From 545063223720505e4da4ef1bdea176d8006bf4f6 Mon Sep 17 00:00:00 2001 From: Christopher Date: Thu, 30 May 2024 03:01:52 +0000 Subject: [PATCH] Removed unnecesary shared memory due to blockreduce using static defined shared memory --- dev/cuda/classifier_fused.cu | 2 +- train_gpt2.cu | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/dev/cuda/classifier_fused.cu b/dev/cuda/classifier_fused.cu index 2125b87..5b6c986 100644 --- a/dev/cuda/classifier_fused.cu +++ b/dev/cuda/classifier_fused.cu @@ -664,7 +664,7 @@ void fused_classifier5(float* dlogits, float* losses, int B, int T, int V, int P, int block_size) { const int N = B * T; const int grid_size = N; - fused_classifier_kernel5<<>>((floatX*)dlogits, (floatX*)losses, NULL, (floatX*)logits, (floatX*)dlosses, targets, B, T, V, P); + fused_classifier_kernel5<<>>((floatX*)dlogits, (floatX*)losses, NULL, (floatX*)logits, (floatX*)dlosses, targets, B, T, V, P); cudaCheck(cudaGetLastError()); } diff --git a/train_gpt2.cu b/train_gpt2.cu index 5918e21..053c2e3 100644 --- a/train_gpt2.cu +++ b/train_gpt2.cu @@ -1899,7 +1899,7 @@ void fused_classifier(Type* logits, Type* losses, const int block_size = 1024; const int N = B * T; const int grid_size = N; - fused_classifier_kernel5<<>>(logits, losses, (floatX*)NULL, dloss, targets, B, T, V, P); + fused_classifier_kernel5<<>>(logits, losses, (floatX*)NULL, dloss, targets, B, T, V, P); cudaCheck(cudaGetLastError()); }