podstack 1.2.0__py3-none-any.whl

podstack_gpu/__init__.py

@@ -0,0 +1,126 @@
+"""
+Podstack GPU SDK - Run code on cloud GPUs.
+
+Example usage:
+    import podstack
+
+    # Define an image with dependencies
+    image = (
+        podstack.Image.debian_slim()
+        .pip_install("torch", "transformers")
+    )
+
+    # Or use a preset environment
+    image = podstack.Image.ml()
+
+    # Or install from requirements.txt
+    image = podstack.Image.debian_slim().pip_install_from_requirements("requirements.txt")
+
+    # Create an app
+    app = podstack.App("my-app")
+
+    # Define a GPU function with full GPU
+    @app.function(gpu="H100", image=image)
+    def train(epochs: int):
+        import torch
+        # Print GPU info at start
+        podstack.print_gpu_status()
+        print(f"Training for {epochs} epochs on {torch.cuda.get_device_name()}")
+        return {"status": "done"}
+
+    # Use fractional GPU for cost savings (25%, 50%, 75%, or 100%)
+    @app.function(gpu="L40S", fraction=25, image=image)
+    def inference(data):
+        # Uses 25% of an L40S GPU - great for inference workloads
+        return model.predict(data)
+
+    # Run remotely with real-time status tracking
+    call = train.spawn(epochs=10)
+    for update in call.stream_status():
+        print(f"Status: {update.status}")
+        if update.queue_position:
+            print(f"  Queue position: {update.queue_position}")
+    result = call.get()
+
+    # Or use the low-level runner
+    # Supports psk_xxx tokens (from cloud.podstack.ai portal) or pjt_xxx tokens (project tokens)
+    gpu = podstack.GPURunner(token="psk_xxx", project_id="xxx")
+    result = gpu.run("print('hello')", gpu="L40S", fraction=50)  # 50% GPU
+
+    # Custom API URL (optional - defaults to PODSTACK_API_URL env var or production URL)
+    gpu = podstack.GPURunner(token="psk_xxx", project_id="xxx", api_url="https://your-api.example.com")
+
+GPU Utilities (use inside your GPU functions):
+    podstack.nvidia_smi()        # Get nvidia-smi output
+    podstack.gpu_info()          # Get GPU info as dict
+    podstack.cuda_available()    # Check if CUDA is available
+    podstack.cuda_device_count() # Get number of GPUs
+    podstack.print_gpu_status()  # Print formatted GPU status
+
+GPU Types:
+    - A10: Entry-level, good for inference
+    - L40, L40S: Mid-range, balanced performance
+    - A100-40G, A100-80G: High-end training
+    - H100: Top-tier performance
+
+Fractional GPU:
+    - fraction=25: 25% of GPU (lowest cost, good for small inference)
+    - fraction=50: 50% of GPU (balanced)
+    - fraction=75: 75% of GPU
+    - fraction=100: Full GPU (default, best for training)
+"""
+
+from .runner import GPURunner, ExecutionResult, StatusUpdate
+from .image import Image, ImageDefinition
+from .app import App, Function, FunctionCall, method
+from .volume import Volume, CloudBucketMount
+from .secret import Secret, SecretDict
+from .utils import (
+    nvidia_smi,
+    gpu_info,
+    cuda_available,
+    cuda_device_count,
+    print_gpu_status,
+)
+from .exceptions import (
+    PodstackError,
+    ExecutionError,
+    AuthenticationError,
+    ValidationError,
+    TimeoutError,
+    InsufficientBalanceError,
+)
+
+__version__ = "0.3.3"
+
+__all__ = [
+    # Core classes
+    "App",
+    "Function",
+    "FunctionCall",
+    "Image",
+    "ImageDefinition",
+    "Volume",
+    "CloudBucketMount",
+    "Secret",
+    "SecretDict",
+    # Low-level runner
+    "GPURunner",
+    "ExecutionResult",
+    "StatusUpdate",
+    # GPU utilities
+    "nvidia_smi",
+    "gpu_info",
+    "cuda_available",
+    "cuda_device_count",
+    "print_gpu_status",
+    # Decorators
+    "method",
+    # Exceptions
+    "PodstackError",
+    "ExecutionError",
+    "AuthenticationError",
+    "ValidationError",
+    "TimeoutError",
+    "InsufficientBalanceError",
+]