upgrade to b6188

[xkong-anaconda]

llama.cpp 0.0.6188

Destination channel: defaults

Links

• PKG-10893
• Upstream repository
• Upstream release b6188
• Upstream diff b6082...b6188
• Relevant dependency PRs:
  • Related to llama-cpp-python 0.3.16 compatibility requirement

CUDA Variants

• Linux CUDA Log
• Windows CUDA Log

Explanation of changes:

This PR upgrades llama.cpp from b6082 to b6188 on the main_b6082 preservation branch. This is a special branch structure to provide b6188 packages for llama-cpp-python 0.3.16 compatibility without affecting the current main feedstock

Why b6188?

• llama-cpp-python 0.3.16 requires llama.cpp b6173-b6238 (API compatibility range)
• The llama_get_kv_self() and related APIs were removed in b6239
• b6188 is a stable release within the compatible range

[Callek]

FWIW looks like there is an issue, our preprod caught:

            {'ctx': {'error': {}},
             'input': ['--suppress-variables',
                       '--skip-existing',
                       '--error-overlinking',
                       '--variants "{skip_cuda_prefect: True}"'],
             'loc': ['body', 'tasks', 14, 'build', 'build_options'],
             'msg': "Value error, Invalid build_options: ['--variants "
                    '"{skip_cuda_prefect: True}"\']. Allowed: [\'--debug\', '
                    "'--error-overdepending', '--error-overlinking', "
                    "'--no-error-overlinking', '--no-include-recipe', "
                    "'--no-test', '--no-verify', '--quiet', '--skip-existing', "
                    "'--suppress-variables']",
             'type': 'value_error'},

(which might be why we don't see anything triggered in prod)

Looks like that exists as part of the abs.yaml in the root of the repo. I've flagged @markan to help determine from a slack thread in our private team channel if thats a regression or a feature of it being flagged in the cli generation code

[xkong-anaconda]

FWIW looks like there is an issue, our preprod caught:

            {'ctx': {'error': {}},
             'input': ['--suppress-variables',
                       '--skip-existing',
                       '--error-overlinking',
                       '--variants "{skip_cuda_prefect: True}"'],
             'loc': ['body', 'tasks', 14, 'build', 'build_options'],
             'msg': "Value error, Invalid build_options: ['--variants "
                    '"{skip_cuda_prefect: True}"\']. Allowed: [\'--debug\', '
                    "'--error-overdepending', '--error-overlinking', "
                    "'--no-error-overlinking', '--no-include-recipe', "
                    "'--no-test', '--no-verify', '--quiet', '--skip-existing', "
                    "'--suppress-variables']",
             'type': 'value_error'},

(which might be why we don't see anything triggered in prod)

Looks like that exists as part of the abs.yaml in the root of the repo. I've flagged @markan to help determine from a slack thread in our private team channel if thats a regression or a feature of it being flagged in the cli generation code

Thank you very much for the help and inestigation.

[cbouss]

What kind of failures were you seeing there? (for metal and cuda)
Do you still have logs?

[xkong-anaconda]

Unfortunately, I don't have the original test failure logs anymore. But it appeared to be related to backend operations not properly working with CUDA. Metal GPU builds fail test-backend-ops with Flash Attention operations producing "not supported [Metal]" errors. and Win failures with values 20-30x higher than tolerance.
This skip is specific to the b6188 downgrade, not intended for the main feedstock
If you need the exact failure details, I could re-run the build without the skip to capture the errors
Let me know if you need me to investigate further!

[cbouss]

Yes, please re-run, as the packages go to the main channel, we better check as much as possible.

You can add this line to the script to capture test-backend-ops logs without failing the whole build:
ctest -L main -C Release --output-on-failure -j${CPU_COUNT} --timeout 900 -R "test-backend-ops" || true