Add CPU registrations to custom operators (#3262)

Summary:
Pull Request resolved: #3262

X-link: facebookresearch/FBGEMM#363

While CPU arguments shouldnt be used for custom cuda kernels, it turns out they sometimes are in production. The outputs will be garbage but doing so seems to be part of the model construction process. This small diff fixes the issue by adding CPU registrations for custom operators. This should enable production use cases without break torch.export support.

Reviewed By: jaconey, jianyuh, jiawenliu64

Differential Revision: D64703788

fbshipit-source-id: c0c8cfb7f0b67c13be10f419c8e3d83991429edb

Author: jwfromm

fbgemm_gpu/experimental/gen_ai/src/comm/car.cpp

@@ -278,4 +278,15 @@ TORCH_LIBRARY_IMPL(fbgemm, CUDA, m) {
   m.impl("two_shot_car_allreduce", two_shot_car_allreduce);
 }
 
+// Though it shouldnt be used, it is useful to define these functions for CPU to
+// accomodate model creation.
+TORCH_LIBRARY_IMPL(fbgemm, CPU, m) {
+  m.impl("nccl_allreduce", nccl_allreduce);
+  m.impl("nccl_allgather", nccl_allgather);
+  m.impl("nccl_alltoall", nccl_alltoall);
+  m.impl("nccl_reducescatter", nccl_reducescatter);
+  m.impl("one_shot_car_allreduce", one_shot_car_allreduce);
+  m.impl("two_shot_car_allreduce", two_shot_car_allreduce);
+}
+
 } // namespace fbgemm_gpu

fbgemm_gpu/experimental/gen_ai/src/quantize/quantize.cpp

@@ -214,4 +214,24 @@ TORCH_LIBRARY_IMPL(fbgemm, CUDA, m) {
 #endif
 }
 
+// Though it should never be used, it still seems helpful to define these
+// functions for CPU to accomodate model creation.
+TORCH_LIBRARY_IMPL(fbgemm, CPU, m) {
+  m.impl("f8f8bf16_blockwise", f8f8bf16_blockwise);
+  m.impl("f8f8bf16_tensorwise", f8f8bf16_tensorwise);
+  m.impl("f8f8bf16_rowwise", f8f8bf16_rowwise);
+  m.impl("quantize_fp8_per_tensor", quantize_fp8_per_tensor);
+  m.impl("quantize_fp8_per_row", quantize_fp8_per_row);
+  m.impl("quantize_fp8_per_col", quantize_fp8_per_col);
+#ifndef USE_ROCM
+  m.impl("i8i8bf16", i8i8bf16);
+  m.impl("f8f8bf16", f8f8bf16);
+  m.impl("f8f8bf16_cublas", f8f8bf16_cublas);
+  m.impl("f8f8bf16_rowwise_batched", f8f8bf16_rowwise_batched);
+  m.impl("f8i4bf16_rowwise", f8i4bf16_rowwise);
+  m.impl("bf16i4bf16_rowwise_batched", bf16i4bf16_rowwise_batched);
+  m.impl("bf16i4bf16_rowwise", bf16i4bf16_rowwise);
+#endif
+}
+
 } // namespace fbgemm_gpu