From: Johannes Gäßler <redacted>
Date: Tue, 6 Aug 2024 15:13:55 +0000 (+0200)
Subject: CUDA: fix padding logic for FP16/FP32 (llama/8884)
X-Git-Tag: upstream/0.0.1642~462
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=9510e3cd55b5b1814cf529aa620de008ee788239;p=pkg%2Fggml%2Fsources%2Fggml

CUDA: fix padding logic for FP16/FP32 (llama/8884)
---

diff --git a/src/ggml-cuda.cu b/src/ggml-cuda.cu
index 68605fff..654f93e8 100644
--- a/src/ggml-cuda.cu
+++ b/src/ggml-cuda.cu
@@ -1501,7 +1501,7 @@ static void ggml_cuda_op_mul_mat(
         }
 
         // If src0 is on a temporary compute buffers (partial offloading) there may be some padding that needs to be cleared:
-        if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
+        if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
             const int64_t nbytes_data    = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
             const int64_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
             CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));