Bug 953101

Summary:	sci-ml/caffe2-2.6.0[rocm] fails to build: device_batched_gemm_multiple_d_xdl_cshuffle_v3.hpp' file not found
Product:	Gentoo Linux	Reporter:	Sv. Lockal <lockalsash>
Component:	Current packages	Assignee:	Tupone Alfredo <tupone>
Status:	RESOLVED FIXED
Severity:	normal	CC:	gentoo, tupone
Priority:	Normal	Keywords:	PullRequest
Version:	unspecified
Hardware:	All
OS:	Linux
See Also:	https://github.com/gentoo/gentoo/pull/41448
Whiteboard:
Package list:		Runtime testing required:	---

Description Sv. Lockal 2025-04-03 18:48:26 UTC

sci-ml/caffe2-2.6.0 with ROCm support fails to build with:

In file included from /var/tmp/portage/sci-ml/caffe2-2.6.0-r2/work/pytorch-2.6.0/aten/src/ATen/native/hip/bgemm_kernels/bgemm_kernel_bf16bf16bf16_128_16x32x64_16x16_1x1_8x16x1_8x16x1_1x16x1x8_4_Intrawave_v1.hip:3:
/var/tmp/portage/sci-ml/caffe2-2.6.0-r2/work/pytorch-2.6.0/aten/src/ATen/native/hip/bgemm_kernels/bgemm_kernel_template.h:11:10: fatal error: 'ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_xdl_cshuffle_v3.hpp' file not found
   11 | #include <ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_xdl_cshuffle_v3.hpp>

Reason: https://github.com/pytorch/pytorch/tree/v2.6.0/third_party depends on not-yet-released version of composable_kernel. 

Latest version at this moment - 6.3.3 - does not include device_batched_gemm_multiple_d_xdl_cshuffle_v3.hpp (this file exists in develop branch, so probably will be released later).

This is one of multiple issues, I'll provide fix in PR soon.

Comment 1 Larry the Git Cow gentoo-dev

2025-04-05 12:55:46 UTC

The bug has been closed via the following commit(s):

https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=0b45f8a6d7fae82791efe2caf4934b25d6138f4f

commit 0b45f8a6d7fae82791efe2caf4934b25d6138f4f
Author:     Sv. Lockal <lockalsash@gmail.com>
AuthorDate: 2025-04-03 18:54:54 +0000
Commit:     Alfredo Tupone <tupone@gentoo.org>
CommitDate: 2025-04-05 12:52:52 +0000

    sci-ml/caffe2: fix build for ROCm
    
    * Revised the fix for `-fclang-abi-compat=17`.
    * Added fixes for libc++
    
    Closes: https://bugs.gentoo.org/953101
    Signed-off-by: Sv. Lockal <lockalsash@gmail.com>
    Closes: https://github.com/gentoo/gentoo/pull/41448
    Signed-off-by: Alfredo Tupone <tupone@gentoo.org>

 sci-ml/caffe2/Manifest                             |  1 +
 ...ffe2-2.6.0-r2.ebuild => caffe2-2.6.0-r3.ebuild} | 64 +++++++++++++---------
 .../files/caffe2-2.6.0-rocm-fix-std-cpp17.patch    | 50 +++++++++++++++++
 3 files changed, 89 insertions(+), 26 deletions(-)