| 203 | } |
| 204 | |
| 205 | T3InferenceComponent::PreparedInput T3InferenceComponent::prepare_conditioning_inputs(const ConditioningInput & input) const { |
| 206 | const int64_t hidden_size = weights_->hidden_size; |
| 207 | const int64_t speaker_embed_size = weights_->speaker_embed_size; |
| 208 | const int64_t num_heads = weights_->perceiver_num_heads; |
| 209 | const int64_t query_tokens = weights_->perceiver_query_tokens; |
| 210 | const auto spkr_enc_weight = engine::assets::tensor_data_to_f32("cond_enc.spkr_enc.weight", weights_->spkr_enc_weight); |
| 211 | const auto spkr_enc_bias = engine::assets::tensor_data_to_f32("cond_enc.spkr_enc.bias", weights_->spkr_enc_bias); |
| 212 | const auto emotion_adv_weight = engine::assets::tensor_data_to_f32( |
| 213 | "cond_enc.emotion_adv_fc.weight", |
| 214 | weights_->emotion_adv_weight); |
| 215 | const auto perceiver_pre_attention_query = engine::assets::tensor_data_to_f32( |
| 216 | "cond_enc.perceiver.pre_attention_query", |
| 217 | weights_->perceiver_pre_attention_query); |
| 218 | const auto perceiver_norm_weight = engine::assets::tensor_data_to_f32( |
| 219 | "cond_enc.perceiver.attn.norm.weight", |
| 220 | weights_->perceiver_norm_weight); |
| 221 | const auto perceiver_norm_bias = engine::assets::tensor_data_to_f32( |
| 222 | "cond_enc.perceiver.attn.norm.bias", |
| 223 | weights_->perceiver_norm_bias); |
| 224 | const auto perceiver_to_q_weight = engine::assets::tensor_data_to_f32( |
| 225 | "cond_enc.perceiver.attn.to_q.weight", |
| 226 | weights_->perceiver_to_q_weight); |
| 227 | const auto perceiver_to_q_bias = engine::assets::tensor_data_to_f32( |
| 228 | "cond_enc.perceiver.attn.to_q.bias", |
| 229 | weights_->perceiver_to_q_bias); |
| 230 | const auto perceiver_to_k_weight = engine::assets::tensor_data_to_f32( |
| 231 | "cond_enc.perceiver.attn.to_k.weight", |
| 232 | weights_->perceiver_to_k_weight); |
| 233 | const auto perceiver_to_k_bias = engine::assets::tensor_data_to_f32( |
| 234 | "cond_enc.perceiver.attn.to_k.bias", |
| 235 | weights_->perceiver_to_k_bias); |
| 236 | const auto perceiver_to_v_weight = engine::assets::tensor_data_to_f32( |
| 237 | "cond_enc.perceiver.attn.to_v.weight", |
| 238 | weights_->perceiver_to_v_weight); |
| 239 | const auto perceiver_to_v_bias = engine::assets::tensor_data_to_f32( |
| 240 | "cond_enc.perceiver.attn.to_v.bias", |
| 241 | weights_->perceiver_to_v_bias); |
| 242 | const auto perceiver_proj_out_weight = engine::assets::tensor_data_to_f32( |
| 243 | "cond_enc.perceiver.attn.proj_out.weight", |
| 244 | weights_->perceiver_proj_out_weight); |
| 245 | const auto perceiver_proj_out_bias = engine::assets::tensor_data_to_f32( |
| 246 | "cond_enc.perceiver.attn.proj_out.bias", |
| 247 | weights_->perceiver_proj_out_bias); |
| 248 | const auto text_embedding_weight = engine::assets::tensor_data_to_f32("text_emb.weight", weights_->text_embedding_weight); |
| 249 | const auto speech_embedding_weight = engine::assets::tensor_data_to_f32("speech_emb.weight", weights_->speech_embedding_weight); |
| 250 | const auto text_position_weight = engine::assets::tensor_data_to_f32( |
| 251 | "text_pos_emb.emb.weight", |
| 252 | weights_->text_position_weight); |
| 253 | const auto speech_position_weight = engine::assets::tensor_data_to_f32( |
| 254 | "speech_pos_emb.emb.weight", |
| 255 | weights_->speech_position_weight); |
| 256 | |
| 257 | const auto cond_prompt_emb = gather_rows( |
| 258 | speech_embedding_weight, |
| 259 | weights_->speech_vocab, |
| 260 | hidden_size, |
| 261 | input.cond_prompt_speech_tokens); |
| 262 | std::vector<int32_t> cond_prompt_positions; |
nothing calls this directly
no test coverage detected