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

Jump to bottom

[core] Make AutoAWQ fused modules compatible with HF transformers #244

Merged
merged 2 commits into from
Dec 11, 2023
Merged

[core] Make AutoAWQ fused modules compatible with HF transformers #244

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
38cd8b2
Update attn.py
younesbelkada Dec 8, 2023
7001b2a
Update attn.py
younesbelkada Dec 8, 2023
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
16 changes: 16 additions & 0 deletions awq/modules/fused/attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,12 +6,22 @@
from awq.modules.fused.cache import WindowedCache
from awq.utils.fused_utils import get_attention_shapes


try:
import ft_inference_engine
FT_INSTALLED = True
except:
FT_INSTALLED = False

HF_NEW_CACHE_FORMAT = False

import transformers
# https://github.com/huggingface/transformers/pull/26681 introduced a new cache format
HF_NEW_CACHE_FORMAT = hasattr(transformers, "cache_utils")
if HF_NEW_CACHE_FORMAT:
from transformers.cache_utils import DynamicCache


class RoPE(nn.Module):
def __init__(self, hidden_size, n_heads, max_seq_len, device):
super(RoPE, self).__init__()
Expand Down Expand Up @@ -223,4 +233,10 @@ def forward(self, hidden_states:torch.Tensor, attention_mask=None, *args, **kwar
# we pass a dummy past kv cache for transformers to be able to retrieve the correct info
# about past key length
past_key_value = [torch.zeros(1, 1, self.start_pos, 1)]

if HF_NEW_CACHE_FORMAT and self.is_hf_transformers:
new_cache = DynamicCache()
new_cache.update(past_key_value[0], past_key_value[0], layer_idx=0)
past_key_value = new_cache

return attn_output, attention_weight, past_key_value