Compile compatibilty for decoder-only models (#32617)

* squash into one commit

* add qwen2-vl for rope standardization

* fix mistral compile

* fix qwen2-vl

* fix-copies
This commit is contained in:
Raushan Turganbay
2024-09-09 10:59:04 +02:00
committed by GitHub
parent eedd21b9e7
commit 65bb284448
37 changed files with 2301 additions and 1367 deletions

View File

@@ -461,6 +461,10 @@ class FalconModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMix
# 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 = FalconRotaryEmbedding(
@@ -468,10 +472,10 @@ class FalconModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMix
max_position_embeddings=config.max_position_embeddings,
base=config.rope_theta,
).to(torch_device)
original_cos_short, original_sin_short = original_rope(x, short_input_length)
original_cos_long, original_sin_long = original_rope(x, long_input_length)
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, :])
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"
@@ -481,14 +485,14 @@ class FalconModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMix
base=config.rope_theta,
scaling_factor=scaling_factor,
).to(torch_device)
linear_cos_short, linear_sin_short = linear_scaling_rope(x, short_input_length)
linear_cos_long, linear_sin_long = linear_scaling_rope(x, long_input_length)
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, :])
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, :])
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
@@ -499,8 +503,8 @@ class FalconModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMix
base=config.rope_theta,
scaling_factor=scaling_factor,
).to(torch_device)
ntk_cos_short, ntk_sin_short = ntk_scaling_rope(x, short_input_length)
ntk_cos_long, ntk_sin_long = ntk_scaling_rope(x, long_input_length)
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):