[Inference/Refactor] Refactor compilation mechanism and unified multi hw (#5613)

* refactor compilation mechanism and unified multi hw

* fix file path bug

* add init.py to make pybind a module to avoid relative path error caused by softlink

* delete duplicated micros

* fix micros bug in gcc

extensions/pybind/optimizer/__init__.py

@@ -0,0 +1,3 @@
+from .fused_optimizer_cuda import FusedOptimizerCudaExtension
+
+__all__ = ["FusedOptimizerCudaExtension"]

extensions/pybind/optimizer/fused_optimizer_cuda.py

@@ -0,0 +1,29 @@
+from ...cuda_extension import _CudaExtension
+from ...utils import get_cuda_cc_flag
+
+
+class FusedOptimizerCudaExtension(_CudaExtension):
+    def __init__(self):
+        super().__init__(name="fused_optim_cuda")
+
+    def sources_files(self):
+        ret = [
+            self.csrc_abs_path(fname)
+            for fname in [
+                "kernel/cuda/multi_tensor_sgd_kernel.cu",
+                "kernel/cuda/multi_tensor_scale_kernel.cu",
+                "kernel/cuda/multi_tensor_adam_kernel.cu",
+                "kernel/cuda/multi_tensor_l2norm_kernel.cu",
+                "kernel/cuda/multi_tensor_lamb_kernel.cu",
+            ]
+        ] + [self.pybind_abs_path("optimizer/optimizer.cpp")]
+        return ret
+
+    def cxx_flags(self):
+        version_dependent_macros = ["-DVERSION_GE_1_1", "-DVERSION_GE_1_3", "-DVERSION_GE_1_5"]
+        return ["-O3"] + version_dependent_macros
+
+    def nvcc_flags(self):
+        extra_cuda_flags = ["-lineinfo"]
+        extra_cuda_flags.extend(get_cuda_cc_flag())
+        return ["-O3", "--use_fast_math"] + extra_cuda_flags + super().nvcc_flags()

extensions/pybind/optimizer/optimizer.cpp

@@ -0,0 +1,49 @@
+// modified from
+// https://github.com/NVIDIA/apex/blob/master/csrc/multi_tensor_adam.cu
+#include <torch/extension.h>
+
+void multi_tensor_scale_cuda(int chunk_size, at::Tensor noop_flag,
+                             std::vector<std::vector<at::Tensor>> tensor_lists,
+                             float scale);
+
+void multi_tensor_sgd_cuda(int chunk_size, at::Tensor noop_flag,
+                           std::vector<std::vector<at::Tensor>> tensor_lists,
+                           float wd, float momentum, float dampening, float lr,
+                           bool nesterov, bool first_run,
+                           bool wd_after_momentum, float scale);
+
+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,
+                            const float div_scale);
+
+void multi_tensor_lamb_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 bias_correction,
+                            const float weight_decay, const int grad_averaging,
+                            const int mode, at::Tensor global_grad_norm,
+                            const float max_grad_norm,
+                            at::optional<bool> use_nvlamb_python);
+
+std::tuple<at::Tensor, at::Tensor> multi_tensor_l2norm_cuda(
+    int chunk_size, at::Tensor noop_flag,
+    std::vector<std::vector<at::Tensor>> tensor_lists,
+    at::optional<bool> per_tensor_python);
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("multi_tensor_scale", &multi_tensor_scale_cuda,
+        "Fused overflow check + scale for a list of contiguous tensors");
+  m.def("multi_tensor_sgd", &multi_tensor_sgd_cuda,
+        "Fused SGD optimizer for list of contiguous tensors");
+  m.def("multi_tensor_adam", &multi_tensor_adam_cuda,
+        "Compute and apply gradient update to parameters for Adam optimizer");
+  m.def("multi_tensor_lamb", &multi_tensor_lamb_cuda,
+        "Computes and apply update for LAMB optimizer");
+  m.def("multi_tensor_l2norm", &multi_tensor_l2norm_cuda,
+        "Computes L2 norm for a list of contiguous tensors");
+}