@aws/ml-container-creator 0.3.0 → 0.4.0

package/src/lib/prompt-runner.js

@@ -18,6 +18,7 @@ import {
     modelLoadStrategyPrompts,
     modelProfilePrompts,
     modulePrompts,
+    benchmarkPrompts,
     infraRegionAndTargetPrompts,
     infraInstancePrompts,
     infraAsyncPrompts,
@@ -252,6 +253,29 @@ export default class PromptRunner {
                 this._autoGpuCount = tpRec.gpuCount;
                 console.log(`   ✓ Auto-set tensor parallelism: TP=${tpRec.tensorParallelism} (${tpRec.gpuCount} GPUs)`);
             }
+
+            // Display capacity type confirmation for selected instance
+            // Requirements: 5.4
+            if (matchingRec && matchingRec.capacityType) {
+                if (matchingRec.capacityType === 'reserved') {
+                    const resType = matchingRec.reservationType === 'capacity-block' ? 'Capacity Block' : 'ODCR';
+                    const endInfo = matchingRec.reservationType === 'capacity-block' && matchingRec.reservationInfo?.endDate
+                        ? `, ends ${new Date(matchingRec.reservationInfo.endDate).toLocaleDateString()}`
+                        : '';
+                    console.log(`   ✓ Using reserved capacity — ${resType} (reservation ${matchingRec.reservationInfo?.reservationId || 'unknown'}${endInfo})`);
+                } else if (matchingRec.capacityType === 'ftp') {
+                    console.log(`   ✓ Using reserved capacity (plan ${matchingRec.ftpInfo?.planName || 'unknown'})`);
+                } else {
+                    const headroom = matchingRec.quotaHeadroom;
+                    console.log(`   ✓ Using on-demand capacity (quota headroom: ${headroom ?? 'unknown'})`);
+                }
+            }
+
+            // Extract reservation ARN from selected instance for deployment config
+            // Requirements: 2.3
+            if (matchingRec && matchingRec.capacityType === 'reserved' && matchingRec.reservationInfo?.reservationArn) {
+                this._selectedCapacityReservationArn = matchingRec.reservationInfo.reservationArn;
+            }
         }
 
         // 3c. Async-specific prompts (only when deploymentTarget === 'async-inference')
@@ -375,6 +399,21 @@ export default class PromptRunner {
             moduleAnswers.includeSampleModel = false;
         }
 
+        // Benchmark prompts — derive includeBenchmark from testTypes selection or CLI flag
+        // Requirements: 1.1, 1.2
+        let benchmarkAnswers = {};
+        if (frameworkAnswers.architecture === 'transformers' || frameworkAnswers.architecture === 'diffusors') {
+            const testTypes = moduleAnswers.testTypes || [];
+            const includeBenchmark = testTypes.includes('sagemaker-ai-automated-benchmarking') ||
+                explicitConfig.includeBenchmark === true ||
+                explicitConfig.includeBenchmark === 'true';
+            benchmarkAnswers.includeBenchmark = includeBenchmark;
+            if (includeBenchmark) {
+                const subAnswers = await this._runPhase(benchmarkPrompts, { ...frameworkAnswers, ...moduleAnswers, includeBenchmark }, explicitConfig, existingConfig);
+                benchmarkAnswers = { ...benchmarkAnswers, ...subAnswers };
+            }
+        }
+
         // Validate instance type against framework requirements (now that framework version is known)
         const finalInstanceType = infraAnswers.customInstanceType || infraAnswers.instanceType;
         if (finalInstanceType && frameworkVersionAnswers.frameworkVersion) {
@@ -416,6 +455,7 @@ export default class PromptRunner {
             ...hfTokenAnswers,
             ...ngcApiKeyAnswers,
             ...moduleAnswers,
+            ...benchmarkAnswers,
             ...projectAnswers,
             ...destinationAnswers,
             buildTimestamp
@@ -435,6 +475,12 @@ export default class PromptRunner {
             combinedAnswers.artifactUri = this._mcpArtifactUri;
         }
 
+        // Flow capacity reservation ARN from instance-sizer selection
+        // Requirements: 2.3
+        if (this._selectedCapacityReservationArn) {
+            combinedAnswers.capacityReservationArn = this._selectedCapacityReservationArn;
+        }
+
         // Validate: non-HF model sources require an artifact URI
         // Without it, the serve script can't download the model at runtime
         // Infer modelSource from model name prefix if not set by MCP
@@ -1036,13 +1082,53 @@ export default class PromptRunner {
                         : '';
 
                     console.log(`   ✓ ${choices.length} compatible instance(s) found${vramInfo}`);
-                    // Display compact recommendation table
-                    for (const rec of recommendations) {
-                        const tp = rec.tensorParallelism > 1 ? ` TP=${rec.tensorParallelism}` : '';
-                        const vram = rec.totalVramGb ? `${rec.totalVramGb}GB` : '?';
-                        const util = rec.utilizationPercent ? `${rec.utilizationPercent}%` : '?';
-                        console.log(`     ${rec === topRec ? '→' : ' '} ${rec.instanceType.padEnd(20)} ${vram.padStart(5)} VRAM  ${util.padStart(4)} util${tp}`);
+
+                    // Check if availability data is present (recommendations have capacityType)
+                    const hasAvailabilityData = recommendations.some(r => r.capacityType);
+
+                    if (hasAvailabilityData) {
+                        // Group by capacityType for display
+                        const reserved = recommendations.filter(r => r.capacityType === 'reserved' || r.capacityType === 'ftp');
+                        const onDemand = recommendations.filter(r => r.capacityType === 'on-demand');
+
+                        if (reserved.length > 0) {
+                            console.log('     ── Reserved Capacity ──');
+                            for (const rec of reserved) {
+                                const tp = rec.tensorParallelism > 1 ? ` TP=${rec.tensorParallelism}` : '';
+                                const vram = rec.totalVramGb ? `${rec.totalVramGb}GB` : '?';
+                                const util = rec.utilizationPercent ? `${rec.utilizationPercent}%` : '?';
+                                const tag = rec.capacityType === 'reserved'
+                                    ? ` [CR] ${rec.reservationInfo?.planName || rec.reservationInfo?.reservationId || ''}`
+                                    : ` [FTP] ${rec.ftpInfo?.planName || ''}`;
+                                console.log(`     ${rec === topRec ? '→' : ' '} ${rec.instanceType.padEnd(20)} ${vram.padStart(5)} VRAM  ${util.padStart(4)} util${tp}${tag}`);
+                            }
+                        }
+
+                        if (onDemand.length > 0) {
+                            console.log('     ── On-Demand ──');
+                            for (const rec of onDemand) {
+                                const tp = rec.tensorParallelism > 1 ? ` TP=${rec.tensorParallelism}` : '';
+                                const vram = rec.totalVramGb ? `${rec.totalVramGb}GB` : '?';
+                                const util = rec.utilizationPercent ? `${rec.utilizationPercent}%` : '?';
+                                const deployed = rec.quotaDeployed;
+                                const quota = rec.quotaLimit;
+                                const tag = quota !== null && quota !== undefined ? ` [Q:${deployed ?? 0}/${quota}]` : '';
+                                console.log(`     ${rec === topRec ? '→' : ' '} ${rec.instanceType.padEnd(20)} ${vram.padStart(5)} VRAM  ${util.padStart(4)} util${tp}${tag}`);
+                            }
+                        }
+                    } else {
+                        // Fallback: display compact recommendation table (no availability data)
+                        for (const rec of recommendations) {
+                            const tp = rec.tensorParallelism > 1 ? ` TP=${rec.tensorParallelism}` : '';
+                            const vram = rec.totalVramGb ? `${rec.totalVramGb}GB` : '?';
+                            const util = rec.utilizationPercent ? `${rec.utilizationPercent}%` : '?';
+                            console.log(`     ${rec === topRec ? '→' : ' '} ${rec.instanceType.padEnd(20)} ${vram.padStart(5)} VRAM  ${util.padStart(4)} util${tp}`);
+                        }
                     }
+                } else if (parsed.metadata?.allFilteredByQuota) {
+                    // All VRAM-compatible instances had zero quota
+                    console.log('   ⚠️ No quota available for compatible instances. Request a quota increase.');
+                    this._instanceSizerMetadata = parsed.metadata || null;
                 } else if (parsed.metadata?.warning) {
                     console.log(`   ⚠️  ${parsed.metadata.warning}`);
                 } else {
@@ -1972,9 +2058,10 @@ export default class PromptRunner {
         '11.4': 'al2-ami-sagemaker-inference-gpu-2-1',
         '11.8': 'al2-ami-sagemaker-inference-gpu-2-1',
         '12.1': 'al2-ami-sagemaker-inference-gpu-3-1',
-        '12.2': 'al2023-ami-sagemaker-inference-gpu-4-1',
-        '12.4': 'al2023-ami-sagemaker-inference-gpu-4-1',
-        '12.6': 'al2023-ami-sagemaker-inference-gpu-4-1'
+        '12.2': 'al2-ami-sagemaker-inference-gpu-3-1',
+        '12.4': 'al2-ami-sagemaker-inference-gpu-3-1',
+        '12.6': 'al2-ami-sagemaker-inference-gpu-3-1',
+        '13.0': 'al2023-ami-sagemaker-inference-gpu-4-1'
     };
 
     /**

package/src/lib/prompts.js

@@ -583,7 +583,7 @@ const modulePrompts = [
         type: 'confirm',
         name: 'includeSampleModel',
         message: 'Include sample Abalone classifier?',
-        default: false,
+        default: true,
         when: (answers) => {
             const architecture = answers.architecture || answers.deploymentConfig?.split('-')[0];
             const backend = answers.backend || answers.deploymentConfig?.split('-').slice(1).join('-');
@@ -622,7 +622,10 @@ const modulePrompts = [
             
             // Transformers and Triton LLM backends only support hosted endpoint tests
             if (architecture === 'transformers') {
-                return ['hosted-model-endpoint'];
+                return ['hosted-model-endpoint', 'sagemaker-ai-automated-benchmarking'];
+            }
+            if (architecture === 'diffusors') {
+                return ['hosted-model-endpoint', 'sagemaker-ai-automated-benchmarking'];
             }
             if (architecture === 'triton' && (backend === 'vllm' || backend === 'tensorrtllm')) {
                 return ['hosted-model-endpoint'];
@@ -635,7 +638,10 @@ const modulePrompts = [
             const backend = answers.backend || answers.deploymentConfig?.split('-').slice(1).join('-');
             
             if (architecture === 'transformers') {
-                return ['hosted-model-endpoint'];
+                return ['hosted-model-endpoint', 'sagemaker-ai-automated-benchmarking'];
+            }
+            if (architecture === 'diffusors') {
+                return ['hosted-model-endpoint', 'sagemaker-ai-automated-benchmarking'];
             }
             if (architecture === 'triton' && (backend === 'vllm' || backend === 'tensorrtllm')) {
                 return ['hosted-model-endpoint'];
@@ -700,7 +706,12 @@ const infraInstancePrompts = [
         when: answers => answers.deploymentTarget === 'realtime-inference' || answers.deploymentTarget === 'async-inference' || answers.deploymentTarget === 'batch-transform' || answers.deploymentTarget === 'hyperpod-eks',
         message: (answers) => {
             const framework = answers.framework || answers.deploymentConfig?.split('-')[0];
-            
+
+            // Skip table when MCP sizer already displayed annotated results
+            if (answers._mcpInstanceChoices && answers._mcpInstanceChoices.length > 0) {
+                return 'Select instance type:';
+            }
+
             const table = new Table({
                 head: [
                     chalk.cyan('Instance Type'),
@@ -1110,6 +1121,56 @@ const baseImagePrompts = [
     }
 ];
 
+/**
+ * Benchmark prompts for SageMaker AI Benchmarking (NVIDIA AIPerf)
+ * Sub-prompts shown when 'sagemaker-ai-automated-benchmarking' is selected in testTypes.
+ * Requirements: 2.1, 2.2, 2.3, 2.4, 2.5
+ */
+const benchmarkPrompts = [
+    {
+        type: 'number',
+        name: 'benchmarkConcurrency',
+        message: 'Concurrent requests for benchmark:',
+        default: 10,
+        when: (answers) => answers.includeBenchmark === true
+    },
+    {
+        type: 'number',
+        name: 'benchmarkInputTokensMean',
+        message: 'Mean input tokens per request:',
+        default: 550,
+        when: (answers) => answers.includeBenchmark === true
+    },
+    {
+        type: 'number',
+        name: 'benchmarkOutputTokensMean',
+        message: 'Mean output tokens per request:',
+        default: 150,
+        when: (answers) => answers.includeBenchmark === true
+    },
+    {
+        type: 'confirm',
+        name: 'benchmarkStreaming',
+        message: 'Enable streaming for benchmark?',
+        default: true,
+        when: (answers) => answers.includeBenchmark === true
+    },
+    {
+        type: 'input',
+        name: 'benchmarkRequestCount',
+        message: 'Total request count (leave empty for service default):',
+        default: '',
+        when: (answers) => answers.includeBenchmark === true
+    },
+    {
+        type: 'input',
+        name: 'benchmarkS3OutputPath',
+        message: 'Benchmark results S3 path (leave empty for auto-created bucket):',
+        default: '',
+        when: (answers) => answers.includeBenchmark === true
+    }
+];
+
 export {
     deploymentConfigPrompts,
     frameworkPrompts, // Deprecated: kept for backward compatibility
@@ -1123,6 +1184,7 @@ export {
     hfTokenPrompts,
     ngcApiKeyPrompts,
     modulePrompts,
+    benchmarkPrompts,
     infrastructurePrompts,
     infraRegionAndTargetPrompts,
     infraInstancePrompts,

package/src/lib/schema-sync.js

@@ -188,6 +188,37 @@ export function loadServiceModel(serviceName, registryPath) {
     return readFileSync(modelPath, 'utf8');
 }
 
+/**
+ * Check whether the SageMaker service model includes the CreateAIBenchmarkJob operation shape.
+ * Used to determine if benchmark parameter validation can be performed.
+ *
+ * @param {string} [registryPath] - Override registry path
+ * @returns {{ available: boolean, reason?: string }}
+ */
+export function hasBenchmarkShape(registryPath) {
+    const regPath = registryPath || getRegistryPath();
+    const modelContent = loadServiceModel('sagemaker', regPath);
+
+    if (!modelContent) {
+        return { available: false, reason: 'SageMaker service model not found in registry' };
+    }
+
+    try {
+        const model = JSON.parse(modelContent);
+        const operations = model.operations || {};
+        const shapes = model.shapes || {};
+
+        // Check for the CreateAIBenchmarkJob operation or its input shape
+        if (operations.CreateAIBenchmarkJob || shapes.CreateAIBenchmarkJobRequest) {
+            return { available: true };
+        }
+
+        return { available: false, reason: 'service model does not include AI Benchmark operations' };
+    } catch {
+        return { available: false, reason: 'Failed to parse SageMaker service model' };
+    }
+}
+
 /**
  * Store a service model in the registry.
  * @param {string} serviceName - Service name (e.g., 'sagemaker')

package/src/lib/template-manager.js

@@ -65,7 +65,7 @@ export default class TemplateManager {
             ],
             buildTargets: ['codebuild'],
             deploymentTargets: ['realtime-inference', 'async-inference', 'batch-transform', 'hyperpod-eks'],
-            testTypes: ['local-model-cli', 'local-model-server', 'hosted-model-endpoint'],
+            testTypes: ['local-model-cli', 'local-model-server', 'hosted-model-endpoint', 'sagemaker-ai-automated-benchmarking'],
             awsRegions: [
                 'us-east-1', 'us-east-2', 'us-west-1', 'us-west-2',
                 'eu-west-1', 'eu-west-2', 'eu-central-1', 'eu-north-1',
@@ -134,6 +134,9 @@ export default class TemplateManager {
 
         // Validate batch transform specific fields
         this._validateBatchTransformConfig();
+
+        // Validate benchmark specific fields
+        this._validateBenchmarkConfig();
         
         // Validate instance type format (ml.*.*) - only for realtime-inference
         if (this.answers.instanceType && this.answers.instanceType !== 'custom') {
@@ -297,6 +300,51 @@ export default class TemplateManager {
         }
     }
 
+    /**
+     * Validates benchmark configuration parameters
+     * @private
+     * @throws {Error} If benchmark configuration is invalid
+     */
+    _validateBenchmarkConfig() {
+        if (!this.answers.includeBenchmark) return;
+
+        // Gate to supported architectures
+        const dc = this.answers.deploymentConfig;
+        const arch = dc ? dc.split('-')[0] : this.answers.architecture;
+        if (arch !== 'transformers' && arch !== 'diffusors') {
+            throw new Error('⚠️  Benchmarking is only supported with transformers and diffusors architectures.');
+        }
+
+        // Gate to supported deployment targets
+        if (this.answers.deploymentTarget === 'hyperpod-eks') {
+            throw new Error('⚠️  Benchmarking is only supported with managed-inference, async-inference, and batch-transform deployment targets');
+        }
+
+        // Validate numeric parameters
+        if (this.answers.benchmarkConcurrency !== undefined) {
+            if (!Number.isInteger(this.answers.benchmarkConcurrency) || this.answers.benchmarkConcurrency < 1) {
+                throw new Error('⚠️  benchmarkConcurrency must be an integer >= 1');
+            }
+        }
+        if (this.answers.benchmarkInputTokensMean !== undefined) {
+            if (!Number.isInteger(this.answers.benchmarkInputTokensMean) || this.answers.benchmarkInputTokensMean < 1) {
+                throw new Error('⚠️  benchmarkInputTokensMean must be an integer >= 1');
+            }
+        }
+        if (this.answers.benchmarkOutputTokensMean !== undefined) {
+            if (!Number.isInteger(this.answers.benchmarkOutputTokensMean) || this.answers.benchmarkOutputTokensMean < 1) {
+                throw new Error('⚠️  benchmarkOutputTokensMean must be an integer >= 1');
+            }
+        }
+
+        // Validate S3 path format
+        if (this.answers.benchmarkS3OutputPath && this.answers.benchmarkS3OutputPath.trim() !== '') {
+            if (!this.answers.benchmarkS3OutputPath.startsWith('s3://')) {
+                throw new Error('⚠️  benchmarkS3OutputPath must start with "s3://". Example: s3://my-bucket/benchmark-results/');
+            }
+        }
+    }
+
     /**
      * Validates GPU instance type requirement for GPU-requiring backends.
      * Called when deploymentConfig is present.

package/src/lib/validate-runner.js

@@ -21,7 +21,7 @@ import SchemaValidationEngine from './schema-validation-engine.js';
 import ServiceModelParser from './service-model-parser.js';
 import CrossCuttingChecker from './cross-cutting-checker.js';
 import HuggingFaceClient from './huggingface-client.js';
-import { getRegistryPath, loadManifest } from './schema-sync.js';
+import { getRegistryPath, loadManifest, hasBenchmarkShape } from './schema-sync.js';
 
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = path.dirname(__filename);
@@ -52,6 +52,115 @@ export function parseDoConfig(configPath) {
     return config;
 }
 
+/**
+ * Validate benchmark parameters against service model constraints.
+ * Called when the CreateAIBenchmarkJob shape is available in the synced schema.
+ *
+ * Validates:
+ * - Concurrency: integer, min 1
+ * - S3OutputLocation: string, starts with s3://
+ * - AIBenchmarkJobName: pattern ^[a-zA-Z0-9](-*[a-zA-Z0-9])*, max 63 chars
+ *
+ * Requirements: 8.1, 8.2, 8.3
+ *
+ * @param {Object} config - Parsed do/config values
+ * @returns {Array<Object>} Array of validation findings
+ */
+export function validateBenchmarkParams(config) {
+    const findings = [];
+
+    // Validate Concurrency (integer, min 1)
+    if (config.BENCHMARK_CONCURRENCY !== null && config.BENCHMARK_CONCURRENCY !== undefined && config.BENCHMARK_CONCURRENCY !== '') {
+        const concurrency = Number(config.BENCHMARK_CONCURRENCY);
+        if (!Number.isInteger(concurrency) || concurrency < 1) {
+            findings.push({
+                severity: 'error',
+                operation: 'CreateAIBenchmarkJob',
+                fieldPath: 'Concurrency',
+                constraint: 'integer >= 1',
+                invalidValue: config.BENCHMARK_CONCURRENCY,
+                remediationHint: 'BENCHMARK_CONCURRENCY must be a positive integer (>= 1)'
+            });
+        }
+    }
+
+    // Validate S3OutputLocation (string, starts with s3://)
+    if (config.BENCHMARK_S3_OUTPUT_PATH !== null && config.BENCHMARK_S3_OUTPUT_PATH !== undefined && config.BENCHMARK_S3_OUTPUT_PATH !== '') {
+        const s3Path = config.BENCHMARK_S3_OUTPUT_PATH;
+        // Skip dynamic shell expressions (e.g., s3://...$(aws ...))
+        if (!s3Path.includes('$(') && !s3Path.startsWith('s3://')) {
+            findings.push({
+                severity: 'error',
+                operation: 'CreateAIBenchmarkJob',
+                fieldPath: 'OutputConfig.S3OutputLocation',
+                constraint: 'must start with s3://',
+                invalidValue: s3Path,
+                remediationHint: 'BENCHMARK_S3_OUTPUT_PATH must start with "s3://". Example: s3://my-bucket/benchmark-results/'
+            });
+        }
+    }
+
+    // Validate AIBenchmarkJobName pattern (^[a-zA-Z0-9](-*[a-zA-Z0-9])*, max 63 chars)
+    if (config.BENCHMARK_JOB_NAME !== null && config.BENCHMARK_JOB_NAME !== undefined && config.BENCHMARK_JOB_NAME !== '') {
+        const jobName = config.BENCHMARK_JOB_NAME;
+        // Skip dynamic shell expressions
+        if (!jobName.includes('$(') && !jobName.includes('${')) {
+            const namePattern = /^[a-zA-Z0-9](-*[a-zA-Z0-9])*$/;
+            if (jobName.length > 63) {
+                findings.push({
+                    severity: 'error',
+                    operation: 'CreateAIBenchmarkJob',
+                    fieldPath: 'AIBenchmarkJobName',
+                    constraint: 'max 63 characters',
+                    invalidValue: jobName,
+                    remediationHint: 'AIBenchmarkJobName must be at most 63 characters'
+                });
+            } else if (!namePattern.test(jobName)) {
+                findings.push({
+                    severity: 'error',
+                    operation: 'CreateAIBenchmarkJob',
+                    fieldPath: 'AIBenchmarkJobName',
+                    constraint: 'pattern: ^[a-zA-Z0-9](-*[a-zA-Z0-9])*',
+                    invalidValue: jobName,
+                    remediationHint: 'AIBenchmarkJobName must start with alphanumeric and contain only alphanumeric characters and hyphens'
+                });
+            }
+        }
+    }
+
+    // Validate input tokens mean (integer, min 1)
+    if (config.BENCHMARK_INPUT_TOKENS_MEAN !== null && config.BENCHMARK_INPUT_TOKENS_MEAN !== undefined && config.BENCHMARK_INPUT_TOKENS_MEAN !== '') {
+        const inputTokens = Number(config.BENCHMARK_INPUT_TOKENS_MEAN);
+        if (!Number.isInteger(inputTokens) || inputTokens < 1) {
+            findings.push({
+                severity: 'error',
+                operation: 'CreateAIWorkloadConfig',
+                fieldPath: 'WorkloadSpec.parameters.prompt_input_tokens_mean',
+                constraint: 'integer >= 1',
+                invalidValue: config.BENCHMARK_INPUT_TOKENS_MEAN,
+                remediationHint: 'BENCHMARK_INPUT_TOKENS_MEAN must be a positive integer (>= 1)'
+            });
+        }
+    }
+
+    // Validate output tokens mean (integer, min 1)
+    if (config.BENCHMARK_OUTPUT_TOKENS_MEAN !== null && config.BENCHMARK_OUTPUT_TOKENS_MEAN !== undefined && config.BENCHMARK_OUTPUT_TOKENS_MEAN !== '') {
+        const outputTokens = Number(config.BENCHMARK_OUTPUT_TOKENS_MEAN);
+        if (!Number.isInteger(outputTokens) || outputTokens < 1) {
+            findings.push({
+                severity: 'error',
+                operation: 'CreateAIWorkloadConfig',
+                fieldPath: 'WorkloadSpec.parameters.output_tokens_mean',
+                constraint: 'integer >= 1',
+                invalidValue: config.BENCHMARK_OUTPUT_TOKENS_MEAN,
+                remediationHint: 'BENCHMARK_OUTPUT_TOKENS_MEAN must be a positive integer (>= 1)'
+            });
+        }
+    }
+
+    return findings;
+}
+
 /**
  * Run the full validation pipeline.
  *
@@ -171,6 +280,20 @@ export async function run(options = {}) {
         }
     }
 
+    // Run benchmark parameter validation (Requirements 8.1, 8.2, 8.3)
+    if (config.BENCHMARK_CONCURRENCY || config.BENCHMARK_INPUT_TOKENS_MEAN ||
+        config.BENCHMARK_OUTPUT_TOKENS_MEAN || config.BENCHMARK_S3_OUTPUT_PATH) {
+        const benchmarkCheck = hasBenchmarkShape(registryPath);
+        if (benchmarkCheck.available) {
+            const benchmarkFindings = validateBenchmarkParams(config);
+            for (const finding of benchmarkFindings) {
+                report.addFinding(finding);
+            }
+        } else {
+            console.log('⚠️  Benchmark validation skipped: service model does not include AI Benchmark operations. Run `bootstrap sync-schemas` to update.');
+        }
+    }
+
     const summary = report.getSummary();
 
     // Load manifest for version info
@@ -213,4 +336,4 @@ export async function run(options = {}) {
     return exitCode;
 }
 
-export default { run, parseDoConfig };
+export default { run, parseDoConfig, validateBenchmarkParams };

package/templates/Dockerfile

@@ -12,6 +12,9 @@
 <% if (framework !== 'transformers') { %>
 FROM <%= baseImage || 'public.ecr.aws/docker/library/python:3.12-slim' %>
 
+# Ensure Python output is unbuffered so SageMaker can capture logs in CloudWatch
+ENV PYTHONUNBUFFERED=1
+
 # Set a docker label to name this project, postpended with the build time
 LABEL project.name="<%= projectName %>-<%= buildTimestamp %>" \
       project.base-name="<%= projectName %>" \
@@ -143,6 +146,9 @@ ARG BASE_IMAGE=<%= baseImage || 'deepjavalibrary/djl-serving:0.36.0-pytorch-gpu'
 
 FROM ${BASE_IMAGE}
 
+# Ensure Python output is unbuffered so SageMaker can capture logs in CloudWatch
+ENV PYTHONUNBUFFERED=1
+
 <% if (comments && comments.chatTemplate) { %>
 <%= comments.chatTemplate %>
 <% } %>
@@ -271,8 +277,9 @@ COPY code/serve /usr/bin/serve_trtllm
 RUN chmod +x /usr/bin/serve_trtllm
 
 # Copy startup script
+COPY code/cuda_compat.sh /usr/bin/cuda_compat.sh
 COPY code/start_server.sh /usr/bin/start_server.sh
-RUN chmod +x /usr/bin/start_server.sh
+RUN chmod +x /usr/bin/start_server.sh /usr/bin/cuda_compat.sh
 
 ENTRYPOINT [ "/usr/bin/start_server.sh" ]
 <% } else if (modelServer === 'lmi' || modelServer === 'djl') { %>
@@ -287,8 +294,9 @@ COPY code/serving.properties /opt/ml/model/serving.properties
 # LMI/DJL containers use their own entrypoint
 # The container will automatically start DJL Serving with the configuration
 <% } else { %>
+COPY code/cuda_compat.sh /usr/bin/cuda_compat.sh
 COPY code/serve /usr/bin/serve
-RUN chmod 777 /usr/bin/serve
+RUN chmod 777 /usr/bin/serve /usr/bin/cuda_compat.sh
 
 <% if (comments && comments.troubleshooting) { %>
 <%= comments.troubleshooting %>

package/templates/code/cuda_compat.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+# CUDA Compatibility Setup
+# Required for SageMaker inference AMIs using NVIDIA Container Toolkit 1.17.4+
+# (al2-ami-sagemaker-inference-gpu-2-1, al2-ami-sagemaker-inference-gpu-3-1,
+#  al2023-ami-sagemaker-inference-gpu-4-1)
+#
+# These AMIs no longer auto-mount CUDA compat libraries. This script detects
+# whether the host NVIDIA driver is older than what the container's CUDA toolkit
+# requires, and adds the compat libraries to LD_LIBRARY_PATH if needed.
+
+_verlt() {
+    [ "$1" = "$2" ] && return 1 || [ "$1" = "$(echo -e "$1\n$2" | sort -V | head -n1)" ]
+}
+
+if [ -f /usr/local/cuda/compat/libcuda.so.1 ]; then
+    CUDA_COMPAT_MAX_DRIVER_VERSION=$(readlink /usr/local/cuda/compat/libcuda.so.1 | cut -d'.' -f 3-)
+    NVIDIA_DRIVER_VERSION=$(sed -n 's/^NVRM.*Kernel Module *\([0-9.]*\).*$/\1/p' /proc/driver/nvidia/version 2>/dev/null || true)
+    if [ -n "$NVIDIA_DRIVER_VERSION" ] && _verlt "$NVIDIA_DRIVER_VERSION" "$CUDA_COMPAT_MAX_DRIVER_VERSION"; then
+        echo "CUDA compat: driver ${NVIDIA_DRIVER_VERSION} < ${CUDA_COMPAT_MAX_DRIVER_VERSION}, adding compat libs"
+        export LD_LIBRARY_PATH=/usr/local/cuda/compat:${LD_LIBRARY_PATH:-}
+    fi
+fi

package/templates/code/serve

@@ -2,6 +2,9 @@
 # Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
 # SPDX-License-Identifier: Apache-2.0
 
+# CUDA compatibility setup (required for newer SageMaker inference AMIs)
+source /usr/bin/cuda_compat.sh 2>/dev/null || true
+
 <% if (modelServer === 'vllm') { %>
 echo "Starting vLLM server"
 <% } else if (modelServer === 'sglang') { %>

package/templates/code/start_server.sh

@@ -2,6 +2,9 @@
 # Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
 # SPDX-License-Identifier: Apache-2.0
 
+# CUDA compatibility setup (required for newer SageMaker inference AMIs)
+source /usr/bin/cuda_compat.sh 2>/dev/null || true
+
 set -e
 
 echo "Starting TensorRT-LLM server on port 8081..."

package/templates/diffusors/Dockerfile

@@ -59,8 +59,9 @@ RUN apt-get update && apt-get install -y --no-install-recommends nginx \
 COPY nginx-diffusors.conf /etc/nginx/nginx.conf
 
 # Copy serve entrypoint and startup scripts
+COPY code/cuda_compat.sh /usr/bin/cuda_compat.sh
 COPY code/serve /usr/bin/serve
-RUN chmod 777 /usr/bin/serve
+RUN chmod 777 /usr/bin/serve /usr/bin/cuda_compat.sh
 
 COPY code/start_server.sh /usr/bin/start_server.sh
 RUN chmod +x /usr/bin/start_server.sh

package/templates/diffusors/serve

@@ -2,6 +2,9 @@
 # Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
 # SPDX-License-Identifier: Apache-2.0
 
+# CUDA compatibility setup (required for newer SageMaker inference AMIs)
+source /usr/bin/cuda_compat.sh 2>/dev/null || true
+
 echo "Starting vLLM-Omni server (diffusion model serving)"
 
 # Resolve model URI prefixes that engines cannot handle natively.