🔴 VLM: compile compatibility (#35724)

* llavas

* add mroe models

* fix `compile_forward` test for all models

* fix copies

* make style

* also doesn't support cache class

* fix some tests

* not copied from

* ci green?

* fix tests

* fix copies

* fix tests

* check with `numel` and remove `item`

* fix copies

* fix copies

* Update src/transformers/models/cohere2/modeling_cohere2.py

Co-authored-by: Arthur <[email protected]>

* opt remove cross attn

* gemma2

* fixup

* fixup

* fix newly added test

* maybe fixed?

* green please?

---------

Co-authored-by: Arthur <[email protected]>

src/transformers/models/blip_2/modeling_blip_2.py

@@ -2016,6 +2016,9 @@ def forward(
 class Blip2ForConditionalGeneration(Blip2PreTrainedModel, GenerationMixin):
     config_class = Blip2Config
     main_input_name = "pixel_values"
+    _supports_cache_class = True
+    _supports_static_cache = True
+    _supports_quantized_cache = False  # not all LM bacbones support (e.g. T5)
 
     def __init__(self, config: Blip2Config):
         super().__init__(config)

src/transformers/models/chameleon/modeling_chameleon.py

@@ -1284,13 +1284,13 @@ def forward(
 
         if pixel_values is not None:
             image_tokens = self.get_image_tokens(pixel_values)
-            n_image_tokens_in_text = (input_ids == self.vocabulary_mapping.image_token_id).sum().item()
-            n_image_features = image_tokens.shape[0] * image_tokens.shape[1]
-            if n_image_tokens_in_text != n_image_features:
+            special_image_mask = input_ids == self.vocabulary_mapping.image_token_id
+            if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_tokens.numel():
+                n_image_tokens_in_text = (input_ids == self.vocabulary_mapping.image_token_id).sum()
+                n_image_features = image_tokens.shape[0] * image_tokens.shape[1]
                 raise ValueError(
                     f"Image features and image tokens do not match: tokens: {n_image_tokens_in_text}, features {n_image_features}"
                 )
-            special_image_mask = input_ids == self.vocabulary_mapping.image_token_id
             image_tokens = image_tokens.to(input_ids.device, input_ids.dtype)
             input_ids = input_ids.masked_scatter(special_image_mask, image_tokens)
 

src/transformers/models/cohere2/modeling_cohere2.py

@@ -25,7 +25,7 @@
 import torch.nn as nn
 
 from ...activations import ACT2FN
-from ...cache_utils import Cache, HybridCache
+from ...cache_utils import Cache, HybridCache, StaticCache
 from ...generation import GenerationMixin
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
@@ -701,7 +701,7 @@ def _update_causal_mask(
 
         dtype, device = input_tensor.dtype, input_tensor.device
         sequence_length = input_tensor.shape[1]
-        if isinstance(past_key_values, HybridCache):
+        if isinstance(past_key_values, (HybridCache, StaticCache)):
             target_length = past_key_values.get_max_cache_shape()
         else:
             target_length = attention_mask.shape[-1] if attention_mask is not None else input_tensor.shape[1]

src/transformers/models/gemma2/modeling_gemma2.py

@@ -25,7 +25,7 @@
 import torch.nn as nn
 
 from ...activations import ACT2FN
-from ...cache_utils import Cache, HybridCache
+from ...cache_utils import Cache, HybridCache, StaticCache
 from ...generation import GenerationMixin
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_outputs import (
@@ -713,7 +713,7 @@ def _update_causal_mask(
 
         dtype, device = input_tensor.dtype, input_tensor.device
         sequence_length = input_tensor.shape[1]
-        if isinstance(past_key_values, HybridCache):
+        if isinstance(past_key_values, (HybridCache, StaticCache)):
             target_length = past_key_values.get_max_cache_shape()
         else:
             target_length = attention_mask.shape[-1] if attention_mask is not None else input_tensor.shape[1]

src/transformers/models/gemma2/modular_gemma2.py

@@ -20,7 +20,7 @@
 import torch.utils.checkpoint
 
 from ...activations import ACT2FN
-from ...cache_utils import Cache, HybridCache
+from ...cache_utils import Cache, HybridCache, StaticCache
 from ...configuration_utils import PretrainedConfig
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_outputs import (
@@ -550,7 +550,7 @@ def _update_causal_mask(
 
         dtype, device = input_tensor.dtype, input_tensor.device
         sequence_length = input_tensor.shape[1]
-        if isinstance(past_key_values, HybridCache):
+        if isinstance(past_key_values, (HybridCache, StaticCache)):
             target_length = past_key_values.get_max_cache_shape()
         else:
             target_length = attention_mask.shape[-1] if attention_mask is not None else input_tensor.shape[1]

src/transformers/models/got_ocr2/configuration_got_ocr2.py

@@ -132,8 +132,6 @@ class GotOcr2Config(PretrainedConfig):
             The config object or dictionary of the vision backbone.
         text_config (`Union[AutoConfig, dict]`, *optional*, defaults to `LlamaConfig`):
             The config object or dictionary of the text backbone.
-        ignore_index (`int`, *optional*, defaults to -100):
-            The ignore index for the loss function.
         image_token_index (`int`, *optional*, defaults to 151859):
             The image token index to encode the image prompt.
         image_seq_length (`int`, *optional*, defaults to 576):
@@ -161,13 +159,11 @@ def __init__(
         self,
         vision_config=None,
         text_config=None,
-        ignore_index=-100,
         image_token_index=151859,
         image_seq_length=576,
         pad_token_id=-1,
         **kwargs,
     ):
-        self.ignore_index = ignore_index
         self.image_token_index = image_token_index
         self.image_seq_length = image_seq_length
         self.pad_token_id = pad_token_id

src/transformers/models/got_ocr2/modeling_got_ocr2.py

@@ -594,6 +594,8 @@ class GotOcr2PreTrainedModel(PreTrainedModel):
     _supports_cache_class = True
     _supports_flash_attn_2 = True
     _supports_sdpa = True
+    _supports_quantized_cache = True
+    _supports_static_cache = True
 
     def _init_weights(self, module):
         # important: this ported version of GotOcr2 isn't meant for training from scratch - only
@@ -748,89 +750,6 @@ def get_image_features(
         image_outputs = self.vision_tower(pixel_values).last_hidden_state
         return self.multi_modal_projector(image_outputs)
 
-    def _merge_input_ids_with_image_features(self, image_features, inputs_embeds, input_ids, attention_mask, labels):
-        num_images, num_image_patches, embed_dim = image_features.shape
-        batch_size, sequence_length = input_ids.shape
-        left_padding = not torch.sum(input_ids[:, -1] == torch.tensor(self.pad_token_id))
-        # 1. Create a mask to know where special image tokens are
-        special_image_token_mask = input_ids == self.config.image_token_index
-        num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1)
-        # Compute the maximum embed dimension
-        max_embed_dim = (num_special_image_tokens.max() * (num_image_patches - 1)) + sequence_length
-        batch_indices, non_image_indices = torch.where(input_ids != self.config.image_token_index)
-
-        # 2. Compute the positions where text should be written
-        # Calculate new positions for text tokens in merged image-text sequence.
-        # `special_image_token_mask` identifies image tokens. Each image token will be replaced by `nb_text_tokens_per_images - 1` text tokens.
-        # `torch.cumsum` computes how each image token shifts subsequent text token positions.
-        # - 1 to adjust for zero-based indexing, as `cumsum` inherently increases indices by one.
-        new_token_positions = torch.cumsum((special_image_token_mask * (num_image_patches - 1) + 1), -1) - 1
-        nb_image_pad = max_embed_dim - 1 - new_token_positions[:, -1]
-        if left_padding:
-            new_token_positions += nb_image_pad[:, None]  # offset for left padding
-        text_to_overwrite = new_token_positions[batch_indices, non_image_indices]
-
-        # 3. Create the full embedding, already padded to the maximum position
-        final_embedding = torch.zeros(
-            batch_size, max_embed_dim, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
-        )
-        final_attention_mask = torch.zeros(
-            batch_size, max_embed_dim, dtype=attention_mask.dtype, device=inputs_embeds.device
-        )
-        if labels is not None:
-            final_labels = torch.full(
-                (batch_size, max_embed_dim), self.config.ignore_index, dtype=input_ids.dtype, device=input_ids.device
-            )
-        # In case the Vision model or the Language model has been offloaded to CPU, we need to manually
-        # set the corresponding tensors into their correct target device.
-        target_device = inputs_embeds.device
-        batch_indices, non_image_indices, text_to_overwrite = (
-            batch_indices.to(target_device),
-            non_image_indices.to(target_device),
-            text_to_overwrite.to(target_device),
-        )
-        attention_mask = attention_mask.to(target_device)
-
-        # 4. Fill the embeddings based on the mask. If we have ["hey" "<image>", "how", "are"]
-        # we need to index copy on [0, 577, 578, 579] for the text and [1:576] for the image features
-        final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[batch_indices, non_image_indices]
-        final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[batch_indices, non_image_indices]
-        if labels is not None:
-            final_labels[batch_indices, text_to_overwrite] = labels[batch_indices, non_image_indices]
-
-        # 5. Fill the embeddings corresponding to the images. Anything that is not `text_positions` needs filling (#29835)
-        image_to_overwrite = torch.full(
-            (batch_size, max_embed_dim), True, dtype=torch.bool, device=inputs_embeds.device
-        )
-        image_to_overwrite[batch_indices, text_to_overwrite] = False
-        if left_padding:
-            image_to_overwrite &= image_to_overwrite.cumsum(-1) - 1 >= nb_image_pad[:, None].to(target_device)
-        else:
-            mask = torch.ones_like(image_to_overwrite, dtype=torch.bool).cumsum(-1) - 1
-            padding_mask = mask <= new_token_positions[:, -1:].to(target_device)
-            image_to_overwrite &= padding_mask
-
-        if image_to_overwrite.sum() != image_features.shape[:-1].numel():
-            raise ValueError(
-                f"The input provided to the model are wrong. The number of image tokens is {torch.sum(special_image_token_mask)} while"
-                f" the number of image given to the model is {num_images}. This prevents correct indexing and breaks batch generation."
-            )
-
-        final_embedding[image_to_overwrite] = image_features.contiguous().reshape(-1, embed_dim).to(target_device)
-        final_attention_mask |= image_to_overwrite
-        position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1)
-
-        # 6. Mask out the embedding at padding positions, as we later use the past_key_value value to determine the non-attended tokens.
-        batch_indices, pad_indices = torch.where(input_ids == self.pad_token_id)
-        indices_to_mask = new_token_positions[batch_indices, pad_indices]
-
-        final_embedding[batch_indices, indices_to_mask] = 0
-
-        if labels is None:
-            final_labels = None
-
-        return final_embedding, final_attention_mask, final_labels, position_ids
-
     @add_start_docstrings_to_model_forward(GOT_OCR2_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=GotOcr2CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
     def forward(

src/transformers/models/got_ocr2/modular_got_ocr2.py

@@ -170,8 +170,6 @@ class GotOcr2Config(PretrainedConfig):
             The config object or dictionary of the vision backbone.
         text_config (`Union[AutoConfig, dict]`, *optional*, defaults to `LlamaConfig`):
             The config object or dictionary of the text backbone.
-        ignore_index (`int`, *optional*, defaults to -100):
-            The ignore index for the loss function.
         image_token_index (`int`, *optional*, defaults to 151859):
             The image token index to encode the image prompt.
         image_seq_length (`int`, *optional*, defaults to 576):
@@ -199,13 +197,11 @@ def __init__(
         self,
         vision_config=None,
         text_config=None,
-        ignore_index=-100,
         image_token_index=151859,
         image_seq_length=576,
         pad_token_id=-1,
         **kwargs,
     ):
-        self.ignore_index = ignore_index
         self.image_token_index = image_token_index
         self.image_seq_length = image_seq_length
         self.pad_token_id = pad_token_id

src/transformers/models/gpt_neox_japanese/modeling_gpt_neox_japanese.py

@@ -51,7 +51,7 @@ class GPTNeoXJapanesePreTrainedModel(PreTrainedModel):
     _skip_keys_device_placement = "past_key_values"
     _supports_cache_class = True
     _supports_quantized_cache = True
-    _supports_static_cache = False  # TODO (fix me): compilation fails due to a stide error?
+    _supports_static_cache = True
 
     def _init_weights(self, module):
         """Initialize the weights"""
@@ -129,8 +129,8 @@ def forward(
 
         cos, sin = position_embeddings
         query, key = apply_rotary_pos_emb(query_rot, key_rot, cos, sin)
-        query = torch.cat((query, query_pass), dim=-1)
-        key = torch.cat((key, key_pass), dim=-1)
+        query = torch.cat((query, query_pass), dim=-1).contiguous()
+        key = torch.cat((key, key_pass), dim=-1).contiguous()
 
         # Cache QKV values
         if layer_past is not None:

src/transformers/models/granitemoe/modeling_granitemoe.py

@@ -1108,6 +1108,7 @@ def forward(
             router_logits=all_router_logits,
         )
 
+    # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask
     def _update_causal_mask(
         self,
         attention_mask: torch.Tensor,
@@ -1116,13 +1117,8 @@ def _update_causal_mask(
         past_key_values: Cache,
         output_attentions: bool,
     ):
-        # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
-        # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
-        # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
-        # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
-
         if self.config._attn_implementation == "flash_attention_2":
-            if attention_mask is not None and 0.0 in attention_mask:
+            if attention_mask is not None and (attention_mask == 0.0).any():
                 return attention_mask
             return None
 
@@ -1143,7 +1139,6 @@ def _update_causal_mask(
                 return None
 
         dtype, device = input_tensor.dtype, input_tensor.device
-        min_dtype = torch.finfo(dtype).min
         sequence_length = input_tensor.shape[1]
         if using_static_cache:
             target_length = past_key_values.get_max_cache_shape()
@@ -1154,25 +1149,17 @@ def _update_causal_mask(
                 else past_seen_tokens + sequence_length + 1
             )
 
-        if attention_mask is not None and attention_mask.dim() == 4:
-            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
-            causal_mask = attention_mask
-        else:
-            causal_mask = torch.full(
-                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
-            )
-            if sequence_length != 1:
-                causal_mask = torch.triu(causal_mask, diagonal=1)
-            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
-            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
-            if attention_mask is not None:
-                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
-                mask_length = attention_mask.shape[-1]
-                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
-                padding_mask = padding_mask == 0
-                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
-                    padding_mask, min_dtype
-                )
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
         if (
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
@@ -1182,6 +1169,7 @@ def _update_causal_mask(
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
             # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
             # Details: https://github.com/pytorch/pytorch/issues/110213
+            min_dtype = torch.finfo(dtype).min
             causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
 
         return causal_mask

src/transformers/models/instructblip/modeling_instructblip.py

@@ -1290,6 +1290,9 @@ def forward(
 class InstructBlipForConditionalGeneration(InstructBlipPreTrainedModel, GenerationMixin):
     config_class = InstructBlipConfig
     main_input_name = "pixel_values"
+    _supports_cache_class = True
+    _supports_static_cache = True
+    _supports_quantized_cache = False  # not all LM bacbones support (e.g. T5)
 
     def __init__(self, config: InstructBlipConfig):
         super().__init__(config)

src/transformers/models/instructblipvideo/modeling_instructblipvideo.py

@@ -1284,6 +1284,9 @@ def forward(
 class InstructBlipVideoForConditionalGeneration(InstructBlipVideoPreTrainedModel, GenerationMixin):
     config_class = InstructBlipVideoConfig
     main_input_name = "pixel_values"
+    _supports_cache_class = True
+    _supports_static_cache = True
+    _supports_quantized_cache = False  # not all LM bacbones support (e.g. T5)
 
     def __init__(self, config: InstructBlipVideoConfig):
         super().__init__(config)

src/transformers/models/llava/configuration_llava.py

@@ -37,8 +37,6 @@ class LlavaConfig(PretrainedConfig):
             The config object or dictionary of the vision backbone.
         text_config (`Union[AutoConfig, dict]`, *optional*, defaults to `LlamaConfig`):
             The config object or dictionary of the text backbone.
-        ignore_index (`int`, *optional*, defaults to -100):
-            The ignore index for the loss function.
         image_token_index (`int`, *optional*, defaults to 32000):
             The image token index to encode the image prompt.
         projector_hidden_act (`str`, *optional*, defaults to `"gelu"`):
@@ -83,7 +81,6 @@ def __init__(
         self,
         vision_config=None,
         text_config=None,
-        ignore_index=-100,
         image_token_index=32000,
         projector_hidden_act="gelu",
         vision_feature_select_strategy="default",
@@ -92,7 +89,6 @@ def __init__(
         multimodal_projector_bias=True,
         **kwargs,
     ):
-        self.ignore_index = ignore_index
         self.image_token_index = image_token_index
         self.projector_hidden_act = projector_hidden_act
         self.image_seq_length = image_seq_length