You write custom CUDA kernels to replace the pytorch operators in the given architecture to get speedups. 

    You have complete freedom to choose the set of operators you want to replace. You may make the decision to replace some operators with custom CUDA kernels and leave others unchanged. You may replace multiple operators with custom implementations, consider operator fusion opportunities (combining multiple operators into a single kernel, for example, combining matmul+relu), or algorithmic changes (such as online softmax). You are only limited by your imagination.


        Here's an example to show you the syntax of inline embedding custom CUDA operators in torch: The example given architecture is: 

        ``` 

        # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import torch
import torch.nn as nn
import torch.nn.functional as F


class Model(nn.Module):
    def __init__(self) -> None:
        super().__init__()

    def forward(self, a, b):
        return a + b


def get_inputs():
    # randomly generate input tensors based on the model architecture
    a = torch.randn(1, 128).cuda()
    b = torch.randn(1, 128).cuda()
    return [a, b]


def get_init_inputs():
    # randomly generate tensors required for initialization based on the model architecture
    return []

        ``` 

        The example new arch with custom CUDA kernels looks like this:
        ```
        # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

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

# Define the custom CUDA kernel for element-wise addition
elementwise_add_source = """
#include <torch/extension.h>
#include <cuda_runtime.h>

__global__ void elementwise_add_kernel(const float* a, const float* b, float* out, int size) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < size) {
        out[idx] = a[idx] + b[idx];
    }
}

torch::Tensor elementwise_add_cuda(torch::Tensor a, torch::Tensor b) {
    auto size = a.numel();
    auto out = torch::zeros_like(a);

    const int block_size = 256;
    const int num_blocks = (size + block_size - 1) / block_size;

    elementwise_add_kernel<<<num_blocks, block_size>>>(a.data_ptr<float>(), b.data_ptr<float>(), out.data_ptr<float>(), size);

    return out;
}
"""

elementwise_add_cpp_source = (
    "torch::Tensor elementwise_add_cuda(torch::Tensor a, torch::Tensor b);"
)

# Compile the inline CUDA code for element-wise addition
elementwise_add = load_inline(
    name="elementwise_add",
    cpp_sources=elementwise_add_cpp_source,
    cuda_sources=elementwise_add_source,
    functions=["elementwise_add_cuda"],
    verbose=True,
    extra_cflags=[""],
    extra_ldflags=[""],
)


class ModelNew(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.elementwise_add = elementwise_add

    def forward(self, a, b):
        return self.elementwise_add.elementwise_add_cuda(a, b)

        ``` 

        
    You are given the following architecture: 

    ```
    # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs a matrix multiplication, adds a value, applies Swish, Tanh, GELU, and Hardtanh activation functions.
    """
    def __init__(self, in_features, out_features, add_value_shape):
        super(Model, self).__init__()
        self.matmul = nn.Linear(in_features, out_features)
        self.add_value = nn.Parameter(torch.randn(add_value_shape)) 

    def forward(self, x):
        x = self.matmul(x)
        x = x + self.add_value
        x = torch.sigmoid(x) * x # Swish
        x = torch.tanh(x)
        x = torch.nn.functional.gelu(x) # GELU
        x = torch.nn.functional.hardtanh(x, min_val=-1, max_val=1) # Hardtanh
        return x

batch_size = 1024
in_features = 8192
out_features = 8192
add_value_shape = (out_features,)

def get_inputs():
    return [torch.rand(batch_size, in_features)]

def get_init_inputs():
    return [in_features, out_features, add_value_shape]
    ```
    
Optimize the architecture named Model with custom CUDA operators! Name your optimized output architecture ModelNew. Output the new code in codeblocks. Please generate real code, NOT pseudocode, make sure the code compiles and is fully functional. Just output the new model code, no other text, and NO testing code! 



This is turn 2 of 6 for this task. Your previous solution (turn 1) was:
```python
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

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

# Fused CUDA kernel:
#   out = hardtanh(gelu(tanh(swish(h + add_value))))
# The GEMM (x @ W^T + bias) is executed with cuBLAS via at::addmm, and the
# entire element-wise chain (bias-add of add_value, Swish, Tanh, GELU,
# Hardtanh) is fused into a single kernel, eliminating ~5 extra global-memory
# round trips and several kernel launches.
cuda_source = r"""
#include <torch/extension.h>
#include <ATen/ATen.h>
#include <ATen/cuda/CUDAContext.h>
#include <cuda_runtime.h>

__device__ __forceinline__ float fused_act(float v) {
    // Swish: v * sigmoid(v) = v / (1 + exp(-v))
    float sv = v / (1.0f + expf(-v));
    // Tanh
    float tv = tanhf(sv);
    // GELU (exact erf formulation, matches F.gelu default)
    float gv = 0.5f * tv * (1.0f + erff(tv * 0.70710678118654752440f));
    // Hardtanh clamp to [-1, 1]
    return fminf(fmaxf(gv, -1.0f), 1.0f);
}

__global__ void fused_epilogue_scalar_kernel(
    const float* __restrict__ h,
    const float* __restrict__ addv,
    float* __restrict__ out,
    int cols)
{
    int col = blockIdx.x * blockDim.x + threadIdx.x;
    int row = blockIdx.y;
    if (col < cols) {
        long idx = (long)row * cols + col;
        out[idx] = fused_act(h[idx] + addv[col]);
    }
}

__global__ void fused_epilogue_vec4_kernel(
    const float4* __restrict__ h,
    const float4* __restrict__ addv,
    float4* __restrict__ out,
    int cols4)
{
    int c4 = blockIdx.x * blockDim.x + threadIdx.x;
    int row = blockIdx.y;
    if (c4 < cols4) {
        long idx = (long)row * cols4 + c4;
        float4 val = h[idx];
        float4 av  = addv[c4];
        val.x = fused_act(val.x + av.x);
        val.y = fused_act(val.y + av.y);
        val.z = fused_act(val.z + av.z);
        val.w = fused_act(val.w + av.w);
        out[idx] = val;
    }
}

torch::Tensor fused_forward_cuda(torch::Tensor x, torch::Tensor weight, torch::Tensor bias, torch::Tensor add_value) {
    TORCH_CHECK(x.is_cuda() && weight.is_cuda() && bias.is_cuda() && add_value.is_cuda(),
                "all inputs must be CUDA tensors");
    TORCH_CHECK(x.scalar_type() == at::kFloat, "only float32 is supported");

    // GEMM with bias via cuBLAS (same numerics as nn.Linear).
    auto h = at::addmm(bias, x, weight.t());  // (batch, out_features)
    auto out = torch::empty_like(h);

    int rows = h.size(0);
    int cols = h.size(1);
    if (rows == 0 || cols == 0) {
        return out;
    }

    auto stream = at::cuda::getCurrentCUDAStream();
    const int threads = 256;

    if ((cols % 4) == 0) {
        int cols4 = cols / 4;
        dim3 grid((cols4 + threads - 1) / threads, rows);
        fused_epilogue_vec4_kernel<<<grid, threads, 0, stream>>>(
            reinterpret_cast<const float4*>(h.data_ptr<float>()),
            reinterpret_cast<const float4*>(add_value.data_ptr<float>()),
            reinterpret_cast<float4*>(out.data_ptr<float>()),
            cols4);
    } else {
        dim3 grid((cols + threads - 1) / threads, rows);
        fused_epilogue_scalar_kernel<<<grid, threads, 0, stream>>>(
            h.data_ptr<float>(),
            add_value.data_ptr<float>(),
            out.data_ptr<float>(),
            cols);
    }
    return out;
}
"""

cpp_source = (
    "torch::Tensor fused_forward_cuda(torch::Tensor x, torch::Tensor weight, "
    "torch::Tensor bias, torch::Tensor add_value);"
)

fused_module = load_inline(
    name="fused_linear_swish_tanh_gelu_hardtanh",
    cpp_sources=cpp_source,
    cuda_sources=cuda_source,
    functions=["fused_forward_cuda"],
    verbose=False,
    extra_cuda_cflags=["-O3"],
)


class ModelNew(nn.Module):
    def __init__(self, in_features, out_features, add_value_shape):
        super(ModelNew, self).__init__()
        self.matmul = nn.Linear(in_features, out_features)
        self.add_value = nn.Parameter(torch.randn(add_value_shape))
        self.fused = fused_module

    def forward(self, x):
        if x.is_cuda and x.dtype == torch.float32:
            return self.fused.fused_forward_cuda(
                x, self.matmul.weight, self.matmul.bias, self.add_value
            )
        # Fallback (e.g. CPU tensors): original operator sequence.
        x = F.linear(x, self.matmul.weight, self.matmul.bias)
        x = x + self.add_value
        x = torch.sigmoid(x) * x  # Swish
        x = torch.tanh(x)
        x = F.gelu(x)  # GELU
        x = F.hardtanh(x, min_val=-1, max_val=1)  # Hardtanh
        return x
```
Result of evaluating it:
The solution is correct on the test inputs. Mean runtime 7.92 ms; the PyTorch reference (with TF32 enabled) takes 1.45 ms; speedup 0.183x.
Profiler, top device kernels of one forward pass of your solution:
| device kernel | total device time (us) | calls |
|---|---|---|
| ampere_sgemm_128x32_tn | 7831.3 | 1 |
| fused_epilogue_vec4_kernel(float4 const*, float4 const*, float4*, int) | 49.7 | 1 |

Write an improved complete ModelNew: fix any error, and if the solution is already correct, make it faster. Output the new code in one ```python codeblock. Just the model code, no other text, and NO testing code.
