aboutsummaryrefslogtreecommitdiff
path: root/src/engine.h
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-16 17:16:14 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-16 17:16:14 +0200
commit3904a33e4416c35a48e8dd7c242a137a2cc17871 (patch)
tree677c7d0a8d272fdc4a089d3bbd2004712938f0ee /src/engine.h
parentcca81e59a9afeebe1237f713e919b5002df3d749 (diff)
downloadffwd-3904a33e4416c35a48e8dd7c242a137a2cc17871.tar.gz
ffwd-3904a33e4416c35a48e8dd7c242a137a2cc17871.zip
cpu: extract late-interaction model into late.c
Diffstat (limited to 'src/engine.h')
-rw-r--r--src/engine.h60
1 files changed, 60 insertions, 0 deletions
diff --git a/src/engine.h b/src/engine.h
new file mode 100644
index 0000000..abdb0c7
--- /dev/null
+++ b/src/engine.h
@@ -0,0 +1,60 @@
+#ifndef EMBED_ENGINE_H
+#define EMBED_ENGINE_H
+
+#include "internal.h"
+#include "alloc.h"
+
+/* Cross-file declarations internal to the CPU engine. These functions are not
+ * part of the public API (embed.h) and carry no EMBED_API visibility; they are
+ * shared between the engine translation units (model loading, the forward pass,
+ * and late interaction) that used to live in one file. The frontends and the
+ * GPU backends do not include this header. */
+
+/* Validate that a safetensor matches an expected dtype/rank/shape and that its
+ * bytes lie within the mapped file. bf16_ok permits BF16 in addition to F32. */
+int tensor_has_supported_shape(const safetensors_file_t *sf,
+ const safetensor_t *t,
+ const char *name,
+ const int64_t *shape,
+ int ndim,
+ int bf16_ok);
+
+/* True if model_dir carries a 1_Dense projection (a late-interaction model). */
+int model_dir_has_late_projection(const char *model_dir);
+
+/* Load (a layer range of) a model. allow_late lets the late loader accept a
+ * 1_Dense snapshot that the pooled loader rejects. layer_end == -1 means all. */
+embed_model_t *
+model_load_range_ex(const char *model_dir, int layer_start, int layer_end, int allow_late);
+
+/* y[seq,out] = x[seq,in] @ w^T, no bias; dispatches F32/BF16 paths. Shared by
+ * the forward pass and the late projection. */
+void linear_nobias_weight(embed_workspace_t *ws,
+ float *y,
+ const float *x,
+ const embed_weight_ref_t *w,
+ int seq_len,
+ int in_dim,
+ int out_dim);
+
+/* Compute packed sequence offsets [batch+1] plus total and max sequence length
+ * over a batch of inputs. require_ids rejects inputs with a NULL id pointer. */
+int build_offsets(const embed_input_t *inputs,
+ int batch,
+ int require_ids,
+ int *offsets,
+ int *total_out,
+ int *max_out);
+
+/* Run the full transformer forward over a packed batch in ws, leaving hidden
+ * states in ws->x. Used by pooled encoding and by late-interaction encoding. */
+int forward_packed_inplace(const embed_model_t *model,
+ embed_workspace_t *ws,
+ const embed_input_t *inputs,
+ int batch,
+ const int *offsets,
+ int total_seq,
+ int max_seq,
+ int apply_final_norm);
+
+#endif /* EMBED_ENGINE_H */