Cordell Bloor:
Hi Ximin,
On 2025-12-01 10:27, Ximin Luo wrote:
Ah, I last downloaded the weights file back in 2021, the upstream project may
have moved on in the meantime. I've uploaded my old one here:
https://people.debian.org/~infinity0/leelaz/
You can drop it into ~/.local/share/leela-zero/best-network then it should work.
Thank you. I think I was trying to use weights from Leela Chess Zero, rather
than from Leela Zero.
I now realize that my environment was not clean enough when I was doing my
initial testing. I've not reproduced the problem you've reported, but I've
noticed that rocm-opencl-icd fails to declare its dependency on libamd-comg3.
I noticed the lack of a direct dependency, but I wasn't sure if it would be
pulled in indirectly via the web of dependencies. If rocm-opencl-cid does not
*directly* depend on libamd-comgr3 at runtime, then it should not declare this
Depends. Unfortunately I do not know enough to judge this one, you guys will
have to figure it out.
What I do know is, the error message I reported appears to be a *direct*
dependency from a comgr-created file
(/tmp/comgr-e99bbb/include/opencl1.2-c.pch) that *directly references*
/usr/lib/llvm-21/lib/clang/21/include/opencl-c-base.h. Now this could be
because it uses some LLVM tool that includes this path as part of how it works;
but the end result is still that comgr3 *directly depends*, at runtime, on
libclang-common-21-dev.
Is there something I need to enter at the "Leela: " prompt to trigger the error
message you described?
No the error occurs as soon as you run leelaz, you don't need to do anything.
Did you forget to *uninstall* libclang-common-21-dev to trigger the error?
~~~~
$ apt policy libclang-common-21-dev
libclang-common-21-dev:
Installed: (none)
[..]
$ leelaz
Using OpenCL batch size of 5
Using 10 thread(s).
RNG seed: 3817374786448165985
Leela Zero 0.17 Copyright (C) 2017-2019 Gian-Carlo Pascutto and contributors
This program comes with ABSOLUTELY NO WARRANTY.
This is free software, and you are welcome to redistribute it
under certain conditions; see the COPYING file for details.
BLAS Core: built-in Eigen 3.4.0 library.
Detecting residual layers...v1...256 channels...40 blocks.
Initializing OpenCL (autodetecting precision).
Detected 1 OpenCL platforms.
Platform version: OpenCL 2.1 AMD-APP (3649.0)
Platform profile: FULL_PROFILE
Platform name: AMD Accelerated Parallel Processing
Platform vendor: Advanced Micro Devices, Inc.
Device ID: 0
Device name: gfx1101
Device type: GPU
Device vendor: Advanced Micro Devices, Inc.
[..]
Selected platform: AMD Accelerated Parallel Processing
Selected device: gfx1101
with OpenCL 2.1 capability.
Half precision compute support: Yes.
Tensor Core support: No.
OpenCL: using fp16/half or tensor core compute support.
Loaded existing SGEMM tuning.
Error building kernels: fatal error: malformed or corrupted precompiled file:
'could not find file '/usr/lib/llvm-21/lib/clang/21/include/opencl-c-base.h'
referenced by AST file '/tmp/comgr-92c557/include/opencl1.2-c.pch''
1 error generated.
Error: Failed to compile source (from CL or HIP source to LLVM IR).
OpenCL: fp16/half or tensor core failed despite driver claiming support.
Falling back to single precision
Detected 1 OpenCL platforms.
[.. same thing again, attempting single precision ..]
Selected platform: AMD Accelerated Parallel Processing
Selected device: gfx1101
with OpenCL 2.1 capability.
Half precision compute support: Yes.
Tensor Core support: No.
Loaded existing SGEMM tuning.
Error building kernels: fatal error: malformed or corrupted precompiled file:
'could not find file '/usr/lib/llvm-21/lib/clang/21/include/opencl-c-base.h'
referenced by AST file '/tmp/comgr-dc51bb/include/opencl1.2-c.pch''
1 error generated.
Error: Failed to compile source (from CL or HIP source to LLVM IR).
terminate called after throwing an instance of 'std::runtime_error'
what(): Error building OpenCL kernels.
Aborted (core dumped) leelaz
exit code 134
~~~~
Best,
X
--
GPG: ed25519/56034877E1F87C35
https://github.com/infinity0/pubkeys.git
