vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Windows nvcc bug workaround

Browse source
This commit is contained in:
Johannes 2023-06-08 10:02:37 +02:00
parent 5c64a0952e
commit 85a9714316
1 changed files with 7 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

7
ggml-cuda.cu
View file

@ -1512,6 +1512,13 @@ 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.
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;