From: xaedes <redacted>
Date: Fri, 21 Apr 2023 15:25:21 +0000 (+0200)
Subject: llama : remember and restore kv cache data pointers (#1104)
X-Git-Tag: gguf-v0.4.0~904
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=8687c1f2581d059cd5b6a9502f89bd343566062a;p=pkg%2Fggml%2Fsources%2Fllama.cpp

llama : remember and restore kv cache data pointers (#1104)

because their value is stored in buf and overwritten by memcpy
---

diff --git a/llama.cpp b/llama.cpp
index 33ee4fbb..0345b61c 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -2092,7 +2092,11 @@ void llama_set_kv_cache(
                          int   n_token_count) {
     // Make sure we have the same kv cache setup
     LLAMA_ASSERT(ctx->model.kv_self.buf.size == n_size);
+    void * k_data = ctx->model.kv_self.k->data; // remember data pointers
+    void * v_data = ctx->model.kv_self.v->data; // because their value is stored in buf and overwritten by memcpy
     memcpy(ctx->model.kv_self.buf.addr, kv_cache, n_size);
+    ctx->model.kv_self.k->data = k_data; // restore correct data pointers
+    ctx->model.kv_self.v->data = v_data;
     ctx->model.kv_self.n = n_token_count;
 }