Avoid the transposed X branch in the Z = X * Y matrix multiplication (#439)

Should make results reproducible for different number of threads and batch sizes
This commit is contained in:
Georgi Gerganov 2023-03-23 23:22:01 +02:00 committed by GitHub
parent 404e1da38e
commit 483bab2e3d
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23

View file

@ -727,11 +727,13 @@ static bool llama_eval_internal(
// V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
struct ggml_tensor * V_trans =
ggml_cpy(ctx0,
ggml_permute(ctx0,
ggml_reshape_3d(ctx0,
ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
n_embd/n_head, n_head, n_past + N),
1, 2, 0, 3);
1, 2, 0, 3),
ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head));
// KQV = transpose(V) * KQ_soft_max
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);