| 268 | } |
| 269 | |
| 270 | runtime::TransformerKVState empty_decoder_state(size_t layers) { |
| 271 | runtime::TransformerKVState state; |
| 272 | state.layers.resize(layers); |
| 273 | return state; |
| 274 | } |
| 275 | |
| 276 | std::vector<ggml_fp16_t> build_causal_prefill_mask_values(int64_t steps) { |
| 277 | const ggml_fp16_t masked = ggml_fp32_to_fp16(-std::numeric_limits<float>::infinity()); |