Skip to content

fix: Flash Attention compatibility check for SM_1xx (RTX 5000 series) #1880

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

Open
wants to merge 3 commits into
base: master
Choose a base branch
from
Open

fix: Flash Attention compatibility check for SM_1xx (RTX 5000 series) #1880

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
45a15c8
improve Flash Attention compatibility check for SM_1xx
Apr 7, 2025
41eccad
Merge branch 'master' into rtx5000_fix
ExtReMLapin Apr 11, 2025
ea1b6b0
fix: Flash Attention compatibility check for SM_1xx (RTX 5000 series)
ice0 Apr 12, 2025
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
17 changes: 10 additions & 7 deletions src/models/model.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -844,16 +844,19 @@ namespace ctranslate2 {
" running independently a model in each device");
}

bool is_sm8x = false;
bool is_sm90 = false;
bool supports_flash_attention = false;
if (device == Device::CUDA) {
int device_id = ctranslate2::get_device_index(ctranslate2::Device::CUDA);
auto dprops = ctranslate2::cuda::get_device_properties(device_id);
is_sm8x = dprops.major == 8 && dprops.minor >= 0;
is_sm90 = dprops.major == 9 && dprops.minor == 0;
}
if (use_flash_attention && (device != Device::CUDA || (!is_sm8x && !is_sm90))) {
throw std::invalid_argument("FlashAttention only supports Ampere GPUs or newer.");
float compute_capability = dprops.major + (dprops.minor / 10.0f);

// Minimum compute capability for Flash Attention is Ampere (8.0)
const float min_flash_attn_compute_capability = 8.0f;
supports_flash_attention = compute_capability >= min_flash_attn_compute_capability;
}

if (use_flash_attention && (device != Device::CUDA || !supports_flash_attention)) {
throw std::invalid_argument("FlashAttention only supports Ampere GPUs (compute capability >= 8.0) or newer.");
}
#endif

Expand Down
Loading