liangsu9988 commited on 13 days ago

Commit

0c2f3c9

verified ·

1 Parent(s): afbdd32

Uploaded using `kernel-builder`.

Browse files

Files changed (20) hide show

build/torch211-cxx11-cu128-x86_64-linux/__init__.py +71 -0
build/torch211-cxx11-cu128-x86_64-linux/_bf16_linear_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/bf16_linear_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 +71 -0
build/torch211-cxx11-cu130-x86_64-linux/_bf16_linear_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/bf16_linear_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 +71 -0
build/torch212-cxx11-cu130-x86_64-linux/_bf16_linear_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/bf16_linear_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 +71 -0
build/torch212-cxx11-cu132-x86_64-linux/_bf16_linear_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/bf16_linear_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 +1,71 @@

+"""FlashRT BF16 decode 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("bf16_decode_gemv_bf16"))
+def _bf16_decode_gemv_fake(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    alpha: float,
+    variant: int,
+    out: torch.Tensor,
+) -> None:
+    k = x.shape[0] if x.dim() == 1 else x.shape[1]
+    if weight.dim() != 2 or weight.shape[1] != k or out.shape != (weight.shape[0],):
+        raise RuntimeError("expected x (K,) or (1,K), weight (N,K), out (N,)")
+    return None
+@torch.library.register_fake(add_op_namespace_prefix("bf16_decode_gemv_unrolled_bf16"))
+def _bf16_decode_gemv_unrolled_fake(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    out: torch.Tensor,
+) -> None:
+    k = x.shape[0] if x.dim() == 1 else x.shape[1]
+    if weight.dim() != 2 or weight.shape[1] != k or out.shape != (weight.shape[0],):
+        raise RuntimeError("expected x (K,) or (1,K), weight (N,K), out (N,)")
+    return None
+def bf16_decode_gemv_bf16(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    *,
+    alpha: float = 1.0,
+    variant: int = 0,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """Compute BF16 M=1 decode GEMV: ``out = x @ weight.T * alpha``."""
+    if out is None:
+        out = torch.empty((weight.shape[0],), device=x.device, dtype=torch.bfloat16)
+    ops.bf16_decode_gemv_bf16(x, weight, float(alpha), int(variant), out)
+    return out
+def bf16_decode_gemv_unrolled_bf16(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    *,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """Compute BF16 M=1 GEMV with the unrolled memory-level-parallel kernel."""
+    if out is None:
+        out = torch.empty((weight.shape[0],), device=x.device, dtype=torch.bfloat16)
+    ops.bf16_decode_gemv_unrolled_bf16(x, weight, out)
+    return out
+__all__ = [
+    "bf16_decode_gemv_bf16",
+    "bf16_decode_gemv_unrolled_bf16",
+]

build/torch211-cxx11-cu128-x86_64-linux/_bf16_linear_gemv_cuda_1683349.abi3.so ADDED Viewed

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

+version https://git-lfs.github.com/spec/v1
+oid sha256:88328fac3bfb2cbd7dfff66ea724593984a203ba7255ab75df5f0d3e7ffcce83
+size 155928

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

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

+import torch
+from . import _bf16_linear_gemv_cuda_1683349
+ops = torch.ops._bf16_linear_gemv_cuda_1683349
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_bf16_linear_gemv_cuda_1683349::{op_name}"

build/torch211-cxx11-cu128-x86_64-linux/bf16_linear_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": "bf16-linear-gemv",
+  "id": "_bf16_linear_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 +1,71 @@

+"""FlashRT BF16 decode 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("bf16_decode_gemv_bf16"))
+def _bf16_decode_gemv_fake(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    alpha: float,
+    variant: int,
+    out: torch.Tensor,
+) -> None:
+    k = x.shape[0] if x.dim() == 1 else x.shape[1]
+    if weight.dim() != 2 or weight.shape[1] != k or out.shape != (weight.shape[0],):
+        raise RuntimeError("expected x (K,) or (1,K), weight (N,K), out (N,)")
+    return None
+@torch.library.register_fake(add_op_namespace_prefix("bf16_decode_gemv_unrolled_bf16"))
+def _bf16_decode_gemv_unrolled_fake(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    out: torch.Tensor,
+) -> None:
+    k = x.shape[0] if x.dim() == 1 else x.shape[1]
+    if weight.dim() != 2 or weight.shape[1] != k or out.shape != (weight.shape[0],):
+        raise RuntimeError("expected x (K,) or (1,K), weight (N,K), out (N,)")
+    return None
+def bf16_decode_gemv_bf16(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    *,
+    alpha: float = 1.0,
+    variant: int = 0,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """Compute BF16 M=1 decode GEMV: ``out = x @ weight.T * alpha``."""
+    if out is None:
+        out = torch.empty((weight.shape[0],), device=x.device, dtype=torch.bfloat16)
+    ops.bf16_decode_gemv_bf16(x, weight, float(alpha), int(variant), out)
+    return out
+def bf16_decode_gemv_unrolled_bf16(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    *,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """Compute BF16 M=1 GEMV with the unrolled memory-level-parallel kernel."""
+    if out is None:
+        out = torch.empty((weight.shape[0],), device=x.device, dtype=torch.bfloat16)
+    ops.bf16_decode_gemv_unrolled_bf16(x, weight, out)
+    return out
+__all__ = [
+    "bf16_decode_gemv_bf16",
+    "bf16_decode_gemv_unrolled_bf16",
+]

build/torch211-cxx11-cu130-x86_64-linux/_bf16_linear_gemv_cuda_1683349.abi3.so ADDED Viewed

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

+version https://git-lfs.github.com/spec/v1
+oid sha256:6a49b37560075e028eb18d463b05af29dd72a2569cc3b2df738afb793363d769
+size 159832

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

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

+import torch
+from . import _bf16_linear_gemv_cuda_1683349
+ops = torch.ops._bf16_linear_gemv_cuda_1683349
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_bf16_linear_gemv_cuda_1683349::{op_name}"

build/torch211-cxx11-cu130-x86_64-linux/bf16_linear_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": "bf16-linear-gemv",
+  "id": "_bf16_linear_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 +1,71 @@

+"""FlashRT BF16 decode 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("bf16_decode_gemv_bf16"))
+def _bf16_decode_gemv_fake(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    alpha: float,
+    variant: int,
+    out: torch.Tensor,
+) -> None:
+    k = x.shape[0] if x.dim() == 1 else x.shape[1]
+    if weight.dim() != 2 or weight.shape[1] != k or out.shape != (weight.shape[0],):
+        raise RuntimeError("expected x (K,) or (1,K), weight (N,K), out (N,)")
+    return None
+@torch.library.register_fake(add_op_namespace_prefix("bf16_decode_gemv_unrolled_bf16"))
+def _bf16_decode_gemv_unrolled_fake(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    out: torch.Tensor,
+) -> None:
+    k = x.shape[0] if x.dim() == 1 else x.shape[1]
+    if weight.dim() != 2 or weight.shape[1] != k or out.shape != (weight.shape[0],):
+        raise RuntimeError("expected x (K,) or (1,K), weight (N,K), out (N,)")
+    return None
+def bf16_decode_gemv_bf16(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    *,
+    alpha: float = 1.0,
+    variant: int = 0,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """Compute BF16 M=1 decode GEMV: ``out = x @ weight.T * alpha``."""
+    if out is None:
+        out = torch.empty((weight.shape[0],), device=x.device, dtype=torch.bfloat16)
+    ops.bf16_decode_gemv_bf16(x, weight, float(alpha), int(variant), out)
+    return out
+def bf16_decode_gemv_unrolled_bf16(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    *,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """Compute BF16 M=1 GEMV with the unrolled memory-level-parallel kernel."""
+    if out is None:
+        out = torch.empty((weight.shape[0],), device=x.device, dtype=torch.bfloat16)
+    ops.bf16_decode_gemv_unrolled_bf16(x, weight, out)
+    return out
+__all__ = [
+    "bf16_decode_gemv_bf16",
+    "bf16_decode_gemv_unrolled_bf16",
+]

build/torch212-cxx11-cu130-x86_64-linux/_bf16_linear_gemv_cuda_1683349.abi3.so ADDED Viewed

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

+version https://git-lfs.github.com/spec/v1
+oid sha256:8e69262a9dc1029a69945311b3e4a1a7437ab2942c5a448c0e63c8ce89ad115c
+size 170344

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

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

+import torch
+from . import _bf16_linear_gemv_cuda_1683349
+ops = torch.ops._bf16_linear_gemv_cuda_1683349
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_bf16_linear_gemv_cuda_1683349::{op_name}"

build/torch212-cxx11-cu130-x86_64-linux/bf16_linear_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": "bf16-linear-gemv",
+  "id": "_bf16_linear_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 +1,71 @@

+"""FlashRT BF16 decode 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("bf16_decode_gemv_bf16"))
+def _bf16_decode_gemv_fake(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    alpha: float,
+    variant: int,
+    out: torch.Tensor,
+) -> None:
+    k = x.shape[0] if x.dim() == 1 else x.shape[1]
+    if weight.dim() != 2 or weight.shape[1] != k or out.shape != (weight.shape[0],):
+        raise RuntimeError("expected x (K,) or (1,K), weight (N,K), out (N,)")
+    return None
+@torch.library.register_fake(add_op_namespace_prefix("bf16_decode_gemv_unrolled_bf16"))
+def _bf16_decode_gemv_unrolled_fake(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    out: torch.Tensor,
+) -> None:
+    k = x.shape[0] if x.dim() == 1 else x.shape[1]
+    if weight.dim() != 2 or weight.shape[1] != k or out.shape != (weight.shape[0],):
+        raise RuntimeError("expected x (K,) or (1,K), weight (N,K), out (N,)")
+    return None
+def bf16_decode_gemv_bf16(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    *,
+    alpha: float = 1.0,
+    variant: int = 0,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """Compute BF16 M=1 decode GEMV: ``out = x @ weight.T * alpha``."""
+    if out is None:
+        out = torch.empty((weight.shape[0],), device=x.device, dtype=torch.bfloat16)
+    ops.bf16_decode_gemv_bf16(x, weight, float(alpha), int(variant), out)
+    return out
+def bf16_decode_gemv_unrolled_bf16(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    *,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """Compute BF16 M=1 GEMV with the unrolled memory-level-parallel kernel."""
+    if out is None:
+        out = torch.empty((weight.shape[0],), device=x.device, dtype=torch.bfloat16)
+    ops.bf16_decode_gemv_unrolled_bf16(x, weight, out)
+    return out
+__all__ = [
+    "bf16_decode_gemv_bf16",
+    "bf16_decode_gemv_unrolled_bf16",
+]

build/torch212-cxx11-cu132-x86_64-linux/_bf16_linear_gemv_cuda_1683349.abi3.so ADDED Viewed

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

+version https://git-lfs.github.com/spec/v1
+oid sha256:c420e09798f32d3e619d2f12e458e51efe43c4105a2013edcffaf5f203359d31
+size 170344

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

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

+import torch
+from . import _bf16_linear_gemv_cuda_1683349
+ops = torch.ops._bf16_linear_gemv_cuda_1683349
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_bf16_linear_gemv_cuda_1683349::{op_name}"

build/torch212-cxx11-cu132-x86_64-linux/bf16_linear_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": "bf16-linear-gemv",
+  "id": "_bf16_linear_gemv_cuda_1683349",
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "12.0a"
+    ]
+  }
+}