      = null
-          instance_count      = 5 # Fallback default (not used when capacity_reservations defined)
+          instance_count      = 2
           gpus_per_instance   = 4 # 4x L4 GPUs
           use_placement_group = false
           architecture        = "x86_64"
@@ -201,16 +201,25 @@ locals {
         "a10g" = {
           instance_type       = "g5.12xlarge"
           instance_types      = null
-          instance_count      = 2
+          instance_count      = 1
           gpus_per_instance   = 4 # 4x A10G GPUs
           use_placement_group = false
           architecture        = "x86_64"
           efa_network_cards   = 1
         }
+        "g7e" = {
+          instance_type       = "g7e.24xlarge"
+          instance_types      = null
+          instance_count      = 2
+          gpus_per_instance   = 4 # 4x RTX PRO 6000 Blackwell GPUs
+          use_placement_group = false
+          architecture        = "x86_64"
+          efa_network_cards   = 2
+        }
         "cpu-arm" = {
           instance_type       = "c7g.8xlarge"
           instance_types      = null
-          instance_count      = 30
+          instance_count      = 10
           gpus_per_instance   = 0
           use_placement_group = false
           architecture        = "arm64"
@@ -219,7 +228,7 @@ locals {
         "cpu-x86" = {
           instance_type       = "c7i.8xlarge"
           instance_types      = null
-          instance_count      = 30
+          instance_count      = 10
           gpus_per_instance   = 0
           use_placement_group = false
           architecture        = "x86_64"
@@ -257,6 +266,7 @@ locals {
       h100 = [
         { key = "cr0", id = "cr-0a3f49b96fe03ca04", instance_count = 4 }, # H100 reservation us-east-2c (p5.48xlarge)
         { key = "cr1", id = null, instance_count = 2 },                   # H100 on-demand (2 instances)
+        { key = "cr2", id = "cr-044bc72b0a6b56062", instance_count = 4 }, # H100 reservation us-east-2a (4 instances)
       ]
       h200 = [
         { key = "cr0", id = "cr-0f6d0766f5d3339e6", instance_count = 2 }, # H200 capacity block (may be expired - keep to prevent ASG destroy)
@@ -264,9 +274,10 @@ locals {
         { key = "cr2", id = null, instance_count = 2 },                   # H200 on-demand (2 instances)
       ]
       b200 = [
-        { key = "cr0", id = "cr-0c366fb8339a10f69", instance_count = 0 }, # B200 reservation (disabled - CR expired)
+        { key = "cr0", id = "cr-0c366fb8339a10f69", instance_count = 0 }, # B200 reservation us-east-2a (disabled - CR freed)
         { key = "cr1", id = "cr-08e7fee0b8dc3de5e", instance_count = 3 }, # B200 reservation (3 instances)
         { key = "cr2", id = null, instance_count = 2 },                   # B200 on-demand (2 instances)
+        { key = "cr3", id = "cr-0f5f6bb30a8fe3c68", instance_count = 2 }, # B200 reservation us-east-2b (2 instances)
       ]
       # T4 and L4 don't have capacity reservations - managed via supported_gpu_types fallback
     }
@@ -292,6 +303,7 @@ locals {
       t4        = "primary"
       l4        = "secondary"
       a10g      = "secondary"
+      g7e       = "secondary"
       "cpu-arm" = "primary"
       "cpu-x86" = "primary"
     }
@@ -307,11 +319,13 @@ locals {
       "cr-0c366fb8339a10f69" = "primary"   # us-east-2a
       "cr-0122dff5e01d566dc" = "secondary" # us-east-2b
       "cr-08e7fee0b8dc3de5e" = "secondary" # us-east-2b
+      "cr-0f5f6bb30a8fe3c68" = "secondary" # us-east-2b
       # H200 capacity reservations
       "cr-0f6d0766f5d3339e6" = "tertiary" # us-east-2c (may be expired - kept to prevent ASG destroy)
       "cr-06c9c978dea756a26" = "tertiary"  # us-east-2c
-      # H100 capacity reservation
+      # H100 capacity reservations
       "cr-0a3f49b96fe03ca04" = "tertiary" # us-east-2c (p5.48xlarge)
+      "cr-044bc72b0a6b56062" = "primary"  # us-east-2a (p5.48xlarge)
       # A100 capacity reservation
       "cr-01cc0f00f28b095af" = "primary" # us-east-2a
     }

{gpu_dev-0.4.1 → gpu_dev-0.5.0}/terraform-gpu-devservers/templates/al2023-user-data.sh

@@ -11,7 +11,7 @@ systemctl disable nodeadm-run.service || true
 systemctl stop nodeadm-config.service || true
 systemctl stop nodeadm-run.service || true
 
-# Install NVIDIA driver 580.82.07 directly on host for CUDA 13 support
+# Install latest NVIDIA driver on host (595.x branch supports CUDA 13.2)
 # GPU Operator will handle toolkit/device-plugin only
 
 # Configure NVIDIA profiling BEFORE driver installation (driver install auto-loads modules)