diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-11 23:40:01 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-14 08:27:10 +0200 |
| commit | 978f4f7969190265e127e8a3d704bb173fbd171c (patch) | |
| tree | 42f3a89e1ecb481ef0e5196a775ea37edb4b98f7 /embed_mlx.h | |
| parent | d9ea2cea7dfa1eb25526a3e9b729beda5ccb006f (diff) | |
| download | ffwd-978f4f7969190265e127e8a3d704bb173fbd171c.tar.gz ffwd-978f4f7969190265e127e8a3d704bb173fbd171c.zip | |
chore: rename project
Diffstat (limited to 'embed_mlx.h')
| -rw-r--r-- | embed_mlx.h | 88 |
1 files changed, 44 insertions, 44 deletions
diff --git a/embed_mlx.h b/embed_mlx.h index be467d0..267c454 100644 --- a/embed_mlx.h +++ b/embed_mlx.h @@ -4,17 +4,17 @@ * Uses Apple's mlx-c to run the transformer on Metal. */ -#ifndef PPLX_EMBED_MLX_H -#define PPLX_EMBED_MLX_H +#ifndef EMBED_MLX_H +#define EMBED_MLX_H /* Public-API export annotation. The libraries are built with - * -fvisibility=hidden, so only declarations carrying PPLX_API are exported + * -fvisibility=hidden, so only declarations carrying EMBED_API are exported * from the shared library; everything else stays internal. */ -#ifndef PPLX_API +#ifndef EMBED_API #if defined(__GNUC__) -#define PPLX_API __attribute__((visibility("default"))) +#define EMBED_API __attribute__((visibility("default"))) #else -#define PPLX_API +#define EMBED_API #endif #endif @@ -23,37 +23,37 @@ #include "embed.h" /* Opaque MLX model context */ -typedef struct pplx_mlx_ctx pplx_mlx_ctx_t; -typedef struct pplx_mlx_late_ctx pplx_mlx_late_ctx_t; -typedef struct pplx_mlx_late_vectors pplx_mlx_late_vectors_t; +typedef struct embed_mlx_ctx embed_mlx_ctx_t; +typedef struct embed_mlx_late_ctx embed_mlx_late_ctx_t; +typedef struct embed_mlx_late_vectors embed_mlx_late_vectors_t; typedef struct { int quantize_bits; /* 0 disables; currently 8 */ int quantize_group_size; /* default 64 when quantize_bits is set */ -} pplx_mlx_options_t; +} embed_mlx_options_t; /* * Load model into MLX arrays from safetensors + config.json. * Returns NULL on error. */ -PPLX_API pplx_mlx_ctx_t *pplx_mlx_load(const char *model_dir); -PPLX_API pplx_mlx_ctx_t *pplx_mlx_load_with_options(const char *model_dir, - const pplx_mlx_options_t *opts); +EMBED_API embed_mlx_ctx_t *embed_mlx_load(const char *model_dir); +EMBED_API embed_mlx_ctx_t *embed_mlx_load_with_options(const char *model_dir, + const embed_mlx_options_t *opts); -PPLX_API void pplx_mlx_free(pplx_mlx_ctx_t *ctx); +EMBED_API void embed_mlx_free(embed_mlx_ctx_t *ctx); /* * Late-interaction MLX path. This is separate from pooled embeddings because * late snapshots produce one projected token vector per input token for MaxSim * scoring, not one pooled document vector. */ -PPLX_API pplx_mlx_late_ctx_t *pplx_mlx_late_load(const char *model_dir); -PPLX_API pplx_mlx_late_ctx_t *pplx_mlx_late_load_with_options( - const char *model_dir, const pplx_mlx_options_t *opts); -PPLX_API void pplx_mlx_late_free(pplx_mlx_late_ctx_t *ctx); +EMBED_API embed_mlx_late_ctx_t *embed_mlx_late_load(const char *model_dir); +EMBED_API embed_mlx_late_ctx_t *embed_mlx_late_load_with_options( + const char *model_dir, const embed_mlx_options_t *opts); +EMBED_API void embed_mlx_late_free(embed_mlx_late_ctx_t *ctx); -PPLX_API const pplx_config_t *pplx_mlx_late_config(const pplx_mlx_late_ctx_t *ctx); -PPLX_API int pplx_mlx_late_token_dim(const pplx_mlx_late_ctx_t *ctx); +EMBED_API const embed_config_t *embed_mlx_late_config(const embed_mlx_late_ctx_t *ctx); +EMBED_API int embed_mlx_late_token_dim(const embed_mlx_late_ctx_t *ctx); /* * Encode token_ids into out_vectors[n_tokens, token_dim]. @@ -61,7 +61,7 @@ PPLX_API int pplx_mlx_late_token_dim(const pplx_mlx_late_ctx_t *ctx); * If normalize is non-zero, every token vector is L2-normalized on device * before copying back to CPU memory. */ -PPLX_API int pplx_mlx_late_encode_tokens(pplx_mlx_late_ctx_t *ctx, +EMBED_API int embed_mlx_late_encode_tokens(embed_mlx_late_ctx_t *ctx, const int *token_ids, int n_tokens, int normalize, float *out_vectors); @@ -72,36 +72,36 @@ PPLX_API int pplx_mlx_late_encode_tokens(pplx_mlx_late_ctx_t *ctx, * outlive ctx. Copying is optional: MaxSim can consume these handles directly * and copy back only final candidate scores. */ -PPLX_API pplx_mlx_late_vectors_t *pplx_mlx_late_encode_tokens_device( - pplx_mlx_late_ctx_t *ctx, const int *token_ids, int n_tokens, +EMBED_API embed_mlx_late_vectors_t *embed_mlx_late_encode_tokens_device( + embed_mlx_late_ctx_t *ctx, const int *token_ids, int n_tokens, int normalize); -PPLX_API void pplx_mlx_late_vectors_free(pplx_mlx_late_vectors_t *vecs); -PPLX_API int pplx_mlx_late_vectors_token_count(const pplx_mlx_late_vectors_t *vecs); -PPLX_API int pplx_mlx_late_vectors_dim(const pplx_mlx_late_vectors_t *vecs); -PPLX_API int pplx_mlx_late_vectors_copy(const pplx_mlx_late_vectors_t *vecs, +EMBED_API void embed_mlx_late_vectors_free(embed_mlx_late_vectors_t *vecs); +EMBED_API int embed_mlx_late_vectors_token_count(const embed_mlx_late_vectors_t *vecs); +EMBED_API int embed_mlx_late_vectors_dim(const embed_mlx_late_vectors_t *vecs); +EMBED_API int embed_mlx_late_vectors_copy(const embed_mlx_late_vectors_t *vecs, float *out_vectors); -PPLX_API pplx_mlx_late_vectors_t *pplx_mlx_late_vectors_concat( - pplx_mlx_late_ctx_t *ctx, - const pplx_mlx_late_vectors_t *const *items, int count); -PPLX_API pplx_mlx_late_vectors_t *pplx_mlx_late_vectors_select( - pplx_mlx_late_ctx_t *ctx, const pplx_mlx_late_vectors_t *vecs, +EMBED_API embed_mlx_late_vectors_t *embed_mlx_late_vectors_concat( + embed_mlx_late_ctx_t *ctx, + const embed_mlx_late_vectors_t *const *items, int count); +EMBED_API embed_mlx_late_vectors_t *embed_mlx_late_vectors_select( + embed_mlx_late_ctx_t *ctx, const embed_mlx_late_vectors_t *vecs, const int *token_indices, int count); -PPLX_API int pplx_mlx_late_maxsim_batch_device( - pplx_mlx_late_ctx_t *ctx, const pplx_mlx_late_vectors_t *query, - const pplx_mlx_late_vectors_t *docs, const int *doc_offsets, +EMBED_API int embed_mlx_late_maxsim_batch_device( + embed_mlx_late_ctx_t *ctx, const embed_mlx_late_vectors_t *query, + const embed_mlx_late_vectors_t *docs, const int *doc_offsets, int docs_count, float *scores); /* * Compute embedding for token_ids[0..n_tokens-1]. * Returns malloc'd float[hidden_size] (caller frees). NULL on error. */ -PPLX_API float *pplx_mlx_embed(pplx_mlx_ctx_t *ctx, const int *token_ids, int n_tokens); +EMBED_API float *embed_mlx_encode(embed_mlx_ctx_t *ctx, const int *token_ids, int n_tokens); /* * Compute one embedding into caller-provided out[hidden_size]. * Returns 0 on success, -1 on error. */ -PPLX_API int pplx_mlx_embed_into(pplx_mlx_ctx_t *ctx, const int *token_ids, +EMBED_API int embed_mlx_encode_into(embed_mlx_ctx_t *ctx, const int *token_ids, int n_tokens, float *out_embedding); /* @@ -111,26 +111,26 @@ PPLX_API int pplx_mlx_embed_into(pplx_mlx_ctx_t *ctx, const int *token_ids, * Padded tokens are masked out of attention keys and mean pooling. * Returns 0 on success, -1 on error. */ -PPLX_API int pplx_mlx_embed_batch(pplx_mlx_ctx_t *ctx, const pplx_input_t *inputs, +EMBED_API int embed_mlx_encode_batch(embed_mlx_ctx_t *ctx, const embed_input_t *inputs, int batch, float *out_embeddings); /* * Run one contextual sequence and pool selected token spans. * Returns 0 on success, -1 on error. */ -PPLX_API int pplx_mlx_embed_spans(pplx_mlx_ctx_t *ctx, const int *token_ids, - int n_tokens, const pplx_span_t *spans, +EMBED_API int embed_mlx_encode_spans(embed_mlx_ctx_t *ctx, const int *token_ids, + int n_tokens, const embed_span_t *spans, int n_spans, float *out_embeddings); /* * Run a padded dense contextual document batch and pool every selected span. * out_embeddings contains spans in document order. */ -PPLX_API int pplx_mlx_embed_spans_batch(pplx_mlx_ctx_t *ctx, - const pplx_context_input_t *inputs, int batch, +EMBED_API int embed_mlx_encode_spans_batch(embed_mlx_ctx_t *ctx, + const embed_context_input_t *inputs, int batch, float *out_embeddings); /* Get the config (so embed_cli.c can read hidden_size etc.) */ -PPLX_API const pplx_config_t *pplx_mlx_config(const pplx_mlx_ctx_t *ctx); +EMBED_API const embed_config_t *embed_mlx_config(const embed_mlx_ctx_t *ctx); -#endif /* PPLX_EMBED_MLX_H */ +#endif /* EMBED_MLX_H */ |
