Skip to content

Commit

Permalink
LLaVA: latency issues (huggingface#34460)
Browse files Browse the repository at this point in the history 
* fix llavas

* code style

* green ci
  • Loading branch information
@zucchini-nlp
zucchini-nlp authored Oct 29, 2024
1 parent a769ed4 commit fe76b60
Show file tree
Hide file tree
Showing 6 changed files with 186 additions and 238 deletions.
127 changes: 60 additions & 67 deletions src/transformers/models/llava/modeling_llava.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -472,76 +472,75 @@ def forward(
(input_ids == self.config.image_token_index).sum(1).max() < self.config.image_seq_length
) or (input_ids.shape[-1] == 1 and pixel_values is not None)

image_features = None
if pixel_values is not None:
image_features = self.get_image_features(
pixel_values=pixel_values,
vision_feature_layer=vision_feature_layer,
vision_feature_select_strategy=vision_feature_select_strategy,
)

if legacy_processing:
logger.warning_once(
"Expanding inputs for image tokens in LLaVa should be done in processing. "
"Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set directly "
"with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
"Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
if legacy_processing:
logger.warning_once(
"Expanding inputs for image tokens in LLaVa should be done in processing. "
"Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set directly "
"with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
"Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
)
# prefill stage vs decoding stage (legacy behavior copied)
if input_ids.shape[1] != 1:
inputs_embeds, attention_mask, labels, position_ids = self._merge_input_ids_with_image_features(
image_features, inputs_embeds, input_ids, attention_mask, labels
)
# prefill stage vs decoding stage (legacy behavior copied)
if input_ids.shape[1] != 1:
inputs_embeds, attention_mask, labels, position_ids = self._merge_input_ids_with_image_features(
image_features, inputs_embeds, input_ids, attention_mask, labels
)
cache_position = torch.arange(attention_mask.shape[1], device=attention_mask.device)
else:
# Retrieve the first layer to inspect the logits and mask out the hidden states
# that are set to 0
first_layer_past_key_value = past_key_values[0][0][:, :, :, 0]

# Sum all dimensions of head_dim (-2) to avoid random errors such as: https://github.com/huggingface/transformers/pull/28032#issuecomment-1863691941
batch_index, non_attended_tokens = torch.where(first_layer_past_key_value.float().sum(-2) == 0)

# Get the target length
target_length = input_ids.shape[1]
past_length = first_layer_past_key_value.shape[-1]

extended_attention_mask = torch.ones(
(attention_mask.shape[0], past_length),
dtype=attention_mask.dtype,
device=attention_mask.device,
)

# Filter out only the tokens that can be un-attended, this can happen
# if one uses Llava + Fused modules where the cache on the
# first iteration is already big enough, or if one passes custom cache
valid_indices = non_attended_tokens < extended_attention_mask.size(-1)
new_batch_index = batch_index[valid_indices]
new_non_attended_tokens = non_attended_tokens[valid_indices]

# Zero-out the places where we don't need to attend
extended_attention_mask[new_batch_index, new_non_attended_tokens] = 0

attention_mask = torch.cat((extended_attention_mask, attention_mask[:, -target_length:]), dim=1)
position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
cache_position = torch.arange(attention_mask.shape[1], device=attention_mask.device)[
-target_length:
]

# TODO: @raushan retain only the new behavior after v4.47
cache_position = torch.arange(attention_mask.shape[1], device=attention_mask.device)
else:
n_image_tokens = (input_ids == self.config.image_token_index).sum(dim=-1)[0].item()
n_image_features = image_features.shape[1]
if n_image_tokens != n_image_features:
raise ValueError(
f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
)
special_image_mask = (
(input_ids == self.config.image_token_index)
.unsqueeze(-1)
.expand_as(inputs_embeds)
.to(inputs_embeds.device)
# Retrieve the first layer to inspect the logits and mask out the hidden states
# that are set to 0
first_layer_past_key_value = past_key_values[0][0][:, :, :, 0]

# Sum all dimensions of head_dim (-2) to avoid random errors such as: https://github.com/huggingface/transformers/pull/28032#issuecomment-1863691941
batch_index, non_attended_tokens = torch.where(first_layer_past_key_value.float().sum(-2) == 0)

# Get the target length
target_length = input_ids.shape[1]
past_length = first_layer_past_key_value.shape[-1]

extended_attention_mask = torch.ones(
(attention_mask.shape[0], past_length),
dtype=attention_mask.dtype,
device=attention_mask.device,
)
image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)

# Filter out only the tokens that can be un-attended, this can happen
# if one uses Llava + Fused modules where the cache on the
# first iteration is already big enough, or if one passes custom cache
valid_indices = non_attended_tokens < extended_attention_mask.size(-1)
new_batch_index = batch_index[valid_indices]
new_non_attended_tokens = non_attended_tokens[valid_indices]

# Zero-out the places where we don't need to attend
extended_attention_mask[new_batch_index, new_non_attended_tokens] = 0

attention_mask = torch.cat((extended_attention_mask, attention_mask[:, -target_length:]), dim=1)
position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
cache_position = torch.arange(attention_mask.shape[1], device=attention_mask.device)[-target_length:]

# TODO: @raushan retain only the new behavior after v4.47
elif image_features is not None:
n_image_tokens = (input_ids == self.config.image_token_index).sum(dim=-1)[0].item()
n_image_features = image_features.shape[1]
if n_image_tokens != n_image_features:
raise ValueError(
f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
)
special_image_mask = (
(input_ids == self.config.image_token_index)
.unsqueeze(-1)
.expand_as(inputs_embeds)
.to(inputs_embeds.device)
)
image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)

outputs = self.language_model(
attention_mask=attention_mask,
Expand Down Expand Up @@ -602,12 +601,6 @@ def prepare_inputs_for_generation(
):
# Overwritten -- in specific circumstances we don't want to forward image inputs to the model

# Trigger the new behavior if we have more than image embeddings seq length tokens for images
legacy_processing = (
input_ids is not None
and (input_ids == self.config.image_token_index).sum(1).max() < self.config.image_seq_length
)

model_inputs = self.language_model.prepare_inputs_for_generation(
input_ids,
past_key_values=past_key_values,
Expand All @@ -618,7 +611,7 @@ def prepare_inputs_for_generation(
**kwargs,
)

if legacy_processing or cache_position[0] == 0:
if cache_position[0] == 0:
# If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
# Otherwise we need pixel values to be passed to model
model_inputs["pixel_values"] = pixel_values
Expand Down
135 changes: 65 additions & 70 deletions src/transformers/models/llava_next/modeling_llava_next.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -846,6 +846,7 @@ def forward(
(input_ids == self.config.image_token_index).sum(1).max() < self.config.image_seq_length
) or (input_ids.shape[-1] == 1 and pixel_values is not None)

image_features = None
if pixel_values is not None and pixel_values.size(0) > 0:
image_features = self.get_image_features(
pixel_values,
Expand All @@ -861,74 +862,73 @@ def forward(
vision_feature_select_strategy=vision_feature_select_strategy,
image_newline=self.image_newline,
)
if legacy_processing:
logger.warning_once(
"Expanding inputs for image tokens in LLaVa-NeXT should be done in processing. "
"Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set directly "
"with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
"Using processors without these attributes in the config is deprecated and will throw an error in v4.47."

if legacy_processing:
logger.warning_once(
"Expanding inputs for image tokens in LLaVa-NeXT should be done in processing. "
"Please add `patch_size` and `vision_feature_select_strategy` to the model's processing config or set directly "
"with `processor.patch_size = {{patch_size}}` and processor.vision_feature_select_strategy = {{vision_feature_select_strategy}}`. "
"Using processors without these attributes in the config is deprecated and will throw an error in v4.47."
)
if input_ids.shape[1] != 1:
inputs_embeds = inputs_embeds.to(image_features.dtype)
inputs_embeds, attention_mask, position_ids, labels, _ = self._merge_input_ids_with_image_features(
image_features,
feature_lens,
inputs_embeds,
input_ids,
attention_mask,
position_ids,
labels=labels,
)
if input_ids.shape[1] != 1:
inputs_embeds = inputs_embeds.to(image_features.dtype)
inputs_embeds, attention_mask, position_ids, labels, _ = self._merge_input_ids_with_image_features(
image_features,
feature_lens,
inputs_embeds,
input_ids,
attention_mask,
position_ids,
labels=labels,
)
cache_position = torch.arange(attention_mask.shape[1], device=attention_mask.device)
else:
# Retrieve the first layer to inspect the logits and mask out the hidden states
# that are set to 0
first_layer_past_key_value = past_key_values[0][0][:, :, :, 0]

# Sum all dimensions of head_dim (-2) to avoid random errors such as: https://github.com/huggingface/transformers/pull/28032#issuecomment-1863691941
batch_index, non_attended_tokens = torch.where(first_layer_past_key_value.float().sum(-2) == 0)

# Get the target length
target_length = input_ids.shape[1]
past_length = first_layer_past_key_value.shape[-1]

extended_attention_mask = torch.ones(
(attention_mask.shape[0], past_length),
dtype=attention_mask.dtype,
device=attention_mask.device,
)
cache_position = torch.arange(attention_mask.shape[1], device=attention_mask.device)
else:
# Retrieve the first layer to inspect the logits and mask out the hidden states
# that are set to 0
first_layer_past_key_value = past_key_values[0][0][:, :, :, 0]

# Filter out only the tokens that can be un-attended, this can happen
# if one uses Llava + Fused modules where the cache on the
# first iteration is already big enough, or if one passes custom cache
valid_indices = non_attended_tokens < extended_attention_mask.size(-1)
new_batch_index = batch_index[valid_indices]
new_non_attended_tokens = non_attended_tokens[valid_indices]

# Zero-out the places where we don't need to attend
extended_attention_mask[new_batch_index, new_non_attended_tokens] = 0
attention_mask = torch.cat((extended_attention_mask, attention_mask[:, -target_length:]), dim=1)
position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
cache_position = torch.arange(attention_mask.shape[1], device=attention_mask.device)[
-target_length:
]
# Sum all dimensions of head_dim (-2) to avoid random errors such as: https://github.com/huggingface/transformers/pull/28032#issuecomment-1863691941
batch_index, non_attended_tokens = torch.where(first_layer_past_key_value.float().sum(-2) == 0)

# TODO: @raushan retain only the new behavior after v4.47
else:
n_image_tokens = (input_ids == self.config.image_token_index).sum().item()
n_image_features = image_features.shape[0]
if n_image_tokens != n_image_features:
raise ValueError(
f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
)
special_image_mask = (
(input_ids == self.config.image_token_index)
.unsqueeze(-1)
.expand_as(inputs_embeds)
.to(inputs_embeds.device)
# Get the target length
target_length = input_ids.shape[1]
past_length = first_layer_past_key_value.shape[-1]

extended_attention_mask = torch.ones(
(attention_mask.shape[0], past_length),
dtype=attention_mask.dtype,
device=attention_mask.device,
)

# Filter out only the tokens that can be un-attended, this can happen
# if one uses Llava + Fused modules where the cache on the
# first iteration is already big enough, or if one passes custom cache
valid_indices = non_attended_tokens < extended_attention_mask.size(-1)
new_batch_index = batch_index[valid_indices]
new_non_attended_tokens = non_attended_tokens[valid_indices]

# Zero-out the places where we don't need to attend
extended_attention_mask[new_batch_index, new_non_attended_tokens] = 0
attention_mask = torch.cat((extended_attention_mask, attention_mask[:, -target_length:]), dim=1)
position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
cache_position = torch.arange(attention_mask.shape[1], device=attention_mask.device)[-target_length:]

# TODO: @raushan retain only the new behavior after v4.47
elif image_features is not None:
n_image_tokens = (input_ids == self.config.image_token_index).sum().item()
n_image_features = image_features.shape[0]
if n_image_tokens != n_image_features:
raise ValueError(
f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
)
image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
special_image_mask = (
(input_ids == self.config.image_token_index)
.unsqueeze(-1)
.expand_as(inputs_embeds)
.to(inputs_embeds.device)
)
image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)

outputs = self.language_model(
attention_mask=attention_mask,
Expand Down Expand Up @@ -990,11 +990,6 @@ def prepare_inputs_for_generation(
):
# Overwritten -- in specific circumstances we don't want to forward image inputs to the model

legacy_processing = (
input_ids is not None
and (input_ids == self.config.image_token_index).sum(1).max() < self.config.image_seq_length
)

model_inputs = self.language_model.prepare_inputs_for_generation(
input_ids,
past_key_values=past_key_values,
Expand All @@ -1007,7 +1002,7 @@ def prepare_inputs_for_generation(

# If we're in cached decoding stage, pixel values should be None because input ids do not contain special image token anymore
# Otherwise we need pixel values to be passed to model
if legacy_processing or cache_position[0] == 0:
if cache_position[0] == 0:
model_inputs["pixel_values"] = pixel_values
model_inputs["image_sizes"] = image_sizes

Expand Down
Loading

0 comments on commit fe76b60

Please sign in to comment.