Vulkan: Fix minor debug mode issues #14899
Conversation
github-actions
bot
added
Vulkan
ggml/src/ggml-vulkan/ggml-vulkan.cpp
Outdated
| @@ -11379,7 +11379,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph * | |||
| tensor_clone = ggml_cpy(ggml_ctx, src_clone[0], src_clone[1]); | |||
| } | |||
| } else if (tensor->op == GGML_OP_SET_ROWS) { | |||
| tensor_clone = ggml_set_rows(ggml_ctx, src_clone[0], src_clone[1]); | |||
| tensor_clone = ggml_set_rows(ggml_ctx, src_clone[0], src_clone[1], src_clone[2]); | |||
There was a problem hiding this comment.
I'm not sure what the parameters should be here. Do we need to clone the dst tensor and pass that as the 'a' parameter?
There was a problem hiding this comment.
I didn't look at it in detail, but this fixed the issue I had with it.
There was a problem hiding this comment.
This is a weird operation because it doesn't keep around it's original 'a' parameter, and I think src_clone[2] will be null. And since it only writes sparsely to the destination we probably need to dup the original contents. So maybe something like:
tensor_clone = ggml_dup(ggml_ctx, dst->view_src);
tensor_clone = ggml_set_rows(ggml_ctx, tensor_clone, src_clone[0], src_clone[1]);
There was a problem hiding this comment.
I checked again, my code just fixed the compile problem, it does not work if actually used. But neither does yours, since the tensor data is still on GPU, so dup does not work. I think manually copying the view tensor to CPU and using it as view_src would work, but that would mean reworking the check_result functions to support this.
I just deactivated SET_ROWS for now to fix the issue, if we need set_rows checks they can be readded at a later point.
0cc4m
force-pushed
the
0cc4m/vulkan-debug-fixes
branch
from
dbf7d78 to
573670b
Compare
No description provided.