[go: nahoru, domu]
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

Cache: Static cache as a standalone object #30476

Merged
merged 12 commits into from
Apr 30, 2024
Merged

Cache: Static cache as a standalone object #30476

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
fa0b05e
working :D
gante Apr 25, 2024
56cf73b
tmp commit
gante Apr 26, 2024
97aa785
working on generate too; make fixup
gante Apr 26, 2024
a6853a1
dbrx
gante Apr 26, 2024
326a1c7
missing in dbrx
gante Apr 26, 2024
9de9556
more dbrx details; gemma
gante Apr 26, 2024
90771df
tmp
gante Apr 26, 2024
cd6aecd
finalized mvp
gante Apr 26, 2024
b8caa5e
test for different cuda compute capabilities
gante Apr 29, 2024
16e01f2
missing doc entry
gante Apr 29, 2024
2d1eab8
tmp commit
gante Apr 30, 2024
bb159d2
final nits
gante Apr 30, 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
Prev Previous commit
final nits
  • Loading branch information
@gante
gante committed Apr 30, 2024
commit bb159d293d15b49329271c851b54a8446c0ce5ca
2 changes: 1 addition & 1 deletion src/transformers/models/llama/modeling_llama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1055,7 +1055,7 @@ def _update_causal_mask(
# For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
# order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
# to infer the attention mask.
past_seen_tokens = cache_position[0] if past_key_values is not None else 0
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
using_static_cache = isinstance(past_key_values, StaticCache)
Copy link
Contributor

Choose a reason for hiding this comment

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

As I understand it, once the StaticCache is initialized, there is no need to pass it in past_key_values argument. That's why additional condition is necessary. Suggestion:
using_static_cache = isinstance(past_key_values, StaticCache) or isinstance( getattr(self.layers[0].self_attn, "past_key_value", None), StaticCache )

Copy link
Member Author

Choose a reason for hiding this comment

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

@poedator This PR changes precisely the assumption you wrote: we will always need to pass the cache, after this PR it is an object that does NOT live inside the model.

This change will make the transformers' team work easier 🤗

if self.config._attn_implementation == "sdpa" and not using_static_cache:
if AttentionMaskConverter._ignore_causal_mask_sdpa(
Expand Down
4 changes: 4 additions & 0 deletions tests/models/llama/test_modeling_llama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,7 @@
import unittest

import pytest
from packaging import version
from parameterized import parameterized

from transformers import LlamaConfig, is_torch_available, set_seed
Expand Down Expand Up @@ -685,6 +686,9 @@ def test_model_13b_greedy_generation(self):
def test_compile_static_cache(self):
Copy link
Collaborator

Choose a reason for hiding this comment

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

should require torch > 2.2

# `torch==2.2` will throw an error on this test (as in other compilation tests), but torch==2.1.2 and torch>2.2
# work as intended. See https://github.com/pytorch/pytorch/issues/121943
if version.parse(torch.__version__) < version.parse("2.3.0"):
self.skipTest("This test requires torch >= 2.3 to run.")

NUM_TOKENS_TO_GENERATE = 40
# Note on `EXPECTED_TEXT_COMPLETION`'s diff: the current value matches the original test if the original test
# was changed to have a cache of 53 tokens (as opposed to 4096), on Ampere GPUs.
Expand Down
Loading