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_internal.h | |
| parent | d9ea2cea7dfa1eb25526a3e9b729beda5ccb006f (diff) | |
| download | ffwd-978f4f7969190265e127e8a3d704bb173fbd171c.tar.gz ffwd-978f4f7969190265e127e8a3d704bb173fbd171c.zip | |
chore: rename project
Diffstat (limited to 'embed_internal.h')
| -rw-r--r-- | embed_internal.h | 72 |
1 files changed, 36 insertions, 36 deletions
diff --git a/embed_internal.h b/embed_internal.h index 53debee..a7365a2 100644 --- a/embed_internal.h +++ b/embed_internal.h @@ -1,5 +1,5 @@ -#ifndef PPLX_EMBED_INTERNAL_H -#define PPLX_EMBED_INTERNAL_H +#ifndef EMBED_INTERNAL_H +#define EMBED_INTERNAL_H #include "embed.h" #include "qwen_safetensors.h" @@ -7,52 +7,52 @@ typedef struct { safetensor_dtype_t dtype; const void *data; -} pplx_weight_ref_t; +} embed_weight_ref_t; typedef struct { - pplx_weight_ref_t wq; /* [q_dim, hidden] */ - pplx_weight_ref_t wk; /* [kv_dim, hidden] */ - pplx_weight_ref_t wv; /* [kv_dim, hidden] */ - pplx_weight_ref_t wo; /* [hidden, q_dim] */ - const float *q_norm; /* [head_dim] */ - const float *k_norm; /* [head_dim] */ - const float *input_norm; /* [hidden] */ - const float *post_attn_norm; /* [hidden] */ - pplx_weight_ref_t gate_proj; /* [intermediate, hidden] */ - pplx_weight_ref_t up_proj; /* [intermediate, hidden] */ - pplx_weight_ref_t down_proj; /* [hidden, intermediate] */ -} pplx_layer_t; + embed_weight_ref_t wq; /* [q_dim, hidden] */ + embed_weight_ref_t wk; /* [kv_dim, hidden] */ + embed_weight_ref_t wv; /* [kv_dim, hidden] */ + embed_weight_ref_t wo; /* [hidden, q_dim] */ + const float *q_norm; /* [head_dim] */ + const float *k_norm; /* [head_dim] */ + const float *input_norm; /* [hidden] */ + const float *post_attn_norm; /* [hidden] */ + embed_weight_ref_t gate_proj; /* [intermediate, hidden] */ + embed_weight_ref_t up_proj; /* [intermediate, hidden] */ + embed_weight_ref_t down_proj; /* [hidden, intermediate] */ +} embed_layer_t; typedef struct { - pplx_weight_ref_t embed_tokens; /* [vocab_size, hidden] */ - pplx_layer_t *layers; /* [n_layers] heap-allocated */ - const float *norm; /* [hidden] */ -} pplx_weights_t; + embed_weight_ref_t embed_tokens; /* [vocab_size, hidden] */ + embed_layer_t *layers; /* [n_layers] heap-allocated */ + const float *norm; /* [hidden] */ +} embed_weights_t; -struct pplx_model { - pplx_config_t config; - pplx_weights_t weights; +struct embed_model { + embed_config_t config; + embed_weights_t weights; void *safetensors; /* multi_safetensors_t*, keeps mmap alive */ int layer_start; /* loaded transformer range, inclusive */ int layer_end; /* loaded transformer range, exclusive */ }; -struct pplx_workspace { - const pplx_model_t *model; /* immutable model this workspace belongs to */ +struct embed_workspace { + const embed_model_t *model; /* immutable model this workspace belongs to */ int buf_seq_cap; - float *x; /* [seq, hidden] */ - float *x_norm; /* [seq, hidden] */ - float *q; /* [seq, q_dim] */ - float *k; /* [seq, kv_dim] */ - float *v; /* [seq, kv_dim] */ - float *attn_out; /* [seq, q_dim] */ - float *proj_out; /* [seq, hidden] */ - float *ffn_gate; /* [seq, intermediate] */ - float *ffn_up; /* [seq, intermediate] */ - float *attn_scores; /* reusable BLAS attention score tiles */ + float *x; /* [seq, hidden] */ + float *x_norm; /* [seq, hidden] */ + float *q; /* [seq, q_dim] */ + float *k; /* [seq, kv_dim] */ + float *v; /* [seq, kv_dim] */ + float *attn_out; /* [seq, q_dim] */ + float *proj_out; /* [seq, hidden] */ + float *ffn_gate; /* [seq, intermediate] */ + float *ffn_up; /* [seq, intermediate] */ + float *attn_scores; /* reusable BLAS attention score tiles */ size_t attn_scores_bytes; - int *offsets; /* [batch + 1] packed sequence offsets */ + int *offsets; /* [batch + 1] packed sequence offsets */ int offsets_cap; /* RoPE cosine/sine cache [n_pos, head_dim], grown lazily. */ @@ -67,4 +67,4 @@ struct pplx_workspace { size_t bf16_widen_count; }; -#endif /* PPLX_EMBED_INTERNAL_H */ +#endif /* EMBED_INTERNAL_H */ |
