Skip to content
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

Fix: bwd may need to first allocate cuda mem for rng_state #1077

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix: bwd may need to first allocate cuda mem for rng_state #1077

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
994ccef
jundaf: may need to first allocate cuda mem for rng_srate
Jul 20, 2024
de1945b
update code
Jul 20, 2024
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
6 changes: 6 additions & 0 deletions csrc/flash_attn/flash_api.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -989,6 +989,7 @@ mha_bwd(const at::Tensor &dout, // batch_size x seqlen_q x num_heads, x head_si

// We use a custom RNG that increases the offset by batch_size * nheads * 32.
int64_t counter_offset = params.b * params.h * 32;
at::Tensor tmp_rng_state;

if ( rng_state.has_value() ) {
params.rng_state = reinterpret_cast<uint64_t*>(rng_state.value().data_ptr());
Expand All @@ -997,6 +998,8 @@ mha_bwd(const at::Tensor &dout, // batch_size x seqlen_q x num_heads, x head_si
std::lock_guard<std::mutex> lock(gen->mutex_);
params.philox_args = gen->philox_cuda_state(counter_offset);
auto seeds = at::cuda::philox::unpack(params.philox_args);
tmp_rng_state = torch::empty({2}, opts.dtype(torch::kInt64));
params.rng_state = reinterpret_cast<uint64_t*>(tmp_rng_state.data_ptr());
params.rng_state[0] = std::get<0>(seeds);
params.rng_state[1] = std::get<1>(seeds);
}
Expand Down Expand Up @@ -1243,6 +1246,7 @@ mha_varlen_bwd(const at::Tensor &dout, // total_q x num_heads, x head_size

// We use a custom RNG that increases the offset by batch_size * nheads * 32.
int64_t counter_offset = params.b * params.h * 32;
at::Tensor tmp_rng_state;

if ( rng_state.has_value() ) {
params.rng_state = reinterpret_cast<uint64_t*>(rng_state.value().data_ptr());
Expand All @@ -1251,6 +1255,8 @@ mha_varlen_bwd(const at::Tensor &dout, // total_q x num_heads, x head_size
std::lock_guard<std::mutex> lock(gen->mutex_);
params.philox_args = gen->philox_cuda_state(counter_offset);
auto seeds = at::cuda::philox::unpack(params.philox_args);
tmp_rng_state = torch::empty({2}, opts.dtype(torch::kInt64));
params.rng_state = reinterpret_cast<uint64_t*>(tmp_rng_state.data_ptr());
params.rng_state[0] = std::get<0>(seeds);
params.rng_state[1] = std::get<1>(seeds);
}
Expand Down