push rocm deepspeed v0.3.13

README.md

@@ -4,8 +4,6 @@
 [![License MIT](https://img.shields.io/badge/License-MIT-blue.svg)](https://github.com/Microsoft/DeepSpeed/blob/master/LICENSE)
 [![Docker Pulls](https://img.shields.io/docker/pulls/deepspeed/deepspeed)](https://hub.docker.com/r/deepspeed/deepspeed)
 
-### 03/2021: DeepSpeed is hiring! Come join us: [SDE 2](https://careers.microsoft.com/us/en/job/1013160/Software-Engineer-2), [Sr. SDE](https://careers.microsoft.com/us/en/job/1017151/Senior-Software-Engineer), [Sr. Researcher](https://careers.microsoft.com/us/en/job/1016440/Senior-Researcher)
-
 [DeepSpeed](https://www.deepspeed.ai/) is a deep learning optimization
 library that makes distributed training easy, efficient, and effective.
 
@@ -33,11 +31,7 @@ information [here](https://innovation.microsoft.com/en-us/exploring-ai-at-scale)
 
 
 # News
-* [2021/04/01] [[DeepSpeed on AzureML] Transformers and CIFAR examples are now available on AzureML GitHub](https://github.com/Azure/azureml-examples/tree/main/workflows/train/deepspeed)
-* [2021/03/30] [[PyTorch Lightning Blog] Accessible Multi-Billion Parameter Model Training with PyTorch Lightning + DeepSpeed](https://medium.com/pytorch-lightning/accessible-multi-billion-parameter-model-training-with-pytorch-lightning-deepspeed-c9333ac3bb59)
-* [2021/03/16] [1-bit Adam v2: NCCL-based implementation and more](https://www.deepspeed.ai/tutorials/onebit-adam/)
 * [2021/03/08] [ZeRO-3 Offload: Scale your models to trillion parameters without code changes while leveraging both CPUs & GPUs](https://www.deepspeed.ai/news/2021/03/07/zero3-offload.html)
-* [2021/01/19] [[🤗Hugging Face Blog] Fit More and Train Faster With ZeRO via DeepSpeed and FairScale](https://huggingface.co/blog/zero-deepspeed-fairscale)
 * [2020/11/12] [Simplified install, JIT compiled ops, PyPI releases, and reduced dependencies](#installation)
 * [2020/11/10] [Efficient and robust compressed training through progressive layer dropping](https://www.deepspeed.ai/news/2020/10/28/progressive-layer-dropping-news.html)
 * [2020/09/10] [DeepSpeed v0.3: Extreme-scale model training for everyone](https://www.microsoft.com/en-us/research/blog/deepspeed-extreme-scale-model-training-for-everyone/)
@@ -45,6 +39,7 @@ information [here](https://innovation.microsoft.com/en-us/exploring-ai-at-scale)
   * [Training a trillion parameters with pipeline parallelism](https://www.deepspeed.ai/news/2020/09/08/pipeline-parallelism.html)
   * [Up to 5x less communication and 3.4x faster training through 1-bit Adam](https://www.deepspeed.ai/news/2020/09/08/onebit-adam-news.html)
   * [10x bigger model training on a single GPU with ZeRO-Offload](https://www.deepspeed.ai/news/2020/09/08/ZeRO-Offload.html)
+* [2020/08/07] [DeepSpeed Microsoft Research Webinar](https://note.microsoft.com/MSR-Webinar-DeepSpeed-Registration-On-Demand.html) is now available on-demand
 
 
 # Table of Contents

\

+"""
+Copyright 2020 The Microsoft DeepSpeed Team
+
+DeepSpeed library
+
+Create a new wheel via the following command: python setup.py bdist_wheel
+
+The wheel will be located at: dist/*.whl
+"""
+
+import os
+import shutil
+import subprocess
+import warnings
+from setuptools import setup, find_packages
+import time
+
+try:
+    import torch
+    from torch.utils.cpp_extension import BuildExtension, CUDAExtension, CppExtension
+    from torch.utils.hipify import hipify_python
+except ImportError:
+    raise ImportError('Unable to import torch, please visit https://pytorch.org/ '
+                      'to see how to properly install torch on your system.')
+###aiss add 
+is_rocm_pytorch = False
+if torch.__version__ >= '1.5':
+    from torch.utils.cpp_extension import ROCM_HOME
+    is_rocm_pytorch = True if ((torch.version.hip is not None) and (ROCM_HOME is not None)) else False
+
+from op_builder import ALL_OPS, get_default_compute_capatabilities
+
+
+def fetch_requirements(path):
+    with open(path, 'r') as fd:
+        return [r.strip() for r in fd.readlines()]
+
+
+install_requires = fetch_requirements('requirements/requirements.txt')
+extras_require = {
+    '1bit_adam': fetch_requirements('requirements/requirements-1bit-adam.txt'),
+    'readthedocs': fetch_requirements('requirements/requirements-readthedocs.txt'),
+    'dev': fetch_requirements('requirements/requirements-dev.txt'),
+}
+###aiss add ################
+if is_rocm_pytorch:
+    print("NOTE: Please manually install torch and torchvision packages for ROCm")
+    install_requires = fetch_requirements('requirements/requirements-rocm.txt')
+
+
+# If MPI is available add 1bit-adam requirements
+if torch.cuda.is_available():
+    if shutil.which('ompi_info') or shutil.which('mpiname'):
+        cupy = f"cupy-cuda{torch.version.cuda.replace('.','')[:3]}"
+        extras_require['1bit_adam'].append(cupy)
+
+# Make an [all] extra that installs all needed dependencies
+all_extras = set()
+for extra in extras_require.items():
+    for req in extra[1]:
+        all_extras.add(req)
+extras_require['all'] = list(all_extras)
+
+cmdclass = {}
+
+# For any pre-installed ops force disable ninja
+cmdclass['build_ext'] = BuildExtension.with_options(use_ninja=False)
+
+TORCH_MAJOR = torch.__version__.split('.')[0]
+TORCH_MINOR = torch.__version__.split('.')[1]
+
+if not torch.cuda.is_available():
+    # Fix to allow docker builds, similar to https://github.com/NVIDIA/apex/issues/486
+    print(
+        "[WARNING] Torch did not find cuda available, if cross-compiling or running with cpu only "
+        "you can ignore this message. Adding compute capability for Pascal, Volta, and Turing "
+        "(compute capabilities 6.0, 6.1, 6.2)")
+    if os.environ.get("TORCH_CUDA_ARCH_LIST", None) is None:
+        os.environ["TORCH_CUDA_ARCH_LIST"] = get_default_compute_capatabilities()
+###########aiss add ######################only need run once for convert
+#if is_rocm_pytorch:
+#    import shutil
+#    this_dir = os.path.dirname(os.path.abspath(__file__))
+#    hipify_python.hipify(project_directory=this_dir, output_directory=this_dir, includes="csrc/*",
+#                                show_detailed=True, is_pytorch_extension=True)
+#    print("cuda file has been transformed to hip format!!!")
+ext_modules = []
+
+# Default to pre-install kernels to false so we rely on JIT
+BUILD_OP_DEFAULT = int(os.environ.get('DS_BUILD_OPS', 0))
+print(f"DS_BUILD_OPS={BUILD_OP_DEFAULT}")
+
+
+def command_exists(cmd):
+    result = subprocess.Popen(f'type {cmd}', stdout=subprocess.PIPE, shell=True)
+    return result.wait() == 0
+
+
+def op_enabled(op_name):
+    assert hasattr(ALL_OPS[op_name], 'BUILD_VAR'), \
+        f"{op_name} is missing BUILD_VAR field"
+    env_var = ALL_OPS[op_name].BUILD_VAR
+    return int(os.environ.get(env_var, BUILD_OP_DEFAULT))
+
+
+install_ops = dict.fromkeys(ALL_OPS.keys(), False)
+for op_name, builder in ALL_OPS.items():
+    op_compatible = builder.is_compatible()
+
+    # If op is compatible update install reqs so it can potentially build/run later
+    if op_compatible:
+        reqs = builder.python_requirements()
+        install_requires += builder.python_requirements()
+
+    # If op install enabled, add builder to extensions
+    if op_enabled(op_name) and op_compatible:
+        install_ops[op_name] = op_enabled(op_name)
+        ext_modules.append(builder.builder())
+
+compatible_ops = {op_name: op.is_compatible() for (op_name, op) in ALL_OPS.items()}
+
+print(f'Install Ops={install_ops}')
+
+# Write out version/git info
+git_hash_cmd = "git rev-parse --short HEAD"
+git_branch_cmd = "git rev-parse --abbrev-ref HEAD"
+if command_exists('git') and 'DS_BUILD_STRING' not in os.environ:
+    try:
+        result = subprocess.check_output(git_hash_cmd, shell=True)
+        git_hash = result.decode('utf-8').strip()
+        result = subprocess.check_output(git_branch_cmd, shell=True)
+        git_branch = result.decode('utf-8').strip()
+    except subprocess.CalledProcessError:
+        git_hash = "unknown"
+        git_branch = "unknown"
+else:
+    git_hash = "unknown"
+    git_branch = "unknown"
+
+# Parse the DeepSpeed version string from version.txt
+version_str = open('version.txt', 'r').read().strip()
+
+# Build specifiers like .devX can be added at install time. Otherwise, add the git hash.
+# example: DS_BUILD_STR=".dev20201022" python setup.py sdist bdist_wheel
+
+# Building wheel for distribution, update version file
+if 'DS_BUILD_STRING' in os.environ:
+    # Build string env specified, probably building for distribution
+    with open('build.txt', 'w') as fd:
+        fd.write(os.environ.get('DS_BUILD_STRING'))
+    version_str += os.environ.get('DS_BUILD_STRING')
+elif os.path.isfile('build.txt'):
+    # build.txt exists, probably installing from distribution
+    with open('build.txt', 'r') as fd:
+        version_str += fd.read().strip()
+else:
+    # None of the above, probably installing from source
+    version_str += f'+{git_hash}'
+
+torch_version = ".".join([TORCH_MAJOR, TORCH_MINOR])
+# Set cuda_version to 0.0 if cpu-only
+cuda_version = "0.0"
+if torch.version.cuda is not None:
+    cuda_version = ".".join(torch.version.cuda.split('.')[:2])
+torch_info = {"version": torch_version, "cuda_version": cuda_version}
+
+print(f"version={version_str}, git_hash={git_hash}, git_branch={git_branch}")
+with open('deepspeed/git_version_info_installed.py', 'w') as fd:
+    fd.write(f"version='{version_str}'\n")
+    fd.write(f"git_hash='{git_hash}'\n")
+    fd.write(f"git_branch='{git_branch}'\n")
+    fd.write(f"installed_ops={install_ops}\n")
+    fd.write(f"compatible_ops={compatible_ops}\n")
+    fd.write(f"torch_info={torch_info}\n")
+
+print(f'install_requires={install_requires}')
+print(f'compatible_ops={compatible_ops}')
+print(f'ext_modules={ext_modules}')
+
+# Parse README.md to make long_description for PyPI page.
+thisdir = os.path.abspath(os.path.dirname(__file__))
+with open(os.path.join(thisdir, 'README.md'), encoding='utf-8') as fin:
+    readme_text = fin.read()
+
+start_time = time.time()
+
+setup(name='deepspeed',
+      version=version_str,
+      description='DeepSpeed library',
+      long_description=readme_text,
+      long_description_content_type='text/markdown',
+      author='DeepSpeed Team',
+      author_email='deepspeed@microsoft.com',
+      url='http://deepspeed.ai',
+      install_requires=install_requires,
+      extras_require=extras_require,
+      packages=find_packages(exclude=["docker",
+                                      "third_party"]),
+      include_package_data=True,
+      scripts=[
+          'bin/deepspeed',
+          'bin/deepspeed.pt',
+          'bin/ds',
+          'bin/ds_ssh',
+          'bin/ds_report',
+          'bin/ds_elastic'
+      ],
+      classifiers=[
+          'Programming Language :: Python :: 3.6',
+          'Programming Language :: Python :: 3.7',
+          'Programming Language :: Python :: 3.8'
+      ],
+      license='MIT',
+      ext_modules=ext_modules,
+      cmdclass=cmdclass)
+
+end_time = time.time()
+print(f'deepspeed build time = {end_time - start_time} secs')

bin/deepspeed

-ds
\ No newline at end of file

bin/deepspeed

+#!/usr/bin/env python
+
+from deepspeed.launcher.runner import main
+
+if __name__ == '__main__':
+    main()

bin/deepspeed.pt

-ds
\ No newline at end of file

bin/deepspeed.pt

+#!/usr/bin/env python
+
+from deepspeed.launcher.runner import main
+
+if __name__ == '__main__':
+    main()

csrc/adam/hip/compat.h

+/* Copyright 2020 The Microsoft DeepSpeed Team
+   Copyright NVIDIA/apex
+   This file is adapted from fused adam in NVIDIA/apex, commit a109f85
+*/
+
+#ifndef TORCH_CHECK
+#define TORCH_CHECK AT_CHECK
+#endif
+
+#ifdef VERSION_GE_1_3
+#define DATA_PTR data_ptr
+#else
+#define DATA_PTR data
+#endif

csrc/adam/hip/custom_hip_kernel.hip

+#include "hip/hip_runtime.h"
+
+
+#include "hip/custom_hip_layers.h"
+
+__global__ void param_update_kernel(const float* input, __half* output, int size)
+{
+    int id = blockIdx.x * blockDim.x + threadIdx.x;
+
+    if (id < size) { output[id] = (__half)input[id]; }
+}
+
+void launch_param_update(const float* input, __half* output, int size, hipStream_t stream)
+{
+    int threads = 1024;
+
+    dim3 grid_dim((size - 1) / threads + 1);
+    dim3 block_dim(threads);
+
+   hipLaunchKernelGGL(( param_update_kernel), dim3(grid_dim), dim3(block_dim), 0, stream, input, output, size);
+}

csrc/adam/hip/fused_adam_frontend.cpp

+#include <torch/extension.h>
+
+void multi_tensor_adam_cuda(int chunk_size,
+                            at::Tensor noop_flag,
+                            std::vector<std::vector<at::Tensor>> tensor_lists,
+                            const float lr,
+                            const float beta1,
+                            const float beta2,
+                            const float epsilon,
+                            const int step,
+                            const int mode,
+                            const int bias_correction,
+                            const float weight_decay);
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m)
+{
+    m.def("multi_tensor_adam",
+          &multi_tensor_adam_cuda,
+          "Compute and apply gradient update to parameters for Adam optimizer");
+}

csrc/adam/hip/multi_tensor_adam.hip

+/* Copyright 2020 The Microsoft DeepSpeed Team
+   Copyright NVIDIA/apex
+   This file is adapted from fused adam in NVIDIA/apex, commit a109f85
+*/
+
+#include <ATen/ATen.h>
+#include <ATen/AccumulateType.h>
+#include <ATen/hip/HIPContext.h>
+#include <ATen/hip/Exceptions.h>
+// Another possibility:
+// #include <torch/all.h>
+
+#include <assert.h>
+
+#include "multi_tensor_apply.cuh"
+#include "type_shim.h"
+
+#define BLOCK_SIZE 512
+#define ILP 4
+
+typedef enum {
+    ADAM_MODE_0 = 0,  // L2 regularization mode
+    ADAM_MODE_1 = 1   // Decoupled weight decay mode(AdamW)
+} adamMode_t;
+
+using MATH_T = float;
+
+template <typename T>
+struct AdamFunctor {
+    __device__ __forceinline__ void operator()(int chunk_size,
+                                               volatile int* noop_gmem,
+                                               TensorListMetadata<4>& tl,
+                                               const float beta1,
+                                               const float beta2,
+                                               const float beta1_correction,
+                                               const float beta2_correction,
+                                               const float epsilon,
+                                               const float lr,
+                                               adamMode_t mode,
+                                               const float decay)
+    {
+        // I'd like this kernel to propagate infs/nans.
+        // if(*noop_gmem == 1)
+        //   return;
+
+        int tensor_loc = tl.block_to_tensor[blockIdx.x];
+
+        // potentially use to pass in list of scalar
+        // int tensor_num = tl.start_tensor_this_launch + tensor_loc;
+
+        int chunk_idx = tl.block_to_chunk[blockIdx.x];
+        int n = tl.sizes[tensor_loc];
+
+        T* g = (T*)tl.addresses[0][tensor_loc];
+        g += chunk_idx * chunk_size;
+
+        T* p = (T*)tl.addresses[1][tensor_loc];
+        p += chunk_idx * chunk_size;
+
+        T* m = (T*)tl.addresses[2][tensor_loc];
+        m += chunk_idx * chunk_size;
+
+        T* v = (T*)tl.addresses[3][tensor_loc];
+        v += chunk_idx * chunk_size;
+
+        n -= chunk_idx * chunk_size;
+
+        // see note in multi_tensor_scale_kernel.cu
+        for (int i_start = 0; i_start < n && i_start < chunk_size; i_start += blockDim.x * ILP) {
+            MATH_T r_g[ILP];
+            MATH_T r_p[ILP];
+            MATH_T r_m[ILP];
+            MATH_T r_v[ILP];
+#pragma unroll
+            for (int ii = 0; ii < ILP; ii++) {
+                int i = i_start + threadIdx.x + ii * blockDim.x;
+                if (i < n && i < chunk_size) {
+                    r_g[ii] = g[i];
+                    r_p[ii] = p[i];
+                    r_m[ii] = m[i];
+                    r_v[ii] = v[i];
+                } else {
+                    r_g[ii] = MATH_T(0);
+                    r_p[ii] = MATH_T(0);
+                    r_m[ii] = MATH_T(0);
+                    r_v[ii] = MATH_T(0);
+                }
+            }
+#pragma unroll
+            for (int ii = 0; ii < ILP; ii++) {
+                if (mode == ADAM_MODE_0) {  // L2
+                    r_g[ii] = r_g[ii] + (decay * r_p[ii]);
+                    r_m[ii] = beta1 * r_m[ii] + (1 - beta1) * r_g[ii];
+                    r_v[ii] = beta2 * r_v[ii] + (1 - beta2) * r_g[ii] * r_g[ii];
+                    MATH_T next_m_unbiased = r_m[ii] / beta1_correction;
+                    MATH_T next_v_unbiased = r_v[ii] / beta2_correction;
+                    MATH_T denom = sqrtf(next_v_unbiased) + epsilon;
+                    MATH_T update = next_m_unbiased / denom;
+                    r_p[ii] = r_p[ii] - (lr * update);
+                } else {  // weight decay
+                    r_m[ii] = beta1 * r_m[ii] + (1 - beta1) * r_g[ii];
+                    r_v[ii] = beta2 * r_v[ii] + (1 - beta2) * r_g[ii] * r_g[ii];
+                    MATH_T next_m_unbiased = r_m[ii] / beta1_correction;
+                    MATH_T next_v_unbiased = r_v[ii] / beta2_correction;
+                    MATH_T denom = sqrtf(next_v_unbiased) + epsilon;
+                    MATH_T update = (next_m_unbiased / denom) + (decay * r_p[ii]);
+                    r_p[ii] = r_p[ii] - (lr * update);
+                }
+            }
+#pragma unroll
+            for (int ii = 0; ii < ILP; ii++) {
+                int i = i_start + threadIdx.x + ii * blockDim.x;
+                if (i < n && i < chunk_size) {
+                    p[i] = r_p[ii];
+                    m[i] = r_m[ii];
+                    v[i] = r_v[ii];
+                }
+            }
+        }
+    }
+};
+
+void multi_tensor_adam_cuda(int chunk_size,
+                            at::Tensor noop_flag,
+                            std::vector<std::vector<at::Tensor>> tensor_lists,
+                            const float lr,
+                            const float beta1,
+                            const float beta2,
+                            const float epsilon,
+                            const int step,
+                            const int mode,
+                            const int bias_correction,
+                            const float weight_decay)
+{
+    using namespace at;
+
+    // Handle bias correction mode
+    float bias_correction1 = 1.0f, bias_correction2 = 1.0f;
+    if (bias_correction == 1) {
+        bias_correction1 = 1 - ::pow(beta1, step);
+        bias_correction2 = 1 - ::pow(beta2, step);
+    }
+
+    // Assume single type across p,g,m1,m2 now
+    DISPATCH_DOUBLE_FLOAT_AND_HALF(tensor_lists[0][0].scalar_type(),
+                                   0,
+                                   "adam",
+                                   multi_tensor_apply<4>(BLOCK_SIZE,
+                                                         chunk_size,
+                                                         noop_flag,
+                                                         tensor_lists,
+                                                         AdamFunctor<scalar_t_0>(),
+                                                         beta1,
+                                                         beta2,
+                                                         bias_correction1,
+                                                         bias_correction2,
+                                                         epsilon,
+                                                         lr,
+                                                         (adamMode_t)mode,
+                                                         weight_decay);)
+
+    AT_CUDA_CHECK(hipGetLastError());
+}

csrc/adam/hip/multi_tensor_apply.cuh

+#include "hip/hip_runtime.h"
+/* Copyright 2020 The Microsoft DeepSpeed Team
+   Copyright NVIDIA/apex
+   This file is adapted from fused adam in NVIDIA/apex, commit a109f85
+*/
+
+#include <ATen/ATen.h>
+#include <ATen/AccumulateType.h>
+#include <ATen/hip/HIPContext.h>
+#include <ATen/hip/Exceptions.h>
+#include <ATen/hip/impl/HIPGuardImplMasqueradingAsCUDA.h>
+#include "compat.h"
+
+#include <assert.h>
+
+// #include <iostream>
+
+// This header is the one-stop shop for all your multi-tensor apply needs.
+
+// TODO:  Kernel arg size limit may be <4KB for some other cards (ie Jetson)
+constexpr int depth_to_max_tensors[5] = {110, 64, 48, 36, 30};
+constexpr int depth_to_max_blocks[5] = {320, 320, 320, 320, 320};
+
+template <int n>
+struct TensorListMetadata {
+    void* addresses[n][depth_to_max_tensors[n - 1]];
+    int sizes[depth_to_max_tensors[n - 1]];
+    unsigned char block_to_tensor[depth_to_max_blocks[n - 1]];
+    int block_to_chunk[depth_to_max_blocks[n - 1]];  // I fear this needs to be a full int.
+    int start_tensor_this_launch;
+};
+
+template <typename T, typename U, typename... ArgTypes>
+__global__ void multi_tensor_apply_kernel(int chunk_size,
+                                          volatile int* noop_flag,
+                                          T tl,
+                                          U callable,
+                                          ArgTypes... args)
+{
+    // Hand the chunk information to the user-supplied functor to process however it likes.
+    callable(chunk_size, noop_flag, tl, args...);
+}
+
+template <int depth, typename T, typename... ArgTypes>
+void multi_tensor_apply(int block_size,
+                        int chunk_size,
+                        const at::Tensor& noop_flag,
+                        const std::vector<std::vector<at::Tensor>>& tensor_lists,
+                        T callable,
+                        ArgTypes... args)
+{
+    TORCH_CHECK(tensor_lists.size() == depth, "tensor_lists.size() != depth");
+    int len0 = tensor_lists[0].size();
+    TORCH_CHECK(len0 > 0, "tensor_lists[0].size() is not > 0");
+    auto ref_device = tensor_lists[0][0].device();
+    TORCH_CHECK(ref_device.type() == at::kCUDA, "expected input to be on cuda");
+    for (int l = 0; l < tensor_lists.size(); l++)  // No range-based for because I need indices
+    {
+        TORCH_CHECK(tensor_lists[l].size() == len0, "Size mismatch among tensor lists");
+        for (int t = 0; t < tensor_lists[l].size(); t++) {
+            // TODO:  Print which tensor fails.
+            bool contiguous_memory = tensor_lists[l][t].is_contiguous();
+#ifdef VERSION_GE_1_5
+            contiguous_memory = (contiguous_memory ||
+                                 tensor_lists[l][t].is_contiguous(at::MemoryFormat::ChannelsLast));
+#endif
+            TORCH_CHECK(contiguous_memory, "A tensor was not contiguous.");
+            TORCH_CHECK(tensor_lists[l][t].device() == ref_device,
+                        "A tensor was not on the same device as the first tensor");
+            TORCH_CHECK(tensor_lists[l][t].numel() == tensor_lists[0][t].numel(), "Size mismatch");
+        }
+    }
+
+    int ntensors = tensor_lists[0].size();
+
+    TensorListMetadata<depth> tl;
+
+    const at::hip::OptionalHIPGuardMasqueradingAsCUDA device_guard(device_of(tensor_lists[0][0]));
+    auto stream = at::hip::getCurrentHIPStreamMasqueradingAsCUDA();
+
+    tl.start_tensor_this_launch = 0;
+    int loc_block_info = 0;
+    int loc_tensor_info = 0;
+    for (int t = 0; t < ntensors; t++) {
+        tl.sizes[loc_tensor_info] = tensor_lists[0][t].numel();
+        for (int d = 0; d < depth; d++)
+            tl.addresses[d][loc_tensor_info] = tensor_lists[d][t].data_ptr();
+        loc_tensor_info++;
+
+        int chunks_this_tensor = (tensor_lists[0][t].numel() + chunk_size - 1) / chunk_size;
+
+        for (int chunk = 0; chunk < chunks_this_tensor; chunk++) {
+            // std::cout << chunks_this_tensor << std::endl;
+            tl.block_to_tensor[loc_block_info] = loc_tensor_info - 1;
+            tl.block_to_chunk[loc_block_info] = chunk;
+            loc_block_info++;
+
+            bool tensors_full = (loc_tensor_info == depth_to_max_tensors[depth - 1] &&
+                                 chunk == chunks_this_tensor - 1);
+            bool blocks_full = (loc_block_info == depth_to_max_blocks[depth - 1]);
+            bool last_chunk = (t == ntensors - 1 && chunk == chunks_this_tensor - 1);
+            if (tensors_full || blocks_full || last_chunk) {
+                // using accscalar_t = acc_type<scalar_t, true>;
+               hipLaunchKernelGGL(( multi_tensor_apply_kernel), dim3(loc_block_info), dim3(block_size), 0, stream, 
+                    chunk_size, noop_flag.DATA_PTR<int>(), tl, callable, args...);
+
+                AT_CUDA_CHECK(hipGetLastError());
+
+                // Reset.  The control flow possibilities here make my brain hurt.
+                loc_block_info = 0;
+                if (chunk == chunks_this_tensor - 1) {
+                    // std::cout << "Hit case 1 " << cond1 << " " << cond2 << " " << cond3 <<
+                    // std::endl;
+                    loc_tensor_info = 0;
+                    tl.start_tensor_this_launch = t + 1;
+                } else {
+                    // std::cout << "Hit case 2 " << cond1 << " " << cond2 << " " << cond3 <<
+                    // std::endl;
+                    tl.sizes[0] = tl.sizes[loc_tensor_info - 1];
+                    for (int d = 0; d < depth; d++)
+                        tl.addresses[d][0] = tl.addresses[d][loc_tensor_info - 1];
+                    loc_tensor_info = 1;
+                    tl.start_tensor_this_launch = t;
+                }
+            }
+        }
+    }
+}

csrc/includes/hip/StopWatch.h

+#pragma once
+#ifdef _WIN32
+#include <windows.h>
+#else
+#include <time.h>
+#endif
+
+#ifdef _WIN32
+
+class Stopwatch {
+private:
+    double m_total_time;
+    LARGE_INTEGER m_start_time;
+
+public:
+    Stopwatch() { m_total_time = 0.0; }
+
+    ~Stopwatch() {}
+
+    void Reset() { m_total_time = 0.0; }
+
+    void Start() { QueryPerformanceCounter(&m_start_time); }
+
+    void Restart()
+    {
+        m_total_time = 0.0;
+        QueryPerformanceCounter(&m_start_time);
+    }
+
+    void Stop()
+    {
+        LARGE_INTEGER frequency;
+        LARGE_INTEGER stop_time;
+        QueryPerformanceFrequency(&frequency);
+        QueryPerformanceCounter(&stop_time);
+        m_total_time +=
+            ((double)(stop_time.QuadPart - m_start_time.QuadPart) / (double)frequency.QuadPart);
+    }
+
+    double GetTimeInSeconds() { return m_total_time; }
+};
+
+#else
+
+class Stopwatch {
+private:
+    double m_total_time;
+    struct timespec m_start_time;
+    bool m_is_started;
+
+public:
+    Stopwatch()
+    {
+        m_total_time = 0.0;
+        m_is_started = false;
+    }
+
+    ~Stopwatch() {}
+
+    void Reset() { m_total_time = 0.0; }
+
+    void Start()
+    {
+        clock_gettime(CLOCK_MONOTONIC, &m_start_time);
+        m_is_started = true;
+    }
+
+    void Restart()
+    {
+        m_total_time = 0.0;
+        clock_gettime(CLOCK_MONOTONIC, &m_start_time);
+        m_is_started = true;
+    }
+
+    void Stop()
+    {
+        if (m_is_started) {
+            m_is_started = false;
+
+            struct timespec end_time;
+            clock_gettime(CLOCK_MONOTONIC, &end_time);
+
+            m_total_time += (double)(end_time.tv_sec - m_start_time.tv_sec) +
+                            (double)(end_time.tv_nsec - m_start_time.tv_nsec) / 1e9;
+        }
+    }
+
+    double GetTimeInSeconds()
+    {
+        if (m_is_started) {
+            Stop();
+            Start();
+        }
+        return m_total_time;
+    }
+};
+
+#endif

csrc/includes/hip/Timer.h

+
+#ifndef __TIMER_H__
+#define __TIMER_H__
+
+#include <hip/hip_runtime.h>
+#include <chrono>
+#include "hip/hip_runtime.h"
+
+class GPUTimer {
+    hipEvent_t start, stop;
+
+public:
+    GPUTimer()
+    {
+        hipEventCreate(&start);
+        hipEventCreate(&stop);
+    }
+    ~GPUTimer()
+    {
+        hipEventDestroy(start);
+        hipEventDestroy(stop);
+    }
+    inline void Record() { hipEventRecord(start); }
+    inline void Elapsed(float& time_elapsed)
+    {
+        hipEventRecord(stop);
+        hipEventSynchronize(stop);
+        hipEventElapsedTime(&time_elapsed, start, stop);
+    }
+};
+
+class CPUTimer {
+    std::chrono::high_resolution_clock::time_point start;
+
+public:
+    CPUTimer() : start(std::chrono::high_resolution_clock::now()) {}
+    inline void Reset() { start = std::chrono::high_resolution_clock::now(); }
+    inline float Elapsed()
+    {
+        auto temp = start;
+        start = std::chrono::high_resolution_clock::now();
+        return (float)(std::chrono::duration_cast<std::chrono::microseconds>(start - temp).count() /
+                       1e3);
+    }
+};
+
+#endif

csrc/includes/hip/context.h

+#pragma once
+
+#include <ATen/hip/HIPContext.h>
+#include <hip/hip_runtime_api.h>
+#include <cassert>
+#include <iostream>
+#include <vector>
+#include "rocblas.h"
+#include "hip/hip_runtime.h"
+#include "hiprand.h"
+#include "gemm_test.h"
+
+#define WARP_SIZE 32
+
+#define CUDA_CHECK(callstr)                                                                    \
+    {                                                                                          \
+        hipError_t error_code = callstr;                                                      \
+        if (error_code != hipSuccess) {                                                       \
+            std::cerr << "CUDA error " << error_code << " at " << __FILE__ << ":" << __LINE__; \
+            assert(0);                                                                         \
+        }                                                                                      \
+    }
+
+#define CUDA_1D_KERNEL_LOOP(i, n) \
+    for (size_t i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); i += blockDim.x * gridDim.x)
+
+#define CUDA_2D_KERNEL_LOOP(i, n, j, m)                                                          \
+    for (size_t i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); i += blockDim.x * gridDim.x) \
+        for (size_t j = blockIdx.y * blockDim.y + threadIdx.y; j < (m); j += blockDim.y * gridDim.y)
+
+#define DS_CUDA_NUM_THREADS 512
+#define DS_MAXIMUM_NUM_BLOCKS 262144
+
+inline int DS_GET_BLOCKS(const int N)
+{
+    return (std::max)(
+        (std::min)((N + DS_CUDA_NUM_THREADS - 1) / DS_CUDA_NUM_THREADS, DS_MAXIMUM_NUM_BLOCKS),
+        // Use at least 1 block, since CUDA does not allow empty block
+        1);
+}
+
+class Context {
+public:
+    Context() : _workspace(nullptr), _seed(42), _curr_offset(0)
+    {
+        hiprandCreateGenerator(&_gen, HIPRAND_RNG_PSEUDO_DEFAULT);
+        hiprandSetPseudoRandomGeneratorSeed(_gen, 123);
+        if (rocblas_create_handle(&_cublasHandle) != rocblas_status_success) {
+            auto message = std::string("Fail to create cublas handle.");
+            std::cerr << message << std::endl;
+            throw std::runtime_error(message);
+        }
+    }
+
+    virtual ~Context()
+    {
+        rocblas_destroy_handle(_cublasHandle);
+        hipFree(_workspace);
+    }
+
+    static Context& Instance()
+    {
+        static Context _ctx;
+        return _ctx;
+    }
+
+    void SetWorkSpace(void* workspace)
+    {
+        if (!workspace) { throw std::runtime_error("Workspace is null."); }
+        _workspace = workspace;
+    }
+
+    void* GetWorkSpace() { return _workspace; }
+
+    hiprandGenerator_t& GetRandGenerator() { return _gen; }
+
+    hipStream_t GetCurrentStream()
+    {
+        // get current pytorch stream.
+        hipStream_t stream = at::hip::getCurrentHIPStreamMasqueradingAsCUDA();
+        return stream;
+    }
+
+    hipStream_t GetNewStream() { return at::hip::getStreamFromPoolMasqueradingAsCUDA(); }
+
+    rocblas_handle GetCublasHandle() { return _cublasHandle; }
+
+    std::pair<uint64_t, uint64_t> IncrementOffset(uint64_t offset_inc)
+    {
+        uint64_t offset = _curr_offset;
+        _curr_offset += offset_inc;
+        return std::pair<uint64_t, uint64_t>(_seed, offset);
+    }
+
+    void SetSeed(uint64_t new_seed) { _seed = new_seed; }
+
+    void TestGemmFP16(bool test_gemm, int batch_size, int seq_len, int head_num, int size_per_head)
+    {
+        // avoid rerun.
+        if (_gemm_algos.size() > 0) return;
+
+        if (test_gemm) {
+            rocblas_handle handle = GetCublasHandle();
+
+            std::unique_ptr<GemmTest<__half>> test_qkv_fw(
+                new GemmTest<__half>(batch_size * seq_len,      // M
+                                     head_num * size_per_head,  // N
+                                     head_num * size_per_head,  // K
+                                     rocblas_operation_transpose,
+                                     rocblas_operation_none,
+                                     handle));
+
+            std::unique_ptr<GemmTest<__half>> test_inter(
+                new GemmTest<__half>(batch_size * seq_len,          // M
+                                     4 * head_num * size_per_head,  // N
+                                     head_num * size_per_head,      // K
+                                     rocblas_operation_transpose,
+                                     rocblas_operation_none,
+                                     handle));
+
+            std::unique_ptr<GemmTest<__half>> test_output(
+                new GemmTest<__half>(batch_size * seq_len,          // M
+                                     head_num * size_per_head,      // N
+                                     4 * head_num * size_per_head,  // K
+                                     rocblas_operation_transpose,
+                                     rocblas_operation_none,
+                                     handle));
+
+            std::unique_ptr<StridedGemmTest<__half>> test_attn_scores(
+                new StridedGemmTest<__half>(batch_size * head_num,  // batch
+                                            seq_len,                // M
+                                            seq_len,                // N
+                                            size_per_head,          // K
+                                            rocblas_operation_transpose,
+                                            rocblas_operation_none,
+                                            handle));
+
+            std::unique_ptr<StridedGemmTest<__half>> test_attn_context(
+                new StridedGemmTest<__half>(batch_size * head_num,  // batch
+                                            size_per_head,          // M
+                                            seq_len,                // N
+                                            seq_len,                // K
+                                            rocblas_operation_none,
+                                            rocblas_operation_none,
+                                            handle));
+
+            _gemm_algos.push_back(test_qkv_fw->TestAlgo(100));
+            _gemm_algos.push_back(test_inter->TestAlgo(100));
+            _gemm_algos.push_back(test_output->TestAlgo(100));
+            _gemm_algos.push_back(test_attn_scores->TestAlgo(100));
+            _gemm_algos.push_back(test_attn_context->TestAlgo(100));
+        } else {
+            // Use default algo.
+            _gemm_algos.push_back(std::array<int, 3>({99, 99, 99}));
+            _gemm_algos.push_back(std::array<int, 3>({99, 99, 99}));
+            _gemm_algos.push_back(std::array<int, 3>({99, 99, 99}));
+            _gemm_algos.push_back(std::array<int, 3>({99, 99, 99}));
+            _gemm_algos.push_back(std::array<int, 3>({99, 99, 99}));
+        }
+    }
+
+    const std::vector<std::array<int, 3>>& GetGemmAlgos() const { return _gemm_algos; }
+
+private:
+    hiprandGenerator_t _gen;
+    rocblas_handle _cublasHandle;
+    void* _workspace;
+    uint64_t _seed;
+    uint64_t _curr_offset;
+    std::vector<std::array<int, 3>> _gemm_algos;
+};

csrc/includes/hip/cpu_adam.h

+#pragma once
+
+#include <cpuid.h>
+#include <hip/hip_fp16.h>
+#include <hip/hip_runtime_api.h>
+#include <stdio.h>
+#include <x86intrin.h>
+#include <cassert>
+#include "context.h"
+#include "rocblas.h"
+#include "hip/hip_runtime.h"
+#include "hiprand.h"
+
+#define CUDA_CHECK(callstr)                                                                    \
+    {                                                                                          \
+        hipError_t error_code = callstr;                                                      \
+        if (error_code != hipSuccess) {                                                       \
+            std::cerr << "CUDA error " << error_code << " at " << __FILE__ << ":" << __LINE__; \
+            assert(0);                                                                         \
+        }                                                                                      \
+    }
+
+#define TILE (1024 * 1024 * 1024)
+
+#if defined(__AVX512__)
+#define SIMD_STORE(a, d) _mm512_storeu_ps(a, d)
+#define SIMD_LOAD(x) _mm512_loadu_ps(x)
+#define SIMD_SET(x) _mm512_set1_ps(x)
+#define SIMD_MUL(x, y) _mm512_mul_ps(x, y)
+#define SIMD_FMA(x, y, c) _mm512_fmadd_ps(x, y, c)
+#define SIMD_SQRT(x) _mm512_sqrt_ps(x)
+#define SIMD_DIV(x, y) _mm512_div_ps(x, y)
+#define SIMD_WIDTH 16
+#else
+#if defined(__AVX256__)
+#define SIMD_STORE(a, d) _mm256_storeu_ps(a, d)
+#define SIMD_LOAD(x) _mm256_loadu_ps(x)
+#define SIMD_SET(x) _mm256_set1_ps(x)
+#define SIMD_MUL(x, y) _mm256_mul_ps(x, y)
+#define SIMD_FMA(x, y, c) _mm256_fmadd_ps(x, y, c)
+#define SIMD_SQRT(x) _mm256_sqrt_ps(x)
+#define SIMD_DIV(x, y) _mm256_div_ps(x, y)
+#define SIMD_WIDTH 8
+#endif
+#endif
+
+class Adam_Optimizer {
+public:
+    Adam_Optimizer(float alpha = 1e-3,
+                   float betta1 = 0.9,
+                   float betta2 = 0.999,
+                   float eps = 1e-8,
+                   float weight_decay = 0,
+                   bool adamw_mode = true)
+        : _alpha(alpha),
+          _betta1(betta1),
+          _betta2(betta2),
+          _eps(eps),
+          _weight_decay(weight_decay),
+          _betta1_t(1.0),
+          _betta2_t(1.0),
+          _step(0),
+          _buf_index(false),
+          _adamw_mode(adamw_mode)
+    {
+        hipHostMalloc((void**)_doubled_buffer, TILE * sizeof(float));
+        hipHostMalloc((void**)(_doubled_buffer + 1), TILE * sizeof(float));
+
+        _streams[0] = Context::Instance().GetCurrentStream();
+        _streams[1] = Context::Instance().GetNewStream();
+    }
+    ~Adam_Optimizer()
+    {
+        hipHostFree(_doubled_buffer[0]);
+        hipHostFree(_doubled_buffer[1]);
+    }
+    void Step(float* _params,
+              float* grads,
+              float* _exp_avg,
+              float* _exp_avg_sq,
+              size_t param_size,
+              __half* dev_param = nullptr);
+    void Step_4(float* _params,
+                float* grads,
+                float* _exp_avg,
+                float* _exp_avg_sa,
+                size_t param_size,
+                __half* dev_param = nullptr);
+    void Step_8(float* _params,
+                float* grads,
+                float* _exp_avg,
+                float* _exp_avg_sq,
+                size_t _param_size,
+                __half* dev_params = nullptr);
+    inline void SynchronizeStreams()
+    {
+        for (int i = 0; i < 2; i++) hipStreamSynchronize(_streams[i]);
+    }
+    inline void IncrementStep(size_t step, float beta1, float beta2)
+    {
+        if (beta1 != _betta1 || beta2 != _betta2) {
+            _step = step;
+            _betta1 = beta1;
+            _betta2 = beta2;
+            _betta1_t = std::pow(_betta1, step);
+            _betta2_t = std::pow(_betta2, step);
+        } else {
+            _step++;
+            if (_step != step) {
+                _betta1_t = std::pow(_betta1, step);
+                _betta2_t = std::pow(_betta2, step);
+                _step = step;
+            } else {
+                _betta1_t *= _betta1;
+                _betta2_t *= _betta2;
+            }
+        }
+    }
+    inline void update_state(float lr, float epsilon, float weight_decay, bool bias_correction)
+    {
+        _alpha = lr;
+        _eps = epsilon;
+        _weight_decay = weight_decay;
+
+        _bias_correction1 = 1.0f;
+        _bias_correction2 = 1.0f;
+        if (bias_correction == 1) {
+            _bias_correction1 = 1 - _betta1_t;
+            _bias_correction2 = 1 / sqrt(1 - _betta2_t);
+        }
+    }
+
+private:
+#if defined(__AVX512__) or defined(__AVX256__)
+    union AVX_Data {
+#if defined(__AVX512__)
+        __m512 data;
+#else
+        __m256 data;
+#endif
+        // float data_f[16];
+    };
+#endif
+
+    float _alpha;
+    float _betta1;
+    float _betta2;
+    float _eps;
+    float _weight_decay;
+
+    float _betta1_t;
+    float _betta2_t;
+    size_t _step;
+
+    float _bias_correction1;
+    float _bias_correction2;
+
+    float* _doubled_buffer[2];
+    bool _buf_index;
+    bool _adamw_mode;
+
+    hipStream_t _streams[2];
+};

csrc/includes/hip/cublas_wrappers.h

+#pragma once
+
+#include <assert.h>
+#include <rocblas.h>
+#include <hip/hip_runtime.h>
+#include <hip/hip_fp16.h>
+#include <hip/hip_runtime.h>
+#ifndef __HIP_PLATFORM_HCC__
+#include <mma.h>
+#endif
+#include <stdio.h>
+
+int cublas_gemm_ex(rocblas_handle handle,
+                   rocblas_operation transa,
+                   rocblas_operation transb,
+                   int m,
+                   int n,
+                   int k,
+                   const float* alpha,
+                   const float* beta,
+                   const float* A,
+                   const float* B,
+                   float* C,
+#ifdef __HIP_PLATFORM_HCC__
+                   rocblas_gemm_algo algo = rocblas_gemm_algo_standard);
+#else
+                   cublasGemmAlgo_t algo = CUBLAS_GEMM_DEFAULT);
+#endif
+
+int cublas_gemm_ex(rocblas_handle handle,
+                   rocblas_operation transa,
+                   rocblas_operation transb,
+                   int m,
+                   int n,
+                   int k,
+                   const float* alpha,
+                   const float* beta,
+                   const __half* A,
+                   const __half* B,
+                   __half* C,
+#ifdef __HIP_PLATFORM_HCC__
+                   rocblas_gemm_algo algo = rocblas_gemm_algo_standard);
+#else
+                   cublasGemmAlgo_t algo = CUBLAS_GEMM_DEFAULT_TENSOR_OP);
+#endif
+
+int cublas_strided_batched_gemm(rocblas_handle handle,
+                                int m,
+                                int n,
+                                int k,
+                                const float* alpha,
+                                const float* beta,
+                                const float* A,
+                                const float* B,
+                                float* C,
+                                rocblas_operation op_A,
+                                rocblas_operation op_B,
+                                int stride_A,
+                                int stride_B,
+                                int stride_C,
+                                int batch,
+#ifdef __HIP_PLATFORM_HCC__
+                                rocblas_gemm_algo algo = rocblas_gemm_algo_standard);
+#else
+                                cublasGemmAlgo_t algo = CUBLAS_GEMM_DEFAULT);
+#endif
+
+int cublas_strided_batched_gemm(rocblas_handle handle,
+                                int m,
+                                int n,
+                                int k,
+                                const float* alpha,
+                                const float* beta,
+                                const __half* A,
+                                const __half* B,
+                                __half* C,
+                                rocblas_operation op_A,
+                                rocblas_operation op_B,
+                                int stride_A,
+                                int stride_B,
+                                int stride_C,
+                                int batch,
+#ifdef __HIP_PLATFORM_HCC__
+                                rocblas_gemm_algo algo = rocblas_gemm_algo_standard);
+#else
+                                cublasGemmAlgo_t algo = CUBLAS_GEMM_DEFAULT_TENSOR_OP);
+#endif

csrc/includes/hip/cublas_wrappers.h.bak

+#pragma once
+
+#include <assert.h>
+#include <rocblas.h>
+#include <hip/hip_runtime.h>
+#include <hip/hip_fp16.h>
+#include <hip/hip_runtime.h>
+#include <mma.h>
+#include <stdio.h>
+
+int cublas_gemm_ex(rocblas_handle handle,
+                   rocblas_operation transa,
+                   rocblas_operation transb,
+                   int m,
+                   int n,
+                   int k,
+                   const float* alpha,
+                   const float* beta,
+                   const float* A,
+                   const float* B,
+                   float* C,
+                   cublasGemmAlgo_t algo = CUBLAS_GEMM_DEFAULT);
+
+int cublas_gemm_ex(rocblas_handle handle,
+                   rocblas_operation transa,
+                   rocblas_operation transb,
+                   int m,
+                   int n,
+                   int k,
+                   const float* alpha,
+                   const float* beta,
+                   const __half* A,
+                   const __half* B,
+                   __half* C,
+                   cublasGemmAlgo_t algo = CUBLAS_GEMM_DEFAULT_TENSOR_OP);
+
+int cublas_strided_batched_gemm(rocblas_handle handle,
+                                int m,
+                                int n,
+                                int k,
+                                const float* alpha,
+                                const float* beta,
+                                const float* A,
+                                const float* B,
+                                float* C,
+                                rocblas_operation op_A,
+                                rocblas_operation op_B,
+                                int stride_A,
+                                int stride_B,
+                                int stride_C,
+                                int batch,
+                                cublasGemmAlgo_t algo = CUBLAS_GEMM_DEFAULT);
+
+int cublas_strided_batched_gemm(rocblas_handle handle,
+                                int m,
+                                int n,
+                                int k,
+                                const float* alpha,
+                                const float* beta,
+                                const __half* A,
+                                const __half* B,
+                                __half* C,
+                                rocblas_operation op_A,
+                                rocblas_operation op_B,
+                                int stride_A,
+                                int stride_B,
+                                int stride_C,
+                                int batch,
+                                cublasGemmAlgo_t algo = CUBLAS_GEMM_DEFAULT_TENSOR_OP);

csrc/includes/hip/custom_hip_layers.h

+#pragma once
+
+#include <hip/hip_runtime.h>
+#include <hip/hip_fp16.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#ifdef __HIP_PLATFORM_HCC__
+#include <hip/hip_cooperative_groups.h>
+#else
+#include <cooperative_groups.h>
+#endif
+#define __HIP_DEVICE_COMPILE__ 1
+#define __HIP_PLATFORM_HCC__
+
+#include <hiprand_kernel.h>
+
+#include "context.h"
+#include "cublas_wrappers.h"
+
+#define MAX_THREADS 1024
+#define THREADS 256
+
+#define MAX_THREAD_STRIDE 32
+#define TILE_DIM 32
+
+// Maximum sequence-length support based on the number of threads (2048) allowed in each block and
+// this MAX is 8K For higher sequence length we need to use higher Max, like for 64K : 32
+#define MAX_THREAD_ITERATIONS 8  // Maximum 8K
+#define MAX_WARP_NUM 32
+
+#define MAX_REGISTERS 256
+
+// Fused bias add with gelu activation
+template <typename T>
+void launch_bias_gelu(const T* input,
+                      const T* bias,
+                      T* output,
+                      int intermediate_size,
+                      int batch_size,
+                      hipStream_t stream);
+
+template <typename T>
+void launch_gelu(const T* input,
+                 T* output,
+                 int intermediate_size,
+                 int batch_size,
+                 hipStream_t stream);
+
+template <typename T>
+void launch_d_gelu(T* d_output,
+                   const T* input,
+                   const T* bias,
+                   int intermediate_size,
+                   int batch_size,
+                   hipStream_t stream);
+
+// Custom fused bias add with layer normalization
+template <typename T>
+void launch_bias_residual_layer_norm(T* vals,
+                                     const T* residual,
+                                     const T* gamma,
+                                     const T* beta,
+                                     float epsilon,
+                                     int batch_size,
+                                     int hidden_dim,
+                                     hipStream_t stream,
+                                     bool preLayerNorm,
+                                     bool training,
+                                     T* vars,
+                                     T* means);
+
+template <typename T>
+void launch_bias_residual_layer_norm(T* vals,
+                                     const T* residual,
+                                     const T* gamma,
+                                     const T* beta,
+                                     float epsilon,
+                                     int batch_size,
+                                     int hidden_dim,
+                                     hipStream_t stream,
+                                     bool preLayerNorm,
+                                     bool training,
+                                     T* vars);
+
+template <typename T>
+void launch_layerNorm_backward_fused_add(const T* out_grad1,
+                                         const T* out_grad2,
+                                         const T* X_data,
+                                         const T* vars,
+                                         const T* means,
+                                         const T* gamma,
+                                         T* gamma_grad,
+                                         T* betta_grad,
+                                         T* inp_grad,
+                                         int batch_size,
+                                         int hidden_dim,
+                                         hipStream_t stream[2]);
+template <typename T>
+void launch_layerNorm_backward_fused_add(const T* out_grad1,
+                                         const T* out_grad2,
+                                         const T* vals_hat,
+                                         const T* vars,
+                                         const T* gamma,
+                                         T* gamma_grad,
+                                         T* betta_grad,
+                                         T* inp_grad,
+                                         int batch_size,
+                                         int hidden_dim,
+                                         hipStream_t stream[2],
+                                         bool invertible = false,
+                                         const T* betta = nullptr);
+
+template <typename T>
+void launch_layerNorm_backward(const T* out_grad,
+                               const T* X_data,
+                               const T* vars,
+                               const T* means,
+                               const T* gamma,
+                               T* gamma_grad,
+                               T* betta_grad,
+                               T* inp_grad,
+                               int batch_size,
+                               int hidden_dim,
+                               hipStream_t stream[2]);
+
+template <typename T>
+void launch_layerNorm_backward(const T* out_grad,
+                               const T* vals_hat,
+                               const T* vars,
+                               const T* gamma,
+                               T* gamma_grad,
+                               T* betta_grad,
+                               T* inp_grad,
+                               int batch_size,
+                               int hidden_dim,
+                               hipStream_t stream[2],
+                               bool invertible = false,
+                               const T* betta = nullptr);
+
+template <typename T>
+void launch_layerNorm_backward_nreversible(const T* out_grad,
+                                           const T* vals,
+                                           const T* out_grad_trans,
+                                           const T* vals_trans,
+                                           const T* means,
+                                           const T* vars,
+                                           const T* gamma,
+                                           T* gamma_grad,
+                                           T* betta_grad,
+                                           T* inp_grad,
+                                           int batch_size,
+                                           int hidden_dim,
+                                           hipStream_t stream[2]);
+
+template <typename T>
+void Transpose(const T* inp_mat, T* out_mat, int rows, int cols, hipStream_t stream);
+
+template <typename T>
+void launch_attn_softmax_backward(T* out_grad,
+                                  const T* soft_inp,
+                                  int batch_size,
+                                  int heads,
+                                  int seq_length,
+                                  hipStream_t stream);
+
+template <typename T>
+void launch_attn_softmax_backward_v2(T* out_grad,
+                                     const T* soft_inp,
+                                     int batch_size,
+                                     int heads,
+                                     int seq_length,
+                                     hipStream_t stream);
+
+// Custom softmax with scaling and attention mask addition
+template <typename T>
+void launch_attn_softmax(T* vals,
+                         const T* attn_mask,
+                         int batch_size,
+                         int heads,
+                         int sequence_length,
+                         hipStream_t stream);
+
+template <typename T>
+void launch_transform_0213(T* output,
+                           const T* vals,
+                           int batch_size,
+                           int seq_length,
+                           int hidden_dim,
+                           int heads,
+                           hipStream_t stream);
+
+// Custom bias add
+template <typename T>
+void launch_bias_add_transform_0213(T* outputs,
+                                    const T* vals,
+                                    const T* bias,
+                                    int batch_size,
+                                    int seq_length,
+                                    int hidden_dim,
+                                    int heads,
+                                    hipStream_t stream,
+                                    int trans_count);
+
+// 4D transform [0, 1, 2, 3] -> [0, 2, 1, 3]
+template <typename T>
+void launch_transform4d_0213(T* out,
+                             const T* in,
+                             int batch_size,
+                             int heads,
+                             int seq_length,
+                             int hidden_dim,
+                             hipStream_t stream,
+                             int trans_count);
+
+template <typename T>
+void launch_dropout(T* vals,
+                    const T* bias,
+                    uint8_t* mask,
+                    int batch,
+                    int dim,
+                    float ratio,
+                    hipStream_t stream);
+
+template <typename T>
+void launch_dropout(T* vals_out,
+                    const T* vals,
+                    uint8_t* mask,
+                    int total_count,
+                    int dim,
+                    float ratio,
+                    hipStream_t stream,
+                    bool bwd = false);
+
+template <typename T>
+void launch_dropout(T* out,
+                    const T* vals,
+                    const T* residual,
+                    const T* bias,
+                    uint8_t* mask,
+                    int batch,
+                    int dim,
+                    float ratio,
+                    hipStream_t stream);
+
+template <typename T>
+void launch_dropout_grad(T* vals, uint8_t* mask, int total_count, float ratio, hipStream_t stream);
+
+template <typename T>
+void launch_dropout_grad(T* vals_out,
+                         const T* vals,
+                         uint8_t* mask,
+                         int total_count,
+                         float ratio,
+                         hipStream_t stream);
+
+template <typename T>
+void launch_fuse_transpose_bias_kernel(const T* inp,
+                                       T* out,
+                                       int rows,
+                                       int cols,
+                                       hipStream_t stream);
+
+void launch_param_update(const float* input, __half* output, int size, hipStream_t stream);

csrc/includes/hip/dropout.h

+#pragma once
+
+#include <hip/hip_runtime.h>
+#include <hip/hip_fp16.h>
+#include <stdio.h>
+
+template <typename T>
+class Dropout {
+public:
+    struct Config {
+        float ratio;
+        uint32_t dim;
+        bool training;
+
+        Config(float r, uint32_t d) : ratio(r), dim(d), training(true) {}
+
+        float RATIO() const { return training ? ratio : 0.0; }
+        inline void SetDim(uint32_t d) { dim = d; }
+    };
+
+    Dropout(const Config& config) : _config(config), _mask(nullptr) {}
+
+    virtual ~Dropout() {}
+
+    void Forward(int bsz, T* out, const T* vals, hipStream_t stream, bool bwd = false)
+    {
+        launch_dropout<T>(
+            out, vals, _mask, bsz * _config.dim, _config.dim, _config.RATIO(), stream, bwd);
+    }
+
+    void ForwardWithBias(int bsz, T* vals, const T* bias, hipStream_t stream)
+    {
+        launch_dropout<T>(vals, bias, _mask, bsz, _config.dim, _config.RATIO(), stream);
+    }
+
+    void ForwardWithBias(int bsz,
+                         T* out,
+                         const T* vals,
+                         const T* residual,
+                         const T* bias,
+                         hipStream_t stream)
+    {
+        launch_dropout<T>(
+            out, vals, residual, bias, _mask, bsz, _config.dim, _config.RATIO(), stream);
+    }
+
+    void Backward(int bsz, T* d_vals, hipStream_t stream)
+    {
+        launch_dropout_grad<T>(d_vals, _mask, bsz * _config.dim, _config.RATIO(), stream);
+    }
+
+    void Backward(int bsz, T* d_vals_out, const T* d_vals, hipStream_t stream)
+    {
+        launch_dropout_grad<T>(
+            d_vals_out, d_vals, _mask, bsz * _config.dim, _config.RATIO(), stream);
+    }
+
+    bool HasDropout() const { return _config.RATIO() > 0.0; }
+
+    void SetTrainingMode(bool training) { _config.training = training; }
+
+    void SetMask(uint8_t* mask)
+    {
+        if (!mask) { throw std::runtime_error("Dropout mask is null."); }
+
+        _mask = mask;
+    }
+
+    Config GetConfig() const { return _config; }
+
+    inline void SetDimension(uint32_t dim) { _config.SetDim(dim); }
+
+private:
+    uint8_t* _mask;
+    Config _config;
+};