Migration to rattler-build, added specific Python tests

conda-forge.yml

@@ -1,5 +1,6 @@
 build_platform:
   osx_arm64: osx_64
+conda_build_tool: rattler-build
 conda_build:
   error_overlinking: true
 conda_forge_output_validation: true

recipe/recipe.yaml

@@ -0,0 +1,98 @@
+schema_version: 1
+
+context:
+  name: llama-cpp-python
+  version: "0.3.32"
+  # NOTE: VERIFY llama_cpp_version before merging!
+  llama_cpp_version: "9851"
+
+package:
+  name: ${{ name|lower }}
+  version: ${{ version }}
+
+source:
+  url: https://pypi.org/packages/source/${{ name[0] }}/${{ name }}/llama_cpp_python-${{ version }}.tar.gz
+  sha256: b06502361770f82eb08b7f1a192eb084b9ead2b88fe32cda8c397a2782eabde6
+  patches:
+    # Asks cdll to look for the library in the path as well.
+    - 0001-Adapt-shared-library-relocation.patch
+
+build:
+  number: 1
+  script:
+    - if: win
+      then:
+        - set "CMAKE_ARGS=%CMAKE_ARGS% -DLLAMA_BUILD=OFF -DLLAVA_BUILD=OFF"
+        - ${{ PYTHON }} -m pip install . -vv
+      else:
+        - export CMAKE_ARGS="${CMAKE_ARGS} -DLLAMA_BUILD=OFF -DLLAVA_BUILD=OFF"
+        - ${{ PYTHON }} -m pip install . -vv
+
+requirements:
+  build:
+    - if: build_platform != target_platform
+      then:
+        - python
+        - cross-python_${{ target_platform }}
+    - ${{ compiler('c') }}
+    - ${{ stdlib('c') }}
+    - ${{ compiler('cxx') }}
+    - cmake
+    - make
+    - pkgconfig
+  host:
+    - python
+    - scikit-build-core >=0.5.1
+    - pip
+  run:
+    - python
+    - typing-extensions >=4.5.0
+    - numpy >=1.20.0
+    - diskcache >=5.6.1
+    - pyyaml >=5.1
+    - llama.cpp ==${{ llama_cpp_version }}
+    # Split into llama-cpp-python-server
+    - uvicorn >=0.22.0
+    - fastapi >=0.100.0
+    - pydantic-settings >=2.0.1
+    - sse-starlette >=1.6.1
+    - starlette-context >=0.3.6,<0.4
+
+tests:
+  - python:
+      imports:
+        - llama_cpp
+      pip_check: true
+  - script:
+      # Exercises the bindings and calls into libllama.
+      - python test.py
+      # Downloads a tiny GGUF model and runs a full load/tokenize/inference/embed check.
+      - if: win
+        then:
+          - curl.exe -L --retry 3 -o tinystories.gguf https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K.gguf
+        else:
+          - curl -L --retry 3 -o tinystories.gguf https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K.gguf
+      - python test_gguf.py tinystories.gguf
+    requirements:
+      run:
+        - curl
+    files:
+      recipe:
+        - test.py
+        - test_gguf.py
+
+about:
+  homepage: https://github.com/abetlen/llama-cpp-python
+  summary: Python bindings for the llama.cpp library
+  license: MIT
+  license_file:
+    - LICENSE.md
+
+extra:
+  recipe-maintainers:
+    - JohanMabille
+    - jjerphan
+    - jonashaag
+    - YYYasin19
+    - sodre
+    - toniher

recipe/test.py

@@ -0,0 +1,4 @@
+import llama_cpp
+print(llama_cpp.__version__)
+print(llama_cpp.llama_backend_init())  # calls into libllama
+llama_cpp.llama_backend_free()

recipe/test_gguf.py

@@ -0,0 +1,51 @@
+#!/usr/bin/env python3
+import sys
+import time
+
+def main():
+    if len(sys.argv) < 2:
+        print(f"Usage: {sys.argv[0]} <path/to/model.gguf>", file=sys.stderr)
+        sys.exit(1)
+
+    model_path = sys.argv[1]
+
+    print("1. Importing llama_cpp ...", end=" ", flush=True)
+    from llama_cpp import Llama
+    print("OK")
+
+    print(f"2. Loading model: {model_path} ...", end=" ", flush=True)
+    t0 = time.time()
+    llm = Llama(model_path=model_path, n_ctx=512, verbose=False)
+    print(f"OK ({time.time() - t0:.1f}s)")
+
+    print("3. Tokenization ...", end=" ", flush=True)
+    tokens = llm.tokenize(b"Hello, world!")
+    assert len(tokens) > 0
+    print(f"OK ({len(tokens)} tokens)")
+
+    print("4. Inference ...", end=" ", flush=True)
+    t0 = time.time()
+    output = llm(
+        "Q: What is the capital of France? A:",
+        max_tokens=16,
+        stop=["Q:", "\n"],
+        echo=False,
+    )
+    elapsed = time.time() - t0
+    text = output["choices"][0]["text"].strip()
+    print(f"OK ({elapsed:.1f}s)")
+    print(f"   Response: {text!r}")
+
+    print("5. Embedding ...", end=" ", flush=True)
+    try:
+        emb_llm = Llama(model_path=model_path, embedding=True, n_ctx=64, verbose=False)
+        vec = emb_llm.embed("test sentence")
+        assert len(vec) > 0
+        print(f"OK (dim={len(vec)})")
+    except Exception as e:
+        print(f"SKIP ({e})")
+
+    print("\nAll checks passed.")
+
+if __name__ == "__main__":
+    main()