CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-04 09:32:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

Windows nvcc workaround (#1753)

Browse Source 
Fix gibberish output on Windows when using CUDA
This commit is contained in:
Johannes Gäßler
2023-06-09 13:58:15 +02:00
committed by GitHub
parent b33dee282f
commit ae9663f188
1 changed files with 9 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
ggml-cuda.cu
View File

@@ -1512,6 +1512,14 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
i01_high = row_high % ne01; i01_high = row_high % ne01;
} }
} }
// There is possibly a bug in the Windows nvcc compiler regarding instruction reordering or optimizing out local variables.
// Removing the first assert or changing the order of the arguments causes the second assert to fail.
// Removing both asserts results in i01_high becoming 0 which in turn results in garbage output.
// The root cause seems to be a problem with i0_offset_high becoming 0 when it should always be >0 (for single GPU).
GGML_ASSERT(i01_low == 0 || g_device_count > 1);
GGML_ASSERT(i01_high == ne01 || g_device_count > 1);
const int64_t i01_diff = i01_high - i01_low; const int64_t i01_diff = i01_high - i01_low;
if (i01_diff == 0) { if (i01_diff == 0) {
continue; continue;
@@ -1727,6 +1735,7 @@ void ggml_cuda_load_data(const char * fname, struct ggml_tensor * tensor, const
row_low -= row_low % GGML_CUDA_DMMV_Y; row_low -= row_low % GGML_CUDA_DMMV_Y;
row_high = id == g_device_count - 1 ? nrows : nrows*g_tensor_split[id + 1]; row_high = id == g_device_count - 1 ? nrows : nrows*g_tensor_split[id + 1];
row_high -= row_high % GGML_CUDA_DMMV_Y; row_high -= row_high % GGML_CUDA_DMMV_Y;
GGML_ASSERT(nrows % GGML_CUDA_DMMV_Y == 0);
} else { } else {
GGML_ASSERT(false); GGML_ASSERT(false);
} }