Build uploaded using `kernels`.

build/torch27-cxx11-cu128-aarch64-linux/activation/__init__.py

@@ -0,0 +1,75 @@
+import torch
+
+from ._ops import ops
+
+from . import layers
+
+
+def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.silu_and_mul(out, x)
+    return out
+
+
+def mul_and_silu(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.mul_and_silu(out, x)
+    return out
+
+
+def gelu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_and_mul(out, x)
+    return out
+
+
+def gelu_tanh_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_tanh_and_mul(out, x)
+    return out
+
+
+def fatrelu_and_mul(out: torch.Tensor, x: torch.Tensor, threshold: float = 0.0) -> None:
+    ops.fatrelu_and_mul(out, x, threshold)
+    return out
+
+
+def gelu(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu(out, x)
+    return out
+
+def silu(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.silu(out, x)
+    return out
+
+
+def gelu_tanh(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_tanh(out, x)
+    return out
+
+
+def gelu_fast(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_fast(out, x)
+    return out
+
+
+def gelu_new(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_new(out, x)
+    return out
+
+
+def gelu_quick(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_quick(out, x)
+    return out
+
+
+__all__ = [
+    "silu_and_mul",
+    "mul_and_silu",
+    "gelu_and_mul",
+    "gelu_tanh_and_mul",
+    "fatrelu_and_mul",
+    "gelu_fast",
+    "gelu_new",
+    "gelu_quick",
+    "gelu_tanh",
+    "silu",
+    "gelu",
+    "layers",
+]

build/{torch28-cxx11-cu129-aarch64-linux/activation/_activation_0c3eb4e_dirty.abi3.so → torch27-cxx11-cu128-aarch64-linux/activation/_activation_320b408.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b9f6a161283a05672bb3ed442990cf8a5ce553625bb482cd31ce514e07cfcf0a
-size 3684504
+oid sha256:34bdeb9ab72686850aef0a16b225b1b956162edb2cf46cba65c5e5b92ae267ae
+size 4207000

build/torch27-cxx11-cu128-aarch64-linux/activation/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _activation_320b408
+ops = torch.ops._activation_320b408
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_activation_320b408::{op_name}"

build/torch27-cxx11-cu128-aarch64-linux/activation/layers.py

@@ -0,0 +1,179 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+
+class SiluAndMul(nn.Module):
+    """An activation function for SwiGLU.
+
+    The function computes x -> silu(x[:d]) * x[d:] where d = x.shape[-1] // 2.
+
+    Shapes:
+        x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.silu_and_mul(out, x)
+        return out
+
+class Silu(nn.Module):
+    """An activation function for SiLU.
+
+    The function computes x -> silu(x).
+
+    Shapes:
+        x: (num_tokens, d) or (batch_size, seq_len, d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        out = torch.empty_like(x)
+        ops.silu(out, x)
+        return out
+
+class Gelu(nn.Module):
+    """An activation function for GELU.
+
+    The function computes x -> gelu(x).
+
+    Shapes:
+        x: (num_tokens, d) or (batch_size, seq_len, d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        out = torch.empty_like(x)
+        ops.gelu(out, x)
+        return out
+
+class GeluTanh(nn.Module):
+    """An activation function for GELU with `tanh` approximation.
+
+    The function computes x -> gelu_tanh(x).
+
+    Shapes:
+        x: (num_tokens, d) or (batch_size, seq_len, d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        out = torch.empty_like(x)
+        ops.gelu_tanh(out, x)
+        return out
+
+
+class MulAndSilu(nn.Module):
+    """An activation function for SwiGLU.
+
+    The function computes x -> x[:d] * silu(x[d:]) where d = x.shape[-1] // 2.
+
+    Shapes:
+        x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.mul_and_silu(out, x)
+        return out
+
+
+class GeluAndMul(nn.Module):
+    """An activation function for GeGLU.
+
+    The function computes x -> GELU(x[:d]) * x[d:] where d = x.shape[-1] // 2.
+
+    Shapes:
+        x: (batch_size, seq_len, 2 * d) or (num_tokens, 2 * d)
+        return: (batch_size, seq_len, d) or (num_tokens, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.gelu_and_mul(out, x)
+        return out
+
+
+class GeluTanhAndMul(nn.Module):
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.gelu_tanh_and_mul(out, x)
+        return out
+
+
+class FatreluAndMul(nn.Module):
+    """An activation function for FATReLU.
+
+    The function computes x -> FATReLU(x[:d]) * x[d:] where
+    d = x.shape[-1] // 2.
+    This is used in openbmb/MiniCPM-S-1B-sft.
+
+    Shapes:
+        x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def __init__(self, threshold: float = 0.0):
+        super().__init__()
+        self.threshold = threshold
+
+    def forward(self, x: torch.Tensor):
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.fatrelu_and_mul(out, x, self.threshold)
+        return out
+
+
+class FastGELU(nn.Module):
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        out = torch.empty_like(x)
+        ops.gelu_fast(out, x)
+        return out
+
+
+class NewGELU(nn.Module):
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        out = torch.empty_like(x)
+        ops.gelu_new(out, x)
+        return out
+
+
+class QuickGELU(nn.Module):
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        out = torch.empty_like(x)
+        ops.gelu_quick(out, x)
+        return out

build/torch28-cxx11-cu129-aarch64-linux/activation/__init__.py

@@ -30,6 +30,20 @@ def fatrelu_and_mul(out: torch.Tensor, x: torch.Tensor, threshold: float = 0.0)
     return out
 
 
+def gelu(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu(out, x)
+    return out
+
+def silu(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.silu(out, x)
+    return out
+
+
+def gelu_tanh(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_tanh(out, x)
+    return out
+
+
 def gelu_fast(out: torch.Tensor, x: torch.Tensor) -> None:
     ops.gelu_fast(out, x)
     return out
@@ -47,11 +61,15 @@ def gelu_quick(out: torch.Tensor, x: torch.Tensor) -> None:
 
 __all__ = [
     "silu_and_mul",
+    "mul_and_silu",
     "gelu_and_mul",
     "gelu_tanh_and_mul",
     "fatrelu_and_mul",
     "gelu_fast",
     "gelu_new",
     "gelu_quick",
+    "gelu_tanh",
+    "silu",
+    "gelu",
     "layers",
 ]

build/torch28-cxx11-cu129-aarch64-linux/activation/_activation_320b408.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3af83bae80c8641200010ba586e5a2cac271fa4fcd344e3532ea7d5094fd7c17
+size 4275744

build/torch28-cxx11-cu129-aarch64-linux/activation/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _activation_0c3eb4e_dirty
-ops = torch.ops._activation_0c3eb4e_dirty
+from . import _activation_320b408
+ops = torch.ops._activation_320b408
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_activation_0c3eb4e_dirty::{op_name}"
+    return f"_activation_320b408::{op_name}"

build/torch28-cxx11-cu129-aarch64-linux/activation/layers.py

@@ -23,6 +23,57 @@ class SiluAndMul(nn.Module):
         ops.silu_and_mul(out, x)
         return out
 
+class Silu(nn.Module):
+    """An activation function for SiLU.
+
+    The function computes x -> silu(x).
+
+    Shapes:
+        x: (num_tokens, d) or (batch_size, seq_len, d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        out = torch.empty_like(x)
+        ops.silu(out, x)
+        return out
+
+class Gelu(nn.Module):
+    """An activation function for GELU.
+
+    The function computes x -> gelu(x).
+
+    Shapes:
+        x: (num_tokens, d) or (batch_size, seq_len, d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        out = torch.empty_like(x)
+        ops.gelu(out, x)
+        return out
+
+class GeluTanh(nn.Module):
+    """An activation function for GELU with `tanh` approximation.
+
+    The function computes x -> gelu_tanh(x).
+
+    Shapes:
+        x: (num_tokens, d) or (batch_size, seq_len, d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        out = torch.empty_like(x)
+        ops.gelu_tanh(out, x)
+        return out
+
 
 class MulAndSilu(nn.Module):
     """An activation function for SwiGLU.

build/torch29-cxx11-cu126-aarch64-linux/activation/__init__.py

@@ -0,0 +1,75 @@
+import torch
+
+from ._ops import ops
+
+from . import layers
+
+
+def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.silu_and_mul(out, x)
+    return out
+
+
+def mul_and_silu(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.mul_and_silu(out, x)
+    return out
+
+
+def gelu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_and_mul(out, x)
+    return out
+
+
+def gelu_tanh_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_tanh_and_mul(out, x)
+    return out
+
+
+def fatrelu_and_mul(out: torch.Tensor, x: torch.Tensor, threshold: float = 0.0) -> None:
+    ops.fatrelu_and_mul(out, x, threshold)
+    return out
+
+
+def gelu(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu(out, x)
+    return out
+
+def silu(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.silu(out, x)
+    return out
+
+
+def gelu_tanh(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_tanh(out, x)
+    return out
+
+
+def gelu_fast(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_fast(out, x)
+    return out
+
+
+def gelu_new(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_new(out, x)
+    return out
+
+
+def gelu_quick(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_quick(out, x)
+    return out
+
+
+__all__ = [
+    "silu_and_mul",
+    "mul_and_silu",
+    "gelu_and_mul",
+    "gelu_tanh_and_mul",
+    "fatrelu_and_mul",
+    "gelu_fast",
+    "gelu_new",
+    "gelu_quick",
+    "gelu_tanh",
+    "silu",
+    "gelu",
+    "layers",
+]

build/torch29-cxx11-cu126-aarch64-linux/activation/_activation_320b408.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9c24e0eb75a09a9fc19e7096276d560226f198617291681c1a18e94002a629e
+size 2963480

build/torch29-cxx11-cu126-aarch64-linux/activation/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _activation_320b408
+ops = torch.ops._activation_320b408
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_activation_320b408::{op_name}"

build/torch29-cxx11-cu126-aarch64-linux/activation/layers.py

@@ -0,0 +1,179 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+
+class SiluAndMul(nn.Module):
+    """An activation function for SwiGLU.
+
+    The function computes x -> silu(x[:d]) * x[d:] where d = x.shape[-1] // 2.
+
+    Shapes:
+        x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.silu_and_mul(out, x)
+        return out
+
+class Silu(nn.Module):
+    """An activation function for SiLU.
+
+    The function computes x -> silu(x).
+
+    Shapes:
+        x: (num_tokens, d) or (batch_size, seq_len, d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        out = torch.empty_like(x)
+        ops.silu(out, x)
+        return out
+
+class Gelu(nn.Module):
+    """An activation function for GELU.
+
+    The function computes x -> gelu(x).
+
+    Shapes:
+        x: (num_tokens, d) or (batch_size, seq_len, d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        out = torch.empty_like(x)
+        ops.gelu(out, x)
+        return out
+
+class GeluTanh(nn.Module):
+    """An activation function for GELU with `tanh` approximation.
+
+    The function computes x -> gelu_tanh(x).
+
+    Shapes:
+        x: (num_tokens, d) or (batch_size, seq_len, d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        out = torch.empty_like(x)
+        ops.gelu_tanh(out, x)
+        return out
+
+
+class MulAndSilu(nn.Module):
+    """An activation function for SwiGLU.
+
+    The function computes x -> x[:d] * silu(x[d:]) where d = x.shape[-1] // 2.
+
+    Shapes:
+        x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.mul_and_silu(out, x)
+        return out
+
+
+class GeluAndMul(nn.Module):
+    """An activation function for GeGLU.
+
+    The function computes x -> GELU(x[:d]) * x[d:] where d = x.shape[-1] // 2.
+
+    Shapes:
+        x: (batch_size, seq_len, 2 * d) or (num_tokens, 2 * d)
+        return: (batch_size, seq_len, d) or (num_tokens, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.gelu_and_mul(out, x)
+        return out
+
+
+class GeluTanhAndMul(nn.Module):
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.gelu_tanh_and_mul(out, x)
+        return out
+
+
+class FatreluAndMul(nn.Module):
+    """An activation function for FATReLU.
+
+    The function computes x -> FATReLU(x[:d]) * x[d:] where
+    d = x.shape[-1] // 2.
+    This is used in openbmb/MiniCPM-S-1B-sft.
+
+    Shapes:
+        x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def __init__(self, threshold: float = 0.0):
+        super().__init__()
+        self.threshold = threshold
+
+    def forward(self, x: torch.Tensor):
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.fatrelu_and_mul(out, x, self.threshold)
+        return out
+
+
+class FastGELU(nn.Module):
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        out = torch.empty_like(x)
+        ops.gelu_fast(out, x)
+        return out
+
+
+class NewGELU(nn.Module):
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        out = torch.empty_like(x)
+        ops.gelu_new(out, x)
+        return out
+
+
+class QuickGELU(nn.Module):
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        out = torch.empty_like(x)
+        ops.gelu_quick(out, x)
+        return out

build/torch29-cxx11-cu128-aarch64-linux/activation/__init__.py

@@ -0,0 +1,75 @@
+import torch
+
+from ._ops import ops
+
+from . import layers
+
+
+def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.silu_and_mul(out, x)
+    return out
+
+
+def mul_and_silu(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.mul_and_silu(out, x)
+    return out
+
+
+def gelu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_and_mul(out, x)
+    return out
+
+
+def gelu_tanh_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_tanh_and_mul(out, x)
+    return out
+
+
+def fatrelu_and_mul(out: torch.Tensor, x: torch.Tensor, threshold: float = 0.0) -> None:
+    ops.fatrelu_and_mul(out, x, threshold)
+    return out
+
+
+def gelu(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu(out, x)
+    return out
+
+def silu(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.silu(out, x)
+    return out
+
+
+def gelu_tanh(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_tanh(out, x)
+    return out
+
+
+def gelu_fast(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_fast(out, x)
+    return out
+
+
+def gelu_new(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_new(out, x)
+    return out
+
+
+def gelu_quick(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_quick(out, x)
+    return out
+
+
+__all__ = [
+    "silu_and_mul",
+    "mul_and_silu",
+    "gelu_and_mul",
+    "gelu_tanh_and_mul",
+    "fatrelu_and_mul",
+    "gelu_fast",
+    "gelu_new",
+    "gelu_quick",
+    "gelu_tanh",
+    "silu",
+    "gelu",
+    "layers",
+]

build/torch29-cxx11-cu128-aarch64-linux/activation/_activation_320b408.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08ee3dfa4d481eaf44ac3c11a0843598c05950f779dba66abd468fecb7839b32
+size 4208760

build/torch29-cxx11-cu128-aarch64-linux/activation/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _activation_320b408
+ops = torch.ops._activation_320b408
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_activation_320b408::{op_name}"

build/torch29-cxx11-cu128-aarch64-linux/activation/layers.py

@@ -0,0 +1,179 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+
+class SiluAndMul(nn.Module):
+    """An activation function for SwiGLU.
+
+    The function computes x -> silu(x[:d]) * x[d:] where d = x.shape[-1] // 2.
+
+    Shapes:
+        x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.silu_and_mul(out, x)
+        return out
+
+class Silu(nn.Module):
+    """An activation function for SiLU.
+
+    The function computes x -> silu(x).
+
+    Shapes:
+        x: (num_tokens, d) or (batch_size, seq_len, d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        out = torch.empty_like(x)
+        ops.silu(out, x)
+        return out
+
+class Gelu(nn.Module):
+    """An activation function for GELU.
+
+    The function computes x -> gelu(x).
+
+    Shapes:
+        x: (num_tokens, d) or (batch_size, seq_len, d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        out = torch.empty_like(x)
+        ops.gelu(out, x)
+        return out
+
+class GeluTanh(nn.Module):
+    """An activation function for GELU with `tanh` approximation.
+
+    The function computes x -> gelu_tanh(x).
+
+    Shapes:
+        x: (num_tokens, d) or (batch_size, seq_len, d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        out = torch.empty_like(x)
+        ops.gelu_tanh(out, x)
+        return out
+
+
+class MulAndSilu(nn.Module):
+    """An activation function for SwiGLU.
+
+    The function computes x -> x[:d] * silu(x[d:]) where d = x.shape[-1] // 2.
+
+    Shapes:
+        x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.mul_and_silu(out, x)
+        return out
+
+
+class GeluAndMul(nn.Module):
+    """An activation function for GeGLU.
+
+    The function computes x -> GELU(x[:d]) * x[d:] where d = x.shape[-1] // 2.
+
+    Shapes:
+        x: (batch_size, seq_len, 2 * d) or (num_tokens, 2 * d)
+        return: (batch_size, seq_len, d) or (num_tokens, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.gelu_and_mul(out, x)
+        return out
+
+
+class GeluTanhAndMul(nn.Module):
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.gelu_tanh_and_mul(out, x)
+        return out
+
+
+class FatreluAndMul(nn.Module):
+    """An activation function for FATReLU.
+
+    The function computes x -> FATReLU(x[:d]) * x[d:] where
+    d = x.shape[-1] // 2.
+    This is used in openbmb/MiniCPM-S-1B-sft.
+
+    Shapes:
+        x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def __init__(self, threshold: float = 0.0):
+        super().__init__()
+        self.threshold = threshold
+
+    def forward(self, x: torch.Tensor):
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.fatrelu_and_mul(out, x, self.threshold)
+        return out
+
+
+class FastGELU(nn.Module):
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        out = torch.empty_like(x)
+        ops.gelu_fast(out, x)
+        return out
+
+
+class NewGELU(nn.Module):
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        out = torch.empty_like(x)
+        ops.gelu_new(out, x)
+        return out
+
+
+class QuickGELU(nn.Module):
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        out = torch.empty_like(x)
+        ops.gelu_quick(out, x)
+        return out

build/torch29-cxx11-cu130-aarch64-linux/activation/__init__.py

@@ -0,0 +1,75 @@
+import torch
+
+from ._ops import ops
+
+from . import layers
+
+
+def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.silu_and_mul(out, x)
+    return out
+
+
+def mul_and_silu(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.mul_and_silu(out, x)
+    return out
+
+
+def gelu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_and_mul(out, x)
+    return out
+
+
+def gelu_tanh_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_tanh_and_mul(out, x)
+    return out
+
+
+def fatrelu_and_mul(out: torch.Tensor, x: torch.Tensor, threshold: float = 0.0) -> None:
+    ops.fatrelu_and_mul(out, x, threshold)
+    return out
+
+
+def gelu(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu(out, x)
+    return out
+
+def silu(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.silu(out, x)
+    return out
+
+
+def gelu_tanh(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_tanh(out, x)
+    return out
+
+
+def gelu_fast(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_fast(out, x)
+    return out
+
+
+def gelu_new(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_new(out, x)
+    return out
+
+
+def gelu_quick(out: torch.Tensor, x: torch.Tensor) -> None:
+    ops.gelu_quick(out, x)
+    return out
+
+
+__all__ = [
+    "silu_and_mul",
+    "mul_and_silu",
+    "gelu_and_mul",
+    "gelu_tanh_and_mul",
+    "fatrelu_and_mul",
+    "gelu_fast",
+    "gelu_new",
+    "gelu_quick",
+    "gelu_tanh",
+    "silu",
+    "gelu",
+    "layers",
+]

build/torch29-cxx11-cu130-aarch64-linux/activation/_activation_320b408.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:73748b54059552f5983322f7dedc36ed349b38ad6fb9318301bb4965b1fe49aa
+size 4094968

build/torch29-cxx11-cu130-aarch64-linux/activation/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _activation_320b408
+ops = torch.ops._activation_320b408
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_activation_320b408::{op_name}"

build/torch29-cxx11-cu130-aarch64-linux/activation/layers.py

@@ -0,0 +1,179 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+
+class SiluAndMul(nn.Module):
+    """An activation function for SwiGLU.
+
+    The function computes x -> silu(x[:d]) * x[d:] where d = x.shape[-1] // 2.
+
+    Shapes:
+        x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.silu_and_mul(out, x)
+        return out
+
+class Silu(nn.Module):
+    """An activation function for SiLU.
+
+    The function computes x -> silu(x).
+
+    Shapes:
+        x: (num_tokens, d) or (batch_size, seq_len, d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        out = torch.empty_like(x)
+        ops.silu(out, x)
+        return out
+
+class Gelu(nn.Module):
+    """An activation function for GELU.
+
+    The function computes x -> gelu(x).
+
+    Shapes:
+        x: (num_tokens, d) or (batch_size, seq_len, d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        out = torch.empty_like(x)
+        ops.gelu(out, x)
+        return out
+
+class GeluTanh(nn.Module):
+    """An activation function for GELU with `tanh` approximation.
+
+    The function computes x -> gelu_tanh(x).
+
+    Shapes:
+        x: (num_tokens, d) or (batch_size, seq_len, d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        out = torch.empty_like(x)
+        ops.gelu_tanh(out, x)
+        return out
+
+
+class MulAndSilu(nn.Module):
+    """An activation function for SwiGLU.
+
+    The function computes x -> x[:d] * silu(x[d:]) where d = x.shape[-1] // 2.
+
+    Shapes:
+        x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.mul_and_silu(out, x)
+        return out
+
+
+class GeluAndMul(nn.Module):
+    """An activation function for GeGLU.
+
+    The function computes x -> GELU(x[:d]) * x[d:] where d = x.shape[-1] // 2.
+
+    Shapes:
+        x: (batch_size, seq_len, 2 * d) or (num_tokens, 2 * d)
+        return: (batch_size, seq_len, d) or (num_tokens, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.gelu_and_mul(out, x)
+        return out
+
+
+class GeluTanhAndMul(nn.Module):
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor):
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.gelu_tanh_and_mul(out, x)
+        return out
+
+
+class FatreluAndMul(nn.Module):
+    """An activation function for FATReLU.
+
+    The function computes x -> FATReLU(x[:d]) * x[d:] where
+    d = x.shape[-1] // 2.
+    This is used in openbmb/MiniCPM-S-1B-sft.
+
+    Shapes:
+        x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
+        return: (num_tokens, d) or (batch_size, seq_len, d)
+    """
+
+    can_torch_compile: bool = True
+
+    def __init__(self, threshold: float = 0.0):
+        super().__init__()
+        self.threshold = threshold
+
+    def forward(self, x: torch.Tensor):
+        d = x.shape[-1] // 2
+        output_shape = x.shape[:-1] + (d,)
+        out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
+        ops.fatrelu_and_mul(out, x, self.threshold)
+        return out
+
+
+class FastGELU(nn.Module):
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        out = torch.empty_like(x)
+        ops.gelu_fast(out, x)
+        return out
+
+
+class NewGELU(nn.Module):
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        out = torch.empty_like(x)
+        ops.gelu_new(out, x)
+        return out
+
+
+class QuickGELU(nn.Module):
+    can_torch_compile: bool = True
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        out = torch.empty_like(x)
+        ops.gelu_quick(out, x)
+        return out