SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

[generate] shape checks in tests compatible with fixed-length caches (+ some minor fixes) (#35993)

Browse Source 
* shape checks compatible with static cache

* add test

* tmp

* manually turn on eager attn when we want to output attn

* typo

* generalize to encoder-decoder models

* force compilation on cpu

* tmp commit

* fix static cache shape checks

* models with odd caches

* fix copies

* shorter cache search loop

* use decoder_past_key_values everywhere

* better test variable names and comments

* signature

* rename _check_outputs into _check_generate_outputs

* add comments

* HybridCache future test note
This commit is contained in:
Joao Gante
2025-02-10 17:50:54 +00:00
committed by GitHub
parent 9510ae39d9
commit be2ac0916a
25 changed files with 379 additions and 917 deletions
Download Patch File Download Diff File Expand all files Collapse all files

69
tests/models/moshi/test_modeling_moshi.py
View File

@@ -575,77 +575,12 @@ class MoshiTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
return config, filtered_inputs_dict
def _check_hidden_states_for_generate(
self, batch_size, hidden_states, min_length, max_length, config, use_cache=False, num_beam_groups=1
):
def _check_generate_outputs(self, output, config, use_cache=False, num_return_sequences=1, num_beams=1):
# Overwrite because the generate method actually alway uses `inputs_embeds` so `use_cache` is always `True`
self.assertIsInstance(hidden_states, tuple)
self.assertListEqual(
[isinstance(iter_hidden_states, tuple) for iter_hidden_states in hidden_states],
[True] * len(hidden_states),
)
self.assertEqual(len(hidden_states), (max_length - min_length) * num_beam_groups)
for idx, iter_hidden_states in enumerate(hidden_states):
seq_len = min_length if idx == 0 else 1
expected_shape = (batch_size * num_beam_groups, seq_len, config.hidden_size)
# check hidden size
self.assertListEqual(
[layer_hidden_states.shape for layer_hidden_states in iter_hidden_states],
[expected_shape] * len(iter_hidden_states),
)
def _check_outputs(self, output, config, use_cache=False, num_return_sequences=1, num_beams=1):
# Overwrite because the generate method actually alway uses `inputs_embeds` so `use_cache` is always `True`
super()._check_outputs(
super()._check_generate_outputs(
output, config, use_cache=True, num_return_sequences=num_return_sequences, num_beams=num_beams
)
def _check_hidden_states_for_generate(
self, batch_size, hidden_states, min_length, max_length, config, use_cache=False, num_beam_groups=1
):
# Overwrite because the generate method actually alway uses `inputs_embeds` so `use_cache` is always `True`
self.assertIsInstance(hidden_states, tuple)
self.assertListEqual(
[isinstance(iter_hidden_states, tuple) for iter_hidden_states in hidden_states],
[True] * len(hidden_states),
)
self.assertEqual(len(hidden_states), (max_length - min_length) * num_beam_groups)
for idx, iter_hidden_states in enumerate(hidden_states):
seq_len = 1
expected_shape = (batch_size * num_beam_groups, seq_len, config.hidden_size)
# check hidden size
self.assertListEqual(
[layer_hidden_states.shape for layer_hidden_states in iter_hidden_states],
[expected_shape] * len(iter_hidden_states),
)
def _check_attentions_for_generate(
self, batch_size, attentions, min_length, max_length, config, use_cache=False, num_beam_groups=1
):
# Overwrite because the generate method actually alway uses `inputs_embeds` so `use_cache` is always `True`
self.assertIsInstance(attentions, tuple)
self.assertListEqual(
[isinstance(iter_attentions, tuple) for iter_attentions in attentions], [True] * len(attentions)
)
self.assertEqual(len(attentions), (max_length - min_length) * num_beam_groups)
for idx, iter_attentions in enumerate(attentions):
tgt_len = 1
src_len = min_length + idx
expected_shape = (
batch_size * num_beam_groups,
config.num_attention_heads,
tgt_len,
src_len,
)
# check attn size
self.assertListEqual(
[layer_attention.shape for layer_attention in iter_attentions], [expected_shape] * len(iter_attentions)
)
def test_initialization(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()