Skip to content

Allow f16/bf16 in custom SDPA #237

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
GregoryComer wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Allow f16/bf16 in custom SDPA #237

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
2 changes: 1 addition & 1 deletion optimum/executorch/attentions/custom_kv_cache.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -174,7 +174,7 @@ def from_legacy_cache(
dtype = legacy_cache.k_cache.dtype

# assert device is None or device == "cpu"
assert dtype is None or dtype == torch.float32
assert dtype is None or dtype in (torch.float32, torch.bfloat16, torch.float16)

# Use the legacy cache's max_seq_len if max_cache_len is not specified
if max_cache_len is None and hasattr(legacy_cache, "max_seq_len"):
Expand Down
4 changes: 0 additions & 4 deletions optimum/executorch/attentions/custom_sdpa.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -81,11 +81,7 @@ def custom_sdpa_with_start_pos_forward(
key = key.transpose(1, 2)
value = value.transpose(1, 2)

# Convert the hell out of the inputs to fp32 and back
input_dtype = query.dtype
query = query.to(torch.float32)
key = key.to(torch.float32)
value = value.to(torch.float32)

# Ignore the causal flag from kwargs but use the one in module
kwargs.pop("is_causal", None)
Expand Down
Loading