gpu-dev 0.5.25__tar.gz → 0.5.27__tar.gz

{gpu_dev-0.5.25 → gpu_dev-0.5.27}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gpu-dev
-Version: 0.5.25
+Version: 0.5.27
 Summary: CLI tool for PyTorch GPU developer server reservations
 Author: PyTorch Team
 Requires-Python: >=3.10

{gpu_dev-0.5.25 → gpu_dev-0.5.27}/cli-tools/gpu-dev-cli/gpu_dev.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gpu-dev
-Version: 0.5.25
+Version: 0.5.27
 Summary: CLI tool for PyTorch GPU developer server reservations
 Author: PyTorch Team
 Requires-Python: >=3.10

{gpu_dev-0.5.25 → gpu_dev-0.5.27}/cli-tools/gpu-dev-cli/gpu_dev.egg-info/SOURCES.txt

@@ -52,6 +52,7 @@ terraform-gpu-devservers/main.tf
 terraform-gpu-devservers/mig-config.tf
 terraform-gpu-devservers/mig-parted-config.yaml
 terraform-gpu-devservers/monitoring.tf
+terraform-gpu-devservers/node-termination-handler.tf
 terraform-gpu-devservers/outputs.tf
 terraform-gpu-devservers/pyproject.toml
 terraform-gpu-devservers/queue.tf

{gpu_dev-0.5.25 → gpu_dev-0.5.27}/cli-tools/gpu-dev-cli/gpu_dev_cli/cli.py

@@ -496,7 +496,7 @@ def main(ctx: click.Context) -> None:
     "--gpu-type",
     "-t",
     type=click.Choice(
-        ["b200", "b200-mig-1g", "b200-mig-2g", "b200-mig-3g", "h200", "h100", "h100-mig-1g", "h100-mig-2g", "h100-mig-3g", "a100", "rtxpro6000", "a10g", "t4", "l4", "t4-small", "cpu-arm", "cpu-x86"], case_sensitive=False
+        ["b300", "b200", "b200-mig-1g", "b200-mig-2g", "b200-mig-3g", "h200", "h100", "h100-mig-1g", "h100-mig-2g", "h100-mig-3g", "a100", "rtxpro6000", "a10g", "t4", "l4", "t4-small", "cpu-arm", "cpu-x86"], case_sensitive=False
     ),
     help="GPU type to reserve. Full GPUs: b200, h200, h100, a100, rtxpro6000, a10g, t4, l4, t4-small. H100 MIG slices: h100-mig-1g (10 GB), h100-mig-2g (20 GB), h100-mig-3g (40 GB). B200 MIG slices (on the mixed B200 node): b200-mig-1g (23 GB), b200-mig-2g (45 GB), b200-mig-3g (90 GB). CPU: cpu-arm, cpu-x86.",
 )
@@ -662,6 +662,7 @@ def reserve(
             "b200-mig-3g": {"max_gpus": 2, "instance_type": "p6-b200.48xlarge"},
             "h200": {"max_gpus": 8, "instance_type": "p5e.48xlarge"},
             "b200": {"max_gpus": 8, "instance_type": "p6-b200.48xlarge"},
+            "b300": {"max_gpus": 8, "instance_type": "p6e-b300.48xlarge"},
             "cpu-arm": {"max_gpus": 0, "instance_type": "c7g.4xlarge"},
             "cpu-x86": {"max_gpus": 0, "instance_type": "c7i.4xlarge"},
         }
@@ -1350,7 +1351,7 @@ def reserve(
         rprint(f"[red]❌ Error: {str(e)}[/red]")
 
 
-_SUBMIT_GPU_TYPES = ["b200", "b200-mig-1g", "b200-mig-2g", "b200-mig-3g", "h200", "h100",
+_SUBMIT_GPU_TYPES = ["b300", "b200", "b200-mig-1g", "b200-mig-2g", "b200-mig-3g", "h200", "h100",
                      "h100-mig-1g", "h100-mig-2g", "h100-mig-3g", "a100", "rtxpro6000",
                      "a10g", "t4", "l4", "t4-small", "cpu-arm", "cpu-x86"]
 
@@ -2719,6 +2720,7 @@ def _show_availability() -> None:
             # GPU architecture mapping (for display)
             gpu_architectures = {
                 "b200": "Blackwell (sm100)",
+                "b300": "Blackwell (sm100)",
                 "h200": "Hopper (sm90)",
                 "h100": "Hopper (sm90)",
                 "a100": "Ampere (sm80)",
@@ -2880,6 +2882,7 @@ def _show_availability_watch(interval: int) -> None:
                         # GPU architecture mapping (for display)
                         gpu_architectures = {
                             "b200": "Blackwell (sm100)",
+                "b300": "Blackwell (sm100)",
                             "h200": "Hopper (sm90)",
                             "h100": "Hopper (sm90)",
                             "a100": "Ampere (sm80)",
@@ -3219,6 +3222,20 @@ def connect(ctx: click.Context, reservation_id: Optional[str]) -> None:
         if "-A" not in ssh_command and "-o ForwardAgent=yes" not in ssh_command:
             ssh_command = ssh_command.replace("ssh ", "ssh -A ", 1)
 
+        # Inject AddKeysToAgent so the first connect from this laptop loads the user\'s
+        # IdentityFile into ssh-agent — without this the forwarded agent is empty on
+        # subsequent pod→pod hops. UseKeychain persists the passphrase across reboots on
+        # macOS; IgnoreUnknown lets Linux SSH ignore the macOS-only option cleanly.
+        # The same options live in ~/.gpu-dev/<id>-sshconfig but ssh only honours them
+        # when the command-line target matches a Host block, which this connect command
+        # bypasses by passing the FQDN directly.
+        if "AddKeysToAgent" not in ssh_command:
+            ssh_command = ssh_command.replace(
+                "ssh ",
+                "ssh -o AddKeysToAgent=yes -o IgnoreUnknown=UseKeychain -o UseKeychain=yes ",
+                1,
+            )
+
         # When running from inside a gpu-dev pod (=GPU_DEV_USER_ID env var set) and the
         # forwarded SSH agent is reachable but empty, the next hop is going to fail with
         # 'Permission denied (publickey)'. Warn upfront so the user knows to ssh-add on

{gpu_dev-0.5.25 → gpu_dev-0.5.27}/cli-tools/gpu-dev-cli/gpu_dev_cli/config.py

@@ -22,6 +22,11 @@ class Config:
             "workspace": "prod",
             "description": "Production environment",
         },
+        "prod-east1": {
+            "region": "us-east-1",
+            "workspace": "prod-east1",
+            "description": "Spot-only us-east-1 environment (T4/L4/CPU)",
+        },
     }
     DEFAULT_ENVIRONMENT = "prod"
 

{gpu_dev-0.5.25 → gpu_dev-0.5.27}/cli-tools/gpu-dev-cli/gpu_dev_cli/reservations.py

@@ -557,6 +557,7 @@ class ReservationManager:
                 "b200-mig-3g": {"max_gpus": 2},
                 "h200": {"max_gpus": 8},
                 "b200": {"max_gpus": 8},
+                "b300": {"max_gpus": 8},
             }
 
             max_gpus_per_node = gpu_configs[gpu_type]["max_gpus"]

{gpu_dev-0.5.25 → gpu_dev-0.5.27}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "gpu-dev"
-version = "0.5.25"
+version = "0.5.27"
 description = "CLI tool for PyTorch GPU developer server reservations"
 authors = [{name = "PyTorch Team"}]
 readme = "cli-tools/gpu-dev-cli/README.md"

{gpu_dev-0.5.25 → gpu_dev-0.5.27}/terraform-gpu-devservers/eks.tf

@@ -189,6 +189,7 @@ locals {
     "h100"     = "h100"
     "h200"     = "h200"
     "b200"     = "b200"
+    "b300"     = "b300"
     "a100"     = "a100"
     "cpu-arm"  = "cpu-arm"
     "cpu-x86"  = "cpu-x86"
@@ -401,11 +402,14 @@ resource "aws_launch_template" "gpu_dev_launch_template" {
     }
   }
 
-  # Conditionally add instance_market_options for capacity block instances (only when capacity reservation exists)
+  # instance_market_options: capacity-block when bound to a reservation, spot when
+  # the workspace's gpu_config has use_spot=true, otherwise on-demand (no block).
+  # Spot is mutually exclusive with capacity reservations — AWS rejects launch templates
+  # carrying both, so the precedence here is CR > spot > on-demand.
   dynamic "instance_market_options" {
-    for_each = each.value.capacity_reservation_id != null ? [1] : []
+    for_each = (each.value.capacity_reservation_id != null || try(each.value.gpu_config.use_spot, false)) ? [1] : []
     content {
-      market_type = "capacity-block"
+      market_type = each.value.capacity_reservation_id != null ? "capacity-block" : "spot"
     }
   }
 

{gpu_dev-0.5.25 → gpu_dev-0.5.27}/terraform-gpu-devservers/lambda/reservation_processor/index.py

@@ -81,6 +81,7 @@ GPU_CONFIG = {
     "h100": {"instance_type": "p5.48xlarge", "max_gpus": 8, "cpus": 192, "memory_gb": 2048, "efa_count": 32},
     "h200": {"instance_type": "p5e.48xlarge", "max_gpus": 8, "cpus": 192, "memory_gb": 2048, "efa_count": 32},
     "b200": {"instance_type": "p6-b200.48xlarge", "max_gpus": 8, "cpus": 192, "memory_gb": 2048, "efa_count": 32},
+    "b300": {"instance_type": "p6e-b300.48xlarge", "max_gpus": 8, "cpus": 192, "memory_gb": 2048, "efa_count": 8},
     "cpu-arm": {"instance_type": "c7g.8xlarge", "max_gpus": 0, "cpus": 32, "memory_gb": 64, "efa_count": 0},
     "cpu-x86": {"instance_type": "c7i.8xlarge", "max_gpus": 0, "cpus": 32, "memory_gb": 64, "efa_count": 0},
 }
@@ -2188,7 +2189,7 @@ def validate_reservation_request(request: dict[str, Any]) -> tuple[bool, str]:
     # Validate GPU type
     valid_gpu_types = ["t4", "l4", "a10g", "rtxpro6000", "t4-small", "a100",
                        "h100", "h100-mig-1g", "h100-mig-2g", "h100-mig-3g",
-                       "h200", "b200", "b200-mig-1g", "b200-mig-2g", "b200-mig-3g",
+                       "h200", "b200", "b300", "b200-mig-1g", "b200-mig-2g", "b200-mig-3g",
                        "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)}"
@@ -2435,6 +2436,7 @@ def update_gpu_availability_table(
             "b200-mig-3g": {"gpus_per_instance": 2},
             "h200": {"gpus_per_instance": 8},
             "b200": {"gpus_per_instance": 8},
+            "b300": {"gpus_per_instance": 8},
         }
 
         gpu_config = gpu_type_configs.get(gpu_type, {"gpus_per_instance": 8})
@@ -6529,6 +6531,7 @@ def get_instance_type_and_gpu_info(k8s_client, pod_name: str) -> tuple[str, str]
             "p5e.48xlarge": "H200",
             "p5en.48xlarge": "H200",
             "p6-b200.48xlarge": "B200",
+            "p6e-b300.48xlarge": "B300",
         }
 
         gpu_type = gpu_type_mapping.get(instance_type, "Unknown")

{gpu_dev-0.5.25 → gpu_dev-0.5.27}/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                     = "0.5.25"
+      LAMBDA_VERSION                     = "0.5.27"
       MIN_CLI_VERSION                    = "0.5.16"
       DISK_CONTENTS_BUCKET               = aws_s3_bucket.disk_contents.bucket
       OPERATIONS_TABLE                   = aws_dynamodb_table.operations.name

{gpu_dev-0.5.25 → gpu_dev-0.5.27}/terraform-gpu-devservers/main.tf

@@ -58,6 +58,13 @@ provider "helm" {
 # Data sources
 data "aws_availability_zones" "available" {
   state = "available"
+  # Exclude Local Zones (e.g. us-east-1-dfw-2a) and Wavelength Zones — EKS control
+  # plane only supports standard AZs. us-east-2 doesn't have Local Zones so the
+  # existing prod workspace was unaffected; us-east-1 has several (dfw, bos, …).
+  filter {
+    name   = "opt-in-status"
+    values = ["opt-in-not-required"]
+  }
 }
 
 data "aws_caller_identity" "current" {}
@@ -315,6 +322,104 @@ locals {
         }
       }
     }
+    # us-east-1 spot-only experimental cluster.
+    # Same provisioning shape as prod (managed via the terraform.workspace switch) but
+    # backed entirely by EC2 Spot — first cheap-and-cheerful environment we can deploy
+    # new instance types into (B300 land here once on-demand quota arrives).
+    "prod-east1" = {
+      aws_region             = "us-east-1"
+      environment            = "prod-east1"
+      domain_name            = "east1.devservers.io"
+      gpu_instance_count     = 1
+      use_self_managed_nodes = true
+      instance_type          = "g4dn.12xlarge"
+      supported_gpu_types = {
+        # 8-GPU spot instances. instance_count=1 means the ASG tries to maintain 1
+        # spot instance per type — if AWS can't grant it (capacity / quota), the ASG
+        # sits at 0 and gpu-dev reservations queue. Bump counts once we see what
+        # actually gets fulfilled in us-east-1.
+        "b300" = {
+          instance_type       = "p6e-b300.48xlarge"
+          instance_types      = null
+          instance_count      = 1
+          gpus_per_instance   = 8
+          use_placement_group = false
+          architecture        = "x86_64"
+          efa_network_cards   = 8
+          use_spot            = true
+        }
+        "b200" = {
+          instance_type       = "p6-b200.48xlarge"
+          instance_types      = null
+          instance_count      = 1
+          gpus_per_instance   = 8
+          use_placement_group = false
+          architecture        = "x86_64"
+          efa_network_cards   = 8
+          use_spot            = true
+        }
+        "h200" = {
+          instance_type       = "p5e.48xlarge"
+          instance_types      = null
+          instance_count      = 1
+          gpus_per_instance   = 8
+          use_placement_group = false
+          architecture        = "x86_64"
+          efa_network_cards   = 16
+          use_spot            = true
+        }
+        "h100" = {
+          instance_type       = "p5.48xlarge"
+          instance_types      = null
+          instance_count      = 1
+          gpus_per_instance   = 8
+          use_placement_group = false
+          architecture        = "x86_64"
+          efa_network_cards   = 32
+          use_spot            = true
+        }
+        "a100" = {
+          instance_type       = "p4d.24xlarge"
+          instance_types      = null
+          instance_count      = 1
+          gpus_per_instance   = 8
+          use_placement_group = false
+          architecture        = "x86_64"
+          efa_network_cards   = 4
+          use_spot            = true
+        }
+        "t4" = {
+          instance_type       = "g4dn.12xlarge"
+          instance_types      = null
+          instance_count      = 1
+          gpus_per_instance   = 4
+          use_placement_group = false
+          architecture        = "x86_64"
+          efa_network_cards   = 0
+          use_spot            = true
+        }
+        "l4" = {
+          instance_type       = "g6.12xlarge"
+          instance_types      = null
+          instance_count      = 1
+          gpus_per_instance   = 4
+          use_placement_group = false
+          architecture        = "x86_64"
+          efa_network_cards   = 1
+          use_spot            = true
+        }
+        "cpu-x86" = {
+          instance_type       = "c7i.8xlarge"
+          instance_types      = null
+          instance_count      = 5
+          gpus_per_instance   = 0
+          use_placement_group = false
+          architecture        = "x86_64"
+          efa_network_cards   = 0
+          use_spot            = true
+        }
+      }
+    }
   }
 
   # Current workspace configuration
@@ -322,6 +427,9 @@ locals {
 
   # Workspace-specific capacity reservations (with manual instance counts)
   capacity_reservations = {
+    "prod-east1" = {
+      # No capacity reservations — this workspace is spot-only.
+    }
     default = {
       # Test environment capacity reservations
       # h100 = [
@@ -366,6 +474,20 @@ locals {
 
   # Workspace-specific GPU type to subnet mappings
   gpu_subnet_assignments = {
+    "prod-east1" = {
+      # All node types land in the primary subnet (us-east-1a). Multi-EFA types
+      # (efa_network_cards > 1) automatically use the private subnet in the same AZ.
+      # Specific instance types may not have capacity in us-east-1a — those ASGs will
+      # sit at 0 until we widen to other AZs, that's expected for beta.
+      b300       = "primary"
+      b200       = "primary"
+      h200       = "primary"
+      h100       = "primary"
+      a100       = "primary"
+      t4         = "primary"
+      l4         = "primary"
+      "cpu-x86"  = "primary"
+    }
     default = {
       # Test environment - T4 nodes in multiple AZs for testing
       t4         = "primary"   # T4 in us-west-1a (primary AZ)
@@ -390,8 +512,27 @@ locals {
     }
   }
 
+  # Subdomain NS delegations to create in *this* workspace's parent zone. Lets
+  # prod (which owns devservers.io) auto-publish NS records pointing at child zones
+  # in other workspaces (prod-east1, future regions) without manual -var flags.
+  # The NS values come from `tofu output devservers_name_servers` in the child
+  # workspace once its hosted zone has been created.
+  prod_subdomain_delegations = {
+    prod = {
+      "east1.devservers.io" = [
+        "ns-1079.awsdns-06.org",
+        "ns-1999.awsdns-57.co.uk",
+        "ns-341.awsdns-42.com",
+        "ns-624.awsdns-14.net",
+      ]
+    }
+  }
+
   # Per-capacity-reservation AZ mappings (overrides gpu_subnet_assignments when CR is used)
   capacity_reservation_azs = {
+    "prod-east1" = {
+      # Empty — no CRs in this workspace.
+    }
     default = {
       "cr-04d3d1d84e127a562" = "secondary" # us-west-1c
     }

gpu_dev-0.5.27/terraform-gpu-devservers/node-termination-handler.tf

@@ -0,0 +1,37 @@
+# AWS Node Termination Handler — graceful drain on spot-interrupt + ASG lifecycle events.
+#
+# IMDS mode (one DaemonSet per node, no SQS / no IAM role) is plenty for our use case:
+# we don't care about queue-processor features (rebalance recommendations, scheduled
+# events). We just want pods to get a clean SIGTERM when AWS sends the 2-minute spot
+# notice via instance metadata, instead of being killed cold.
+#
+# Tolerates everything so it runs on the GPU nodes that have nvidia.com/gpu:NoSchedule.
+
+resource "helm_release" "aws_node_termination_handler" {
+  name             = "aws-node-termination-handler"
+  repository       = "https://aws.github.io/eks-charts"
+  chart            = "aws-node-termination-handler"
+  namespace        = "kube-system"
+  # No version pin — chart versions advance frequently and my first guess (0.27.1)
+  # didn't exist. helm picks current latest stable. Add a pin once we hit a regression.
+  cleanup_on_fail  = true
+
+  values = [yamlencode({
+    enableSpotInterruptionDraining = true
+    enableScheduledEventDraining   = true
+    enableRebalanceMonitoring      = true
+    enableRebalanceDraining        = false # warning only; rebalance recommendations are too noisy
+    nodeSelector = {
+      "kubernetes.io/os" = "linux"
+    }
+    tolerations = [
+      { operator = "Exists" }, # tolerate every taint; we want NTH on every node, including GPU nodes
+    ]
+    resources = {
+      requests = { cpu = "50m", memory = "64Mi" }
+      limits   = { cpu = "100m", memory = "128Mi" }
+    }
+  })]
+
+  depends_on = [aws_eks_cluster.gpu_dev_cluster]
+}

{gpu_dev-0.5.25 → gpu_dev-0.5.27}/terraform-gpu-devservers/route53.tf

@@ -51,6 +51,19 @@ resource "aws_route53_record" "manual_subdomain_delegation" {
   records = var.subdomain_ns_records
 }
 
+# Auto-published NS delegations for child workspaces. Iterates prod_subdomain_delegations
+# (defined in main.tf) for the current workspace and creates an NS record per entry in
+# the parent zone — so `tofu apply` in prod automatically wires up east1.devservers.io
+# (and any future region) without -var flags.
+resource "aws_route53_record" "workspace_subdomain_delegations" {
+  for_each = local.effective_domain_name != "" && !local.is_subdomain ? try(local.prod_subdomain_delegations[terraform.workspace], {}) : {}
+  zone_id  = data.aws_route53_zone.parent[0].zone_id
+  name     = each.key
+  type     = "NS"
+  ttl      = 300
+  records  = each.value
+}
+
 # Use appropriate hosted zone (subdomain if created, otherwise parent)
 locals {
   hosted_zone_id = local.is_subdomain ? aws_route53_zone.subdomain[0].zone_id : (local.effective_domain_name != "" ? data.aws_route53_zone.parent[0].zone_id : "")