```python
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.cpp_extension import load_inline


_cpp_source = r"""
#include <torch/extension.h>

torch::Tensor fused_sub_mul_relu(
    torch::Tensor input, double subtract_value, double multiply_value);
"""

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

__device__ __forceinline__ float activate(float x, float sub, float mul) {
    float y = __fmul_rn(__fsub_rn(x, sub), mul);
    return y < 0.0f ? 0.0f : y;
}

__global__ void fused_sub_mul_relu_kernel(
    const float* __restrict__ input,
    float* __restrict__ output,
    int64_t count,
    float sub,
    float mul
) {
    const int64_t vector_count = count / 4;
    const int64_t thread =
        static_cast<int64_t>(blockIdx.x) * blockDim.x + threadIdx.x;
    const int64_t stride =
        static_cast<int64_t>(gridDim.x) * blockDim.x;

    const float4* input4 = reinterpret_cast<const float4*>(input);
    float4* output4 = reinterpret_cast<float4*>(output);

    for (int64_t i = thread; i < vector_count; i += stride) {
        float4 v = input4[i];
        v.x = activate(v.x, sub, mul);
        v.y = activate(v.y, sub, mul);
        v.z = activate(v.z, sub, mul);
        v.w = activate(v.w, sub, mul);
        output4[i] = v;
    }

    const int64_t tail = vector_count * 4 + thread;
    if (tail < count) {
        output[tail] = activate(input[tail], sub, mul);
    }
}

torch::Tensor fused_sub_mul_relu(
    torch::Tensor input,
    double subtract_value,
    double multiply_value
) {
    TORCH_CHECK(input.is_cuda(), "input must be a CUDA tensor");
    TORCH_CHECK(input.scalar_type() == at::kFloat, "input must be float32");
    TORCH_CHECK(input.is_contiguous(), "input must be contiguous");

    const c10::cuda::CUDAGuard device_guard(input.device());
    auto output = torch::empty_like(input);
    const int64_t count = input.numel();

    if (count == 0) {
        return output;
    }

    constexpr int threads = 256;
    const int64_t work = (count + 3) / 4;
    const int blocks = static_cast<int>(
        std::min<int64_t>((work + threads - 1) / threads, 4096));

    fused_sub_mul_relu_kernel<<<
        blocks, threads, 0, at::cuda::getCurrentCUDAStream()
    >>>(
        input.data_ptr<float>(),
        output.data_ptr<float>(),
        count,
        static_cast<float>(subtract_value),
        static_cast<float>(multiply_value)
    );
    C10_CUDA_KERNEL_LAUNCH_CHECK();
    return output;
}
"""

_fused_ops = load_inline(
    name="linear_sub_mul_relu_fused_cuda",
    cpp_sources=_cpp_source,
    cuda_sources=_cuda_source,
    functions=["fused_sub_mul_relu"],
    extra_cflags=["-O3"],
    extra_cuda_cflags=["-O3"],
    verbose=False,
)


class ModelNew(nn.Module):
    def __init__(
        self, in_features, out_features, subtract_value, multiply_value
    ):
        super().__init__()
        self.linear = nn.Linear(in_features, out_features)
        self.subtract_value = subtract_value
        self.multiply_value = multiply_value

    def forward(self, x):
        x = F.linear(x, self.linear.weight, self.linear.bias)
        if x.is_cuda and x.dtype == torch.float32 and x.is_contiguous():
            return _fused_ops.fused_sub_mul_relu(
                x, float(self.subtract_value), float(self.multiply_value)
            )
        return torch.relu(
            (x - self.subtract_value) * self.multiply_value
        )
```