Performance Class — pytorch Architecture

Source Code

benchmarks/dynamo/genai_layers/utils.py lines 24–44

class Performance:
    # Benchmark setting usually the shape of the input tensor
    setting: str

    # Latency in milliseconds
    latency: float

    # Number of  memory access in bytes
    memory_bytes: float

    # Memory bandwidth in GB/s
    memory_bandwidth: float = 0.0

    # Compute intensity in FLOPs/byte
    compute_intensity: float = 0.0

    def __post_init__(self):
        self.memory_bandwidth = self.memory_bytes / (self.latency / 1000) / 1e9

    def __str__(self):
        return f"setting: {self.setting}, latency: {self.latency} ms, memory bandwidth: {self.memory_bandwidth} GB/s"