ik-llama-cpp-python 0.1.0__tar.gz → 0.1.2__tar.gz

{ik_llama_cpp_python-0.1.0 → ik_llama_cpp_python-0.1.2}/.github/workflows/publish.yml

@@ -119,10 +119,11 @@ jobs:
             CMAKE_BUILD_PARALLEL_LEVEL=4
             CXXFLAGS="-include cstdint -mcmodel=medium"
             CFLAGS="-mcmodel=medium"
+            CUDAFLAGS="-Xcompiler=-mcmodel=medium"
           CIBW_BEFORE_BUILD: >-
             sed -i 's/name = "ik-llama-cpp-python"/name = "ik-llama-cpp-python-cuda"/' pyproject.toml
           CIBW_CONFIG_SETTINGS: >-
-            cmake.args=-DGGML_NATIVE=OFF;-DBUILD_SHARED_LIBS=ON;-DGGML_CUDA=ON;-DGGML_AVX2=ON;-DGGML_FMA=ON;-DGGML_F16C=ON
+            cmake.args=-DGGML_NATIVE=OFF;-DBUILD_SHARED_LIBS=ON;-DGGML_CUDA=ON;-DGGML_AVX2=ON;-DGGML_FMA=ON;-DGGML_F16C=ON;-DCMAKE_CUDA_ARCHITECTURES=89-real
           CIBW_REPAIR_WHEEL_COMMAND_LINUX: >-
             bash -c 'TMPWHL=$(mktemp -d) && unzip -q {wheel} -d "$TMPWHL" && LD_LIBRARY_PATH="$TMPWHL/ik_llama_cpp/lib:$LD_LIBRARY_PATH" auditwheel repair -w {dest_dir} {wheel} --exclude libcuda.so.1 --exclude libnvidia-ml.so.1 && rm -rf "$TMPWHL"'
 

{ik_llama_cpp_python-0.1.0 → ik_llama_cpp_python-0.1.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ik-llama-cpp-python
-Version: 0.1.0
+Version: 0.1.2
 Summary: Python bindings for ik_llama.cpp — high-performance llama.cpp fork
 License-Expression: MIT
 Classifier: Programming Language :: Python :: 3

{ik_llama_cpp_python-0.1.0 → ik_llama_cpp_python-0.1.2}/ik_llama_cpp/_ctypes_api.py

@@ -342,6 +342,12 @@ def llama_batch_free(batch: llama_batch) -> None: ...
 def llama_decode(ctx: int, batch: llama_batch) -> int: ...
 
 
+# -- KV cache --
+
+@_cfunc("llama_kv_cache_clear", [ctypes.c_void_p], None)
+def llama_kv_cache_clear(ctx: int) -> None: ...
+
+
 # -- Logits --
 
 @_cfunc("llama_get_logits_ith", [ctypes.c_void_p, ctypes.c_int32], ctypes.POINTER(ctypes.c_float))

{ik_llama_cpp_python-0.1.0 → ik_llama_cpp_python-0.1.2}/ik_llama_cpp/_internals.py

@@ -105,6 +105,9 @@ class IkContext:
     def decode(self, batch: C.llama_batch) -> int:
         return C.llama_decode(self._ctx, batch)
 
+    def kv_cache_clear(self):
+        C.llama_kv_cache_clear(self._ctx)
+
     def get_logits(self, idx: int = -1):
         return C.llama_get_logits_ith(self._ctx, idx)
 

{ik_llama_cpp_python-0.1.0 → ik_llama_cpp_python-0.1.2}/ik_llama_cpp/llama.py

@@ -105,6 +105,7 @@ class IkLlama:
         top_p: float = 0.95,
     ) -> list[int]:
         """Generate tokens from a prompt token list. Returns generated token ids."""
+        self._context.kv_cache_clear()
         self._context.perf_reset()
 
         n_ubatch = self._context._n_ubatch

{ik_llama_cpp_python-0.1.0 → ik_llama_cpp_python-0.1.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "scikit_build_core.build"
 
 [project]
 name = "ik-llama-cpp-python"
-version = "0.1.0"
+version = "0.1.2"
 description = "Python bindings for ik_llama.cpp — high-performance llama.cpp fork"
 readme = "README.md"
 license = "MIT"