Uploaded using `kernel-builder`.

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

@@ -1,9 +1,9 @@
 import torch
-from . import _paged_attention_cuda_86f75d9
-ops = torch.ops._paged_attention_cuda_86f75d9
+from . import _paged_attention_cuda_2567cd6
+ops = torch.ops._paged_attention_cuda_2567cd6
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_paged_attention_cuda_86f75d9::{op_name}"
+    return f"_paged_attention_cuda_2567cd6::{op_name}"

build/torch210-cxx11-cu126-aarch64-linux/{_paged_attention_cuda_86f75d9.abi3.so → _paged_attention_cuda_2567cd6.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ccb30dbb13f2f96b9dbb69415bdaa8edca7caaf9b4ff54fd003c0ca7c143d432
+oid sha256:9e3b1c9befa49206a2bd8d5e5cdf577c72d12621018a909013811ccf68142772
 size 140013424

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

@@ -1,6 +1,6 @@
 {
   "name": "paged-attention",
-  "id": "_paged_attention_cuda_86f75d9",
+  "id": "_paged_attention_cuda_2567cd6",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

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

@@ -1,9 +1,9 @@
 import torch
-from . import _paged_attention_cuda_86f75d9
-ops = torch.ops._paged_attention_cuda_86f75d9
+from . import _paged_attention_cuda_2567cd6
+ops = torch.ops._paged_attention_cuda_2567cd6
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_paged_attention_cuda_86f75d9::{op_name}"
+    return f"_paged_attention_cuda_2567cd6::{op_name}"

build/torch210-cxx11-cu128-aarch64-linux/{_paged_attention_cuda_86f75d9.abi3.so → _paged_attention_cuda_2567cd6.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c7577cfb5ab4acf981e99dcd288d4c5b416bddac1add8724ead08ade73b7432c
+oid sha256:53dfe24d07a74182292d12d6d6376644305e1fe715f7b58f101e324a28b9d1f1
 size 167603936

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

@@ -1,6 +1,6 @@
 {
   "name": "paged-attention",
-  "id": "_paged_attention_cuda_86f75d9",
+  "id": "_paged_attention_cuda_2567cd6",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

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

@@ -1,9 +1,9 @@
 import torch
-from . import _paged_attention_cuda_86f75d9
-ops = torch.ops._paged_attention_cuda_86f75d9
+from . import _paged_attention_cuda_2567cd6
+ops = torch.ops._paged_attention_cuda_2567cd6
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_paged_attention_cuda_86f75d9::{op_name}"
+    return f"_paged_attention_cuda_2567cd6::{op_name}"

build/torch210-cxx11-cu130-aarch64-linux/{_paged_attention_cuda_86f75d9.abi3.so → _paged_attention_cuda_2567cd6.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a68308a8e0157354f879eb6d6442b0cff86fe740f12a7ee0d0eebd80643d5d61
+oid sha256:b8f25d8a0b39bb6710dded0e031c96fc2be3d615690e3bc95126d620f5a5c756
 size 86069336

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

@@ -1,6 +1,6 @@
 {
   "name": "paged-attention",
-  "id": "_paged_attention_cuda_86f75d9",
+  "id": "_paged_attention_cuda_2567cd6",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

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

@@ -1,9 +1,9 @@
 import torch
-from . import _paged_attention_cuda_86f75d9
-ops = torch.ops._paged_attention_cuda_86f75d9
+from . import _paged_attention_cuda_2567cd6
+ops = torch.ops._paged_attention_cuda_2567cd6
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_paged_attention_cuda_86f75d9::{op_name}"
+    return f"_paged_attention_cuda_2567cd6::{op_name}"

build/torch211-cxx11-cu126-aarch64-linux/{_paged_attention_cuda_86f75d9.abi3.so → _paged_attention_cuda_2567cd6.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5318f7d4226f2b7e4d4683e6ad576edcd03a07f898a5455942779a2e97083131
+oid sha256:16d77f814330ce56502186eb42be3ac21b1e862f254ba57f4eee212c85c17886
 size 140009304

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

@@ -1,6 +1,6 @@
 {
   "name": "paged-attention",
-  "id": "_paged_attention_cuda_86f75d9",
+  "id": "_paged_attention_cuda_2567cd6",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

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

@@ -1,9 +1,9 @@
 import torch
-from . import _paged_attention_cuda_86f75d9
-ops = torch.ops._paged_attention_cuda_86f75d9
+from . import _paged_attention_cuda_2567cd6
+ops = torch.ops._paged_attention_cuda_2567cd6
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_paged_attention_cuda_86f75d9::{op_name}"
+    return f"_paged_attention_cuda_2567cd6::{op_name}"

build/torch211-cxx11-cu128-aarch64-linux/{_paged_attention_cuda_86f75d9.abi3.so → _paged_attention_cuda_2567cd6.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:501f09fe11ea475737175f7fcab105412acdfdba36dab0dd0ccc6d67094a1046
+oid sha256:60277de4e122424d6e5430299c37543b75f62b29c977189cf65b9449eeb1a705
 size 167599784

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

@@ -1,6 +1,6 @@
 {
   "name": "paged-attention",
-  "id": "_paged_attention_cuda_86f75d9",
+  "id": "_paged_attention_cuda_2567cd6",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

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

@@ -1,9 +1,9 @@
 import torch
-from . import _paged_attention_cuda_86f75d9
-ops = torch.ops._paged_attention_cuda_86f75d9
+from . import _paged_attention_cuda_2567cd6
+ops = torch.ops._paged_attention_cuda_2567cd6
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_paged_attention_cuda_86f75d9::{op_name}"
+    return f"_paged_attention_cuda_2567cd6::{op_name}"

build/torch211-cxx11-cu130-aarch64-linux/{_paged_attention_cuda_86f75d9.abi3.so → _paged_attention_cuda_2567cd6.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b347d4485444ece322c18ec55ffc10d389d8fdb52b638d36bc963b60036b442d
+oid sha256:dc3151992ff276d6413339e601a6c708475185d6f1e0349dc93fc25890542347
 size 86065184

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

@@ -1,6 +1,6 @@
 {
   "name": "paged-attention",
-  "id": "_paged_attention_cuda_86f75d9",
+  "id": "_paged_attention_cuda_2567cd6",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

build/torch212-cxx11-cu126-aarch64-linux/__init__.py

@@ -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/torch212-cxx11-cu126-aarch64-linux/_custom_ops.py

@@ -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/torch212-cxx11-cu126-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _paged_attention_cuda_2567cd6
+ops = torch.ops._paged_attention_cuda_2567cd6
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_paged_attention_cuda_2567cd6::{op_name}"

build/torch212-cxx11-cu126-aarch64-linux/_paged_attention_cuda_2567cd6.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a3102d6efd314c44aff63ccff2c1f38cc6bfd0c54c30ee6dfa9d98d3b2a6e67
+size 140012416

build/torch212-cxx11-cu126-aarch64-linux/metadata.json

@@ -0,0 +1,20 @@
+{
+  "name": "paged-attention",
+  "id": "_paged_attention_cuda_2567cd6",
+  "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/torch212-cxx11-cu126-aarch64-linux/paged_attention/__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/torch212-cxx11-cu126-aarch64-linux/platforms.py

@@ -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
+)

build/torch212-cxx11-cu130-aarch64-linux/__init__.py

@@ -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/torch212-cxx11-cu130-aarch64-linux/_custom_ops.py

@@ -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/torch212-cxx11-cu130-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _paged_attention_cuda_2567cd6
+ops = torch.ops._paged_attention_cuda_2567cd6
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_paged_attention_cuda_2567cd6::{op_name}"

build/torch212-cxx11-cu130-aarch64-linux/_paged_attention_cuda_2567cd6.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b181e6a863e320fe528e28d30b0d4dd3bb92e8d4a54f7a92c25627966dc8f32b
+size 86068296

build/torch212-cxx11-cu130-aarch64-linux/metadata.json

@@ -0,0 +1,21 @@
+{
+  "name": "paged-attention",
+  "id": "_paged_attention_cuda_2567cd6",
+  "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/torch212-cxx11-cu130-aarch64-linux/paged_attention/__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/torch212-cxx11-cu130-aarch64-linux/platforms.py

@@ -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
+)

build/torch212-cxx11-cu132-aarch64-linux/__init__.py

@@ -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/torch212-cxx11-cu132-aarch64-linux/_custom_ops.py

@@ -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/torch212-cxx11-cu132-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _paged_attention_cuda_2567cd6
+ops = torch.ops._paged_attention_cuda_2567cd6
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_paged_attention_cuda_2567cd6::{op_name}"

build/torch212-cxx11-cu132-aarch64-linux/_paged_attention_cuda_2567cd6.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffc6f41ff69cf6222fac5e183b60bd1fc371addbc570f795daf3371d57710282
+size 84956984

build/torch212-cxx11-cu132-aarch64-linux/metadata.json

@@ -0,0 +1,21 @@
+{
+  "name": "paged-attention",
+  "id": "_paged_attention_cuda_2567cd6",
+  "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/torch212-cxx11-cu132-aarch64-linux/paged_attention/__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/torch212-cxx11-cu132-aarch64-linux/platforms.py

@@ -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
+)