Merge pull request #579 from InfiniTensor/issue/578

issue/578: 添加 `empty_like` 并在 `silu` 的参数为 `None` 时默认调用 `ntops` 的 `silu`

.github/workflows/ruff.yml

+name: Ruff
+on: [push, pull_request]
+jobs:
+  ruff:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: chartboost/ruff-action@v1
+        with:
+          src: './python/'
+      - uses: chartboost/ruff-action@v1
+        with:
+          src: './python/'
+          args: format --check

python/infinicore/__init__.py

 import contextlib
 
+import infinicore.nn as nn
 from infinicore.device import device
 from infinicore.dtype import (
     bfloat16,
@@ -33,6 +34,7 @@ from infinicore.ops.rearrange import rearrange
 from infinicore.tensor import (
     Tensor,
     empty,
+    empty_like,
     from_blob,
     ones,
     strided_empty,
@@ -40,12 +42,13 @@ from infinicore.tensor import (
     zeros,
 )
 
-from infinicore import nn as nn
-
 __all__ = [
+    # Modules.
+    "nn",
     # Classes.
     "device",
     "dtype",
+    "Tensor",
     # Data Types.
     "bfloat16",
     "bool",
@@ -75,6 +78,7 @@ __all__ = [
     "matmul",
     "rearrange",
     "empty",
+    "empty_like",
     "from_blob",
     "ones",
     "strided_empty",

python/infinicore/nn/functional.py

 import infinicore
 from infinicore.lib import _infinicore
+from infinicore.tensor import Tensor
 
 __all__ = ["causal_softmax", "rms_norm", "silu", "swiglu"]
 
 
-def causal_softmax(
-        input: infinicore.Tensor,
-        out=None
-) -> infinicore.Tensor:
-    r"""Apply a causal softmax function.
-    """
+def causal_softmax(input: Tensor, out=None) -> Tensor:
+    r"""Apply a causal softmax function."""
 
     if out is None:
-        return infinicore.Tensor(_infinicore.causal_softmax(input._underlying))
+        return Tensor(_infinicore.causal_softmax(input._underlying))
 
     _infinicore.causal_softmax_(out._underlying, input._underlying)
 
@@ -20,49 +17,50 @@ def causal_softmax(
 
 
 def rms_norm(
-        input: infinicore.Tensor,
+    input: Tensor,
     normalized_shape: list[int],
-        weight: infinicore.Tensor,
+    weight: Tensor,
     eps: float = 1e-5,
-        out=None
-) -> infinicore.Tensor:
-    r"""Apply Root Mean Square Layer Normalization.
-    """
+    *,
+    out=None,
+) -> Tensor:
+    r"""Apply Root Mean Square Layer Normalization."""
 
-    assert normalized_shape == weight.shape, "normalized_shape does not match weight.shape."
+    assert normalized_shape == weight.shape, (
+        "normalized_shape does not match weight.shape."
+    )
 
     if out is None:
-        return infinicore.Tensor(
-            _infinicore.rms_norm(input._underlying, weight._underlying, eps)
-        )
+        return Tensor(_infinicore.rms_norm(input._underlying, weight._underlying, eps))
 
     _infinicore.rms_norm_(out._underlying, input._underlying, weight._underlying, eps)
 
     return out
 
 
-def silu(input: infinicore.Tensor, inplace: bool = False, out=None) -> infinicore.Tensor:
-    r"""Apply the Sigmoid Linear Unit (SiLU) function, element-wise.
-    """
+def silu(input: Tensor, inplace: bool = False, *, out=None) -> Tensor:
+    r"""Apply the Sigmoid Linear Unit (SiLU) function, element-wise."""
+
+    if infinicore.use_ntops and input.device.type in ("cuda", "musa") and out is None:
+        return infinicore.ntops.torch.silu(input, inplace=inplace)
 
     if inplace:
         _infinicore.silu_(input._underlying, input._underlying)
         return input
 
     if out is None:
-        return infinicore.Tensor(_infinicore.silu(input._underlying))
+        return Tensor(_infinicore.silu(input._underlying))
 
     _infinicore.silu_(out._underlying, input._underlying)
 
     return out
 
 
-def swiglu(input: infinicore.Tensor, other: infinicore.Tensor, out=None):
-    r"""Apply the Swish-Gated Linear Unit (SwiGLU) function, element-wise.
-    """
+def swiglu(input: Tensor, other: Tensor, *, out=None):
+    r"""Apply the Swish-Gated Linear Unit (SwiGLU) function, element-wise."""
 
     if out is None:
-        return infinicore.Tensor(_infinicore.swiglu(input._underlying, other._underlying))
+        return Tensor(_infinicore.swiglu(input._underlying, other._underlying))
 
     _infinicore.swiglu_(out._underlying, input._underlying, other._underlying)
 

python/infinicore/tensor.py

@@ -93,6 +93,16 @@ def empty(size, *, dtype=None, device=None, pin_memory=False):
     )
 
 
+def empty_like(input, *, dtype=None, device=None):
+    if dtype is None:
+        dtype = input.dtype
+
+    if device is None:
+        device = input.device
+
+    return empty(input.size(), dtype=dtype, device=device)
+
+
 def strided_empty(size, strides, *, dtype=None, device=None, pin_memory=False):
     return Tensor(
         _infinicore.strided_empty(