@gpuai/sdk 0.3.2 → 0.3.4

package/README.md

@@ -1,4 +1,4 @@
-# @gpuai/sdk@0.3.2
+# @gpuai/sdk@0.3.4
 
 A TypeScript SDK client for the api.gpu.ai API.
 
@@ -196,7 +196,7 @@ and is automatically generated by the
 [OpenAPI Generator](https://openapi-generator.tech) project:
 
 - API version: `1.0.0`
-- Package version: `0.3.2`
+- Package version: `0.3.4`
 - Generator version: `7.24.0`
 - Build package: `org.openapitools.codegen.languages.TypeScriptFetchClientCodegen`
 

package/dist/esm/models/CreateInstanceRequest.d.ts

@@ -64,7 +64,7 @@ export interface CreateInstanceRequest {
      */
     viewedPricePerHour?: number;
     /**
-     * Deploy a curated application template (e.g. comfyui, vllm). When set, gpu_type may be omitted — the cheapest eligible GPU is auto-selected. For templates serving a Hugging Face model via env.MODEL (vllm, sglang), the model's published weights size is checked against the GPU's VRAM at submit time: an oversized model on an explicit gpu_type returns 422 `model_too_large` (auto-select instead raises its VRAM floor to the estimate, and 422s only when no available GPU fits). Gated/private/unknown repos are never blocked — the deploy proceeds and the 202 carries a `warnings` entry.
+     * Deploy a curated application template (e.g. comfyui, vllm). When set, gpu_type may be omitted — the cheapest eligible GPU is auto-selected. For templates serving a Hugging Face model via env.MODEL (vllm, sglang), the model's published weights size is checked against the GPU's VRAM at submit time: an oversized model on an explicit gpu_type returns 422 `model_too_large` (auto-select instead raises its VRAM floor to the estimate, and 422s only when no available GPU fits). Gated/private/unknown repos are never blocked — the deploy proceeds and the 202 carries a `warnings` entry. Separately, an explicit gpu_type whose VRAM is below the template's own `min_vram_gb` returns 422 `gpu_vram_too_small`, naming the floor, the type's VRAM and the compatible GPU types. Auto-select is unaffected — it already filters by the floor.
      * @type {string}
      * @memberof CreateInstanceRequest
      */
@@ -88,13 +88,19 @@ export interface CreateInstanceRequest {
      */
     diskGb?: number;
     /**
-     * Per-deploy env overrides; only keys the template marks user_overridable are accepted (others 422). Secret values are redacted from operation metadata.
+     * Environment variables. With `template_id`: per-deploy overrides — only keys the template marks user_overridable are accepted (others 422), and secret values are redacted from operation metadata. With `image`: plain container env vars — keys must be POSIX names and the platform's reserved namespace (`PUBLIC_KEY`, `GPUAI_*`) is refused (422).
      * @type {{ [key: string]: string; }}
      * @memberof CreateInstanceRequest
      */
     env?: {
         [key: string]: string;
     };
+    /**
+     * Launch from your own container image, e.g. `pytorch/pytorch:2.4.1-cuda12.4-cudnn9-devel` or `ghcr.io/acme/trainer:v1`. The image must be publicly pullable (private registries are not yet supported) and provide linux/amd64 — both are verified synchronously before anything is provisioned: an unknown image returns 422 `image_unavailable` and a registry outage 502 `registry_unreachable`, with no operation created and nothing billed. The launch is pinned to the exact manifest digest verified, so a tag re-push cannot change what boots. After SSH/tunnel setup the image's own ENTRYPOINT/CMD is started (its output is in /var/log/gpuai-entrypoint.log on the instance); an image with neither is a plain SSH environment. Placement narrows to container-native capacity — a gpu_type only available on VM capacity returns 422 `image_unavailable`. Mutually exclusive with `template_id` and `environment` (422). A malformed reference (bad charset, localhost/private-IP registry) returns 400.
+     * @type {string}
+     * @memberof CreateInstanceRequest
+     */
+    image?: string;
 }
 /**
  * @export

package/dist/esm/models/CreateInstanceRequest.js

@@ -51,6 +51,7 @@ export function CreateInstanceRequestFromJSONTyped(json, ignoreDiscriminator) {
         'offeringId': json['offering_id'] == null ? undefined : json['offering_id'],
         'diskGb': json['disk_gb'] == null ? undefined : json['disk_gb'],
         'env': json['env'] == null ? undefined : json['env'],
+        'image': json['image'] == null ? undefined : json['image'],
     };
 }
 export function CreateInstanceRequestToJSON(json) {
@@ -74,5 +75,6 @@ export function CreateInstanceRequestToJSONTyped(value, ignoreDiscriminator = fa
         'offering_id': value['offeringId'],
         'disk_gb': value['diskGb'],
         'env': value['env'],
+        'image': value['image'],
     };
 }

package/dist/esm/models/Instance.d.ts

@@ -76,6 +76,12 @@ export interface Instance {
      * @memberof Instance
      */
     diskGb?: number;
+    /**
+     * The customer-supplied container image this instance was launched from, digest-pinned at verification time (e.g. `ghcr.io/acme/trainer:v1@sha256:…`). Present only on custom-image launches; template and certified launches never expose their images here.
+     * @type {string}
+     * @memberof Instance
+     */
+    image?: string;
     /**
      * Why the instance is in the error state, e.g. "provisioning stalled: the instance started but never became reachable". Present only on errored instances.
      * @type {string}

package/dist/esm/models/Instance.js

@@ -72,6 +72,7 @@ export function InstanceFromJSONTyped(json, ignoreDiscriminator) {
         'pricePerHour': json['price_per_hour'],
         'connection': json['connection'] == null ? undefined : InstanceConnectionFromJSON(json['connection']),
         'diskGb': json['disk_gb'] == null ? undefined : json['disk_gb'],
+        'image': json['image'] == null ? undefined : json['image'],
         'statusReason': json['status_reason'] == null ? undefined : json['status_reason'],
         'createdAt': (new Date(json['created_at'])),
         'readyAt': json['ready_at'] == null ? undefined : (new Date(json['ready_at'])),
@@ -98,6 +99,7 @@ export function InstanceToJSONTyped(value, ignoreDiscriminator = false) {
         'price_per_hour': value['pricePerHour'],
         'connection': InstanceConnectionToJSON(value['connection']),
         'disk_gb': value['diskGb'],
+        'image': value['image'],
         'status_reason': value['statusReason'],
         'created_at': value['createdAt'].toISOString(),
         'ready_at': value['readyAt'] == null ? value['readyAt'] : value['readyAt'].toISOString(),

package/dist/esm/models/Problem.d.ts

@@ -67,6 +67,7 @@ export declare const ProblemCodeEnum: {
     readonly DiskUnavailable: "disk_unavailable";
     readonly DiskExceedsOffering: "disk_exceeds_offering";
     readonly ModelTooLarge: "model_too_large";
+    readonly GpuVramTooSmall: "gpu_vram_too_small";
     readonly PriceChanged: "price_changed";
     readonly InsufficientBalance: "insufficient_balance";
     readonly QuotaExceeded: "quota_exceeded";

package/dist/esm/models/Problem.js

@@ -26,6 +26,7 @@ export const ProblemCodeEnum = {
     DiskUnavailable: 'disk_unavailable',
     DiskExceedsOffering: 'disk_exceeds_offering',
     ModelTooLarge: 'model_too_large',
+    GpuVramTooSmall: 'gpu_vram_too_small',
     PriceChanged: 'price_changed',
     InsufficientBalance: 'insufficient_balance',
     QuotaExceeded: 'quota_exceeded',

package/dist/models/CreateInstanceRequest.d.ts

@@ -64,7 +64,7 @@ export interface CreateInstanceRequest {
      */
     viewedPricePerHour?: number;
     /**
-     * Deploy a curated application template (e.g. comfyui, vllm). When set, gpu_type may be omitted — the cheapest eligible GPU is auto-selected. For templates serving a Hugging Face model via env.MODEL (vllm, sglang), the model's published weights size is checked against the GPU's VRAM at submit time: an oversized model on an explicit gpu_type returns 422 `model_too_large` (auto-select instead raises its VRAM floor to the estimate, and 422s only when no available GPU fits). Gated/private/unknown repos are never blocked — the deploy proceeds and the 202 carries a `warnings` entry.
+     * Deploy a curated application template (e.g. comfyui, vllm). When set, gpu_type may be omitted — the cheapest eligible GPU is auto-selected. For templates serving a Hugging Face model via env.MODEL (vllm, sglang), the model's published weights size is checked against the GPU's VRAM at submit time: an oversized model on an explicit gpu_type returns 422 `model_too_large` (auto-select instead raises its VRAM floor to the estimate, and 422s only when no available GPU fits). Gated/private/unknown repos are never blocked — the deploy proceeds and the 202 carries a `warnings` entry. Separately, an explicit gpu_type whose VRAM is below the template's own `min_vram_gb` returns 422 `gpu_vram_too_small`, naming the floor, the type's VRAM and the compatible GPU types. Auto-select is unaffected — it already filters by the floor.
      * @type {string}
      * @memberof CreateInstanceRequest
      */
@@ -88,13 +88,19 @@ export interface CreateInstanceRequest {
      */
     diskGb?: number;
     /**
-     * Per-deploy env overrides; only keys the template marks user_overridable are accepted (others 422). Secret values are redacted from operation metadata.
+     * Environment variables. With `template_id`: per-deploy overrides — only keys the template marks user_overridable are accepted (others 422), and secret values are redacted from operation metadata. With `image`: plain container env vars — keys must be POSIX names and the platform's reserved namespace (`PUBLIC_KEY`, `GPUAI_*`) is refused (422).
      * @type {{ [key: string]: string; }}
      * @memberof CreateInstanceRequest
      */
     env?: {
         [key: string]: string;
     };
+    /**
+     * Launch from your own container image, e.g. `pytorch/pytorch:2.4.1-cuda12.4-cudnn9-devel` or `ghcr.io/acme/trainer:v1`. The image must be publicly pullable (private registries are not yet supported) and provide linux/amd64 — both are verified synchronously before anything is provisioned: an unknown image returns 422 `image_unavailable` and a registry outage 502 `registry_unreachable`, with no operation created and nothing billed. The launch is pinned to the exact manifest digest verified, so a tag re-push cannot change what boots. After SSH/tunnel setup the image's own ENTRYPOINT/CMD is started (its output is in /var/log/gpuai-entrypoint.log on the instance); an image with neither is a plain SSH environment. Placement narrows to container-native capacity — a gpu_type only available on VM capacity returns 422 `image_unavailable`. Mutually exclusive with `template_id` and `environment` (422). A malformed reference (bad charset, localhost/private-IP registry) returns 400.
+     * @type {string}
+     * @memberof CreateInstanceRequest
+     */
+    image?: string;
 }
 /**
  * @export

package/dist/models/CreateInstanceRequest.js

@@ -59,6 +59,7 @@ function CreateInstanceRequestFromJSONTyped(json, ignoreDiscriminator) {
         'offeringId': json['offering_id'] == null ? undefined : json['offering_id'],
         'diskGb': json['disk_gb'] == null ? undefined : json['disk_gb'],
         'env': json['env'] == null ? undefined : json['env'],
+        'image': json['image'] == null ? undefined : json['image'],
     };
 }
 function CreateInstanceRequestToJSON(json) {
@@ -82,5 +83,6 @@ function CreateInstanceRequestToJSONTyped(value, ignoreDiscriminator = false) {
         'offering_id': value['offeringId'],
         'disk_gb': value['diskGb'],
         'env': value['env'],
+        'image': value['image'],
     };
 }

package/dist/models/Instance.d.ts

@@ -76,6 +76,12 @@ export interface Instance {
      * @memberof Instance
      */
     diskGb?: number;
+    /**
+     * The customer-supplied container image this instance was launched from, digest-pinned at verification time (e.g. `ghcr.io/acme/trainer:v1@sha256:…`). Present only on custom-image launches; template and certified launches never expose their images here.
+     * @type {string}
+     * @memberof Instance
+     */
+    image?: string;
     /**
      * Why the instance is in the error state, e.g. "provisioning stalled: the instance started but never became reachable". Present only on errored instances.
      * @type {string}

package/dist/models/Instance.js

@@ -80,6 +80,7 @@ function InstanceFromJSONTyped(json, ignoreDiscriminator) {
         'pricePerHour': json['price_per_hour'],
         'connection': json['connection'] == null ? undefined : (0, InstanceConnection_1.InstanceConnectionFromJSON)(json['connection']),
         'diskGb': json['disk_gb'] == null ? undefined : json['disk_gb'],
+        'image': json['image'] == null ? undefined : json['image'],
         'statusReason': json['status_reason'] == null ? undefined : json['status_reason'],
         'createdAt': (new Date(json['created_at'])),
         'readyAt': json['ready_at'] == null ? undefined : (new Date(json['ready_at'])),
@@ -106,6 +107,7 @@ function InstanceToJSONTyped(value, ignoreDiscriminator = false) {
         'price_per_hour': value['pricePerHour'],
         'connection': (0, InstanceConnection_1.InstanceConnectionToJSON)(value['connection']),
         'disk_gb': value['diskGb'],
+        'image': value['image'],
         'status_reason': value['statusReason'],
         'created_at': value['createdAt'].toISOString(),
         'ready_at': value['readyAt'] == null ? value['readyAt'] : value['readyAt'].toISOString(),

package/dist/models/Problem.d.ts

@@ -67,6 +67,7 @@ export declare const ProblemCodeEnum: {
     readonly DiskUnavailable: "disk_unavailable";
     readonly DiskExceedsOffering: "disk_exceeds_offering";
     readonly ModelTooLarge: "model_too_large";
+    readonly GpuVramTooSmall: "gpu_vram_too_small";
     readonly PriceChanged: "price_changed";
     readonly InsufficientBalance: "insufficient_balance";
     readonly QuotaExceeded: "quota_exceeded";

package/dist/models/Problem.js

@@ -34,6 +34,7 @@ exports.ProblemCodeEnum = {
     DiskUnavailable: 'disk_unavailable',
     DiskExceedsOffering: 'disk_exceeds_offering',
     ModelTooLarge: 'model_too_large',
+    GpuVramTooSmall: 'gpu_vram_too_small',
     PriceChanged: 'price_changed',
     InsufficientBalance: 'insufficient_balance',
     QuotaExceeded: 'quota_exceeded',

package/docs/CreateInstanceRequest.md

@@ -19,6 +19,7 @@ Name | Type
 `offeringId` | string
 `diskGb` | number
 `env` | { [key: string]: string; }
+`image` | string
 
 ## Example
 
@@ -40,6 +41,7 @@ const example = {
   "offeringId": null,
   "diskGb": 200,
   "env": null,
+  "image": ghcr.io/acme/trainer:v1,
 } satisfies CreateInstanceRequest
 
 console.log(example)

package/docs/Instance.md

@@ -16,6 +16,7 @@ Name | Type
 `pricePerHour` | number
 `connection` | [InstanceConnection](InstanceConnection.md)
 `diskGb` | number
+`image` | string
 `statusReason` | string
 `createdAt` | Date
 `readyAt` | Date
@@ -40,6 +41,7 @@ const example = {
   "pricePerHour": null,
   "connection": null,
   "diskGb": null,
+  "image": null,
   "statusReason": null,
   "createdAt": null,
   "readyAt": null,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gpuai/sdk",
-  "version": "0.3.2",
+  "version": "0.3.4",
   "description": "TypeScript SDK for the GPU.ai Public Developer API",
   "author": "GPU.ai Support <support@gpu.ai>",
   "repository": {

package/src/models/CreateInstanceRequest.ts

@@ -68,7 +68,7 @@ export interface CreateInstanceRequest {
      */
     viewedPricePerHour?: number;
     /**
-     * Deploy a curated application template (e.g. comfyui, vllm). When set, gpu_type may be omitted — the cheapest eligible GPU is auto-selected. For templates serving a Hugging Face model via env.MODEL (vllm, sglang), the model's published weights size is checked against the GPU's VRAM at submit time: an oversized model on an explicit gpu_type returns 422 `model_too_large` (auto-select instead raises its VRAM floor to the estimate, and 422s only when no available GPU fits). Gated/private/unknown repos are never blocked — the deploy proceeds and the 202 carries a `warnings` entry.
+     * Deploy a curated application template (e.g. comfyui, vllm). When set, gpu_type may be omitted — the cheapest eligible GPU is auto-selected. For templates serving a Hugging Face model via env.MODEL (vllm, sglang), the model's published weights size is checked against the GPU's VRAM at submit time: an oversized model on an explicit gpu_type returns 422 `model_too_large` (auto-select instead raises its VRAM floor to the estimate, and 422s only when no available GPU fits). Gated/private/unknown repos are never blocked — the deploy proceeds and the 202 carries a `warnings` entry. Separately, an explicit gpu_type whose VRAM is below the template's own `min_vram_gb` returns 422 `gpu_vram_too_small`, naming the floor, the type's VRAM and the compatible GPU types. Auto-select is unaffected — it already filters by the floor.
      * @type {string}
      * @memberof CreateInstanceRequest
      */
@@ -92,11 +92,17 @@ export interface CreateInstanceRequest {
      */
     diskGb?: number;
     /**
-     * Per-deploy env overrides; only keys the template marks user_overridable are accepted (others 422). Secret values are redacted from operation metadata.
+     * Environment variables. With `template_id`: per-deploy overrides — only keys the template marks user_overridable are accepted (others 422), and secret values are redacted from operation metadata. With `image`: plain container env vars — keys must be POSIX names and the platform's reserved namespace (`PUBLIC_KEY`, `GPUAI_*`) is refused (422).
      * @type {{ [key: string]: string; }}
      * @memberof CreateInstanceRequest
      */
     env?: { [key: string]: string; };
+    /**
+     * Launch from your own container image, e.g. `pytorch/pytorch:2.4.1-cuda12.4-cudnn9-devel` or `ghcr.io/acme/trainer:v1`. The image must be publicly pullable (private registries are not yet supported) and provide linux/amd64 — both are verified synchronously before anything is provisioned: an unknown image returns 422 `image_unavailable` and a registry outage 502 `registry_unreachable`, with no operation created and nothing billed. The launch is pinned to the exact manifest digest verified, so a tag re-push cannot change what boots. After SSH/tunnel setup the image's own ENTRYPOINT/CMD is started (its output is in /var/log/gpuai-entrypoint.log on the instance); an image with neither is a plain SSH environment. Placement narrows to container-native capacity — a gpu_type only available on VM capacity returns 422 `image_unavailable`. Mutually exclusive with `template_id` and `environment` (422). A malformed reference (bad charset, localhost/private-IP registry) returns 400.
+     * @type {string}
+     * @memberof CreateInstanceRequest
+     */
+    image?: string;
 }
 
 
@@ -143,6 +149,7 @@ export function CreateInstanceRequestFromJSONTyped(json: any, ignoreDiscriminato
         'offeringId': json['offering_id'] == null ? undefined : json['offering_id'],
         'diskGb': json['disk_gb'] == null ? undefined : json['disk_gb'],
         'env': json['env'] == null ? undefined : json['env'],
+        'image': json['image'] == null ? undefined : json['image'],
     };
 }
 
@@ -170,6 +177,7 @@ export function CreateInstanceRequestToJSONTyped(value?: CreateInstanceRequest |
         'offering_id': value['offeringId'],
         'disk_gb': value['diskGb'],
         'env': value['env'],
+        'image': value['image'],
     };
 }
 

package/src/models/Instance.ts

@@ -87,6 +87,12 @@ export interface Instance {
      * @memberof Instance
      */
     diskGb?: number;
+    /**
+     * The customer-supplied container image this instance was launched from, digest-pinned at verification time (e.g. `ghcr.io/acme/trainer:v1@sha256:…`). Present only on custom-image launches; template and certified launches never expose their images here.
+     * @type {string}
+     * @memberof Instance
+     */
+    image?: string;
     /**
      * Why the instance is in the error state, e.g. "provisioning stalled: the instance started but never became reachable". Present only on errored instances.
      * @type {string}
@@ -186,6 +192,7 @@ export function InstanceFromJSONTyped(json: any, ignoreDiscriminator: boolean):
         'pricePerHour': json['price_per_hour'],
         'connection': json['connection'] == null ? undefined : InstanceConnectionFromJSON(json['connection']),
         'diskGb': json['disk_gb'] == null ? undefined : json['disk_gb'],
+        'image': json['image'] == null ? undefined : json['image'],
         'statusReason': json['status_reason'] == null ? undefined : json['status_reason'],
         'createdAt': (new Date(json['created_at'])),
         'readyAt': json['ready_at'] == null ? undefined : (new Date(json['ready_at'])),
@@ -216,6 +223,7 @@ export function InstanceToJSONTyped(value?: Instance | null, ignoreDiscriminator
         'price_per_hour': value['pricePerHour'],
         'connection': InstanceConnectionToJSON(value['connection']),
         'disk_gb': value['diskGb'],
+        'image': value['image'],
         'status_reason': value['statusReason'],
         'created_at': value['createdAt'].toISOString(),
         'ready_at': value['readyAt'] == null ? value['readyAt'] : value['readyAt'].toISOString(),

package/src/models/Problem.ts

@@ -73,6 +73,7 @@ export const ProblemCodeEnum = {
     DiskUnavailable: 'disk_unavailable',
     DiskExceedsOffering: 'disk_exceeds_offering',
     ModelTooLarge: 'model_too_large',
+    GpuVramTooSmall: 'gpu_vram_too_small',
     PriceChanged: 'price_changed',
     InsufficientBalance: 'insufficient_balance',
     QuotaExceeded: 'quota_exceeded',