]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commitdiff
fix server sampling: top k sampler first (#1977)
authoranon998 <redacted>
Sun, 25 Jun 2023 08:48:36 +0000 (08:48 +0000)
committerGitHub <redacted>
Sun, 25 Jun 2023 08:48:36 +0000 (10:48 +0200)
Co-authored-by: anon <redacted>
examples/server/server.cpp

index de22d301342d6bd4dcb2fceec93e645809fc0c19..79df5e84762cdd0a44f8911c439a1ce6a5c86662 100644 (file)
@@ -325,10 +325,10 @@ struct llama_server_context {
                     id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
                 } else {
                     // Temperature sampling
+                    llama_sample_top_k(ctx, &candidates_p, top_k, 1);
                     llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
                     llama_sample_typical(ctx, &candidates_p, typical_p, 1);
                     llama_sample_top_p(ctx, &candidates_p, top_p, 1);
-                    llama_sample_top_k(ctx, &candidates_p, top_k, 1);
                     llama_sample_temperature(ctx, &candidates_p, temp);
                     id = llama_sample_token(ctx, &candidates_p);
                 }