```python
import math
import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline

_cpp_source = """
torch::Tensor fused_activation_cuda(torch::Tensor x, torch::Tensor bias);
"""

_cuda_source = r"""
#include <torch/extension.h>
#include <ATen/AccumulateType.h>
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>
#include <c10/cuda/CUDAException.h>
#include <cuda_runtime.h>

template <typename T>
__device__ __forceinline__ T activate(T x, T bias) {
    T y = (x > T(0) ? x : T(0.2) * x) + bias;
    y = y < T(-1) ? T(-1) : y;
    y = y > T(1) ? T(1) : y;
    return T(0.5) * y *
           (T(1) + erf(y * T(0.70710678118654752440)));
}

template <>
__device__ __forceinline__ float activate(float x, float bias) {
    float y = (x > 0.0f ? x : 0.2f * x) + bias;
    y = y < -1.0f ? -1.0f : y;
    y = y > 1.0f ? 1.0f : y;
    return 0.5f * y * (1.0f + erff(y * 0.70710678118654752440f));
}

__global__ void fused_float4_kernel(
    const float4* __restrict__ x,
    const float* __restrict__ bias,
    float4* __restrict__ out,
    int64_t count,
    int64_t spatial_vectors,
    int64_t channels
) {
    for (int64_t i = int64_t(blockIdx.x) * blockDim.x + threadIdx.x;
         i < count;
         i += int64_t(blockDim.x) * gridDim.x) {
        float b = bias[(i / spatial_vectors) % channels];
        float4 v = x[i];
        v.x = activate(v.x, b);
        v.y = activate(v.y, b);
        v.z = activate(v.z, b);
        v.w = activate(v.w, b);
        out[i] = v;
    }
}

template <typename scalar_t>
__global__ void fused_scalar_kernel(
    const scalar_t* __restrict__ x,
    const scalar_t* __restrict__ bias,
    scalar_t* __restrict__ out,
    int64_t count,
    int64_t spatial,
    int64_t channels
) {
    using acc_t = at::acc_type<scalar_t, true>;
    for (int64_t i = int64_t(blockIdx.x) * blockDim.x + threadIdx.x;
         i < count;
         i += int64_t(blockDim.x) * gridDim.x) {
        acc_t v = static_cast<acc_t>(x[i]);
        acc_t b = static_cast<acc_t>(bias[(i / spatial) % channels]);
        out[i] = static_cast<scalar_t>(activate(v, b));
    }
}

torch::Tensor fused_activation_cuda(torch::Tensor x, torch::Tensor bias) {
    TORCH_CHECK(x.is_cuda() && bias.is_cuda(), "Expected CUDA tensors");
    TORCH_CHECK(x.device() == bias.device(), "Device mismatch");
    TORCH_CHECK(x.scalar_type() == bias.scalar_type(), "Dtype mismatch");
    TORCH_CHECK(x.dim() == 5 && x.is_contiguous(), "Expected contiguous NCDHW");
    TORCH_CHECK(bias.is_contiguous() && bias.numel() == x.size(1),
                "Expected one bias value per channel");

    const c10::cuda::CUDAGuard guard(x.device());
    auto out = torch::empty_like(x);
    int64_t count = x.numel();
    if (count == 0) return out;

    int64_t spatial = x.size(2) * x.size(3) * x.size(4);
    int64_t channels = x.size(1);
    constexpr int threads = 256;
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();

    if (x.scalar_type() == at::kFloat && spatial % 4 == 0) {
        int64_t vectors = count / 4;
        int blocks = static_cast<int>(
            std::min<int64_t>((vectors + threads - 1) / threads, 65535));
        fused_float4_kernel<<<blocks, threads, 0, stream>>>(
            reinterpret_cast<const float4*>(x.data_ptr<float>()),
            bias.data_ptr<float>(),
            reinterpret_cast<float4*>(out.data_ptr<float>()),
            vectors, spatial / 4, channels);
    } else {
        int blocks = static_cast<int>(
            std::min<int64_t>((count + threads - 1) / threads, 65535));
        AT_DISPATCH_FLOATING_TYPES(
            x.scalar_type(), "fused_activation", [&] {
                fused_scalar_kernel<scalar_t><<<blocks, threads, 0, stream>>>(
                    x.data_ptr<scalar_t>(), bias.data_ptr<scalar_t>(),
                    out.data_ptr<scalar_t>(), count, spatial, channels);
            });
    }
    C10_CUDA_KERNEL_LAUNCH_CHECK();
    return out;
}
"""

_fused_ops = load_inline(
    name="conv3d_leaky_add_clamp_gelu_fused_v1",
    cpp_sources=_cpp_source,
    cuda_sources=_cuda_source,
    functions=["fused_activation_cuda"],
    extra_cflags=["-O3"],
    extra_cuda_cflags=["-O3", "--fmad=false"],
    verbose=False,
)


class _FusedActivation(torch.autograd.Function):
    @staticmethod
    def forward(ctx, x, bias):
        ctx.save_for_backward(x, bias)
        return _fused_ops.fused_activation_cuda(x, bias)

    @staticmethod
    def backward(ctx, grad_output):
        x, bias = ctx.saved_tensors
        preclamp = torch.where(x > 0, x, 0.2 * x) + bias
        z = preclamp.clamp(-1.0, 1.0)
        gelu_grad = (
            0.5 * (1.0 + torch.erf(z / math.sqrt(2.0)))
            + z * torch.exp(-0.5 * z.square()) / math.sqrt(2.0 * math.pi)
        )
        grad = grad_output * gelu_grad
        grad = torch.where(
            (preclamp >= -1.0) & (preclamp <= 1.0), grad, 0.0
        )
        grad_x = (
            grad * torch.where(x > 0, 1.0, 0.2)
            if ctx.needs_input_grad[0]
            else None
        )
        grad_bias = (
            grad.sum(dim=(0, 2, 3, 4)).reshape_as(bias)
            if ctx.needs_input_grad[1]
            else None
        )
        return grad_x, grad_bias


class ModelNew(nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size, sum_tensor_shape):
        super().__init__()
        self.conv = nn.Conv3d(in_channels, out_channels, kernel_size)
        self.sum_tensor = nn.Parameter(torch.randn(sum_tensor_shape))

    def forward(self, x):
        x = self.conv(x)
        bias = self.sum_tensor

        if (
            x.is_cuda
            and x.dtype in (torch.float32, torch.float64)
            and bias.dtype == x.dtype
            and tuple(bias.shape) == (x.shape[1], 1, 1, 1)
        ):
            return _FusedActivation.apply(x.contiguous(), bias.contiguous())

        x = torch.nn.functional.leaky_relu(x, negative_slope=0.2)
        x = torch.clamp(x + bias, min=-1.0, max=1.0)
        return torch.nn.functional.gelu(x)
```