diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-10 08:24:39 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-14 08:27:03 +0200 |
| commit | 2c37727e653d2340d8977dd5db4addc767dd81ea (patch) | |
| tree | 578ad6affac511f847a346a0efc6d40ebf6a279f /pplx_embed_internal.h | |
| parent | ec417bcb52f2aa8e24aeec312e8ddbf3a0798056 (diff) | |
| download | ffwd-2c37727e653d2340d8977dd5db4addc767dd81ea.tar.gz ffwd-2c37727e653d2340d8977dd5db4addc767dd81ea.zip | |
gitignore: track new binary and library names
Diffstat (limited to 'pplx_embed_internal.h')
| -rw-r--r-- | pplx_embed_internal.h | 64 |
1 files changed, 0 insertions, 64 deletions
diff --git a/pplx_embed_internal.h b/pplx_embed_internal.h deleted file mode 100644 index 1e7e4d0..0000000 --- a/pplx_embed_internal.h +++ /dev/null @@ -1,64 +0,0 @@ -#ifndef PPLX_EMBED_INTERNAL_H -#define PPLX_EMBED_INTERNAL_H - -#include "pplx_embed.h" -#include "qwen_safetensors.h" - -typedef struct { - safetensor_dtype_t dtype; - const void *data; -} pplx_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; - -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; - -struct pplx_model { - pplx_config_t config; - pplx_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 */ - 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 */ - size_t attn_scores_bytes; - - int *offsets; /* [batch + 1] packed sequence offsets */ - int offsets_cap; - - /* RoPE cosine/sine cache [n_pos, head_dim], grown lazily. */ - float *rope_cos; - float *rope_sin; - int rope_cache_cap; -}; - -#endif /* PPLX_EMBED_INTERNAL_H */ |
