aboutsummaryrefslogtreecommitdiff
path: root/embed_cuda.h
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-10 08:24:39 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-14 08:27:03 +0200
commit2c37727e653d2340d8977dd5db4addc767dd81ea (patch)
tree578ad6affac511f847a346a0efc6d40ebf6a279f /embed_cuda.h
parentec417bcb52f2aa8e24aeec312e8ddbf3a0798056 (diff)
downloadffwd-2c37727e653d2340d8977dd5db4addc767dd81ea.tar.gz
ffwd-2c37727e653d2340d8977dd5db4addc767dd81ea.zip
gitignore: track new binary and library names
Diffstat (limited to 'embed_cuda.h')
-rw-r--r--embed_cuda.h30
1 files changed, 30 insertions, 0 deletions
diff --git a/embed_cuda.h b/embed_cuda.h
new file mode 100644
index 0000000..badff9c
--- /dev/null
+++ b/embed_cuda.h
@@ -0,0 +1,30 @@
+/*
+ * pplx_embed_cuda.h - CUDA/cuBLAS backend for pplx-embed inference.
+ */
+
+#ifndef PPLX_EMBED_CUDA_H
+#define PPLX_EMBED_CUDA_H
+
+#include "pplx_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_cuda_ctx_t *pplx_cuda_load(const char *model_dir);
+void pplx_cuda_free(pplx_cuda_ctx_t *ctx);
+
+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,
+ int n_tokens, float *out_embedding);
+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,
+ int batch, float *out_embeddings);
+int pplx_cuda_embed_spans_batch(pplx_cuda_ctx_t *ctx,
+ const pplx_context_input_t *inputs,
+ int batch, float *out_embeddings);
+
+#endif /* PPLX_EMBED_CUDA_H */