Downgrate to b6188

[xkong-anaconda]

llama.cpp 0.0.6188

Destination channel: defaults

Links

• Upstream repository
• Upstream changelog/diff
• Relevant dependency PRs:
  • llama-cpp-python-feedstock PR (pending - requires this package)

Explanation of changes:

• Downgrade to b6188 for compatibility with llama-cpp-python 0.3.16

  • The llama_get_kv_self() API and related llama_kv_self_* functions were removed in b6239 (PR
    #15472)
  • llama-cpp-python 0.3.16 requires these APIs (compatible range: b6173-b6238)

• Regenerated patches for b6188:

  • increase-nmse-tolerance.patch - Updated line numbers (5 hunks instead of 7)
  • increase-nmse-tolerance-aarch64.patch - Updated line numbers
  • mkl.patch - Updated for unquoted CMake variable syntax
  • metal_gpu_selection.patch - Targets ggml-metal.m (b6188 structure)

• Removed patches not needed for b6188:

  • disable-metal-bf16.patch - BF16 is OFF by default in b6188
  • disable-metal-flash-attention.patch - Not needed with BF16 disabled