From c8732dfa6f763332962d97ff040af156e24a9e62 Mon Sep 17 00:00:00 2001 From: Louis Del Valle <92354925+nero-dv@users.noreply.github.com> Date: Wed, 10 May 2023 22:05:18 -0500 Subject: [PATCH] Update sub_quadratic_attention.py 1. Determine the number of query chunks. 2. Calculate the final shape of the res tensor. 3. Initialize the tensor with the calculated shape and dtype, (same dtype as the input tensors, usually) Can initialize the tensor as a zero-filled tensor with the correct shape and dtype, then compute the attention scores for each query chunk and fill the corresponding slice of tensor. --- modules/sub_quadratic_attention.py | 21 +++++++++++++++------ 1 file changed, 15 insertions(+), 6 deletions(-) diff --git a/modules/sub_quadratic_attention.py b/modules/sub_quadratic_attention.py index 055953236..f80c1600a 100644 --- a/modules/sub_quadratic_attention.py +++ b/modules/sub_quadratic_attention.py @@ -202,13 +202,22 @@ def efficient_dot_product_attention( value=value, ) - # TODO: maybe we should use torch.empty_like(query) to allocate storage in-advance, - # and pass slices to be mutated, instead of torch.cat()ing the returned slices - res = torch.cat([ - compute_query_chunk_attn( + # slices of res tensor are mutable, modifications made + # to the slices will affect the original tensor. + # if output of compute_query_chunk_attn function has same number of + # dimensions as input query tensor, we initialize tensor like this: + num_query_chunks = int(np.ceil(q_tokens / query_chunk_size)) + query_shape = get_query_chunk(0).shape + res_shape = (query_shape[0], query_shape[1] * num_query_chunks, *query_shape[2:]) + res_dtype = get_query_chunk(0).dtype + res = torch.zeros(res_shape, dtype=res_dtype) + + for i in range(num_query_chunks): + attn_scores = compute_query_chunk_attn( query=get_query_chunk(i * query_chunk_size), key=key, value=value, - ) for i in range(math.ceil(q_tokens / query_chunk_size)) - ], dim=1) + ) + res[:, i * query_chunk_size:(i + 1) * query_chunk_size, :] = attn_scores + return res