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

RoPE models: add numerical sanity-check test for RoPE scaling (#29808)

Browse Source 
* add hard rope scaling test

* make fixup

* quick rope scaling tests

* add copy statements
This commit is contained in:
Joao Gante
2024-03-28 11:25:50 +00:00
committed by GitHub
parent aac7099c92
commit 441de62f49
6 changed files with 435 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

70
tests/models/llama/test_modeling_llama.py
View File

@@ -51,6 +51,11 @@ if is_torch_available():
LlamaModel,
LlamaTokenizer,
)
from transformers.models.llama.modeling_llama import (
LlamaDynamicNTKScalingRotaryEmbedding,
LlamaLinearScalingRotaryEmbedding,
LlamaRotaryEmbedding,
)
class LlamaModelTester:
@@ -370,7 +375,7 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
pass
@parameterized.expand([("linear",), ("dynamic",)])
def test_model_rope_scaling(self, scaling_type):
def test_model_rope_scaling_from_config(self, scaling_type):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
short_input = ids_tensor([1, 10], config.vocab_size)
long_input = ids_tensor([1, int(config.max_position_embeddings * 1.5)], config.vocab_size)
@@ -400,6 +405,69 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
# The output should be different for long inputs
self.assertFalse(torch.allclose(original_long_output, scaled_long_output, atol=1e-5))
def test_model_rope_scaling(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
hidden_size = config.hidden_size
num_heads = config.num_attention_heads
head_dim = hidden_size // num_heads
scaling_factor = 10
short_input_length = 10
long_input_length = int(config.max_position_embeddings * 1.5)
# Inputs
x = torch.randn(1, dtype=torch.float32, device=torch_device) # used exlusively to get the dtype and the device
position_ids_short = torch.arange(short_input_length, dtype=torch.long, device=torch_device)
position_ids_short = position_ids_short.unsqueeze(0)
position_ids_long = torch.arange(long_input_length, dtype=torch.long, device=torch_device)
position_ids_long = position_ids_long.unsqueeze(0)
# Sanity check original RoPE
original_rope = LlamaRotaryEmbedding(
head_dim,
max_position_embeddings=config.max_position_embeddings,
base=config.rope_theta,
).to(torch_device)
original_cos_short, original_sin_short = original_rope(x, position_ids_short)
original_cos_long, original_sin_long = original_rope(x, position_ids_long)
torch.testing.assert_close(original_cos_short, original_cos_long[:, :short_input_length, :])
torch.testing.assert_close(original_sin_short, original_sin_long[:, :short_input_length, :])
# Sanity check linear RoPE scaling
# New position "x" should match original position with index "x/scaling_factor"
linear_scaling_rope = LlamaLinearScalingRotaryEmbedding(
head_dim,
max_position_embeddings=config.max_position_embeddings,
base=config.rope_theta,
scaling_factor=scaling_factor,
).to(torch_device)
linear_cos_short, linear_sin_short = linear_scaling_rope(x, position_ids_short)
linear_cos_long, linear_sin_long = linear_scaling_rope(x, position_ids_long)
torch.testing.assert_close(linear_cos_short, linear_cos_long[:, :short_input_length, :])
torch.testing.assert_close(linear_sin_short, linear_sin_long[:, :short_input_length, :])
for new_position in range(0, long_input_length, scaling_factor):
original_position = int(new_position // scaling_factor)
torch.testing.assert_close(linear_cos_long[:, new_position, :], original_cos_long[:, original_position, :])
torch.testing.assert_close(linear_sin_long[:, new_position, :], original_sin_long[:, original_position, :])
# Sanity check Dynamic NTK RoPE scaling
# Scaling should only be observed after a long input is fed. We can observe that the frequencies increase
# with scaling_factor (or that `inv_freq` decreases)
ntk_scaling_rope = LlamaDynamicNTKScalingRotaryEmbedding(
head_dim,
max_position_embeddings=config.max_position_embeddings,
base=config.rope_theta,
scaling_factor=scaling_factor,
).to(torch_device)
ntk_cos_short, ntk_sin_short = ntk_scaling_rope(x, position_ids_short)
ntk_cos_long, ntk_sin_long = ntk_scaling_rope(x, position_ids_long)
torch.testing.assert_close(ntk_cos_short, original_cos_short)
torch.testing.assert_close(ntk_sin_short, original_sin_short)
with self.assertRaises(AssertionError):
torch.testing.assert_close(ntk_cos_long, original_cos_long)
with self.assertRaises(AssertionError):
torch.testing.assert_close(ntk_sin_long, original_sin_long)
self.assertTrue((ntk_scaling_rope.inv_freq <= original_rope.inv_freq).all())
@require_flash_attn
@require_torch_gpu
@require_bitsandbytes