Skip to content

Fix issue with corrupted output on long generation for GPT #2344

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
andrewchernyh wants to merge 2 commits into from
Closed

Fix issue with corrupted output on long generation for GPT #2344

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d9e34ed
Fix issue with corrupted output on long generation for GPT
andrewchernyh Sep 22, 2022
13f57ce
Update MAX_OUT_TOKES to 2048 as in GPT-J
andrewchernyh Sep 22, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions csrc/transformer/inference/csrc/pt_binding.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -107,7 +107,7 @@ void allocate_workspace(size_t hidden_dim,
unsigned num_layers,
size_t head_size = 128)
{
size_t _workSpaceSize = 16 * (hidden_dim * batch_size * max_seq_len) +
size_t _workSpaceSize = 20 * (hidden_dim * batch_size * max_seq_len) +
(num_layers * batch_size * max_seq_len * hidden_dim * 2); // KV-cache
Context::Instance().GenWorkSpace(_workSpaceSize * sizeof(T));
}
Expand Down Expand Up @@ -451,7 +451,7 @@ std::vector<at::Tensor> ds_softmax_context(at::Tensor& query_key_value,
auto kv_cache = workspace + offset + (hidden_dim / heads) * (is_prompt ? 0 : soft_len - 1);
size_t value_offset = bsz * MAX_OUT_TOKES * hidden_dim;

T* temp_buf = (T*)output.data_ptr() + at::numel(output);
T* temp_buf = (T*)kv_cache+value_offset+value_offset;
launch_bias_add_transform_0213<T>((T*)query_cont,
kv_cache,
kv_cache + value_offset,
Expand Down
2 changes: 1 addition & 1 deletion csrc/transformer/inference/includes/inference_cuda_layers.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,7 +21,7 @@ Copyright 2022 The Microsoft DeepSpeed Team
#include <cassert>
#include <iostream>

#define MAX_OUT_TOKES 128
#define MAX_OUT_TOKES 2048
#define MAX_WARP_NUM 32
#define WARP_SIZE 32

Expand Down