diff options
| -rw-r--r-- | Makefile | 6 | ||||
| -rw-r--r-- | embed_cuda.h | 29 |
2 files changed, 23 insertions, 12 deletions
@@ -136,7 +136,7 @@ cuda: # -arch=native targets the GPU present at build time (we always build CUDA # on the machine that runs it), avoiding first-load PTX JIT. embed_cuda.o: embed_cuda.cu embed_cuda.h embed_internal.h embed.h - $(NVCC) -O3 -std=c++17 -arch=native -Xcompiler -fPIC -DUSE_BLAS -DUSE_OPENBLAS -DUSE_CUDA -I/usr/include/openblas -I$(CUDA_HOME)/include -x cu -c -o $@ $< + $(NVCC) -O3 -std=c++17 -arch=native -Xcompiler -fPIC,-fvisibility=hidden -DUSE_BLAS -DUSE_OPENBLAS -DUSE_CUDA -I/usr/include/openblas -I$(CUDA_HOME)/include -x cu -c -o $@ $< # ============================================================================= # Shared library (built by every backend target from the same objects) @@ -197,7 +197,7 @@ test: ./tests/test_late $(CC) $(TEST_CC_FLAGS) $(TEST_BLAS_CFLAGS) -o tests/cli_under_test \ $(TEST_CLI_BIN_SRCS) -lm -lpthread $(TEST_BLAS_LDFLAGS) - $(CC) $(TEST_CC_FLAGS) -o tests/test_cli tests/test_cli.c + $(CC) $(TEST_CC_FLAGS) -o tests/test_cli tests/test_cli.c -lm ./tests/test_cli ./tests/cli_under_test $(CC) $(TEST_CC_FLAGS) $(TEST_BLAS_CFLAGS) $(CJSON_CFLAGS) -Ideps/ae \ -o tests/test_server $(TEST_SERVER_SRCS) \ @@ -255,7 +255,7 @@ coverage: LLVM_PROFILE_FILE=$(COV_DIR)/late.profraw ./tests/test_late $(CC) $(TEST_CC_FLAGS) $(COV_FLAGS) $(TEST_BLAS_CFLAGS) -o tests/cli_under_test \ $(TEST_CLI_BIN_SRCS) -lm -lpthread $(TEST_BLAS_LDFLAGS) - $(CC) $(TEST_CC_FLAGS) -o tests/test_cli tests/test_cli.c + $(CC) $(TEST_CC_FLAGS) -o tests/test_cli tests/test_cli.c -lm LLVM_PROFILE_FILE=$(COV_DIR)/cli_%p.profraw ./tests/test_cli ./tests/cli_under_test $(CC) $(TEST_CC_FLAGS) $(COV_FLAGS) $(TEST_BLAS_CFLAGS) $(CJSON_CFLAGS) -Ideps/ae \ -o tests/test_server $(TEST_SERVER_SRCS) \ diff --git a/embed_cuda.h b/embed_cuda.h index 8ebe6e2..3c9011d 100644 --- a/embed_cuda.h +++ b/embed_cuda.h @@ -5,25 +5,36 @@ #ifndef PPLX_EMBED_CUDA_H #define PPLX_EMBED_CUDA_H +/* Public-API export annotation. The libraries are built with + * -fvisibility=hidden, so only declarations carrying PPLX_API are exported + * from the shared library; everything else stays internal. */ +#ifndef PPLX_API +#if defined(__GNUC__) +#define PPLX_API __attribute__((visibility("default"))) +#else +#define PPLX_API +#endif +#endif + #include "embed.h" typedef struct pplx_cuda_ctx pplx_cuda_ctx_t; -int pplx_cuda_set_fast_gemm(const char *mode); -int pplx_cuda_set_weights_bf16(int on); +PPLX_API int pplx_cuda_set_fast_gemm(const char *mode); +PPLX_API int pplx_cuda_set_weights_bf16(int on); -pplx_cuda_ctx_t *pplx_cuda_load(const char *model_dir); -void pplx_cuda_free(pplx_cuda_ctx_t *ctx); +PPLX_API pplx_cuda_ctx_t *pplx_cuda_load(const char *model_dir); +PPLX_API void pplx_cuda_free(pplx_cuda_ctx_t *ctx); -const pplx_config_t *pplx_cuda_config(const pplx_cuda_ctx_t *ctx); +PPLX_API const pplx_config_t *pplx_cuda_config(const pplx_cuda_ctx_t *ctx); -int pplx_cuda_embed_into(pplx_cuda_ctx_t *ctx, const int *token_ids, +PPLX_API int pplx_cuda_embed_into(pplx_cuda_ctx_t *ctx, const int *token_ids, int n_tokens, float *out_embedding); -float *pplx_cuda_embed(pplx_cuda_ctx_t *ctx, const int *token_ids, +PPLX_API float *pplx_cuda_embed(pplx_cuda_ctx_t *ctx, const int *token_ids, int n_tokens); -int pplx_cuda_embed_batch(pplx_cuda_ctx_t *ctx, const pplx_input_t *inputs, +PPLX_API int pplx_cuda_embed_batch(pplx_cuda_ctx_t *ctx, const pplx_input_t *inputs, int batch, float *out_embeddings); -int pplx_cuda_embed_spans_batch(pplx_cuda_ctx_t *ctx, +PPLX_API int pplx_cuda_embed_spans_batch(pplx_cuda_ctx_t *ctx, const pplx_context_input_t *inputs, int batch, float *out_embeddings); |
