Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-01 00:39:00 +01:00
Code Issues Packages Projects Releases Wiki Activity

test-backend-ops : make experts more evenly probable (test_moe)

Browse Source
This commit is contained in:
Georgi Gerganov 2023-12-10 15:27:41 +02:00
parent b0b83dd9e2
commit 54ba263410
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
tests/test-backend-ops.cpp
View File

@ -1172,7 +1172,7 @@ struct test_moe : public test_case {
ggml_tensor * cur = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_tokens); ggml_tensor * cur = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_tokens);
ggml_tensor * logits = ggml_mul_mat(ctx, ffn_gate_inp, cur); // [n_tokens, num_experts] ggml_tensor * logits = ggml_mul_mat(ctx, ffn_gate_inp, cur); // [n_tokens, num_experts]
ggml_tensor * probs = ggml_soft_max(ctx, logits); // [n_tokens, num_experts] ggml_tensor * probs = ggml_soft_max_ext(ctx, logits, nullptr, 1.0f/sqrtf(n_embd)); // [n_tokens, num_experts]
// select experts // select experts
ggml_tensor * selected_experts = ggml_top_k(ctx, probs, n_experts_per_tok); // [n_tokens, num_experts_per_tok] ggml_tensor * selected_experts = ggml_top_k(ctx, probs, n_experts_per_tok); // [n_tokens, num_experts_per_tok]