Kernels:

kernels-community
/

paged-attention

Trusted publisher

Kernel card Files Files and versions

xet

Community

kernels-bot commited on Apr 15

Commit

b0506d9

verified ·

1 Parent(s): 1988537

Uploaded using `kernel-builder`.

Browse files

Files changed (11) hide show

build/torch210-metal-aarch64-darwin/_ops.py +3 -3
build/torch210-metal-aarch64-darwin/{_paged_attention_metal_c6404f1.abi3.so → _paged_attention_metal_0278991.abi3.so} +2 -2
build/torch210-metal-aarch64-darwin/metadata.json +4 -1
build/torch210-metal-aarch64-darwin/paged_attention/__init__.py +2 -2
build/torch211-metal-aarch64-darwin/__init__.py +21 -0
build/torch211-metal-aarch64-darwin/_custom_ops.py +173 -0
build/torch211-metal-aarch64-darwin/_ops.py +9 -0
build/torch211-metal-aarch64-darwin/_paged_attention_metal_0278991.abi3.so +3 -0
build/torch211-metal-aarch64-darwin/metadata.json +8 -0
build/torch211-metal-aarch64-darwin/paged_attention/__init__.py +26 -0
build/torch211-metal-aarch64-darwin/platforms.py +92 -0

build/torch210-metal-aarch64-darwin/_ops.py CHANGED Viewed

@@ -1,9 +1,9 @@
 import torch
-from . import _paged_attention_metal_c6404f1
-ops = torch.ops._paged_attention_metal_c6404f1
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_paged_attention_metal_c6404f1::{op_name}"

 import torch
+from . import _paged_attention_metal_0278991
+ops = torch.ops._paged_attention_metal_0278991
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_paged_attention_metal_0278991::{op_name}"

build/torch210-metal-aarch64-darwin/{_paged_attention_metal_c6404f1.abi3.so → _paged_attention_metal_0278991.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:63b99454be1d50260beb973bd80b5fb2af61ac288505840e164461ce9239b0ed
-size 14893560

 version https://git-lfs.github.com/spec/v1
+oid sha256:7e911d1cfa77e7c8995b2a7e911f7bd53ee83c825b5cbb6e3f1a132dca3b53e3
+size 14893480

build/torch210-metal-aarch64-darwin/metadata.json CHANGED Viewed

@@ -1,5 +1,8 @@
 {
   "version": 1,
   "license": "Apache-2.0",
-  "python-depends": []
 }

 {
   "version": 1,
   "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "metal"
+  }
 }

build/torch210-metal-aarch64-darwin/paged_attention/__init__.py CHANGED Viewed

@@ -1,10 +1,10 @@
 import ctypes
 import sys
-import importlib
 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

 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

build/torch211-metal-aarch64-darwin/__init__.py ADDED Viewed

	@@ -0,0 +1,21 @@

+from ._custom_ops import (
+    convert_fp8,
+    copy_blocks,
+    paged_attention_v1,
+    paged_attention_v2,
+    reshape_and_cache,
+    reshape_and_cache_flash,
+    swap_blocks,
+)
+from ._ops import ops
+__all__ = [
+    "convert_fp8",
+    "copy_blocks",
+    "ops",
+    "paged_attention_v1",
+    "paged_attention_v2",
+    "reshape_and_cache",
+    "reshape_and_cache_flash",
+    "swap_blocks",
+]

build/torch211-metal-aarch64-darwin/_custom_ops.py ADDED Viewed

	@@ -0,0 +1,173 @@

+from typing import List, Optional
+import torch
+from ._ops import ops
+# page attention ops
+def paged_attention_v1(
+    out: torch.Tensor,
+    query: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    num_kv_heads: int,
+    scale: float,
+    block_tables: torch.Tensor,
+    seq_lens: torch.Tensor,
+    block_size: int,
+    max_seq_len: int,
+    alibi_slopes: Optional[torch.Tensor],
+    kv_cache_dtype: str,
+    k_scale: float,
+    v_scale: float,
+    tp_rank: int = 0,
+    blocksparse_local_blocks: int = 0,
+    blocksparse_vert_stride: int = 0,
+    blocksparse_block_size: int = 64,
+    blocksparse_head_sliding_step: int = 0,
+) -> None:
+    ops.paged_attention_v1(
+        out,
+        query,
+        key_cache,
+        value_cache,
+        num_kv_heads,
+        scale,
+        block_tables,
+        seq_lens,
+        block_size,
+        max_seq_len,
+        alibi_slopes,
+        kv_cache_dtype,
+        k_scale,
+        v_scale,
+        tp_rank,
+        blocksparse_local_blocks,
+        blocksparse_vert_stride,
+        blocksparse_block_size,
+        blocksparse_head_sliding_step,
+    )
+def paged_attention_v2(
+    out: torch.Tensor,
+    exp_sum: torch.Tensor,
+    max_logits: torch.Tensor,
+    tmp_out: torch.Tensor,
+    query: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    num_kv_heads: int,
+    scale: float,
+    block_tables: torch.Tensor,
+    seq_lens: torch.Tensor,
+    block_size: int,
+    max_seq_len: int,
+    alibi_slopes: Optional[torch.Tensor],
+    kv_cache_dtype: str,
+    k_scale: float,
+    v_scale: float,
+    tp_rank: int = 0,
+    blocksparse_local_blocks: int = 0,
+    blocksparse_vert_stride: int = 0,
+    blocksparse_block_size: int = 64,
+    blocksparse_head_sliding_step: int = 0,
+) -> None:
+    ops.paged_attention_v2(
+        out,
+        exp_sum,
+        max_logits,
+        tmp_out,
+        query,
+        key_cache,
+        value_cache,
+        num_kv_heads,
+        scale,
+        block_tables,
+        seq_lens,
+        block_size,
+        max_seq_len,
+        alibi_slopes,
+        kv_cache_dtype,
+        k_scale,
+        v_scale,
+        tp_rank,
+        blocksparse_local_blocks,
+        blocksparse_vert_stride,
+        blocksparse_block_size,
+        blocksparse_head_sliding_step,
+    )
+def reshape_and_cache(
+    key: torch.Tensor,
+    value: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    slot_mapping: torch.Tensor,
+    kv_cache_dtype: str,
+    k_scale: float,
+    v_scale: float,
+) -> None:
+    ops.reshape_and_cache(
+        key,
+        value,
+        key_cache,
+        value_cache,
+        slot_mapping,
+        kv_cache_dtype,
+        k_scale,
+        v_scale,
+    )
+def reshape_and_cache_flash(
+    key: torch.Tensor,
+    value: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    slot_mapping: torch.Tensor,
+    kv_cache_dtype: str,
+    k_scale: torch.Tensor,
+    v_scale: torch.Tensor,
+) -> None:
+    ops.reshape_and_cache_flash(
+        key,
+        value,
+        key_cache,
+        value_cache,
+        slot_mapping,
+        kv_cache_dtype,
+        k_scale,
+        v_scale,
+    )
+def copy_blocks(
+    key_caches: List[torch.Tensor],
+    value_caches: List[torch.Tensor],
+    block_mapping: torch.Tensor,
+) -> None:
+    ops.copy_blocks(key_caches, value_caches, block_mapping)
+def swap_blocks(
+    src: torch.Tensor, dst: torch.Tensor, block_mapping: torch.Tensor
+) -> None:
+    ops.swap_blocks(src, dst, block_mapping)
+def convert_fp8(
+    output: torch.Tensor, input: torch.Tensor, scale: float = 1.0, kv_dtype: str = "fp8"
+) -> None:
+    ops.convert_fp8(output, input, scale, kv_dtype)
+__all__ = [
+    "convert_fp8",
+    "paged_attention_v1",
+    "paged_attention_v2",
+    "reshape_and_cache",
+    "copy_blocks",
+]

build/torch211-metal-aarch64-darwin/_ops.py ADDED Viewed

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

+import torch
+from . import _paged_attention_metal_0278991
+ops = torch.ops._paged_attention_metal_0278991
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_paged_attention_metal_0278991::{op_name}"

build/torch211-metal-aarch64-darwin/_paged_attention_metal_0278991.abi3.so ADDED Viewed

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

+version https://git-lfs.github.com/spec/v1
+oid sha256:273c7d18a6b88dd0fdedfd050f43532d7cca70e1f6f3da95e9c8fa4272a82cf2
+size 14893480

build/torch211-metal-aarch64-darwin/metadata.json ADDED Viewed

	@@ -0,0 +1,8 @@

+{
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "metal"
+  }
+}

build/torch211-metal-aarch64-darwin/paged_attention/__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-metal-aarch64-darwin/platforms.py ADDED Viewed

	@@ -0,0 +1,92 @@

+import os
+import random
+from abc import ABC, abstractmethod
+from functools import lru_cache, wraps
+from typing import Callable, ParamSpec, TypeVar
+import numpy as np
+import torch
+IS_ROCM = torch.version.hip is not None
+IS_MPS = torch.backends.mps.is_available()
+class Platform(ABC):
+    @classmethod
+    def seed_everything(cls, seed: int) -> None:
+        """
+        Set the seed of each random module.
+        `torch.manual_seed` will set seed on all devices.
+        Loosely based on: https://github.com/Lightning-AI/pytorch-lightning/blob/2.4.0/src/lightning/fabric/utilities/seed.py#L20
+        """
+        random.seed(seed)
+        np.random.seed(seed)
+        torch.manual_seed(seed)
+    @abstractmethod
+    def get_device_name(self, device_id: int = 0) -> str: ...
+    @abstractmethod
+    def is_cuda(self) -> bool: ...
+    @abstractmethod
+    def is_rocm(self) -> bool: ...
+    @abstractmethod
+    def is_mps(self) -> bool: ...
+class CudaPlatform(Platform):
+    @classmethod
+    @lru_cache(maxsize=8)
+    def get_device_name(cls, device_id: int = 0) -> str:
+        return torch.cuda.get_device_name(0)
+    def is_cuda(self) -> bool:
+        return True
+    def is_rocm(self) -> bool:
+        return False
+    def is_mps(self) -> bool:
+        return False
+class RocmPlatform(Platform):
+    @classmethod
+    @lru_cache(maxsize=8)
+    def get_device_name(cls, device_id: int = 0) -> str:
+        return torch.cuda.get_device_name(device_id)
+    def is_cuda(self) -> bool:
+        return False
+    def is_rocm(self) -> bool:
+        return True
+    def is_mps(self) -> bool:
+        return False
+class MpsPlatform(Platform):
+    @classmethod
+    @lru_cache(maxsize=8)
+    def get_device_name(cls, device_id: int = 0) -> str:
+        return torch.cuda.get_device_name(device_id)
+    def is_cuda(self) -> bool:
+        return False
+    def is_rocm(self) -> bool:
+        return False
+    def is_mps(self) -> bool:
+        return True
+current_platform = (
+    RocmPlatform() if IS_ROCM else
+    MpsPlatform() if IS_MPS else
+    CudaPlatform() if torch.cuda.is_available() else
+    None
+)