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

Support FP8 model fallback KVCache to bfloat16 #1505

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

Support FP8 model fallback KVCache to bfloat16 #1505

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ed64add
Support FP8 model fallback kvcache to bfloat16
changwangss Nov 20, 2024
0863e51
improve
changwangss Nov 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
8 changes: 6 additions & 2 deletions optimum/habana/transformers/models/llama/modeling_llama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -628,10 +628,14 @@ def pre_attn_forward(
else:
if past_key_value is None:
past_key = torch.zeros(
key_states.shape, dtype=self.get_k_proj_weight_dtype(), device=key_states.device
key_states.shape,
dtype=torch.bfloat16 if isinstance(self.k_cache, KVCache) else self.get_k_proj_weight_dtype(),
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why not use the function?
The default value is:
self.k_proj.weight.dtype

Copy link
Contributor Author

@changwangss changwangss Nov 21, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

for recipes FP8Config(allowlist={"types": ["Linear"], "names": []}, blocklist=blocklist = {"types": [], "names": []}), self.k_proj.weight.dtype is torch.float8_e4m3fn, but the past_key dtype should be torch.bfloat16

device=key_states.device
)
past_value = torch.zeros(
key_states.shape, dtype=self.get_k_proj_weight_dtype(), device=key_states.device
key_states.shape,
dtype=torch.bfloat16 if isinstance(self.v_cache, KVCache) else self.get_k_proj_weight_dtype(),
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

same

device=key_states.device
)
# Return list instead of tuple
past_key_value = [past_key, past_value]
Expand Down