mirror of
https://github.com/ollama/ollama.git
synced 2026-07-23 09:10:53 -05:00
1fd1ccf7ad
Update llama.cpp to pick up upstream Laguna implementation and remove Ollama's local Laguna implementation. Retain a narrow Metal-only scaling workaround for routed-MoE prompt overflow. Translate older Ollama GGUF attention-gate and SWA metadata names so existing models continue to load.
43 lines
2.1 KiB
Diff
43 lines
2.1 KiB
Diff
diff --git a/src/models/laguna.cpp b/src/models/laguna.cpp
|
|
index 1d705440e..7e708b144 100644
|
|
--- a/src/models/laguna.cpp
|
|
+++ b/src/models/laguna.cpp
|
|
@@ -275,16 +275,35 @@ llama_model_laguna::graph::graph(const llama_model & model, const llm_graph_para
|
|
if ((uint32_t)il >= hparams.n_layer_dense_lead) {
|
|
// MoE: sigmoid routing + score-correction bias + sum-norm +
|
|
// routed_scaling_factor (all handled by build_moe_ffn).
|
|
+ ggml_tensor * up_scale = nullptr;
|
|
+ float expert_weights_scale = hparams.expert_weights_scale;
|
|
+
|
|
+#if defined(GGML_USE_METAL)
|
|
+ const ggml_type down_type = model.layers[il].ffn_down_exps->type;
|
|
+ if (n_tokens >= 32 && (ggml_is_quantized(down_type) || down_type == GGML_TYPE_F16)) {
|
|
+ // At 32 prompt tokens, Metal switches MUL_MAT_ID from its
|
|
+ // range-safe matrix-vector kernel to FP16 matrix tiles.
|
|
+ // Laguna's routed SwiGLU activations can overflow those tiles.
|
|
+ // Generation uses one token and does not need this workaround.
|
|
+ constexpr float down_input_scale = 1.0f / 256.0f;
|
|
+ up_scale = ggml_fill(ctx0,
|
|
+ model.layers[il].ffn_exp_probs_b, down_input_scale);
|
|
+ expert_weights_scale /= down_input_scale;
|
|
+ }
|
|
+#endif
|
|
+
|
|
ggml_tensor * moe_out = build_moe_ffn(cur,
|
|
model.layers[il].ffn_gate_inp,
|
|
model.layers[il].ffn_up_exps,
|
|
model.layers[il].ffn_gate_exps,
|
|
model.layers[il].ffn_down_exps,
|
|
model.layers[il].ffn_exp_probs_b,
|
|
n_expert, n_expert_used,
|
|
LLM_FFN_SILU,
|
|
hparams.expert_weights_norm,
|
|
- hparams.expert_weights_scale,
|
|
+ expert_weights_scale,
|
|
(llama_expert_gating_func_type) hparams.expert_gating_func,
|
|
- il);
|
|
+ il,
|
|
+ nullptr, nullptr,
|
|
+ up_scale);
|
|
cb(moe_out, "ffn_moe_out", il);
|