RMSNormForward Class — pytorch Architecture

Source Code

benchmarks/dynamo/genai_layers/kernels.py lines 356–445

class RMSNormForward(BenchmarkKernel):
    def __init__(self, script_args):
        super().__init__(script_args)
        self.available_backends = ["eager", "compiled", "quack", "liger"]

    def get_shapes(self) -> tuple[tuple[int, ...], ...]:
        return (
            (32768, 256),
            (32768, 512),
            (32768, 1024),
            (32768, 2048),
            (32768, 4096),
            (32768, 8192),
            (32768, 16384),
            (32768, 32768),
            (32768, 65536),
            (16384, 131072),
            (8192, 262144),
        ) + extra_shapes_for_norm

    def get_memory_bytes(self, args, kwargs) -> int:
        x, w = args
        M, N = x.shape
        return 2 * M * N * x.dtype.itemsize + N * w.dtype.itemsize

    def rms_norm_ref(self, x, w):
        x_f32 = x.float()
        return (
            x_f32
            * torch.rsqrt(torch.mean(x_f32.square(), dim=-1, keepdim=True) + 1e-6)
            * w
        ).to(x.dtype)

    def eager(self, args, kwargs=None) -> Any:
        if kwargs is not None:
            raise AssertionError(f"Expected kwargs to be None, but got {kwargs}")
        x, w = args
        return lambda: self.rms_norm_ref(x, w)

    def compiled(self, args, kwargs=None) -> Any:
        if kwargs is not None:
            raise AssertionError(f"Expected kwargs to be None, but got {kwargs}")
        x, w = args

        # Mark batch size as dynamic for realistic workload
        torch._dynamo.mark_dynamic(x, 0)

        compiled_rms_norm = torch.compile(
            self.rms_norm_ref, mode=self.compile_mode, fullgraph=True
        )
        return lambda: compiled_rms_norm(x, w)

    def quack(self, args, kwargs=None) -> Any:
        # Note: only supper weight with float32 dtype
        from quack.rmsnorm import _rmsnorm_fwd

        x, w = args
        y = torch.empty_like(x)

        def quack_fwd():
            _rmsnorm_fwd(
                x,
                w,
                out=y,
                bias=None,
                rstd=None,
                residual=None,
                residual_out=None,
                eps=1e-6,
            )
            return y

        return quack_fwd

    def liger(self, args, kwargs) -> Any:
        from liger_kernel.transformers.rms_norm import LigerRMSNorm

        x, w = args
        M, N = x.shape
        liger_rmsnorm = LigerRMSNorm(hidden_size=N, eps=1e-6).cuda()
        liger_rmsnorm.weight.data.copy_(w)
        return lambda: liger_rmsnorm(x)

    def benchmark(self):
        for M, N in self.get_shapes():
            print(f"Tensor dimensions: [{M}, {N}]")
            torch_dtype = cutlass_torch.dtype(cutlass.BFloat16)
            x = torch.randn(M, N, device="cuda", dtype=torch_dtype)
            w = torch.randn(N, device="cuda", dtype=torch.float32)
            self.benchmark_single_shape((x, w), setting=f"shape: [{M}, {N}]")