diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-20 20:10:56 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-21 05:57:19 +0200 |
| commit | 6b969d25ce136eaf695222bd1a10bff1ecac0220 (patch) | |
| tree | c23afb50c28b9deaddae5c6526a74ee11cf2d865 /libffwd/model_internal.h | |
| parent | d4bdb8dae753462f621507d5fc3582e0e53da35c (diff) | |
| download | ffwd-6b969d25ce136eaf695222bd1a10bff1ecac0220.tar.gz ffwd-6b969d25ce136eaf695222bd1a10bff1ecac0220.zip | |
all: split Makefile and reorg library/tools
Diffstat (limited to 'libffwd/model_internal.h')
| -rw-r--r-- | libffwd/model_internal.h | 76 |
1 files changed, 76 insertions, 0 deletions
diff --git a/libffwd/model_internal.h b/libffwd/model_internal.h new file mode 100644 index 0000000..97de064 --- /dev/null +++ b/libffwd/model_internal.h @@ -0,0 +1,76 @@ +#ifndef FFWD_MODEL_INTERNAL_H +#define FFWD_MODEL_INTERNAL_H + +#include "model_types.h" +#include "alloc.h" + +/* Cross-file function declarations private to the CPU model code. The companion + * to model_types.h: that header holds the shared struct definitions, this one the + * shared functions. None are part of the public API (ffwd.h) and none carry + * FFWD_API visibility; they are shared between the 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 either 16-bit float format + * (BF16 or F16) 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); + +/* Build a weight ref from a located tensor: F32/BF16 borrow the mmap, F16 is + * widened to an owned F32 buffer (no fused F16 kernel). Shared by the model + * loader and the late-interaction projection loader; the owner frees ref.data + * when ref.owned is set. Returns ref.data == NULL on allocation failure. */ +ffwd_weight_ref_t weight_ref_from_tensor(const safetensors_file_t *sf, const safetensor_t *t); + +/* 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. */ +ffwd_model_t * +model_load_range_ex(const char *model_dir, int layer_start, int layer_end, int allow_late); + +/* Workspace scratch growth, lazy per forward call. Defined in model.c beside the + * workspace lifecycle (ffwd_workspace_nbytes mirrors their sizing); called from + * the forward pass in forward.c. */ +int ensure_buffers(ffwd_workspace_t *ws, const ffwd_config_t *c, int seq); +void ensure_attention_scores(ffwd_workspace_t *ws, const int *offsets, int batch); +int ensure_rope_cache(ffwd_workspace_t *ws, const ffwd_config_t *cfg, int n_pos); +int ensure_offsets(ffwd_workspace_t *ws, int batch); + +/* 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(ffwd_workspace_t *ws, + float *y, + const float *x, + const ffwd_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 ffwd_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 ffwd_model_t *model, + ffwd_workspace_t *ws, + const ffwd_input_t *inputs, + int batch, + const int *offsets, + int total_seq, + int max_seq, + int apply_final_norm); + +#endif /* FFWD_MODEL_INTERNAL_H */ |
