From: Georgi Gerganov <redacted>
Date: Fri, 29 Sep 2023 16:05:18 +0000 (+0300)
Subject: train : fix KQ_pos allocation (#3392)
X-Git-Tag: gguf-v0.4.0~9
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=bc34dd4f5b5a7c10ae3ed85a265ce6f2ed2fab79;p=pkg%2Fggml%2Fsources%2Fllama.cpp

train : fix KQ_pos allocation (#3392)

* train : fix KQ_pos allocation

* make sure KQ_pos is not reallocated in finetune

---------

Co-authored-by: xaedes <redacted>
---

diff --git a/examples/finetune/finetune.cpp b/examples/finetune/finetune.cpp
index b61165fb..8ca1874d 100644
--- a/examples/finetune/finetune.cpp
+++ b/examples/finetune/finetune.cpp
@@ -626,7 +626,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
 
     // KQ_pos - contains the positions
     struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
-    {
+    ggml_allocr_alloc(alloc, KQ_pos);
+    if (!ggml_allocr_is_measure(alloc)) {
         int * data = (int *) KQ_pos->data;
         for (int i = 0; i < N; ++i) {
             data[i] = n_past + i;
@@ -786,6 +787,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
     GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
     ggml_allocr_alloc(alloc, t36->grad);
+    // KQ_pos
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
 
     // make sure base model tensors data cannot be used in viewable operations
     ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, one));
diff --git a/examples/train-text-from-scratch/train-text-from-scratch.cpp b/examples/train-text-from-scratch/train-text-from-scratch.cpp
index 5043f32d..be693b3a 100644
--- a/examples/train-text-from-scratch/train-text-from-scratch.cpp
+++ b/examples/train-text-from-scratch/train-text-from-scratch.cpp
@@ -334,7 +334,8 @@ static struct ggml_tensor * llama_build_train_graphs(
 
     // KQ_pos - contains the positions
     struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
-    {
+    ggml_allocr_alloc(alloc, KQ_pos);
+    if (!ggml_allocr_is_measure(alloc)) {
         int * data = (int *) KQ_pos->data;
         for (int i = 0; i < N; ++i) {
             data[i] = n_past + i;