Fix XPU and ROCm issues for Torch 2.13

AGENTS.md

@@ -1,3 +1,64 @@
+# General instructions that apply to all kernels
+
+## Imports
+
+All imports in a kernel (files inside `torch-ext/<kernel_name>`) of other
+modules within the same kernel must be relative. For example:
+
+```python
+# Incorrect:
+from activation.activations import silu_and_mul
+
+# Correct:
+from .activations import silu_and_mul
+```
+
+## `_ops` module
+
+The `_ops` module (`_ops.py`) is generated by the build system in
+`torch-ext/<kernel_name>`. This module contains the following:
+
+- An `ops` variable that is assigned the Torch ops for the kernel
+  (`ops = torch.ops.<kernel_ops_name>`).
+- A function `def add_op_namespace_prefix(op_name: str) -> str`,
+  which returns `f"<kernel_ops_name>::{op_name}"`.
+
+## Registering Torch operators
+
+All Torch operators that are registered must have a prefix that is unique
+to the kernel and is generated by the build system. Under no circumstances,
+the prefix should be fixed. The aforementioned `add_op_namespace_prefix`
+is used to prepend this prefix. For instance, consider this registration
+with `custom_op`:
+
+```python
+# Incorrect:
+@torch.library.custom_op("scattermoe::scatter2scatter", mutates_args={"output"})
+
+# Correct:
+@torch.library.custom_op(add_op_namespace_prefix("scatter2scatter"), mutates_args={"output"})
+```
+
+Similarly, an op must always be prefixed using the kernel-unique prefix:
+
+```python
+# Incorrect:
+@torch.library.custom_op("_flash_attn_forward", mutates_args=(), device_types="cuda")
+
+# Correct:
+@torch.library.custom_op(add_op_namespace_prefix("_flash_attn_forward"), mutates_args=(), device_types="cuda")
+```
+
+The same applies for other registrations of ops, such as fake ops:
+
+```python
+# Incorrect:
+@register_fake("moe::single_marlin_gemm_moe")
+
+# Correct:
+@register_fake(add_op_namespace_prefix("single_marlin_gemm_moe"))
+```
+
 # Kernel-specific instructions
 
 ## flash-attn3