From c10704d01e21e3dbe4d6ca1026ebff85349dd239 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Sun, 27 Aug 2023 18:55:41 +0300
Subject: [PATCH] llama : fix MPI threads (close #2827)

---
 llama.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/llama.cpp b/llama.cpp
index 0bb8fcd6e..72d2d1de0 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -2845,7 +2845,6 @@ static bool llama_eval_internal(
 
     GGML_ASSERT(n_tokens > 0);
     GGML_ASSERT(n_past >= 0);
-    GGML_ASSERT(n_threads > 0);
     // TODO: keep the values of n_batch and n_ctx
     // GGML_ASSERT(n_tokens <= n_batch);
     // GGML_ASSERT(n_past + n_tokens <= n_ctx);
@@ -2856,6 +2855,8 @@ static bool llama_eval_internal(
     ggml_mpi_eval_init(lctx.ctx_mpi, &n_tokens, &n_past, &n_threads);
 #endif
 
+    GGML_ASSERT(n_threads > 0);
+
     const int N = n_tokens;
 
     const auto & model   = lctx.model;