Skip to content

Commit 5a3823c

Browse files
gujinghuigujinghui
authored
Reorg and update torch patches for PyTorch 1.10 (#1695)
Signed-off-by: Gu, Jinghui <jinghui.gu@intel.com>
1 parent 2e13f97 commit 5a3823c

File tree

37 files changed

+245
-1697
lines changed

37 files changed

+245
-1697
lines changed

‎torch_patches/0001-add-SparseXPU-to-dispatch-key-set-25.patch‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,7 @@
1-
From c01f049d638a48e2ac01ded9baaca3fb8589107c Mon Sep 17 00:00:00 2001
1+
From b27bdfe9139a84c762aab173f29991cd1b2aa9ab Mon Sep 17 00:00:00 2001
22
From: majing921201 <jing1.ma@intel.com>
33
Date: Fri, 11 Feb 2022 10:16:42 +0800
4-
Subject: [PATCH 01/36] add SparseXPU to dispatch key set (#25)
4+
Subject: [PATCH 01/28] add SparseXPU to dispatch key set (#25)
55

66
Signed-off-by: Ma, Jing1 <jing1.ma@intel.com>
77
---

‎torch_patches/0002-Add-QuantizedXPU-to-DispatchKeySet-28.patch‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,7 @@
1-
From a642474c362816032917045b138adaa185d1bf0b Mon Sep 17 00:00:00 2001
1+
From 815d3da1ef69a241269d9f1b51e6ac03b0303286 Mon Sep 17 00:00:00 2001
22
From: xiaolil1 <xiaoli.liu@intel.com>
33
Date: Tue, 15 Feb 2022 14:01:53 +0800
4-
Subject: [PATCH 02/36] Add QuantizedXPU to DispatchKeySet (#28)
4+
Subject: [PATCH 02/28] Add QuantizedXPU to DispatchKeySet (#28)
55

66
---
77
c10/core/DispatchKeySet.h | 1 +

‎torch_patches/0003-channels-last-1d-for-rebase-pytorch-1.10-29.patch‎

Lines changed: 0 additions & 512 deletions
This file was deleted.

torch_patches/0006-enable-channels-last-for-xpu-group-norm-32.patch renamed to torch_patches/0003-enable-channels-last-for-xpu-group-norm-32.patch

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,7 @@
1-
From 206c68c039d43655648a0f8cf3c07025355a3fdf Mon Sep 17 00:00:00 2001
1+
From a7b5ef1005cbed1ebe74a411386b7bb1dd31b65e Mon Sep 17 00:00:00 2001
22
From: KevinYuk <kevin.yu@intel.com>
33
Date: Wed, 2 Mar 2022 10:45:52 +0800
4-
Subject: [PATCH 06/36] enable channels last for xpu group norm (#32)
4+
Subject: [PATCH 03/28] enable channels last for xpu group norm (#32)
55

66
---
77
aten/src/ATen/native/group_norm.cpp | 2 +-

torch_patches/0009-upgrade-protobuf-from-tag-v3.13.0.1-d1eca4e4-to-v3.1.patch renamed to torch_patches/0004-upgrade-protobuf-from-tag-v3.13.0.1-d1eca4e4-to-v3.1.patch

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,7 @@
1-
From c9fdbae991eabd295a7fd8576453e702e023c7f0 Mon Sep 17 00:00:00 2001
1+
From 0aced766f1ed27e4d3265c01d72e2512e0d7da39 Mon Sep 17 00:00:00 2001
22
From: Guo Yejun <yejun.guo@intel.com>
33
Date: Mon, 21 Mar 2022 18:56:25 +0800
4-
Subject: [PATCH 09/36] upgrade protobuf from tag: v3.13.0.1 (d1eca4e4) to
4+
Subject: [PATCH 04/28] upgrade protobuf from tag: v3.13.0.1 (d1eca4e4) to
55
v3.18.2 (6c6b0778) (#34)
66

77
due to the vulnerability discovered at

torch_patches/0004-JIT-disable-decomposition-of-BN-LN-for-XPU-31.patch renamed to torch_patches/0005-adjust-conditions-to-enable-jit-decomposition-pass-o.patch

Lines changed: 18 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -1,23 +1,36 @@
1-
From fac1bb693309722b047abf8c13efc81548927fbf Mon Sep 17 00:00:00 2001
1+
From e3e3a887f1e2fe1a2c5ceb1530b1c79b2a5e9beb Mon Sep 17 00:00:00 2001
22
From: Feng Yuan <feng1.yuan@intel.com>
3-
Date: Thu, 24 Feb 2022 15:23:23 +0800
4-
Subject: [PATCH 04/36] JIT: disable decomposition of BN/LN for XPU (#31)
3+
Date: Mon, 21 Mar 2022 14:23:28 -0700
4+
Subject: [PATCH 05/28] adjust conditions to enable jit decomposition pass only
5+
for GPU device (#73637)
56

7+
Summary:
68
Signed-off-by: Feng Yuan <feng1.yuan@intel.com>
9+
10+
Pull Request resolved: https://github.com/pytorch/pytorch/pull/73637
11+
12+
Reviewed By: cpuhrsch
13+
14+
Differential Revision: D34931641
15+
16+
Pulled By: davidberard98
17+
18+
fbshipit-source-id: 849ace475eb9038912e902462f0eb2cdebc14ada
19+
(cherry picked from commit ae61c4fba20345c0122f24e02c661b0da29f0ee4)
720
---
821
torch/csrc/jit/passes/decompose_ops.cpp | 2 +-
922
1 file changed, 1 insertion(+), 1 deletion(-)
1023

1124
diff --git a/torch/csrc/jit/passes/decompose_ops.cpp b/torch/csrc/jit/passes/decompose_ops.cpp
12-
index 0706c9c14a..e2db04b0aa 100644
25+
index 0706c9c14a..4223c72658 100644
1326
--- a/torch/csrc/jit/passes/decompose_ops.cpp
1427
+++ b/torch/csrc/jit/passes/decompose_ops.cpp
1528
@@ -42,7 +42,7 @@ bool isDecomposableNorm(Node* normalize_op) {
1629
auto device = input->type()->expectRef<TensorType>().device();
1730
// As of now, we do the decomposition for batchnorm/layernorm on GPU device
1831
// only
1932
- if (!device || (*device).is_cpu()) {
20-
+ if (!device || (*device).is_cpu() || (*device).is_xpu()) {
33+
+ if (!device || !(*device).is_cuda()) {
2134
return false;
2235
}
2336

torch_patches/0005-Add-xpu-for-ddp-fix-the-index-error-in-bert-training.patch renamed to torch_patches/0006-Add-xpu-for-ddp-fix-the-index-error-in-bert-training.patch

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,7 @@
1-
From 916d02cf3318858f56c68d7ec6c336eea7f75e87 Mon Sep 17 00:00:00 2001
1+
From 9a60b998821f8f5ea8547ea28a80c2be1f74c875 Mon Sep 17 00:00:00 2001
22
From: zhuhong61 <95205772+zhuhong61@users.noreply.github.com>
33
Date: Tue, 1 Mar 2022 14:24:08 +0800
4-
Subject: [PATCH 05/36] Add xpu for ddp (fix the index error in bert training)
4+
Subject: [PATCH 06/28] Add xpu for ddp (fix the index error in bert training)
55
(#27)
66

77
---

torch_patches/0011-add-python-API-to-print-all-operators-that-have-kern.patch renamed to torch_patches/0007-add-python-API-to-print-all-operators-that-have-kern.patch

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,7 @@
1-
From ca942ce28515a826d70a515ddeebd10076ae0b6a Mon Sep 17 00:00:00 2001
1+
From e28446e89b1484c0c7fb360bfae038e2b46006ae Mon Sep 17 00:00:00 2001
22
From: zhuhong61 <95205772+zhuhong61@users.noreply.github.com>
33
Date: Thu, 14 Apr 2022 11:25:28 +0800
4-
Subject: [PATCH 11/36] add python API to print all operators that have kernels
4+
Subject: [PATCH 07/28] add python API to print all operators that have kernels
55
registered to a particular DispatchKey (#63575) (#38)
66

77
Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/63575

torch_patches/0012-More-distutils.version.LooseVersion-changes-69947-43.patch renamed to torch_patches/0008-More-distutils.version.LooseVersion-changes-69947-43.patch

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,7 @@
1-
From cb9f926ad257791545466cc6be1774df19828ef3 Mon Sep 17 00:00:00 2001
1+
From 377e0459f6fe5a455ccefb635d250a40445f01dc Mon Sep 17 00:00:00 2001
22
From: Kim <kimyangbaochen@vip.qq.com>
33
Date: Mon, 18 Apr 2022 16:30:24 +0800
4-
Subject: [PATCH 12/36] More distutils.version.LooseVersion changes (#69947)
4+
Subject: [PATCH 08/28] More distutils.version.LooseVersion changes (#69947)
55
(#43)
66

77
Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/69947

torch_patches/0014-Fix-pytorch-issue-in-sparce_coo_tensor-46.patch renamed to torch_patches/0009-Fix-issue-in-sparce_coo_tensor-only-supporting-CUDA-.patch

Lines changed: 12 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -1,8 +1,16 @@
1-
From 5d95de97cb99bf541a85f89212b8417ee1970017 Mon Sep 17 00:00:00 2001
2-
From: chengjunlu <chengjun.lu@intel.com>
3-
Date: Mon, 9 May 2022 09:20:27 +0800
4-
Subject: [PATCH 14/36] Fix pytorch issue in sparce_coo_tensor (#46)
1+
From 09a779feae2c14ef309e5057229cce4ed7bbf5df Mon Sep 17 00:00:00 2001
2+
From: johnlu <chengjun.lu@intel.com>
3+
Date: Tue, 26 Apr 2022 00:56:10 +0000
4+
Subject: [PATCH 09/28] Fix issue in sparce_coo_tensor only supporting CUDA
5+
device.
56

7+
## Motivation
8+
The at::native::_validate_sparse_coo_tensor_args only supports checking the indices on CUDA device and CPU device. To extend the function to support more device type.
9+
10+
## Solution
11+
Copy the indices to the CPU to validate the correctness.
12+
Pull Request resolved: https://github.com/pytorch/pytorch/pull/76293
13+
Approved by: https://github.com/mrshenli
614
---
715
aten/src/ATen/native/sparse/SparseTensor.cpp | 2 +-
816
1 file changed, 1 insertion(+), 1 deletion(-)

0 commit comments

Comments
 (0)