llama.cpp f74e2fa for NVIDIA Volta (sm_70), CUDA 12.9 — released 2026-10-20 with the Q5_1 caveat stated
0b864c5 verified | diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp | |
| index 2bc3a29..84c043b 100644 | |
| --- a/src/llama-graph.cpp | |
| +++ b/src/llama-graph.cpp | |
| ggml_tensor * llm_graph_context::build_lora_mm( | |
| const float adapter_scale = lora.second; | |
| const float scale = lw->get_scale(lora.first->alpha, adapter_scale); | |
| + if (scale == 0.0f) { | |
| + continue; // a resident adapter at scale 0 contributes nothing; skip its matmuls (Atlas 2026-10-16) | |
| + } | |
| ggml_tensor * ab_cur = ggml_mul_mat( | |
| ctx0, lw->b, | |
| ggml_tensor * llm_graph_context::build_lora_mm_id( | |
| const float alpha = lora.first->alpha; | |
| const float rank = (float) lw->b->ne[0]; | |
| const float scale = alpha ? lora.second * alpha / rank : lora.second; | |
| + if (scale == 0.0f) { | |
| + continue; // skip a zero-scale adapter (Atlas 2026-10-16) | |
| + } | |
| ggml_tensor * ab_cur = ggml_mul_mat_id( | |
| ctx0, lw->b, | |
| ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd, float to | |
| const float adapter_scale = lora.second; | |
| const float scale = lw->get_scale(lora.first->alpha, adapter_scale); | |
| + if (scale == 0.0f) { | |
| + continue; // skip a zero-scale adapter (Atlas 2026-10-16) | |
| + } | |
| ggml_tensor * inpL_delta = ggml_scale(ctx0, ggml_mul_mat( | |
| ctx0, lw->b, // non-transposed lora_b | |