aparler/c-llm
1
0
Fork 0
mirror of https://github.com/karpathy/llm.c.git synced 2026-07-28 20:35:09 -04:00
Code Issues Projects Releases Packages Wiki Activity Actions

Fix mem leak

Browse source
This commit is contained in:
Aleksa Gordic 2024-06-02 21:28:42 +02:00
parent a26041af12
commit 30d81cce15
1 changed files with 2 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

3
train_gpt2.cu
View file

@ -2554,7 +2554,7 @@ void gpt2_backward(GPT2 *model, int* inputs) {
model->grads_memory = malloc_and_point_parameters(&model->grads, model->param_elements, model->param_sizeof);
// we're going to be clever for the activations backward pass. we don't need to exactly
// mirror the forward pass activations and we will save memory.
size_t bw_act_sizes[NUM_ACTIVATION_TENSORS];
size_t bw_act_sizes[NUM_BACKWARD_TENSORS];
fill_in_grad_act_sizes(bw_act_sizes, model->batch_size, model->seq_len, model->config);
// count up and allocate the space
model->num_grad_acts = 0;
@ -2985,6 +2985,7 @@ void load_state(int* step, GPT2* model, DataLoader* loader, const char* filename
cudaCheck(cudaMemcpy(model->m_memory, cpu_buffer, shard_num_parameters * sizeof(float), cudaMemcpyHostToDevice));
freadCheck(cpu_buffer, sizeof(float), shard_num_parameters, state_file);
cudaCheck(cudaMemcpy(model->v_memory, cpu_buffer, shard_num_parameters * sizeof(float), cudaMemcpyHostToDevice));
free(cpu_buffer);
}
// ----------------------------------------------------------------------------