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

Llama: RoPE refactor (#32135)

Browse Source 
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
This commit is contained in:
Joao Gante
2024-07-23 10:42:55 +01:00
committed by GitHub
parent 5a4a76edb7
commit 2e113422b3
20 changed files with 830 additions and 285 deletions
Download Patch File Download Diff File Expand all files Collapse all files

78
tests/models/llama/test_modeling_llama.py
View File

@@ -51,12 +51,7 @@ if is_torch_available():
LlamaModel,
LlamaTokenizer,
)
from transformers.models.llama.modeling_llama import (
LlamaDynamicNTKScalingRotaryEmbedding,
LlamaLinearScalingRotaryEmbedding,
LlamaRotaryEmbedding,
LlamaYarnScalingRotaryEmbedding,
)
from transformers.models.llama.modeling_llama import LlamaLinearScalingRotaryEmbedding, LlamaRotaryEmbedding
class LlamaModelTester:
@@ -431,9 +426,6 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
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)
@@ -446,11 +438,7 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
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_rope = LlamaRotaryEmbedding(config=config).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, :])
@@ -458,12 +446,8 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
# 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)
config.rope_scaling = {"type": "linear", "factor": scaling_factor}
linear_scaling_rope = LlamaRotaryEmbedding(config=config).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, :])
@@ -476,12 +460,8 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
# 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)
config.rope_scaling = {"type": "dynamic", "factor": scaling_factor}
ntk_scaling_rope = LlamaRotaryEmbedding(config=config).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)
@@ -493,12 +473,9 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
self.assertTrue((ntk_scaling_rope.inv_freq <= original_rope.inv_freq).all())
# Sanity check Yarn RoPE scaling
yarn_scaling_rope = LlamaYarnScalingRotaryEmbedding(
head_dim,
max_position_embeddings=config.max_position_embeddings,
base=config.rope_theta,
scaling_factor=scaling_factor,
).to(torch_device)
# Scaling should be over the entire input
config.rope_scaling = {"type": "yarn", "factor": scaling_factor}
yarn_scaling_rope = LlamaRotaryEmbedding(config=config).to(torch_device)
yarn_cos_short, yarn_sin_short = yarn_scaling_rope(x, position_ids_short)
yarn_cos_long, yarn_sin_long = yarn_scaling_rope(x, position_ids_long)
torch.testing.assert_close(yarn_cos_short, yarn_cos_long[:, :short_input_length, :])
@@ -512,6 +489,43 @@ class LlamaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
with self.assertRaises(AssertionError):
torch.testing.assert_close(yarn_sin_long, original_sin_long)
def test_rope_class_retrocompatibility(self):
# Delete me when we remove compatibility for the old API :)
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
scaling_factor = 10
short_input_length = 10
long_input_length = int(config.max_position_embeddings * 1.5)
config.rope_scaling = {"type": "linear", "factor": 10}
# 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)
# Old API -- under the hood, "type": "linear" is set and `LlamaRotaryEmbedding` is called
old_api_rope = LlamaLinearScalingRotaryEmbedding(
config.hidden_size // config.num_attention_heads,
max_position_embeddings=config.max_position_embeddings,
base=config.rope_theta,
scaling_factor=scaling_factor,
).to(torch_device)
old_cos_short, old_sin_short = old_api_rope(x, position_ids_short)
old_cos_long, old_sin_long = old_api_rope(x, position_ids_long)
# New API
config.rope_scaling = {"type": "linear", "factor": scaling_factor}
new_api_rope = LlamaRotaryEmbedding(config=config).to(torch_device)
new_cos_short, new_sin_short = new_api_rope(x, position_ids_short)
new_cos_long, new_sin_long = new_api_rope(x, position_ids_long)
# The results should match
torch.testing.assert_close(old_cos_short, new_cos_short)
torch.testing.assert_close(old_sin_short, new_sin_short)
torch.testing.assert_close(old_cos_long, new_cos_long)
torch.testing.assert_close(old_sin_long, new_sin_long)
@require_flash_attn
@require_torch_gpu
@require_bitsandbytes