llama : fix expert weighting in the FFN

YellowRoseCx · Dec 9, 2023 · 8b185b7 · 8b185b7
1 parent 7ea3695
commit 8b185b7
Showing 1 changed file with 7 additions and 5 deletions.
diff --git a/llama.cpp b/llama.cpp
@@ -4250,11 +4250,13 @@ struct llm_build_context {
  ggml_tensor * probs = ggml_soft_max(ctx0, logits); // [n_tokens, num_experts]
 
  // select experts
- ggml_tensor * selected_experts = ggml_top_k(ctx0, probs, n_experts_per_tok); // [n_tokens, num_experts_per_tok]
- //ggml_tensor * weights = ggml_get_rows(ctx0, probs, selected_experts); // [n_tokens, num_experts_per_tok, 1]
- ggml_tensor * weights = ggml_get_rows(ctx0,
- ggml_reshape_3d(ctx0, probs, 1, n_experts, n_tokens), selected_experts);
- weights = ggml_div(ctx0, weights, ggml_sum_rows(ctx0, weights)); // [n_tokens, num_experts_per_tok, 1]
+ ggml_tensor * selected_experts = ggml_top_k(ctx0, probs, n_experts_per_tok); // [n_tokens, num_experts_per_tok]
+ ggml_tensor * weights =
+ ggml_reshape_2d(ctx0,
+ ggml_get_rows(ctx0,
+ ggml_reshape_3d(ctx0, probs, 1, n_experts, n_tokens), selected_experts),
+ n_experts_per_tok, n_tokens); // [n_tokens, num_experts_per_tok]
+ weights = ggml_div(ctx0, weights, ggml_sum_rows(ctx0, weights)); // [n_tokens, num_experts_per_tok]
 
  // compute expert outputs
  ggml_tensor * moe_out;