Build uploaded using `kernels`.

build/torch210-cxx11-cpu-x86_64-linux/{_flash_attn2_cpu_9f0ed09.abi3.so → _flash_attn2_cpu_abda3a0.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:45757a80c809dcbf1a8c75bde0c42dcba171960901f1c085699036d908a81c20
+oid sha256:98686f498c9fba8e4ff02197c83bbe4222be43dec5268ea3f4a1b51b50eee64a
 size 1942496

build/torch210-cxx11-cpu-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _flash_attn2_cpu_9f0ed09
-ops = torch.ops._flash_attn2_cpu_9f0ed09
+from . import _flash_attn2_cpu_abda3a0
+ops = torch.ops._flash_attn2_cpu_abda3a0
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_flash_attn2_cpu_9f0ed09::{op_name}"
+    return f"_flash_attn2_cpu_abda3a0::{op_name}"

build/torch210-cxx11-cpu-x86_64-linux/flash_attn_interface.py

@@ -33,8 +33,12 @@ def _get_device():
 
 
 def _get_block_size_n(device, head_dim, is_dropout, is_causal):
-    # This should match the block sizes in the CUDA kernel
     assert head_dim <= 256
+
+    if device.type == "xpu":
+        return 64
+
+    # This should match the block sizes in the CUDA kernel
     major, minor = torch.cuda.get_device_capability(device)
     is_sm8x = major == 8 and minor > 0  # Only include sm86 and sm89, exclude sm80 (A100)
     is_sm80 = major == 8 and minor == 0

build/torch210-cxx11-cu126-x86_64-linux/{_flash_attn2_cuda_9f0ed09.abi3.so → _flash_attn2_cuda_abda3a0.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:039e336f68c3efaa02ef0b103a2607b2a544d68a46fa0165e2433464f26223a3
+oid sha256:d009f413f2eb3e85146168654e36cbc61c3733856b03bd743285b63e4084b713
 size 448709016

build/torch210-cxx11-cu126-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _flash_attn2_cuda_9f0ed09
-ops = torch.ops._flash_attn2_cuda_9f0ed09
+from . import _flash_attn2_cuda_abda3a0
+ops = torch.ops._flash_attn2_cuda_abda3a0
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_flash_attn2_cuda_9f0ed09::{op_name}"
+    return f"_flash_attn2_cuda_abda3a0::{op_name}"

build/torch210-cxx11-cu126-x86_64-linux/flash_attn_interface.py

@@ -33,8 +33,12 @@ def _get_device():
 
 
 def _get_block_size_n(device, head_dim, is_dropout, is_causal):
-    # This should match the block sizes in the CUDA kernel
     assert head_dim <= 256
+
+    if device.type == "xpu":
+        return 64
+
+    # This should match the block sizes in the CUDA kernel
     major, minor = torch.cuda.get_device_capability(device)
     is_sm8x = major == 8 and minor > 0  # Only include sm86 and sm89, exclude sm80 (A100)
     is_sm80 = major == 8 and minor == 0

build/torch210-cxx11-cu128-x86_64-linux/{_flash_attn2_cuda_9f0ed09.abi3.so → _flash_attn2_cuda_abda3a0.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ac175faf91cfb9cd9827985bae32380035cb9f880f0bbb702e7f045eee90ae0a
+oid sha256:109d07ab749761b297b27783dd9ff702dedacd2b7cdb7a57a67e2cf7ac574516
 size 1037803408

build/torch210-cxx11-cu128-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _flash_attn2_cuda_9f0ed09
-ops = torch.ops._flash_attn2_cuda_9f0ed09
+from . import _flash_attn2_cuda_abda3a0
+ops = torch.ops._flash_attn2_cuda_abda3a0
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_flash_attn2_cuda_9f0ed09::{op_name}"
+    return f"_flash_attn2_cuda_abda3a0::{op_name}"

build/torch210-cxx11-cu128-x86_64-linux/flash_attn_interface.py

@@ -33,8 +33,12 @@ def _get_device():
 
 
 def _get_block_size_n(device, head_dim, is_dropout, is_causal):
-    # This should match the block sizes in the CUDA kernel
     assert head_dim <= 256
+
+    if device.type == "xpu":
+        return 64
+
+    # This should match the block sizes in the CUDA kernel
     major, minor = torch.cuda.get_device_capability(device)
     is_sm8x = major == 8 and minor > 0  # Only include sm86 and sm89, exclude sm80 (A100)
     is_sm80 = major == 8 and minor == 0

build/torch210-cxx11-cu130-x86_64-linux/{_flash_attn2_cuda_9f0ed09.abi3.so → _flash_attn2_cuda_abda3a0.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e67e00b79ea3625b0ec32a083544d1808cce682dfe593a7212c525a292d1764f
+oid sha256:c7d4d237f20ea46768cde96cc0c982809482b2f874aeb0678d4139d9d8663664
 size 1009055088

build/torch210-cxx11-cu130-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _flash_attn2_cuda_9f0ed09
-ops = torch.ops._flash_attn2_cuda_9f0ed09
+from . import _flash_attn2_cuda_abda3a0
+ops = torch.ops._flash_attn2_cuda_abda3a0
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_flash_attn2_cuda_9f0ed09::{op_name}"
+    return f"_flash_attn2_cuda_abda3a0::{op_name}"

build/torch210-cxx11-cu130-x86_64-linux/flash_attn_interface.py

@@ -33,8 +33,12 @@ def _get_device():
 
 
 def _get_block_size_n(device, head_dim, is_dropout, is_causal):
-    # This should match the block sizes in the CUDA kernel
     assert head_dim <= 256
+
+    if device.type == "xpu":
+        return 64
+
+    # This should match the block sizes in the CUDA kernel
     major, minor = torch.cuda.get_device_capability(device)
     is_sm8x = major == 8 and minor > 0  # Only include sm86 and sm89, exclude sm80 (A100)
     is_sm80 = major == 8 and minor == 0

build/{torch29-cxx11-xpu20252-x86_64-linux/_flash_attn2_xpu_9f0ed09.abi3.so → torch210-cxx11-xpu20253-x86_64-linux/_flash_attn2_xpu_abda3a0.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:262841040bd11d2ea11f317107fdc9484d864db0377b423eb9007f4c8a7eb74f
-size 13923672
+oid sha256:d4fa4fb6b42430c26695ce71f24cfd3562287538e838bd4fd951d1fc238f0a29
+size 15742736

build/torch210-cxx11-xpu20253-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _flash_attn2_xpu_9f0ed09
-ops = torch.ops._flash_attn2_xpu_9f0ed09
+from . import _flash_attn2_xpu_abda3a0
+ops = torch.ops._flash_attn2_xpu_abda3a0
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_flash_attn2_xpu_9f0ed09::{op_name}"
+    return f"_flash_attn2_xpu_abda3a0::{op_name}"

build/torch210-cxx11-xpu20253-x86_64-linux/flash_attn_interface.py

@@ -33,8 +33,12 @@ def _get_device():
 
 
 def _get_block_size_n(device, head_dim, is_dropout, is_causal):
-    # This should match the block sizes in the CUDA kernel
     assert head_dim <= 256
+
+    if device.type == "xpu":
+        return 64
+
+    # This should match the block sizes in the CUDA kernel
     major, minor = torch.cuda.get_device_capability(device)
     is_sm8x = major == 8 and minor > 0  # Only include sm86 and sm89, exclude sm80 (A100)
     is_sm80 = major == 8 and minor == 0

build/torch29-cxx11-cpu-x86_64-linux/{_flash_attn2_cpu_9f0ed09.abi3.so → _flash_attn2_cpu_abda3a0.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7313920d802d0477ca3d4144bc3b11e3c4761ea0e42b55a9b1b0b05567d23f71
+oid sha256:f48a5de0256294c0baf17e17ff3eafa7c818a6b70c869d6eec74cdff07b994f1
 size 1932200

build/torch29-cxx11-cpu-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _flash_attn2_cpu_9f0ed09
-ops = torch.ops._flash_attn2_cpu_9f0ed09
+from . import _flash_attn2_cpu_abda3a0
+ops = torch.ops._flash_attn2_cpu_abda3a0
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_flash_attn2_cpu_9f0ed09::{op_name}"
+    return f"_flash_attn2_cpu_abda3a0::{op_name}"

build/torch29-cxx11-cpu-x86_64-linux/flash_attn_interface.py

@@ -33,8 +33,12 @@ def _get_device():
 
 
 def _get_block_size_n(device, head_dim, is_dropout, is_causal):
-    # This should match the block sizes in the CUDA kernel
     assert head_dim <= 256
+
+    if device.type == "xpu":
+        return 64
+
+    # This should match the block sizes in the CUDA kernel
     major, minor = torch.cuda.get_device_capability(device)
     is_sm8x = major == 8 and minor > 0  # Only include sm86 and sm89, exclude sm80 (A100)
     is_sm80 = major == 8 and minor == 0

build/torch29-cxx11-cu126-x86_64-linux/{_flash_attn2_cuda_9f0ed09.abi3.so → _flash_attn2_cuda_abda3a0.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4d422421456bf5ac34486ee898a7a6aaea7fff2edda3bce062d0283f69806275
+oid sha256:9cb31caa7d395686694813cff9742e6a04c492af9bdebf80fd043bb98410d6d2
 size 448648752

build/torch29-cxx11-cu126-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _flash_attn2_cuda_9f0ed09
-ops = torch.ops._flash_attn2_cuda_9f0ed09
+from . import _flash_attn2_cuda_abda3a0
+ops = torch.ops._flash_attn2_cuda_abda3a0
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_flash_attn2_cuda_9f0ed09::{op_name}"
+    return f"_flash_attn2_cuda_abda3a0::{op_name}"

build/torch29-cxx11-cu126-x86_64-linux/flash_attn_interface.py

@@ -33,8 +33,12 @@ def _get_device():
 
 
 def _get_block_size_n(device, head_dim, is_dropout, is_causal):
-    # This should match the block sizes in the CUDA kernel
     assert head_dim <= 256
+
+    if device.type == "xpu":
+        return 64
+
+    # This should match the block sizes in the CUDA kernel
     major, minor = torch.cuda.get_device_capability(device)
     is_sm8x = major == 8 and minor > 0  # Only include sm86 and sm89, exclude sm80 (A100)
     is_sm80 = major == 8 and minor == 0

build/torch29-cxx11-cu128-x86_64-linux/{_flash_attn2_cuda_9f0ed09.abi3.so → _flash_attn2_cuda_abda3a0.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:29b22b8c1bbd77125b6c82aca5fecfe0416d2f116be7b1e1a4638f76fe542a2e
+oid sha256:d58e62ead713b97bf86390e13e331edb26dd00d83c34ac7e9a5e95968493fa11
 size 1037644632

build/torch29-cxx11-cu128-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _flash_attn2_cuda_9f0ed09
-ops = torch.ops._flash_attn2_cuda_9f0ed09
+from . import _flash_attn2_cuda_abda3a0
+ops = torch.ops._flash_attn2_cuda_abda3a0
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_flash_attn2_cuda_9f0ed09::{op_name}"
+    return f"_flash_attn2_cuda_abda3a0::{op_name}"

build/torch29-cxx11-cu128-x86_64-linux/flash_attn_interface.py

@@ -33,8 +33,12 @@ def _get_device():
 
 
 def _get_block_size_n(device, head_dim, is_dropout, is_causal):
-    # This should match the block sizes in the CUDA kernel
     assert head_dim <= 256
+
+    if device.type == "xpu":
+        return 64
+
+    # This should match the block sizes in the CUDA kernel
     major, minor = torch.cuda.get_device_capability(device)
     is_sm8x = major == 8 and minor > 0  # Only include sm86 and sm89, exclude sm80 (A100)
     is_sm80 = major == 8 and minor == 0

build/torch29-cxx11-cu130-x86_64-linux/{_flash_attn2_cuda_9f0ed09.abi3.so → _flash_attn2_cuda_abda3a0.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c9e000e77b2d5f5b8554c1ba0e1edfc173bd19d904b95eede3f9cc7ecefbcf89
+oid sha256:52ec533d0801a9757a28bbe0531535a4386d84058a4eb8a2b6a721040eb5fb61
 size 1009019192

build/torch29-cxx11-cu130-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _flash_attn2_cuda_9f0ed09
-ops = torch.ops._flash_attn2_cuda_9f0ed09
+from . import _flash_attn2_cuda_abda3a0
+ops = torch.ops._flash_attn2_cuda_abda3a0
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_flash_attn2_cuda_9f0ed09::{op_name}"
+    return f"_flash_attn2_cuda_abda3a0::{op_name}"

build/torch29-cxx11-cu130-x86_64-linux/flash_attn_interface.py

@@ -33,8 +33,12 @@ def _get_device():
 
 
 def _get_block_size_n(device, head_dim, is_dropout, is_causal):
-    # This should match the block sizes in the CUDA kernel
     assert head_dim <= 256
+
+    if device.type == "xpu":
+        return 64
+
+    # This should match the block sizes in the CUDA kernel
     major, minor = torch.cuda.get_device_capability(device)
     is_sm8x = major == 8 and minor > 0  # Only include sm86 and sm89, exclude sm80 (A100)
     is_sm80 = major == 8 and minor == 0

build/{torch210-cxx11-xpu20253-x86_64-linux/_flash_attn2_xpu_9f0ed09.abi3.so → torch29-cxx11-xpu20252-x86_64-linux/_flash_attn2_xpu_abda3a0.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:192cfb14df060c1fd913c7a6e5a588da44dec07d0a9adf156ed377d860c8d3c2
-size 15436096
+oid sha256:6a96bf55da7a5422a5ca6f06e2deb147962be6d78db5e9569541e8257460ee19
+size 14205096

build/torch29-cxx11-xpu20252-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _flash_attn2_xpu_9f0ed09
-ops = torch.ops._flash_attn2_xpu_9f0ed09
+from . import _flash_attn2_xpu_abda3a0
+ops = torch.ops._flash_attn2_xpu_abda3a0
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_flash_attn2_xpu_9f0ed09::{op_name}"
+    return f"_flash_attn2_xpu_abda3a0::{op_name}"

build/torch29-cxx11-xpu20252-x86_64-linux/flash_attn_interface.py

@@ -33,8 +33,12 @@ def _get_device():
 
 
 def _get_block_size_n(device, head_dim, is_dropout, is_causal):
-    # This should match the block sizes in the CUDA kernel
     assert head_dim <= 256
+
+    if device.type == "xpu":
+        return 64
+
+    # This should match the block sizes in the CUDA kernel
     major, minor = torch.cuda.get_device_capability(device)
     is_sm8x = major == 8 and minor > 0  # Only include sm86 and sm89, exclude sm80 (A100)
     is_sm80 = major == 8 and minor == 0