@aws/ml-container-creator 0.2.6 → 0.4.0

package/servers/instance-sizer/lib/model-resolver.js

@@ -214,12 +214,16 @@ const resolveModelMetadata = async (modelName, options = {}) => {
     const catalogEntry = catalogLookup(modelName, catalog)
 
     if (catalogEntry) {
-        return {
-            parameterCount: catalogEntry.parameterCount,
-            dtype: catalogEntry.defaultDtype,
-            architecture: catalogEntry.architecture,
-            maxPositionEmbeddings: catalogEntry.maxPositionEmbeddings,
-            source: 'catalog'
+        // Only use catalog entry if it has a usable parameterCount for VRAM estimation.
+        // If parameterCount is missing, fall through to HuggingFace API (tier 2).
+        if (catalogEntry.parameterCount) {
+            return {
+                parameterCount: catalogEntry.parameterCount,
+                dtype: catalogEntry.defaultDtype,
+                architecture: catalogEntry.architecture,
+                maxPositionEmbeddings: catalogEntry.maxPositionEmbeddings,
+                source: 'catalog'
+            }
         }
     }
 

package/servers/instance-sizer/lib/quota-resolver.js

@@ -0,0 +1,368 @@
+// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+// SPDX-License-Identifier: Apache-2.0
+
+/**
+ * Quota Resolver
+ *
+ * Queries AWS APIs to determine account-level quota headroom, capacity
+ * reservations, and Flexible Training Plans for SageMaker instance types.
+ * Used in discover mode to filter and prioritize instance recommendations.
+ *
+ * All methods degrade gracefully — API failures return null and log to stderr.
+ */
+
+import { ServiceQuotasClient, ListServiceQuotasCommand } from '@aws-sdk/client-service-quotas'
+import { SageMakerClient, ListEndpointsCommand, ListTrainingPlansCommand } from '@aws-sdk/client-sagemaker'
+
+// ── Constants ────────────────────────────────────────────────────────────────
+
+const SAGEMAKER_SERVICE_CODE = 'sagemaker'
+const DEFAULT_TIMEOUT_MS = 5000
+const DEFAULT_CACHE_TTL_MS = 300000 // 5 minutes
+const QUOTA_NAME_PATTERN = /^(ml\.[a-z0-9]+\.[a-z0-9]+) for endpoint usage$/
+
+// ── Logging ──────────────────────────────────────────────────────────────────
+
+function log(message) {
+    process.stderr.write(`[quota-resolver] ${message}\n`)
+}
+
+// ── QuotaResolver Class ──────────────────────────────────────────────────────
+
+class QuotaResolver {
+    /**
+     * @param {string} region - AWS region to query
+     * @param {object} [options={}]
+     * @param {number} [options.timeout=5000] - Timeout per API call in ms
+     * @param {number} [options.cacheTtl=300000] - Cache TTL in ms (default 5 min)
+     */
+    constructor(region, options = {}) {
+        this.region = region
+        this.timeout = options.timeout || DEFAULT_TIMEOUT_MS
+        this.cacheTtl = options.cacheTtl || DEFAULT_CACHE_TTL_MS
+        this.cache = new Map()
+
+        const clientConfig = {
+            region: this.region,
+            requestHandler: {
+                requestTimeout: this.timeout
+            }
+        }
+
+        this.quotasClient = new ServiceQuotasClient(clientConfig)
+        this.sagemakerClient = new SageMakerClient(clientConfig)
+    }
+
+    /**
+     * Check cache for a key. Returns cached value if within TTL, else null.
+     * @param {string} key - Cache key
+     * @returns {*|null} Cached value or null
+     */
+    _getCached(key) {
+        const entry = this.cache.get(key)
+        if (!entry) return null
+        if (Date.now() - entry.timestamp > this.cacheTtl) {
+            this.cache.delete(key)
+            return null
+        }
+        return entry.value
+    }
+
+    /**
+     * Store a value in the cache with current timestamp.
+     * @param {string} key - Cache key
+     * @param {*} value - Value to cache
+     */
+    _setCache(key, value) {
+        this.cache.set(key, { value, timestamp: Date.now() })
+    }
+
+    /**
+     * Parse a SageMaker quota name to extract the instance type.
+     * Expected pattern: "ml.<family>.<size> for endpoint usage"
+     *
+     * @param {string} quotaName - Quota name from Service Quotas API
+     * @returns {string|null} Instance type or null if pattern doesn't match
+     */
+    _parseQuotaName(quotaName) {
+        const match = quotaName.match(QUOTA_NAME_PATTERN)
+        return match ? match[1] : null
+    }
+
+    /**
+     * Get quota headroom for a list of instance types.
+     *
+     * Queries Service Quotas for SageMaker endpoint instance limits and
+     * ListEndpoints to count currently deployed instances per type.
+     * Headroom = quota limit - deployed count.
+     *
+     * @param {string[]} instanceTypes - Instance types to check (e.g., ['ml.g5.xlarge'])
+     * @returns {Promise<Map|null>} Map: instanceType → { quota, deployed, headroom }, or null on failure
+     */
+    async getQuotaHeadroom(instanceTypes) {
+        const cacheKey = 'quotaHeadroom'
+        const cached = this._getCached(cacheKey)
+        if (cached) return cached
+
+        try {
+            const [quotaMap, deployedMap] = await Promise.allSettled([
+                this._fetchServiceQuotas(),
+                this._fetchDeployedCounts()
+            ])
+
+            const quotas = quotaMap.status === 'fulfilled' ? quotaMap.value : null
+            const deployed = deployedMap.status === 'fulfilled' ? deployedMap.value : null
+
+            if (!quotas) {
+                return null
+            }
+
+            const result = new Map()
+            const deployedCounts = deployed || new Map()
+
+            for (const instanceType of instanceTypes) {
+                const quota = quotas.get(instanceType)
+                if (quota != null) {
+                    const deployedCount = deployedCounts.get(instanceType) || 0
+                    const headroom = quota - deployedCount
+                    result.set(instanceType, {
+                        quota,
+                        deployed: deployedCount,
+                        headroom
+                    })
+                }
+            }
+
+            this._setCache(cacheKey, result)
+            return result
+        } catch (err) {
+            if (err.name === 'AccessDeniedException' || err.Code === 'AccessDeniedException') {
+                log(`AccessDenied: insufficient permissions for quota queries — skipping`)
+            } else if (err.name === 'ThrottlingException' || err.Code === 'ThrottlingException') {
+                log(`Throttled: Service Quotas API rate limit hit — skipping`)
+            } else {
+                log(`Failed to get quota headroom: ${err.message}`)
+            }
+            return null
+        }
+    }
+
+    /**
+     * Fetch all SageMaker service quotas for endpoint instance types.
+     * Paginates through all results.
+     *
+     * @returns {Promise<Map>} Map: instanceType → quota limit (number)
+     */
+    async _fetchServiceQuotas() {
+        const quotaMap = new Map()
+        let nextToken = undefined
+
+        do {
+            const command = new ListServiceQuotasCommand({
+                ServiceCode: SAGEMAKER_SERVICE_CODE,
+                ...(nextToken && { NextToken: nextToken })
+            })
+
+            const response = await this.quotasClient.send(command)
+
+            for (const quota of (response.Quotas || [])) {
+                const instanceType = this._parseQuotaName(quota.QuotaName || '')
+                if (instanceType && quota.Value != null) {
+                    quotaMap.set(instanceType, quota.Value)
+                }
+            }
+
+            nextToken = response.NextToken
+        } while (nextToken)
+
+        return quotaMap
+    }
+
+    /**
+     * Fetch currently deployed endpoint instances and count per type.
+     * Paginates through all endpoints.
+     *
+     * @returns {Promise<Map>} Map: instanceType → deployed count
+     */
+    async _fetchDeployedCounts() {
+        const deployedMap = new Map()
+        let nextToken = undefined
+
+        do {
+            const command = new ListEndpointsCommand({
+                StatusEquals: 'InService',
+                ...(nextToken && { NextToken: nextToken })
+            })
+
+            const response = await this.sagemakerClient.send(command)
+
+            for (const endpoint of (response.Endpoints || [])) {
+                // ListEndpoints returns endpoint summaries; instance type info
+                // is in the ProductionVariants. We count each endpoint as 1
+                // instance of its configured type. For more accurate counts,
+                // DescribeEndpoint would be needed, but that's too many API calls.
+                // The endpoint name often encodes the instance type, but the
+                // reliable approach is to count endpoints and map via config.
+                // For now, we track endpoint counts by checking production variants
+                // if available, otherwise skip.
+                if (endpoint.ProductionVariants) {
+                    for (const variant of endpoint.ProductionVariants) {
+                        if (variant.InstanceType) {
+                            const current = deployedMap.get(variant.InstanceType) || 0
+                            const count = variant.CurrentInstanceCount || 1
+                            deployedMap.set(variant.InstanceType, current + count)
+                        }
+                    }
+                }
+            }
+
+            nextToken = response.NextToken
+        } while (nextToken)
+
+        return deployedMap
+    }
+
+    /**
+     * Get active Training Plan reservations for inference endpoints.
+     *
+     * Queries ListTrainingPlans for active plans with TargetResources=endpoint.
+     * These are SageMaker-managed capacity reservations that can be referenced
+     * via MlReservationArn in CreateEndpointConfig.
+     *
+     * ⚠️  EXPERIMENTAL: Training Plans for inference is a newer feature.
+     *
+     * @returns {Promise<Map|null>} Map: instanceType → { planName, planArn, remainingCapacity, startDate, endDate }, or null on failure
+     */
+    async getCapacityReservations() {
+        const cacheKey = 'capacityReservations'
+        const cached = this._getCached(cacheKey)
+        if (cached) return cached
+
+        try {
+            const result = new Map()
+            let nextToken = undefined
+
+            do {
+                const command = new ListTrainingPlansCommand({
+                    StatusEquals: 'Active',
+                    ...(nextToken && { NextToken: nextToken })
+                })
+
+                const response = await this.sagemakerClient.send(command)
+                const now = new Date()
+
+                for (const plan of (response.TrainingPlanSummaries || [])) {
+                    // Only include plans targeting inference endpoints
+                    const targetResources = plan.TargetResources || []
+                    if (!targetResources.includes('endpoint')) continue
+
+                    const instanceType = plan.InstanceType || plan.ReservedCapacityInstanceType
+                    if (!instanceType) continue
+
+                    const planArn = plan.TrainingPlanArn
+                    const planName = plan.TrainingPlanName || 'unknown'
+                    const remainingCapacity = plan.AvailableInstanceCount
+                        ?? plan.RemainingCapacity
+                        ?? plan.TotalInstanceCount
+                        ?? 0
+                    const startDate = plan.StartTime || null
+                    const endDate = plan.EndTime || plan.ExpirationTime || null
+
+                    // Skip plans outside their time window
+                    if (startDate && new Date(startDate) > now) continue
+                    if (endDate && new Date(endDate) < now) continue
+
+                    // Only include if there's remaining capacity
+                    if (remainingCapacity <= 0) continue
+
+                    result.set(instanceType, {
+                        planName,
+                        planArn,
+                        type: 'training-plan',
+                        count: remainingCapacity,
+                        startDate: startDate ? (startDate instanceof Date ? startDate.toISOString() : startDate) : null,
+                        endDate: endDate ? (endDate instanceof Date ? endDate.toISOString() : endDate) : null
+                    })
+                }
+
+                nextToken = response.NextToken
+            } while (nextToken)
+
+            this._setCache(cacheKey, result)
+            return result
+        } catch (err) {
+            if (err.name === 'AccessDeniedException' || err.Code === 'AccessDeniedException') {
+                log(`AccessDenied: insufficient permissions for training plan queries — skipping`)
+            } else if (err.name === 'ValidationException') {
+                log(`ListTrainingPlans not available in region ${this.region} — skipping`)
+            } else if (err.name === 'ThrottlingException' || err.Code === 'ThrottlingException') {
+                log(`Throttled: ListTrainingPlans rate limit hit — skipping`)
+            } else {
+                log(`Failed to get capacity reservations: ${err.message}`)
+            }
+            return null
+        }
+    }
+
+    /**
+     * Get active Flexible Training Plans with remaining capacity.
+     *
+     * Calls ListTrainingPlans with status filter for active plans and
+     * extracts instance types and remaining capacity from each plan.
+     *
+     * @returns {Promise<Map|null>} Map: instanceType → { planName, remainingCapacity, expiresAt }, or null on failure
+     */
+    async getTrainingPlans() {
+        const cacheKey = 'trainingPlans'
+        const cached = this._getCached(cacheKey)
+        if (cached) return cached
+
+        try {
+            const result = new Map()
+            let nextToken = undefined
+
+            do {
+                const command = new ListTrainingPlansCommand({
+                    StatusEquals: 'Active',
+                    ...(nextToken && { NextToken: nextToken })
+                })
+
+                const response = await this.sagemakerClient.send(command)
+
+                for (const plan of (response.TrainingPlanSummaries || [])) {
+                    const instanceType = plan.InstanceType || plan.ReservedCapacityInstanceType
+                    const planName = plan.TrainingPlanName || plan.TrainingPlanArn || 'unknown'
+                    const remainingCapacity = plan.AvailableInstanceCount
+                        ?? plan.RemainingCapacity
+                        ?? plan.TotalInstanceCount
+                        ?? 0
+                    const expiresAt = plan.EndTime || plan.ExpirationTime || null
+
+                    if (instanceType && remainingCapacity > 0) {
+                        result.set(instanceType, {
+                            planName,
+                            remainingCapacity,
+                            expiresAt
+                        })
+                    }
+                }
+
+                nextToken = response.NextToken
+            } while (nextToken)
+
+            this._setCache(cacheKey, result)
+            return result
+        } catch (err) {
+            if (err.name === 'AccessDeniedException' || err.Code === 'AccessDeniedException') {
+                log(`AccessDenied: insufficient permissions for training plan queries — skipping`)
+            } else if (err.name === 'ValidationException') {
+                log(`ListTrainingPlans not available in region ${this.region} — skipping`)
+            } else {
+                log(`Failed to get training plans: ${err.message}`)
+            }
+            return null
+        }
+    }
+}
+
+export { QuotaResolver, QUOTA_NAME_PATTERN, SAGEMAKER_SERVICE_CODE, DEFAULT_TIMEOUT_MS, DEFAULT_CACHE_TTL_MS }

package/servers/instance-sizer/package.json

@@ -10,6 +10,8 @@
     "test": "node test.js"
   },
   "dependencies": {
+    "@aws-sdk/client-sagemaker": "^3.700.0",
+    "@aws-sdk/client-service-quotas": "^3.700.0",
     "@modelcontextprotocol/sdk": "^1.0.0"
   }
 }