diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-16 16:27:47 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-16 16:27:47 +0200 |
| commit | d1ba1ebc06ef0d62ca932388fae83e6992594719 (patch) | |
| tree | f962207b779016593258f990348216999a58766a /src/internal.h | |
| parent | be20dfda0f836808dccfc14db244592e41ba28bc (diff) | |
| download | ffwd-d1ba1ebc06ef0d62ca932388fae83e6992594719.tar.gz ffwd-d1ba1ebc06ef0d62ca932388fae83e6992594719.zip | |
chore: move sources to src/ with proper names, drop qwen_ prefix
Diffstat (limited to 'src/internal.h')
| -rw-r--r-- | src/internal.h | 110 |
1 files changed, 110 insertions, 0 deletions
diff --git a/src/internal.h b/src/internal.h new file mode 100644 index 0000000..cdc02a1 --- /dev/null +++ b/src/internal.h @@ -0,0 +1,110 @@ +#ifndef EMBED_INTERNAL_H +#define EMBED_INTERNAL_H + +#include "embed.h" +#include "safetensors.h" + +typedef struct { + safetensor_dtype_t dtype; + const void *data; +} embed_weight_ref_t; + +typedef struct { + 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_bias; /* [q_dim], optional */ + const float *k_bias; /* [kv_dim], optional */ + const float *v_bias; /* [kv_dim], optional */ + 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] */ + /* BERT family (all NULL for Qwen3). The attention projections reuse + * wq/wk/wv/wo and q/k/v_bias; the two FFN matrices reuse up_proj + * (intermediate.dense) and down_proj (output.dense); the two block + * LayerNorm weights reuse input_norm (post-attention) and post_attn_norm + * (post-FFN). These fields add the biases those reused slots lack. */ + const float *o_bias; /* attention output dense bias [hidden] */ + const float *attn_ln_bias; /* post-attention LayerNorm bias [hidden] */ + const float *ffn_inter_bias; /* intermediate.dense bias [intermediate] */ + const float *ffn_out_bias; /* output.dense bias [hidden] */ + const float *ffn_ln_bias; /* post-FFN LayerNorm bias [hidden] */ +} embed_layer_t; + +typedef struct { + embed_weight_ref_t embed_tokens; /* [vocab_size, hidden] */ + embed_layer_t *layers; /* [n_layers] heap-allocated */ + const float *norm; /* [hidden]; NULL for BERT (no final norm) */ + /* BERT family (NULL for Qwen3): learned absolute position embeddings, + * token-type (segment) embeddings, and the embedding LayerNorm. */ + const float *position_embeddings; /* [max_position_embeddings, hidden] */ + const float *token_type_embeddings; /* [2, hidden] */ + const float *embed_ln_w; /* embedding LayerNorm weight [hidden] */ + const float *embed_ln_b; /* embedding LayerNorm bias [hidden] */ +} embed_weights_t; + +typedef void (*embed_attention_fn)(float *out, + const float *Q, + const float *K, + const float *V, + const int *offsets, + int batch, + int n_heads, + int n_kv_heads, + int head_dim, + float scale, + float *scratch, + size_t scratch_bytes); + +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 */ + embed_attention_fn attention; +}; + +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 */ + 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; + + /* F32 view of one BF16 weight matrix for SGEMM dispatch, grown to the + * largest projection. Workspace-owned so concurrent workspaces on one + * model never share conversion scratch. */ + float *bf16_widen; + size_t bf16_widen_count; +}; + +/* Internal accessors for the late-interaction model. GPU backends build a device + * context from an already-loaded CPU late model because the model dir is a late + * snapshot that the pooled loader rejects; these expose the base model and the + * projection weight without making struct embed_late_model public. */ +embed_model_t *embed_late_model_base(const embed_late_model_t *model); +const embed_weight_ref_t *embed_late_model_projection(const embed_late_model_t *model); + +#endif /* EMBED_INTERNAL_H */ |
