Build uploaded using `kernels`.

build/torch210-cxx11-cu126-aarch64-linux/__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/torch210-cxx11-cu126-aarch64-linux/_activation_cuda_5e1630d.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53cb1adecf7bb00650edb28c861b149c48729739ca1c2a6bae39fe52e22657bb
+size 3228128

build/torch210-cxx11-cu126-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _activation_cuda_5e1630d
+ops = torch.ops._activation_cuda_5e1630d
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_activation_cuda_5e1630d::{op_name}"

build/torch210-cxx11-cu126-aarch64-linux/activation/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+
+
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+
+
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch210-cxx11-cu126-aarch64-linux/layers.py

@@ -0,0 +1,201 @@
+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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        out = torch.empty_like(x)
+        ops.gelu_quick(out, x)
+        return out

build/torch210-cxx11-cu126-aarch64-linux/metadata.json

@@ -0,0 +1,18 @@
+{
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "7.0",
+      "7.2",
+      "7.5",
+      "8.0",
+      "8.6",
+      "8.7",
+      "8.9",
+      "9.0+PTX"
+    ]
+  }
+}

build/torch210-cxx11-cu128-aarch64-linux/__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/torch210-cxx11-cu128-aarch64-linux/_activation_cuda_5e1630d.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9eac4ae53a85546234a483c48c9bfcb717d16bb434b1f9723909fb838d366cb3
+size 4538960

build/torch210-cxx11-cu128-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _activation_cuda_5e1630d
+ops = torch.ops._activation_cuda_5e1630d
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_activation_cuda_5e1630d::{op_name}"

build/torch210-cxx11-cu128-aarch64-linux/activation/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+
+
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+
+
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch210-cxx11-cu128-aarch64-linux/layers.py

@@ -0,0 +1,201 @@
+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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        out = torch.empty_like(x)
+        ops.gelu_quick(out, x)
+        return out

build/torch210-cxx11-cu128-aarch64-linux/metadata.json

@@ -0,0 +1,21 @@
+{
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "10.1",
+      "12.0+PTX",
+      "7.0",
+      "7.2",
+      "7.5",
+      "8.0",
+      "8.6",
+      "8.7",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch210-cxx11-cu130-aarch64-linux/__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/torch210-cxx11-cu130-aarch64-linux/_activation_cuda_5e1630d.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1be21db6774214a4e1b58290f71e2e5cd0146af7d0646220ed4c0873d959b7e2
+size 4293520

build/torch210-cxx11-cu130-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _activation_cuda_5e1630d
+ops = torch.ops._activation_cuda_5e1630d
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_activation_cuda_5e1630d::{op_name}"

build/torch210-cxx11-cu130-aarch64-linux/activation/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+
+
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+
+
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch210-cxx11-cu130-aarch64-linux/layers.py

@@ -0,0 +1,201 @@
+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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        out = torch.empty_like(x)
+        ops.gelu_quick(out, x)
+        return out

build/torch210-cxx11-cu130-aarch64-linux/metadata.json

@@ -0,0 +1,19 @@
+{
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "11.0",
+      "12.0+PTX",
+      "7.5",
+      "8.0",
+      "8.6",
+      "8.7",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch211-cxx11-cu126-aarch64-linux/__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/torch211-cxx11-cu126-aarch64-linux/_activation_cuda_5e1630d.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7015eb787a6fbbc2142ae85a8e169de810f27e09650870845d317305fa668eda
+size 3224336

build/torch211-cxx11-cu126-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _activation_cuda_5e1630d
+ops = torch.ops._activation_cuda_5e1630d
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_activation_cuda_5e1630d::{op_name}"

build/torch211-cxx11-cu126-aarch64-linux/activation/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+
+
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+
+
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch211-cxx11-cu126-aarch64-linux/layers.py

@@ -0,0 +1,201 @@
+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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        out = torch.empty_like(x)
+        ops.gelu_quick(out, x)
+        return out

build/torch211-cxx11-cu126-aarch64-linux/metadata.json

@@ -0,0 +1,18 @@
+{
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "7.0",
+      "7.2",
+      "7.5",
+      "8.0",
+      "8.6",
+      "8.7",
+      "8.9",
+      "9.0+PTX"
+    ]
+  }
+}

build/torch211-cxx11-cu128-aarch64-linux/__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/torch211-cxx11-cu128-aarch64-linux/_activation_cuda_5e1630d.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:074f2ccbabf93779cef6f1c81167dbc6076a6787e059354128f4c84993d17b6b
+size 4535168

build/torch211-cxx11-cu128-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _activation_cuda_5e1630d
+ops = torch.ops._activation_cuda_5e1630d
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_activation_cuda_5e1630d::{op_name}"

build/torch211-cxx11-cu128-aarch64-linux/activation/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+
+
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+
+
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch211-cxx11-cu128-aarch64-linux/layers.py

@@ -0,0 +1,201 @@
+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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        out = torch.empty_like(x)
+        ops.gelu_quick(out, x)
+        return out

build/torch211-cxx11-cu128-aarch64-linux/metadata.json

@@ -0,0 +1,21 @@
+{
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "10.1",
+      "12.0+PTX",
+      "7.0",
+      "7.2",
+      "7.5",
+      "8.0",
+      "8.6",
+      "8.7",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch211-cxx11-cu130-aarch64-linux/__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/torch211-cxx11-cu130-aarch64-linux/_activation_cuda_5e1630d.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e9bedacf7611cdde0c229d46ec84b159c62154e4053cd21db492118d8ccddf8
+size 4289720

build/torch211-cxx11-cu130-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _activation_cuda_5e1630d
+ops = torch.ops._activation_cuda_5e1630d
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_activation_cuda_5e1630d::{op_name}"

build/torch211-cxx11-cu130-aarch64-linux/activation/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+
+
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+
+
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch211-cxx11-cu130-aarch64-linux/layers.py

@@ -0,0 +1,201 @@
+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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        out = torch.empty_like(x)
+        ops.gelu_quick(out, x)
+        return out

build/torch211-cxx11-cu130-aarch64-linux/metadata.json

@@ -0,0 +1,19 @@
+{
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "11.0",
+      "12.0+PTX",
+      "7.5",
+      "8.0",
+      "8.6",
+      "8.7",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch29-cxx11-cu129-aarch64-linux/__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-cu129-aarch64-linux/_activation_cuda_5e1630d.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:62c3d96bdf09677af2537ac4a9f4cf67a241c6bd4a2888771faaa9e16c0973f4
+size 4538112

build/torch29-cxx11-cu129-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _activation_cuda_5e1630d
+ops = torch.ops._activation_cuda_5e1630d
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_activation_cuda_5e1630d::{op_name}"

build/torch29-cxx11-cu129-aarch64-linux/activation/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+
+
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+
+
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch29-cxx11-cu129-aarch64-linux/layers.py

@@ -0,0 +1,201 @@
+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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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):
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        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:
+        if not x.is_contiguous():
+            x = x.contiguous()
+        out = torch.empty_like(x)
+        ops.gelu_quick(out, x)
+        return out

build/torch29-cxx11-cu129-aarch64-linux/metadata.json

@@ -0,0 +1,21 @@
+{
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "10.1",
+      "12.0+PTX",
+      "7.0",
+      "7.2",
+      "7.5",
+      "8.0",
+      "8.6",
+      "8.7",
+      "8.9",
+      "9.0"
+    ]
+  }
+}