Build (aarch64)

build/torch26-cxx11-cu126-aarch64-linux/paged_attention/__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/torch26-cxx11-cu126-aarch64-linux/paged_attention/_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/torch26-cxx11-cu126-aarch64-linux/paged_attention/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _paged_attention_dde9676
+ops = torch.ops._paged_attention_dde9676
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_paged_attention_dde9676::{op_name}"

build/torch26-cxx11-cu126-aarch64-linux/paged_attention/_paged_attention_dde9676.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b58e9a829c6f648f5dae9dfd29db85c3bceef48d6de3cc55861262343205c44b
+size 75255256

build/torch26-cxx11-cu126-aarch64-linux/paged_attention/platforms.py

@@ -0,0 +1,62 @@
+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
+
+
+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: ...
+
+
+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
+
+
+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
+
+
+current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()

build/torch26-cxx98-cu126-aarch64-linux/paged_attention/__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/torch26-cxx98-cu126-aarch64-linux/paged_attention/_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/torch26-cxx98-cu126-aarch64-linux/paged_attention/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _paged_attention_dde9676
+ops = torch.ops._paged_attention_dde9676
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_paged_attention_dde9676::{op_name}"

build/torch26-cxx98-cu126-aarch64-linux/paged_attention/_paged_attention_dde9676.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b5291420043170980483291edb8630c9ee4330cad0e8c80483348e9f674cec0
+size 75248048

build/torch26-cxx98-cu126-aarch64-linux/paged_attention/platforms.py

@@ -0,0 +1,62 @@
+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
+
+
+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: ...
+
+
+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
+
+
+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
+
+
+current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()

build/torch27-cxx11-cu126-aarch64-linux/paged_attention/__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/torch27-cxx11-cu126-aarch64-linux/paged_attention/_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/torch27-cxx11-cu126-aarch64-linux/paged_attention/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _paged_attention_dde9676
+ops = torch.ops._paged_attention_dde9676
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_paged_attention_dde9676::{op_name}"

build/torch27-cxx11-cu126-aarch64-linux/paged_attention/_paged_attention_dde9676.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6376f5fd010f324f63912ff624b9802a086ba71c391dea5019b8de7a1e8de7d8
+size 75255456

build/torch27-cxx11-cu126-aarch64-linux/paged_attention/platforms.py

@@ -0,0 +1,62 @@
+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
+
+
+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: ...
+
+
+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
+
+
+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
+
+
+current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()

build/torch27-cxx11-cu128-aarch64-linux/paged_attention/__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/torch27-cxx11-cu128-aarch64-linux/paged_attention/_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/torch27-cxx11-cu128-aarch64-linux/paged_attention/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _paged_attention_dde9676
+ops = torch.ops._paged_attention_dde9676
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_paged_attention_dde9676::{op_name}"

build/torch27-cxx11-cu128-aarch64-linux/paged_attention/_paged_attention_dde9676.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86c24c6ed593b28ff7fcd07129e08b21ce9d9de27779c65ab215be0ba7a96312
+size 83905640

build/torch27-cxx11-cu128-aarch64-linux/paged_attention/platforms.py

@@ -0,0 +1,62 @@
+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
+
+
+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: ...
+
+
+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
+
+
+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
+
+
+current_platform = RocmPlatform() if IS_ROCM else CudaPlatform()