gpu-dev 0.5.6__tar.gz → 0.5.8__tar.gz

{gpu_dev-0.5.6 → gpu_dev-0.5.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gpu-dev
-Version: 0.5.6
+Version: 0.5.8
 Summary: CLI tool for PyTorch GPU developer server reservations
 Author: PyTorch Team
 Requires-Python: >=3.10

{gpu_dev-0.5.6 → gpu_dev-0.5.8}/cli-tools/gpu-dev-cli/gpu_dev.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gpu-dev
-Version: 0.5.6
+Version: 0.5.8
 Summary: CLI tool for PyTorch GPU developer server reservations
 Author: PyTorch Team
 Requires-Python: >=3.10

{gpu_dev-0.5.6 → gpu_dev-0.5.8}/cli-tools/gpu-dev-cli/gpu_dev_cli/cli.py

@@ -161,11 +161,8 @@ def _show_single_reservation(connection_info: dict) -> None:
     gpu_type = connection_info.get("gpu_type", "Unknown")
     instance_type = connection_info.get("instance_type", "unknown")
 
-    # Format GPU information
-    if gpu_type != "Unknown" and gpu_type != "unknown":
-        gpu_info = f"{gpu_count}x {gpu_type}"
-    else:
-        gpu_info = f"{gpu_count} GPU(s)"
+    # Format GPU information (MIG-aware)
+    gpu_info = _format_gpu_display(gpu_count, gpu_type)
 
     # Format timestamps - only show launched_at (started time), not created time
     launched_at = connection_info.get("launched_at", "N/A")
@@ -2042,11 +2039,8 @@ def cancel(
                 status = reservation.get("status", "unknown")
                 created_at = reservation.get("created_at", "N/A")
 
-                # Format GPU information
-                if gpu_type and gpu_type not in ["unknown", "Unknown"]:
-                    gpu_display = f"{gpu_count}x {gpu_type.upper()}"
-                else:
-                    gpu_display = str(gpu_count)
+                # Format GPU information (MIG-aware)
+                gpu_display = _format_gpu_display(gpu_count, gpu_type)
 
                 # Format created_at
                 created_formatted = "N/A"
@@ -2379,6 +2373,28 @@ def show(ctx: click.Context, reservation_id: Optional[str]) -> None:
         rprint(f"[red]❌ Error: {str(e)}[/red]")
 
 
+
+def _format_gpu_display(gpu_count, gpu_type):
+    """Render a friendly '{N}× {type}' string for reservation listings.
+
+    For MIG slice SKUs, surface the slice memory + the underlying physical type
+    (e.g. '2× 10GB H100 (MIG)') instead of the raw 'H100-MIG-1G' identifier.
+    """
+    if not gpu_type or str(gpu_type).lower() in ("unknown", ""):
+        return f"{gpu_count} GPU(s)"
+    gt_lower = str(gpu_type).lower()
+    mig_friendly = {
+        "h100-mig-1g": "10GB H100 (MIG)",
+        "h100-mig-2g": "20GB H100 (MIG)",
+        "h100-mig-3g": "40GB H100 (MIG)",
+        "h100-mig-4g": "40GB H100 (MIG)",
+        "h100-mig-7g": "80GB H100 (MIG)",
+    }
+    if gt_lower in mig_friendly:
+        return f"{gpu_count}× {mig_friendly[gt_lower]}"
+    return f"{gpu_count}x {str(gpu_type).upper()}"
+
+
 def _show_availability() -> None:
     """Shared function to show GPU availability"""
     try:

{gpu_dev-0.5.6 → gpu_dev-0.5.8}/cli-tools/gpu-dev-cli/gpu_dev_cli/interactive.py

@@ -64,6 +64,18 @@ def select_gpu_type_interactive(
         if "-mig-" not in gt
     }
 
+    # Aggregate MIG slice availability so we can hint it on the h100 row of this picker.
+    mig_total_available = sum(
+        int(info.get("available", 0))
+        for gt, info in (availability_info or {}).items()
+        if gt.startswith("h100-mig-")
+    )
+    mig_total_capacity = sum(
+        int(info.get("total", 0))
+        for gt, info in (availability_info or {}).items()
+        if gt.startswith("h100-mig-")
+    )
+
     # Display availability table first
     console.print("\n[cyan]🖥️  GPU Availability:[/cyan]")
     table = Table()
@@ -132,6 +144,8 @@ def select_gpu_type_interactive(
             )
             if queue_length > 0:
                 choice_label += f" - {queue_length} in queue"
+            if gpu_type == "h100" and mig_total_capacity > 0:
+                choice_label += f" — also {mig_total_available}/{mig_total_capacity} MIG slices"
 
             choices.append(questionary.Choice(title=choice_label, value=gpu_type))
 

{gpu_dev-0.5.6 → gpu_dev-0.5.8}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "gpu-dev"
-version = "0.5.6"
+version = "0.5.8"
 description = "CLI tool for PyTorch GPU developer server reservations"
 authors = [{name = "PyTorch Team"}]
 readme = "cli-tools/gpu-dev-cli/README.md"

{gpu_dev-0.5.6 → gpu_dev-0.5.8}/terraform-gpu-devservers/docker/Dockerfile

@@ -132,13 +132,32 @@ RUN mkdir -p /run/sshd /var/run/sshd && \
 # Create SSH config
 COPY ssh_config /etc/ssh/sshd_config
 
-# Set up npm global directory for dev user and install Claude CLI
+# Install Claude Code SYSTEM-WIDE via the official native installer (as root).
+# Lives in /opt/claude with a /usr/local/bin/claude symlink, survives every reservation
+# regardless of persistent-disk state, and can't be shadowed by stale npm installs on
+# user disks. Image rebuilds are the only update path — controlled, reproducible.
+USER root
+RUN curl -fsSL https://claude.ai/install.sh | HOME=/opt/claude bash || echo "Claude native install failed (non-fatal at build time)"
+RUN if [ -e /opt/claude/.local/bin/claude ]; then \
+        ln -sf /opt/claude/.local/bin/claude /usr/local/bin/claude; \
+        chmod -R a+rX /opt/claude; \
+    fi
+
+# Set up npm global directory for dev user (kept for ad-hoc dev-installed CLIs).
 USER dev
 WORKDIR /home/dev
-
 RUN mkdir -p ~/.npm-global && \
-    npm config set prefix ~/.npm-global && \
-    npm install -g @anthropic-ai/claude-code || echo "Claude CLI install failed, will retry at runtime"
+    npm config set prefix ~/.npm-global
+
+# OpenAI Codex CLI installed SYSTEM-WIDE as root (parallels the Claude install above).
+# Auth is per-user: either `export OPENAI_API_KEY=sk-…` in their shell, or `codex login`
+# for the browser/ChatGPT-account flow. AWS Bedrock now serves Codex (announced 2026-04-28
+# at openai.com/index/openai-on-aws) — once we have early access, swap in the Bedrock-backed
+# install + flip an env var; users won't need their own keys anymore.
+USER root
+RUN npm install -g --prefix /usr/local @openai/codex || echo "Codex CLI install failed (non-fatal at build time)"
+
+USER dev
 
 # Install oh-my-zsh for dev user
 RUN sh -c "$(curl -fsSL https://raw.githubusercontent.com/ohmyzsh/ohmyzsh/master/tools/install.sh)" "" --unattended

{gpu_dev-0.5.6 → gpu_dev-0.5.8}/terraform-gpu-devservers/docker/bashrc_ext

@@ -24,3 +24,17 @@ check_warnings() {
 
 # Run warning check before every command prompt
 PROMPT_COMMAND="check_warnings; $PROMPT_COMMAND"
+
+# Auto-cleanup of deprecated ANTHROPIC_MODEL pinning. An older docker image
+# hardcoded ANTHROPIC_MODEL=us.anthropic.claude-sonnet-4-20250514-v1:0 in
+# .shell_env, which then got cached on persistent disks. We unset it in the
+# current session and strip the line from disk so it doesn't come back.
+# Self-healing — once cleaned, the case statement no longer matches.
+case "${ANTHROPIC_MODEL:-}" in
+    *sonnet-4-20250514*)
+        unset ANTHROPIC_MODEL
+        if [ -f "$HOME/.shell_env" ]; then
+            sed -i '/ANTHROPIC_MODEL.*sonnet-4-20250514/d' "$HOME/.shell_env" 2>/dev/null || true
+        fi
+        ;;
+esac

{gpu_dev-0.5.6 → gpu_dev-0.5.8}/terraform-gpu-devservers/docker/zshrc_ext

@@ -25,3 +25,17 @@ check_warnings() {
 
 # Run warning check before every command prompt (zsh hook)
 precmd() { check_warnings }
+
+# Auto-cleanup of deprecated ANTHROPIC_MODEL pinning. An older docker image
+# hardcoded ANTHROPIC_MODEL=us.anthropic.claude-sonnet-4-20250514-v1:0 in
+# .shell_env, which then got cached on persistent disks. We unset it in the
+# current session and strip the line from disk so it doesn't come back.
+# Self-healing — once cleaned, the case statement no longer matches.
+case "${ANTHROPIC_MODEL:-}" in
+    *sonnet-4-20250514*)
+        unset ANTHROPIC_MODEL
+        if [ -f "$HOME/.shell_env" ]; then
+            sed -i '/ANTHROPIC_MODEL.*sonnet-4-20250514/d' "$HOME/.shell_env" 2>/dev/null || true
+        fi
+        ;;
+esac

{gpu_dev-0.5.6 → gpu_dev-0.5.8}/terraform-gpu-devservers/lambda/availability_updater/index.py

@@ -24,6 +24,45 @@ RESERVATIONS_TABLE = os.environ.get("RESERVATIONS_TABLE", "pytorch-gpu-dev-reser
 SUPPORTED_GPU_TYPES = json.loads(os.environ["SUPPORTED_GPU_TYPES"])
 
 
+def _parse_expires_at(value):
+    """Parse the reservations table's `expires_at` field to a unix epoch (int).
+
+    DDB stores it as either an ISO-8601 datetime string ("2026-05-02T00:12:03.674845") OR
+    occasionally a numeric epoch — handle both. Returns None if unparseable.
+    """
+    if value is None:
+        return None
+    # Numeric (Decimal/int/float) → epoch seconds directly.
+    if not isinstance(value, str):
+        try:
+            return int(float(value))
+        except (ValueError, TypeError):
+            return None
+    s = value.strip()
+    if not s:
+        return None
+    # ISO-8601 first (the actual production format).
+    try:
+        from datetime import datetime, timezone
+        # `fromisoformat` accepts microseconds; tolerate optional 'Z' suffix.
+        if s.endswith("Z"):
+            s2 = s[:-1] + "+00:00"
+        else:
+            s2 = s
+        dt = datetime.fromisoformat(s2)
+        if dt.tzinfo is None:
+            # Convention in this codebase: timestamps written via datetime.utcnow().isoformat() are UTC.
+            dt = dt.replace(tzinfo=timezone.utc)
+        return int(dt.timestamp())
+    except (ValueError, TypeError):
+        pass
+    # Numeric-as-string fallback.
+    try:
+        return int(float(s))
+    except (ValueError, TypeError):
+        return None
+
+
 def get_gpu_resource_name(gpu_type: str) -> str:
     return SUPPORTED_GPU_TYPES.get(gpu_type, {}).get("k8s_resource", "nvidia.com/gpu")
 
@@ -76,6 +115,13 @@ def handler(event: Dict[str, Any], context: Any) -> Dict[str, Any]:
                 logger.error(f"=== Failed to update availability for {gpu_type}: {gpu_error} ===")
                 # Continue with other GPU types
 
+        # Best-effort: delete stale rows for SKUs no longer in SUPPORTED_GPU_TYPES
+        # (e.g. after a GPU type rename like g7e -> rtxpro6000).
+        try:
+            cleanup_stale_availability_rows()
+        except Exception as cleanup_err:
+            logger.warning(f"Stale-row cleanup failed: {cleanup_err}")
+
         return {
             "statusCode": 200,
             "body": json.dumps(
@@ -535,9 +581,8 @@ def compute_size_etas(v1, gpu_type, node_label_value, resource_name, gpus_per_in
             continue
         if pod_name not in pod_to_info:
             continue
-        try:
-            ts = int(float(expires_at))
-        except (ValueError, TypeError):
+        ts = _parse_expires_at(expires_at)
+        if ts is None:
             continue
         node_name, gpus = pod_to_info[pod_name]
         node_state[node_name]["expirations"].append((ts, gpus))
@@ -594,3 +639,29 @@ def compute_size_etas(v1, gpu_type, node_label_value, resource_name, gpus_per_in
 
     return etas
 
+
+def cleanup_stale_availability_rows():
+    """Delete rows in the availability table whose gpu_type isn't in SUPPORTED_GPU_TYPES.
+
+    Triggered on every Lambda invocation. Idempotent. Used to garbage-collect renamed
+    SKUs (e.g. g7e -> rtxpro6000) that would otherwise linger as zero rows.
+    """
+    table = dynamodb.Table(AVAILABILITY_TABLE)
+    valid_keys = set(SUPPORTED_GPU_TYPES.keys())
+    last_key = None
+    deleted = []
+    while True:
+        kwargs = {"ProjectionExpression": "gpu_type"}
+        if last_key:
+            kwargs["ExclusiveStartKey"] = last_key
+        resp = table.scan(**kwargs)
+        for item in resp.get("Items", []):
+            gt = item.get("gpu_type")
+            if gt and gt not in valid_keys:
+                table.delete_item(Key={"gpu_type": gt})
+                deleted.append(gt)
+        last_key = resp.get("LastEvaluatedKey")
+        if not last_key:
+            break
+    if deleted:
+        logger.info(f"Deleted {len(deleted)} stale availability rows: {deleted}")

{gpu_dev-0.5.6 → gpu_dev-0.5.8}/terraform-gpu-devservers/lambda/reservation_processor/index.py

@@ -4030,7 +4030,9 @@ def create_pod(
                 client.V1Container(
                     name="gpu-dev",
                     image=container_image,
-                    image_pull_policy="IfNotPresent",
+                    # :latest is a moving tag — always re-pull so a new docker image
+                    # rolled out to ECR is picked up by every fresh reservation.
+                    image_pull_policy="Always",
                     **({
                         "command": ["/bin/bash"],
                         "args": [
@@ -4443,6 +4445,34 @@ EOF_ZSHRC_EXT
                         done
                         echo "[STARTUP] ✓ Shell extension sourcing configured"
 
+                        # Surgically remove the deprecated ANTHROPIC_MODEL pinning that older docker
+                        # images baked into shell_env. claude-sonnet-4-20250514 is being deprecated
+                        # by Anthropic, and an old hardcoded value lingers on persistent disks.
+                        # Idempotent — strips only the matching line, leaves any user-customized value alone.
+                        if [ -f /home/dev/.shell_env ] && grep -q "ANTHROPIC_MODEL.*sonnet-4-20250514" /home/dev/.shell_env 2>/dev/null; then
+                            echo "[STARTUP] Removing deprecated ANTHROPIC_MODEL=sonnet-4-20250514 line from .shell_env"
+                            sed -i '/ANTHROPIC_MODEL.*sonnet-4-20250514/d' /home/dev/.shell_env || true
+                        fi
+
+                        # Remove the legacy npm-based Claude install (~/.npm-global/bin/claude).
+                        # Older docker images ran `npm install -g @anthropic-ai/claude-code` and that
+                        # binary still lingers on persistent disks where it shadows the system-wide
+                        # /usr/local/bin/claude (because $HOME/.npm-global/bin precedes /usr/local/bin
+                        # on PATH). The system-wide install is kept current via image rebuilds; the
+                        # user's own ~/.local/bin/claude (from `claude install` in their home, or
+                        # Claude's self-update mechanism) is left intact so users can opt into newer
+                        # versions ahead of our image refresh.
+                        if [ -e /home/dev/.npm-global/bin/claude ]; then
+                            echo "[STARTUP] Removing legacy npm-installed claude at /home/dev/.npm-global/bin/claude"
+                            rm -f /home/dev/.npm-global/bin/claude || true
+                        fi
+                        # Also drop the npm package files for @anthropic-ai/claude-code so npm doesn't
+                        # think it's still installed on next `npm list -g`.
+                        if [ -d /home/dev/.npm-global/lib/node_modules/@anthropic-ai/claude-code ]; then
+                            echo "[STARTUP] Removing legacy @anthropic-ai/claude-code npm package files"
+                            rm -rf /home/dev/.npm-global/lib/node_modules/@anthropic-ai/claude-code || true
+                        fi
+
                         # Fix ownership - recursive only for new disks (fast, empty disk)
                         # For existing disks, only fix the specific files we just created/modified
                         if [ "$CREATE_SH_ENV" = "true" ]; then
@@ -6233,7 +6263,13 @@ def should_use_persistent_disk(user_id: str, current_reservation_id: str) -> boo
 
 
 def get_instance_type_and_gpu_info(k8s_client, pod_name: str) -> tuple[str, str]:
-    """Get instance type and GPU type from the node where pod is scheduled"""
+    """Get instance type and GPU type from the node where pod is scheduled.
+
+    For MIG slice pods, the SKU is derived from the pod's resource request
+    (nvidia.com/mig-Ng.NNgb) rather than the host instance type, so the DDB
+    record reflects the actual partition the user reserved instead of the
+    physical card.
+    """
     try:
         v1 = client.CoreV1Api(k8s_client)
 
@@ -6250,7 +6286,27 @@ def get_instance_type_and_gpu_info(k8s_client, pod_name: str) -> tuple[str, str]
             "node.kubernetes.io/instance-type", "unknown"
         )
 
-        # Map instance type to GPU type
+        # If the pod requests a MIG slice resource, the GPU type is the slice SKU.
+        # Map back to the canonical SKU stored elsewhere in this code (matches CLI flag).
+        mig_resource_to_sku = {
+            "nvidia.com/mig-1g.10gb": "h100-mig-1g",
+            "nvidia.com/mig-1g.20gb": "h100-mig-1g",  # memory-doubled variant, same SKU bucket
+            "nvidia.com/mig-2g.20gb": "h100-mig-2g",
+            "nvidia.com/mig-3g.40gb": "h100-mig-3g",
+            "nvidia.com/mig-4g.40gb": "h100-mig-4g",
+            "nvidia.com/mig-7g.80gb": "h100-mig-7g",
+        }
+        if pod.spec.containers:
+            for c in pod.spec.containers:
+                reqs = (c.resources.requests if c.resources and c.resources.requests else {}) or {}
+                for r_name, sku in mig_resource_to_sku.items():
+                    if reqs.get(r_name):
+                        logger.info(
+                            f"Pod {pod_name} on {node_name} requests {r_name} -> MIG SKU {sku}"
+                        )
+                        return instance_type, sku
+
+        # Map instance type to GPU type for non-MIG (full GPU) pods
         gpu_type_mapping = {
             "g4dn.4xlarge": "T4",
             "g4dn.8xlarge": "T4",
@@ -6261,6 +6317,7 @@ def get_instance_type_and_gpu_info(k8s_client, pod_name: str) -> tuple[str, str]
             "g6.12xlarge": "L4",
             "g6.16xlarge": "L4",
             "g6.24xlarge": "L4",
+            "g7e.24xlarge": "rtxpro6000",
             "p4d.24xlarge": "A100",
             "p5.48xlarge": "H100",
             "p5e.48xlarge": "H200",

{gpu_dev-0.5.6 → gpu_dev-0.5.8}/terraform-gpu-devservers/lambda.tf

@@ -180,7 +180,7 @@ 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                     = "0.5.6"
+      LAMBDA_VERSION                     = "0.5.8"
       MIN_CLI_VERSION                    = "0.5.5"
       DISK_CONTENTS_BUCKET               = aws_s3_bucket.disk_contents.bucket
       OPERATIONS_TABLE                   = aws_dynamodb_table.operations.name