TPU Optimizer
Multi-Agent AI & HPC Workload Analyzer • TPU Code Generator • GKE & TPU VM Runner • AI Runtime Debugger
📦 Active Target: Ready ☁️ Project: mb-development-447000 🪣 Bucket: gs://mb-development-447000-tpu-optimizer-artifacts

Step 1: Add a GitHub Repository Link OR Upload Files / Folder

Paste any GitHub repository URL (e.g. https://github.com/LLNL/RAJAPerf) or upload files (PDF, YAML, README, code) or a project folder, then click Analyze Input & Generate Summary.

Quick Presets:
🔗 Option A: GitHub Repository Link
Enter any public or private GitHub repository URL to scan its files, README.md, and kernels
📂 Option B: Upload Files (PDF, YAML, Text, Code) or Entire Folder
Upload single files or an entire local folder to analyze alongside or instead of a GitHub link
📄 📁
Drag & Drop Files or Folder Here
Supports .pdf, .md, .txt, .yaml, .json, .py, .cc, .cpp, .cu, CMakeLists.txt
🗂️ Input Queue (0 local files staged)
Click any staged file below to inspect or edit its content before running analysis
File Path Type Size
Ready: Enter a GitHub URL (e.g. https://github.com/LLNL/RAJAPerf) or upload files, then click the button below.

Step 2: Use-Case Summary & Benchmark Metrics

Analyzed source details, architectural summary, and benchmark metrics with verified source citations.

Detected Project / Workload
—
—
Detected Languages & Frameworks
—
—
Numeric Precision
—
Hardware Execution Format
Communication / Parallelism
—
Multi-Chip ICI Pattern
📋 Use-Case Summary (From Code, README.md & Uploaded Files)
Synthesized from repository tree and documentation

📖 Key README.md, PDF & Repository Highlights:
    🔬 Inline Metrics Extracted From Files:
    🏆 Benchmark & XProf Hardware Metrics
    Reference benchmark performance and XProf hardware counter targets
    Verified Source
    📌 Benchmark Source & Citation
    Source Link:
    ⚡ Workload Benchmark Metrics:
    🛰️ Target TPU XProf Hardware Counters:

    Step 3: Recommended Path to TPU & Equivalent Runnable Code

    Evaluated across JAX, Torch-TPU, Pallas, Native TPU-C (tpucc), and vLLM-TPU / MaxText.

    🧭 Recommended Path to TPU & Framework Comparison
    Click any framework row below to view its runnable code on the right
    STRONG TPU FIT
    Framework Path Best Use Case Expected Efficiency Code
    1. JAX / OpenXLA (shard_map) SPMD Distributed Compute & Collectives 62–67% Util
    2. Torch-TPU (torch_xla) PyTorch SPMD Migration 54–60% Util
    3. JAX Pallas (pltpu.VMEM) Explicit VMEM SRAM Stencil / Attention Kernels 68–78% Kernel Eff.
    4. Native TPU-C (tpucc C++20) C++ HPC / VREG (`v1024f32`) / SparseCore / PGAS 85%+ VPU/SC Peak
    5. vLLM-TPU / MaxText High-Throughput LLM Serving & Batching Paged KV-Cache
    📊 Detailed Metrics for Selected Path
    Active: JAX
    🎛️ Optional: Adjust Batch / Problem Size & TPU Slice Chips (What-If Simulator)
    💻 Equivalent Runnable TPU Code Widget
    Complete runnable implementation tailored to your analyzed workload
    JAX / OpenXLA tpu_workload_jax.py 
    
          
    

          
    
            Ready to stage to GCS & execute on GKE or Cloud TPU VMs
            
    
              
              
            
    
          
    
        
    
      
    
    
    

    
    
    
    
    
      
    
        
    
          
    Step 4: GCP Project Login, GCS Bucket Selection & Direct Execution on GKE or Cloud TPU VMs
    
          
    Authenticate to your GCP project, select a Cloud Storage (GCS) bucket, and run your generated TPU code directly on a GKE Cluster or Cloud TPU VM.
    
        
    
        
      
    

      
    
        
        
    
          
    
            
    
              
    🔐 1. GCP Project Login & GCS Bucket Selection
    
              
    Connect to your GCP project and select the GCS bucket for staging code & traces
    
            
    
            Authenticated
          
    

          
    
            
    
              
              
            
    
            
    
              
              
            
    
          
    

          

          

          
    
            
          
    

          
    
            
            
          
    

          
    
            
            
          
    
        
    

        
        
    
          
    
            
    
              
    ⚡ 2. Select Execution Target & Run Code
    
              
    Choose between your GKE Cluster TPU Node Pool or Cloud TPU VMs
    
            
    
          
    

          
    
            
            
          
    

          
    
            
    
              
    
                
                
              
    
              
    
                
                
              
    
            
    
          
    

          

          
    
            
    
              
              
            
    
            
    
              
              
            
    
          
    

          
        
    
      
    

      
    
        
    
          
    
            
    
              
    📡 Execution Console & Telemetry Logs
    
              
    Staged GCS URI: Ready
    
            
    
            Idle
          
    
          
    Click "Stage Code to Selected GCS Bucket & Execute Workload Now" above to stage and dispatch your TPU workload.
    
        
    

        
    
          
    
            
    
              
    📜 Generated GKE Job YAML & Runner CLI
    
              
    Ready-to-run manifest pinned to your selected bucket & cluster/VM
    
            
    
            
          
    
          
    
        
    
      
    
    
    

    
    
    
    
    
      
    
        
    
          
    Step 5: AI Runtime Debugger & Troubleshooter
    
          
    Select your AI model (Antigravity, Gemini 4 Carbon, or Anthropic Claude 4.6) to diagnose runtime errors, XLA compiler warnings, or VMEM spills and auto-patch your code.
    
        
    
        
    
          
          
        
    
      
    

      
    
        Load Sample Issue:
        
        
        
        
      
    

      
    
        
    
          
    
            
    
              
    📝 Code & Runtime Error / XProf Log Input
    
              
    Paste your code and any runtime traceback or XProf warning
    
            
    
          
    

          
    
            
            
          
    

          
    
            
            
          
    

          
    
            
            
          
    

          
        
    

        
    
          
    
            
    
              
    
                
    🔬 Root-Cause Diagnosis & Patched Code
    
                
    Engine: Antigravity Agent
    
              
    
              Ready