diff options
| author | Georgi Gerganov <ggerganov@gmail.com> | 2023-09-29 19:05:18 +0300 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2023-09-29 19:05:18 +0300 |
| commit | bc34dd4f5b5a7c10ae3ed85a265ce6f2ed2fab79 (patch) | |
| tree | 361b5fc14a3d49f7e4fb68efbbd28cc9dc1608e9 /examples | |
| parent | 2777a84be429401a2b7d33c2b6a4ada1f0776f1b (diff) | |
train : fix KQ_pos allocation (#3392)
* train : fix KQ_pos allocation
* make sure KQ_pos is not reallocated in finetune
---------
Co-authored-by: xaedes <xaedes@gmail.com>
Diffstat (limited to 'examples')
| -rw-r--r-- | examples/finetune/finetune.cpp | 5 | ||||
| -rw-r--r-- | examples/train-text-from-scratch/train-text-from-scratch.cpp | 3 |
2 files changed, 6 insertions, 2 deletions
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; |