From 96e5dc95397cd9038212dc69cbe0384b24640dc3 Mon Sep 17 00:00:00 2001 From: Onur Gündüz Date: Mon, 8 Jun 2026 05:14:12 +0200 Subject: cuda: add initial CUDA embedding backend --- pplx_embed.c | 58 +--------------------------------------------------------- 1 file changed, 1 insertion(+), 57 deletions(-) (limited to 'pplx_embed.c') diff --git a/pplx_embed.c b/pplx_embed.c index 2c0ade6..4dc358e 100644 --- a/pplx_embed.c +++ b/pplx_embed.c @@ -1,4 +1,5 @@ #include "pplx_embed.h" +#include "pplx_embed_internal.h" #include "qwen_asr_kernels.h" #include "qwen_asr_safetensors.h" @@ -9,63 +10,6 @@ #include #include -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; -}; - struct pplx_late_model { pplx_model_t *base; safetensors_file_t *dense_sf; -- cgit v1.3.1