liangsu9988 commited on 19 days ago

Commit

16ecd37

verified ·

1 Parent(s): aab237a

Uploaded using `kernel-builder`.

Browse files

Files changed (20) hide show

build/torch211-cxx11-cu128-x86_64-linux/__init__.py +99 -0
build/torch211-cxx11-cu128-x86_64-linux/_grouped_moe_gemv_cuda_1683349.abi3.so +3 -0
build/torch211-cxx11-cu128-x86_64-linux/_ops.py +9 -0
build/torch211-cxx11-cu128-x86_64-linux/grouped_moe_gemv/__init__.py +26 -0
build/torch211-cxx11-cu128-x86_64-linux/metadata.json +13 -0
build/torch211-cxx11-cu130-x86_64-linux/__init__.py +99 -0
build/torch211-cxx11-cu130-x86_64-linux/_grouped_moe_gemv_cuda_1683349.abi3.so +3 -0
build/torch211-cxx11-cu130-x86_64-linux/_ops.py +9 -0
build/torch211-cxx11-cu130-x86_64-linux/grouped_moe_gemv/__init__.py +26 -0
build/torch211-cxx11-cu130-x86_64-linux/metadata.json +13 -0
build/torch212-cxx11-cu130-x86_64-linux/__init__.py +99 -0
build/torch212-cxx11-cu130-x86_64-linux/_grouped_moe_gemv_cuda_1683349.abi3.so +3 -0
build/torch212-cxx11-cu130-x86_64-linux/_ops.py +9 -0
build/torch212-cxx11-cu130-x86_64-linux/grouped_moe_gemv/__init__.py +26 -0
build/torch212-cxx11-cu130-x86_64-linux/metadata.json +13 -0
build/torch212-cxx11-cu132-x86_64-linux/__init__.py +99 -0
build/torch212-cxx11-cu132-x86_64-linux/_grouped_moe_gemv_cuda_1683349.abi3.so +3 -0
build/torch212-cxx11-cu132-x86_64-linux/_ops.py +9 -0
build/torch212-cxx11-cu132-x86_64-linux/grouped_moe_gemv/__init__.py +26 -0
build/torch212-cxx11-cu132-x86_64-linux/metadata.json +13 -0

build/torch211-cxx11-cu128-x86_64-linux/__init__.py ADDED Viewed

	@@ -0,0 +1,99 @@

+"""FlashRT grouped MoE GEMV kernels."""
+from __future__ import annotations
+from typing import Optional
+import torch
+from ._ops import add_op_namespace_prefix, ops
+@torch.library.register_fake(add_op_namespace_prefix("w4a16_decode_gemv_bf16"))
+def _w4a16_decode_gemv_fake(
+    x_bf16: torch.Tensor,
+    weight_packed: torch.Tensor,
+    sfb: torch.Tensor,
+    alpha: float,
+    out: torch.Tensor,
+) -> None:
+    k = x_bf16.shape[0] if x_bf16.dim() == 1 else x_bf16.shape[1]
+    if weight_packed.dim() != 2 or weight_packed.shape[1] != k // 2 or out.shape != (weight_packed.shape[0],):
+        raise RuntimeError("expected x (K,) or (1,K), weight_packed (N,K/2), out (N,)")
+    return None
+@torch.library.register_fake(add_op_namespace_prefix("grouped_w4a16_gemv_bf16"))
+def _grouped_w4a16_gemv_fake(
+    activations: torch.Tensor,
+    weight_stack: torch.Tensor,
+    sfb_stack: torch.Tensor,
+    alpha_stack: torch.Tensor,
+    expert_idx: torch.Tensor,
+    w_stride: int,
+    sfb_stride: int,
+    out: torch.Tensor,
+) -> None:
+    if activations.dim() != 2 or out.dim() != 2 or out.shape[0] != activations.shape[0]:
+        raise RuntimeError("expected activations (slots,K), out (slots,N)")
+    if expert_idx.shape != (activations.shape[0],):
+        raise RuntimeError("expert_idx must have shape (slots,)")
+    return None
+def w4a16_decode_gemv_bf16(
+    x_bf16: torch.Tensor,
+    weight_packed: torch.Tensor,
+    sfb: torch.Tensor,
+    *,
+    alpha: float = 1.0,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    if out is None:
+        out = torch.empty((weight_packed.shape[0],), device=x_bf16.device, dtype=torch.bfloat16)
+    ops.w4a16_decode_gemv_bf16(x_bf16, weight_packed, sfb, float(alpha), out)
+    return out
+def grouped_w4a16_gemv_bf16(
+    activations: torch.Tensor,
+    weight_stack: torch.Tensor,
+    sfb_stack: torch.Tensor,
+    alpha_stack: torch.Tensor,
+    expert_idx: torch.Tensor,
+    *,
+    n: int,
+    w_stride: Optional[int] = None,
+    sfb_stride: Optional[int] = None,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """Run one W4A16 GEMV per routed slot.
+    `weight_stack` is a flat expert stack. `w_stride` and `sfb_stride` are byte
+    strides between experts; by default `w_stride = n * K / 2`.
+    """
+    k = activations.shape[1]
+    if out is None:
+        out = torch.empty((activations.shape[0], int(n)), device=activations.device, dtype=torch.bfloat16)
+    if w_stride is None:
+        w_stride = int(n) * k // 2
+    if sfb_stride is None:
+        raise RuntimeError("sfb_stride must be provided because swizzled SF size is layout-dependent")
+    ops.grouped_w4a16_gemv_bf16(
+        activations,
+        weight_stack,
+        sfb_stack,
+        alpha_stack,
+        expert_idx,
+        int(w_stride),
+        int(sfb_stride),
+        out,
+    )
+    return out
+__all__ = [
+    "grouped_w4a16_gemv_bf16",
+    "w4a16_decode_gemv_bf16",
+]

build/torch211-cxx11-cu128-x86_64-linux/_grouped_moe_gemv_cuda_1683349.abi3.so ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:e1393d4970efc36091e142b61f250e978caf3744de67fd93e872867c6dc1595f
+size 173904

build/torch211-cxx11-cu128-x86_64-linux/_ops.py ADDED Viewed

	@@ -0,0 +1,9 @@

+import torch
+from . import _grouped_moe_gemv_cuda_1683349
+ops = torch.ops._grouped_moe_gemv_cuda_1683349
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_grouped_moe_gemv_cuda_1683349::{op_name}"

build/torch211-cxx11-cu128-x86_64-linux/grouped_moe_gemv/__init__.py ADDED Viewed

	@@ -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-x86_64-linux/metadata.json ADDED Viewed

	@@ -0,0 +1,13 @@

+{
+  "name": "grouped-moe-gemv",
+  "id": "_grouped_moe_gemv_cuda_1683349",
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "12.0a"
+    ]
+  }
+}

build/torch211-cxx11-cu130-x86_64-linux/__init__.py ADDED Viewed

	@@ -0,0 +1,99 @@

+"""FlashRT grouped MoE GEMV kernels."""
+from __future__ import annotations
+from typing import Optional
+import torch
+from ._ops import add_op_namespace_prefix, ops
+@torch.library.register_fake(add_op_namespace_prefix("w4a16_decode_gemv_bf16"))
+def _w4a16_decode_gemv_fake(
+    x_bf16: torch.Tensor,
+    weight_packed: torch.Tensor,
+    sfb: torch.Tensor,
+    alpha: float,
+    out: torch.Tensor,
+) -> None:
+    k = x_bf16.shape[0] if x_bf16.dim() == 1 else x_bf16.shape[1]
+    if weight_packed.dim() != 2 or weight_packed.shape[1] != k // 2 or out.shape != (weight_packed.shape[0],):
+        raise RuntimeError("expected x (K,) or (1,K), weight_packed (N,K/2), out (N,)")
+    return None
+@torch.library.register_fake(add_op_namespace_prefix("grouped_w4a16_gemv_bf16"))
+def _grouped_w4a16_gemv_fake(
+    activations: torch.Tensor,
+    weight_stack: torch.Tensor,
+    sfb_stack: torch.Tensor,
+    alpha_stack: torch.Tensor,
+    expert_idx: torch.Tensor,
+    w_stride: int,
+    sfb_stride: int,
+    out: torch.Tensor,
+) -> None:
+    if activations.dim() != 2 or out.dim() != 2 or out.shape[0] != activations.shape[0]:
+        raise RuntimeError("expected activations (slots,K), out (slots,N)")
+    if expert_idx.shape != (activations.shape[0],):
+        raise RuntimeError("expert_idx must have shape (slots,)")
+    return None
+def w4a16_decode_gemv_bf16(
+    x_bf16: torch.Tensor,
+    weight_packed: torch.Tensor,
+    sfb: torch.Tensor,
+    *,
+    alpha: float = 1.0,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    if out is None:
+        out = torch.empty((weight_packed.shape[0],), device=x_bf16.device, dtype=torch.bfloat16)
+    ops.w4a16_decode_gemv_bf16(x_bf16, weight_packed, sfb, float(alpha), out)
+    return out
+def grouped_w4a16_gemv_bf16(
+    activations: torch.Tensor,
+    weight_stack: torch.Tensor,
+    sfb_stack: torch.Tensor,
+    alpha_stack: torch.Tensor,
+    expert_idx: torch.Tensor,
+    *,
+    n: int,
+    w_stride: Optional[int] = None,
+    sfb_stride: Optional[int] = None,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """Run one W4A16 GEMV per routed slot.
+    `weight_stack` is a flat expert stack. `w_stride` and `sfb_stride` are byte
+    strides between experts; by default `w_stride = n * K / 2`.
+    """
+    k = activations.shape[1]
+    if out is None:
+        out = torch.empty((activations.shape[0], int(n)), device=activations.device, dtype=torch.bfloat16)
+    if w_stride is None:
+        w_stride = int(n) * k // 2
+    if sfb_stride is None:
+        raise RuntimeError("sfb_stride must be provided because swizzled SF size is layout-dependent")
+    ops.grouped_w4a16_gemv_bf16(
+        activations,
+        weight_stack,
+        sfb_stack,
+        alpha_stack,
+        expert_idx,
+        int(w_stride),
+        int(sfb_stride),
+        out,
+    )
+    return out
+__all__ = [
+    "grouped_w4a16_gemv_bf16",
+    "w4a16_decode_gemv_bf16",
+]

build/torch211-cxx11-cu130-x86_64-linux/_grouped_moe_gemv_cuda_1683349.abi3.so ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:d4d5f5b0bb26387a375b49bd17940ec6650e627d34434a448445b68f02cbc870
+size 163912

build/torch211-cxx11-cu130-x86_64-linux/_ops.py ADDED Viewed

	@@ -0,0 +1,9 @@

+import torch
+from . import _grouped_moe_gemv_cuda_1683349
+ops = torch.ops._grouped_moe_gemv_cuda_1683349
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_grouped_moe_gemv_cuda_1683349::{op_name}"

build/torch211-cxx11-cu130-x86_64-linux/grouped_moe_gemv/__init__.py ADDED Viewed

	@@ -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-x86_64-linux/metadata.json ADDED Viewed

	@@ -0,0 +1,13 @@

+{
+  "name": "grouped-moe-gemv",
+  "id": "_grouped_moe_gemv_cuda_1683349",
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "12.0a"
+    ]
+  }
+}

build/torch212-cxx11-cu130-x86_64-linux/__init__.py ADDED Viewed

	@@ -0,0 +1,99 @@

+"""FlashRT grouped MoE GEMV kernels."""
+from __future__ import annotations
+from typing import Optional
+import torch
+from ._ops import add_op_namespace_prefix, ops
+@torch.library.register_fake(add_op_namespace_prefix("w4a16_decode_gemv_bf16"))
+def _w4a16_decode_gemv_fake(
+    x_bf16: torch.Tensor,
+    weight_packed: torch.Tensor,
+    sfb: torch.Tensor,
+    alpha: float,
+    out: torch.Tensor,
+) -> None:
+    k = x_bf16.shape[0] if x_bf16.dim() == 1 else x_bf16.shape[1]
+    if weight_packed.dim() != 2 or weight_packed.shape[1] != k // 2 or out.shape != (weight_packed.shape[0],):
+        raise RuntimeError("expected x (K,) or (1,K), weight_packed (N,K/2), out (N,)")
+    return None
+@torch.library.register_fake(add_op_namespace_prefix("grouped_w4a16_gemv_bf16"))
+def _grouped_w4a16_gemv_fake(
+    activations: torch.Tensor,
+    weight_stack: torch.Tensor,
+    sfb_stack: torch.Tensor,
+    alpha_stack: torch.Tensor,
+    expert_idx: torch.Tensor,
+    w_stride: int,
+    sfb_stride: int,
+    out: torch.Tensor,
+) -> None:
+    if activations.dim() != 2 or out.dim() != 2 or out.shape[0] != activations.shape[0]:
+        raise RuntimeError("expected activations (slots,K), out (slots,N)")
+    if expert_idx.shape != (activations.shape[0],):
+        raise RuntimeError("expert_idx must have shape (slots,)")
+    return None
+def w4a16_decode_gemv_bf16(
+    x_bf16: torch.Tensor,
+    weight_packed: torch.Tensor,
+    sfb: torch.Tensor,
+    *,
+    alpha: float = 1.0,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    if out is None:
+        out = torch.empty((weight_packed.shape[0],), device=x_bf16.device, dtype=torch.bfloat16)
+    ops.w4a16_decode_gemv_bf16(x_bf16, weight_packed, sfb, float(alpha), out)
+    return out
+def grouped_w4a16_gemv_bf16(
+    activations: torch.Tensor,
+    weight_stack: torch.Tensor,
+    sfb_stack: torch.Tensor,
+    alpha_stack: torch.Tensor,
+    expert_idx: torch.Tensor,
+    *,
+    n: int,
+    w_stride: Optional[int] = None,
+    sfb_stride: Optional[int] = None,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """Run one W4A16 GEMV per routed slot.
+    `weight_stack` is a flat expert stack. `w_stride` and `sfb_stride` are byte
+    strides between experts; by default `w_stride = n * K / 2`.
+    """
+    k = activations.shape[1]
+    if out is None:
+        out = torch.empty((activations.shape[0], int(n)), device=activations.device, dtype=torch.bfloat16)
+    if w_stride is None:
+        w_stride = int(n) * k // 2
+    if sfb_stride is None:
+        raise RuntimeError("sfb_stride must be provided because swizzled SF size is layout-dependent")
+    ops.grouped_w4a16_gemv_bf16(
+        activations,
+        weight_stack,
+        sfb_stack,
+        alpha_stack,
+        expert_idx,
+        int(w_stride),
+        int(sfb_stride),
+        out,
+    )
+    return out
+__all__ = [
+    "grouped_w4a16_gemv_bf16",
+    "w4a16_decode_gemv_bf16",
+]

build/torch212-cxx11-cu130-x86_64-linux/_grouped_moe_gemv_cuda_1683349.abi3.so ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:f7b735196f5e18c071afa5416f9eca8db4e2ef9f27189ffab025c52e323457fc
+size 174448

build/torch212-cxx11-cu130-x86_64-linux/_ops.py ADDED Viewed

	@@ -0,0 +1,9 @@

+import torch
+from . import _grouped_moe_gemv_cuda_1683349
+ops = torch.ops._grouped_moe_gemv_cuda_1683349
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_grouped_moe_gemv_cuda_1683349::{op_name}"

build/torch212-cxx11-cu130-x86_64-linux/grouped_moe_gemv/__init__.py ADDED Viewed

	@@ -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/torch212-cxx11-cu130-x86_64-linux/metadata.json ADDED Viewed

	@@ -0,0 +1,13 @@

+{
+  "name": "grouped-moe-gemv",
+  "id": "_grouped_moe_gemv_cuda_1683349",
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "12.0a"
+    ]
+  }
+}

build/torch212-cxx11-cu132-x86_64-linux/__init__.py ADDED Viewed

	@@ -0,0 +1,99 @@

+"""FlashRT grouped MoE GEMV kernels."""
+from __future__ import annotations
+from typing import Optional
+import torch
+from ._ops import add_op_namespace_prefix, ops
+@torch.library.register_fake(add_op_namespace_prefix("w4a16_decode_gemv_bf16"))
+def _w4a16_decode_gemv_fake(
+    x_bf16: torch.Tensor,
+    weight_packed: torch.Tensor,
+    sfb: torch.Tensor,
+    alpha: float,
+    out: torch.Tensor,
+) -> None:
+    k = x_bf16.shape[0] if x_bf16.dim() == 1 else x_bf16.shape[1]
+    if weight_packed.dim() != 2 or weight_packed.shape[1] != k // 2 or out.shape != (weight_packed.shape[0],):
+        raise RuntimeError("expected x (K,) or (1,K), weight_packed (N,K/2), out (N,)")
+    return None
+@torch.library.register_fake(add_op_namespace_prefix("grouped_w4a16_gemv_bf16"))
+def _grouped_w4a16_gemv_fake(
+    activations: torch.Tensor,
+    weight_stack: torch.Tensor,
+    sfb_stack: torch.Tensor,
+    alpha_stack: torch.Tensor,
+    expert_idx: torch.Tensor,
+    w_stride: int,
+    sfb_stride: int,
+    out: torch.Tensor,
+) -> None:
+    if activations.dim() != 2 or out.dim() != 2 or out.shape[0] != activations.shape[0]:
+        raise RuntimeError("expected activations (slots,K), out (slots,N)")
+    if expert_idx.shape != (activations.shape[0],):
+        raise RuntimeError("expert_idx must have shape (slots,)")
+    return None
+def w4a16_decode_gemv_bf16(
+    x_bf16: torch.Tensor,
+    weight_packed: torch.Tensor,
+    sfb: torch.Tensor,
+    *,
+    alpha: float = 1.0,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    if out is None:
+        out = torch.empty((weight_packed.shape[0],), device=x_bf16.device, dtype=torch.bfloat16)
+    ops.w4a16_decode_gemv_bf16(x_bf16, weight_packed, sfb, float(alpha), out)
+    return out
+def grouped_w4a16_gemv_bf16(
+    activations: torch.Tensor,
+    weight_stack: torch.Tensor,
+    sfb_stack: torch.Tensor,
+    alpha_stack: torch.Tensor,
+    expert_idx: torch.Tensor,
+    *,
+    n: int,
+    w_stride: Optional[int] = None,
+    sfb_stride: Optional[int] = None,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """Run one W4A16 GEMV per routed slot.
+    `weight_stack` is a flat expert stack. `w_stride` and `sfb_stride` are byte
+    strides between experts; by default `w_stride = n * K / 2`.
+    """
+    k = activations.shape[1]
+    if out is None:
+        out = torch.empty((activations.shape[0], int(n)), device=activations.device, dtype=torch.bfloat16)
+    if w_stride is None:
+        w_stride = int(n) * k // 2
+    if sfb_stride is None:
+        raise RuntimeError("sfb_stride must be provided because swizzled SF size is layout-dependent")
+    ops.grouped_w4a16_gemv_bf16(
+        activations,
+        weight_stack,
+        sfb_stack,
+        alpha_stack,
+        expert_idx,
+        int(w_stride),
+        int(sfb_stride),
+        out,
+    )
+    return out
+__all__ = [
+    "grouped_w4a16_gemv_bf16",
+    "w4a16_decode_gemv_bf16",
+]

build/torch212-cxx11-cu132-x86_64-linux/_grouped_moe_gemv_cuda_1683349.abi3.so ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:d1e7570b56c4dddb95cd4c18015c02517beedc8eee8678fced97250cdc9bb2cc
+size 174448

build/torch212-cxx11-cu132-x86_64-linux/_ops.py ADDED Viewed

	@@ -0,0 +1,9 @@

+import torch
+from . import _grouped_moe_gemv_cuda_1683349
+ops = torch.ops._grouped_moe_gemv_cuda_1683349
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_grouped_moe_gemv_cuda_1683349::{op_name}"

build/torch212-cxx11-cu132-x86_64-linux/grouped_moe_gemv/__init__.py ADDED Viewed

	@@ -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/torch212-cxx11-cu132-x86_64-linux/metadata.json ADDED Viewed

	@@ -0,0 +1,13 @@

+{
+  "name": "grouped-moe-gemv",
+  "id": "_grouped_moe_gemv_cuda_1683349",
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "12.0a"
+    ]
+  }
+}