| 1971 | .build(ctx, x, {weights.lm_head, std::nullopt}); |
| 1972 | logits_ = logits.tensor; |
| 1973 | ggml_set_output(logits_); |
| 1974 | ggml_build_forward_expand(graph_, logits_); |
| 1975 | buffer_ = ggml_backend_alloc_ctx_tensors(ctx_.get(), runtime_->backend()); |
| 1976 | if (buffer_ == nullptr) { |
| 1977 | throw std::runtime_error("failed to allocate ACE-Step planner decode graph"); |
| 1978 | } |
| 1979 | attention_mask_values_.assign( |
| 1980 | static_cast<size_t>(cache_steps_ + 1), |
| 1981 | ggml_fp32_to_fp16(-std::numeric_limits<float>::infinity())); |
nothing calls this directly
no test coverage detected