# coding=utf-8 # Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ PyTorch Phi-4-MM model.""" import math import warnings from typing import List, Optional, Tuple, Union import numpy as np import torch import torch.utils.checkpoint from torch import nn from torch.nn import CrossEntropyLoss from transformers.activations import ACT2FN from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache from transformers.generation import GenerationMixin from transformers.modeling_attn_mask_utils import AttentionMaskConverter from transformers.modeling_flash_attention_utils import _flash_attention_forward from transformers.modeling_outputs import ( BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast, TokenClassifierOutput, ) from transformers.modeling_utils import PreTrainedModel from transformers.utils import ( add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, is_flash_attn_greater_or_equal_2_10, logging, replace_return_docstrings, ) from transformers import AutoConfig, AutoModelForCausalLM, PretrainedConfig from .configuration_phi4mm import Phi4MMConfig from .processing_phi4mm import InputMode from .vision_siglip_navit import get_siglip_vision_model from .speech_conformer_encoder import ConformerEncoder logger = logging.get_logger(__name__) _CHECKPOINT_FOR_DOC = "TBA" _CONFIG_FOR_DOC = "Phi4MMConfig" # Special token ids _IMAGE_SPECIAL_TOKEN_ID = 200010 # '<|endoftext10|>', or we can better name it (in `tokenizer_config.json`) _AUDIO_SPECIAL_TOKEN_ID = 200011 # '<|endoftext11|>' _COMPATIBLE_IMAGE_SPECIAL_TOKEN_ID_RANGE = [-9999, -1] # For backward compatibility _COMPATIBLE_AUDIO_SPECIAL_TOKEN_ID_RANGE = [float('-inf'), -10000] # For backward compatibility class Phi4MMImageEmbedding(nn.Module): """Image embedding.""" def __init__(self, config: PretrainedConfig, **kwargs) -> None: super().__init__() # n_embed or hidden_size hidden_size = config.n_embd if hasattr(config, 'n_embd') else config.hidden_size if hasattr(config, 'embd_pdrop') or hasattr(config, 'embed_pdrop'): embd_drop = config.embd_pdrop if hasattr(config, 'embd_pdrop') else config.embed_pdrop self.drop = nn.Dropout(embd_drop) else: self.drop = None logger.info(f"create image tower {config.img_processor}") enable_gradient_checkpointing = kwargs.get('enable_gradient_checkpointing', False) # Load SigLIP model self.img_processor = get_siglip_vision_model( _flash_attn_2_enabled=config._attn_implementation == 'flash_attention_2' ) pe_weight = self.img_processor.embeddings.position_embedding.weight L, D = pe_weight.size() H = int(math.sqrt(L)) assert H**2 == L if H % 2 != 0: #and kwargs.get('image_token_compression_cls', None) is None: self.img_processor_padding = nn.ReflectionPad2d((0, 1, 0, 1)) H += 1 image_dim_out = D # ((448/14)//2)**2 self.num_img_tokens = (H//2)**2 self.base_feat_height_target = H if enable_gradient_checkpointing: self.img_processor.encoder.gradient_checkpointing = True self.image_dim_out = image_dim_out self.img_sizes = None self.image_attention_mask = None # global_gn and sub_gn for hd transform, serves as line separator self.use_hd_transform = kwargs.get('use_hd_transform', False) self.with_learnable_separator = kwargs.get('with_learnable_separator', False) self.hd_transform_order = kwargs.get('hd_transform_order', 'glb_sub') self.freeze_img_processor = kwargs.get('freeze_img_processor', False) self.crop_size = kwargs.get('crop_size', 336) logger.info(f'freeze_img_processor = {self.freeze_img_processor}') # image token compression self.image_token_compression_cls = kwargs.get('image_token_compression_cls', None) if self.image_token_compression_cls == 'avg_pool_2d': self.image_token_compression = nn.AvgPool2d(kernel_size=2, stride=2) self.base_feat_height_reduction = 1 self.base_feat_height_target = self.base_feat_height_target // 2 elif self.image_token_compression_cls is None: self.image_token_compression = None self.base_feat_height_reduction = 2 else: raise NotImplementedError(f'image_token_compression_cls = {self.image_token_compression_cls}, not implemented') # with_hd_transform and with_learnable_separator should have same value assert self.use_hd_transform == self.with_learnable_separator, 'use_hd_transform and with_learnable_separator should have same value' if self.with_learnable_separator: assert self.use_hd_transform, 'learnable separator is only for hd transform' # 1024 * 4, merge spatial to channel dimension self.glb_GN = nn.Parameter(torch.zeros([1, 1, self.image_dim_out * self.base_feat_height_reduction**2])) self.sub_GN = nn.Parameter(torch.zeros([1, 1, 1, self.image_dim_out * self.base_feat_height_reduction**2])) logger.info(f'learnable separator enabled for hd transform, hd_transform_order = {self.hd_transform_order}') projection_cls = kwargs.get('projection_cls', 'linear') if projection_cls == 'linear': self.img_projection = nn.Linear(image_dim_out, hidden_size) elif projection_cls == 'mlp' and self.use_hd_transform: dim_projection = hidden_size depth = 2 layers = [nn.Linear(image_dim_out * self.base_feat_height_reduction**2, dim_projection)] for _ in range(1, depth): layers.extend([nn.GELU(), nn.Linear(dim_projection, dim_projection)]) self.img_projection = nn.Sequential(*layers) elif projection_cls == 'mlp': # follow llava-v1.5's implementation # (do not use image_projection and image_proj_norm) dim_projection = hidden_size depth = 2 layers = [nn.Linear(image_dim_out, dim_projection)] for _ in range(1, depth): layers.extend([nn.GELU(), nn.Linear(dim_projection, dim_projection)]) self.img_projection = nn.Sequential(*layers) else: raise NotImplementedError(f'projection_cls = {projection_cls}, not implemented') self.vocab_size = config.vocab_size self.img_features = None if isinstance(config.img_processor, dict): self.layer_idx = config.img_processor.get('layer_idx', -2) self.type_feature = config.img_processor.get('type_feature', 'patch') else: self.layer_idx = -2 self.type_feature = 'patch' def set_img_features(self, img_features: torch.FloatTensor) -> None: self.img_features = img_features def set_img_sizes(self, img_sizes: torch.LongTensor) -> None: self.img_sizes = img_sizes def set_img_attn_mask(self, image_attention_mask: torch.FloatTensor) -> None: self.image_attention_mask = image_attention_mask def get_img_features(self, img_embeds: torch.FloatTensor, attention_mask=None) -> torch.FloatTensor: LAYER_IDX = self.layer_idx TYPE_FEATURE = self.type_feature if self.freeze_img_processor: with torch.no_grad(): if attention_mask is not None: img_processor_output = self.img_processor(img_embeds, output_hidden_states=True, patch_attention_mask=attention_mask) else: img_processor_output = self.img_processor(img_embeds, output_hidden_states=True) img_feature = img_processor_output.hidden_states[LAYER_IDX] else: if attention_mask is not None: img_processor_output = self.img_processor(img_embeds, output_hidden_states=True, patch_attention_mask=attention_mask) else: img_processor_output = self.img_processor(img_embeds, output_hidden_states=True) img_feature = img_processor_output.hidden_states[LAYER_IDX] if TYPE_FEATURE == "patch": patch_feature = img_feature if self.image_token_compression is not None: # reshape to 2D tensor width = int(math.sqrt(patch_feature.size(1))) patch_feature = patch_feature.view(-1, width, width, patch_feature.size(-1)) # convert to NCHW patch_feature = patch_feature.permute(0, 3, 1, 2) if getattr(self, 'img_processor_padding', None) is not None: patch_feature = self.img_processor_padding(patch_feature) patch_feature = self.image_token_compression(patch_feature) # convert to NHWC patch_feature = patch_feature.permute(0, 2, 3, 1) patch_feature = patch_feature.view(-1, patch_feature.size(1) * patch_feature.size(2), patch_feature.size(-1)) elif getattr(self, 'img_processor_padding', None) is not None: width = int(math.sqrt(patch_feature.size(1))) patch_feature = patch_feature.view(-1, width, width, patch_feature.size(-1)) # convert to NCHW patch_feature = patch_feature.permute(0, 3, 1, 2) patch_feature = self.img_processor_padding(patch_feature) # convert to NHWC patch_feature = patch_feature.permute(0, 2, 3, 1) patch_feature = patch_feature.view(-1, patch_feature.size(1) * patch_feature.size(2), patch_feature.size(-1)) return patch_feature if TYPE_FEATURE == "cls_patch": if self.image_token_compression is not None: # reshape to 2D tensor patch_feature = img_feature[:, 1:] cls_feature = img_feature[:, 0] width = math.sqrt(patch_feature.size(1)) patch_feature = patch_feature.view(-1, width, width, patch_feature.size(-1)) patch_feature = self.image_token_compression(patch_feature) patch_feature = patch_feature.view(-1, patch_feature.size(-2) * patch_feature.size(-1)) img_feature = torch.cat([cls_feature, patch_feature], dim=1) return img_feature logger.info(f'processed img feature size = {img_feature.size()}') raise NotImplementedError def spatiotemporal_pool(self, x, num_img_tokens, batch_size=1, T=1): if self.image_pos_embed is not None: x = x.view(batch_size * T, -1, x.shape[-1]) num_tokens = x.shape[-2] h, w = int(num_tokens ** 0.5), int(num_tokens ** 0.5) assert h * w == num_tokens, 'only support square feature maps for now' x = x.view(batch_size * T, h, w, x.shape[-1]) pos_embed = self.image_pos_embed(x) x = x + pos_embed x = x.view(batch_size, T * h * w, x.shape[-1]) if self.visual_temporal_embed is not None: visual_temporal_embed = self.visual_temporal_embed(x.view(batch_size, T, -1, x.shape[-1])[:, :, 0]) x = x.view(batch_size, T, -1, x.shape[-1]) + visual_temporal_embed.view(1, T, 1, x.shape[-1]) new_x = [] # [bsz, T * H' * W', C] -> [bsz, T, C] spatial_avg_pool_x = x.view(batch_size, T, -1, x.shape[-1]).mean(dim=2) new_x.append(spatial_avg_pool_x) # [bsz, T * H' * W', C] -> [bsz, H'*W', C] temporal_avg_pool_x = x.view(batch_size, T, -1, x.shape[-1]).mean(dim=1) new_x.append(temporal_avg_pool_x) x = torch.cat(new_x, dim=1).view(-1, self.image_dim_out) num_img_tokens += T return x, num_img_tokens def forward(self, input_ids: torch.LongTensor, input_embeds: torch.FloatTensor, image_sizes=None, **kwargs) -> torch.FloatTensor: if isinstance(input_ids, tuple): # # pipeline parallel input_ids, input_embeds = input_ids img_embeds = input_embeds if image_sizes is None and 'image_sizes' in kwargs: image_sizes = kwargs['image_sizes'] img_sizes = image_sizes if self.img_features is not None: img_embeds = self.img_features.clone() self.img_features = None if self.img_sizes is not None: img_sizes = self.img_sizes dtype = self.img_processor.embeddings.patch_embedding.weight.dtype if img_embeds is not None: # convert to bf16 img_embeds = img_embeds.to(dtype) if self.image_attention_mask is not None: image_attention_mask = self.image_attention_mask.clone() self.image_attention_mask = None elif 'image_attention_mask' in kwargs: image_attention_mask = kwargs['image_attention_mask'] else: image_attention_mask = None input_shape = input_ids.size() input_ids = input_ids.view(-1, input_shape[-1]) with torch.no_grad(): positions = torch.nonzero(input_ids == _IMAGE_SPECIAL_TOKEN_ID, as_tuple=False) positions_tuple = torch.nonzero(input_ids == _IMAGE_SPECIAL_TOKEN_ID, as_tuple=True) # logger.info(f'position size: {positions.size()} ...') fake_image_forward = False select = False hd_transform = False if isinstance(self.img_projection, nn.Sequential): target_device = self.img_projection[0].bias.device target_dtype = self.img_projection[0].bias.dtype else: # It's a single nn.Linear layer target_device = self.img_projection.bias.device target_dtype = self.img_projection.bias.dtype num_img_tokens = self.num_img_tokens if len(positions.tolist()) > 0: if self.use_hd_transform and img_sizes is not None and len(img_sizes): hd_transform = True assert img_embeds.ndim == 5, f'(branch 1) img_embeds size: {img_embeds.size()}, expect 5D tensor for hd transform' # img_embeds: (num_images, max_num_crops, 3, H, W) # img_sizes: (num_images, 2).view(1, -1) bs = img_embeds.shape[0] # Nx(HW)xC if image_attention_mask is not None and len(image_attention_mask) > 0: img_features = self.get_img_features(img_embeds.flatten(0, 1), attention_mask=image_attention_mask.type(torch.BoolTensor).flatten(0,1).to(target_device)) else: img_features = self.get_img_features(img_embeds.flatten(0, 1)) base_feat_height_target = self.base_feat_height_target base_resolution = self.crop_size base_feat_height_reduction = self.base_feat_height_reduction base_feat_height = base_feat_width = int(np.sqrt(img_features.shape[1])) assert base_feat_height == base_feat_height_target and base_feat_width == base_feat_height_target, f'base_feat_height: {base_feat_height}, base_feat_width: {base_feat_width}, expect {base_feat_height_target} features for hd transform' # bs x max_num_crops x (24x24) x C img_features = img_features.view(bs, -1, base_feat_height * base_feat_width, self.image_dim_out) C = self.image_dim_out H = base_feat_height output_imgs = [] output_len = [] # training is tensor, inference is list if isinstance(img_sizes, torch.Tensor): img_sizes = img_sizes.view(-1, 2) for _bs in range(bs): h, w = img_sizes[_bs] h = h // base_resolution w = w // base_resolution B_ = h * w # 1 x (24x24) x 1024 global_img_feature = img_features[_bs, :1] # 1 x 12 x 12 x 4096 glb_img = global_img_feature.reshape(1,H,H,C).reshape(1,H//base_feat_height_reduction,base_feat_height_reduction,H//base_feat_height_reduction,base_feat_height_reduction,C).contiguous().permute(0,1,3,2,4,5).reshape(1,H//base_feat_height_reduction,H//base_feat_height_reduction,base_feat_height_reduction*base_feat_height_reduction*C).contiguous() temp_glb_GN = self.sub_GN.repeat(1, H//base_feat_height_reduction, 1, 1) # 1 x 156 x 4096 glb_img = torch.cat([glb_img, temp_glb_GN], dim=2).reshape(1,-1,base_feat_height_reduction*base_feat_height_reduction*C) # (max_num_crops-1) x (12x12) x C sub_img = img_features[_bs, 1:] # 16x574x1024 # get rid of padding sub_img sub_img = sub_img[:B_] # (num_crops, 12, 2, 12, 2, 1024) -> (num_crops, 12, 12, 2, 2, 1024) -> (num_crops, 12*12, 4*1024) sub_img = sub_img.reshape(B_,H,H,C).reshape(B_,H//base_feat_height_reduction,base_feat_height_reduction,H//base_feat_height_reduction,base_feat_height_reduction,C).contiguous().permute(0,1,3,2,4,5).reshape(B_,-1,base_feat_height_reduction*base_feat_height_reduction*C).contiguous() sub_img = sub_img.reshape(1, h, w, base_feat_height // base_feat_height_reduction, base_feat_width // base_feat_height_reduction, -1).permute(0,1,3,2,4,5).reshape(1,h*base_feat_height//base_feat_height_reduction,w*base_feat_width//base_feat_height_reduction,base_feat_height_reduction*base_feat_height_reduction*C) if image_attention_mask is not None and len(image_attention_mask) > 0: reshaped_image_attention_mask = image_attention_mask[_bs,1:B_+1,0::2,0::2].reshape(1, h, w, base_feat_height // base_feat_height_reduction, base_feat_width // base_feat_height_reduction).permute(0,1,3,2,4).reshape(1,h*base_feat_height//base_feat_height_reduction,w*base_feat_width//base_feat_height_reduction) useful_height = int(reshaped_image_attention_mask[0,:,0].sum().item()) useful_width = int(reshaped_image_attention_mask[0,0,:].sum().item()) sub_img = sub_img[:,:useful_height, :useful_width] temp_sub_GN = self.sub_GN.repeat(1, useful_height, 1, 1) temp_len = int(image_attention_mask[_bs,:B_+1,0::2,0::2].sum().item()) + (useful_height+1) + base_feat_height//base_feat_height_reduction else: temp_sub_GN = self.sub_GN.repeat(1, h*base_feat_height//base_feat_height_reduction, 1, 1) temp_len = int((h*w+1)*self.num_img_tokens+ 1 + (h+1)*base_feat_height//base_feat_height_reduction) sub_img = torch.cat([sub_img, temp_sub_GN], dim=2).reshape(1,-1,base_feat_height_reduction*base_feat_height_reduction*C) # (1, num_img_tokens, 1024*4) # glb + sub if self.hd_transform_order == 'glb_sub': output_imgs.append(torch.cat([glb_img, self.glb_GN, sub_img], dim=1)) elif self.hd_transform_order == 'sub_glb': output_imgs.append(torch.cat([sub_img, self.glb_GN, glb_img], dim=1)) else: raise NotImplementedError(f'hd_transform_order = {self.hd_transform_order}, not implemented') #temp_len = int((h*w+1)*144 + 1 + (h+1)*12) assert temp_len == output_imgs[-1].shape[1], f'temp_len: {temp_len}, output_imgs[-1].shape[1]: {output_imgs[-1].shape[1]}' output_len.append(temp_len) num_img_tokens = output_len img_set_tensor = [] for _output_img in output_imgs: img_feature_proj = self.img_projection(_output_img.to(target_device).to(target_dtype)) img_set_tensor.append(img_feature_proj) #logger.info(f'img_embeds size: {img_embeds.size()}, image sizes: {img_sizes} loading time {datetime.now() - start_time}') #assert sum(num_img_tokens) == len(g_values), f'(branch 1) sum(num_img_tokens): {sum(num_img_tokens)}, g_values size: {len(g_values)}, g_values {g_values}' else: raise NotImplementedError select = True else: # # create a fake image tensor # # TODO: need define image size for different vision model if self.training: img_embeds = torch.zeros(1, 3, self.crop_size, self.crop_size, dtype=target_dtype, device=input_ids.device) tt = ( self.get_img_features(img_embeds) .to(target_device) .to(target_dtype) .reshape(-1, 1024) ) if self.use_hd_transform: img_set_tensor = self.img_projection(tt.reshape(-1, self.image_dim_out*self.base_feat_height_reduction**2) * self.glb_GN[0] * self.sub_GN[0, 0]) else: img_set_tensor = self.img_projection(tt) # adapted visual features. fake_image_forward = True # we use the token embedding layer from the huggingface model, this is REQUIRED to make sure we are using the loaded weights. hidden_states = kwargs['wte'](input_ids) if select: if hd_transform: # new implementation without in-place operation # Ref: https://huggingface.co/microsoft/Phi-3.5-vision-instruct/blob/4a0d683eba9f1d0cbfb6151705d1ee73c25a80ca/modeling_phi3_v.py#L233 # Ref: https://pytorch.org/docs/stable/generated/torch.Tensor.index_put.html # Ref: https://pytorch.org/docs/stable/generated/torch.Tensor.index_put_.html#torch.Tensor.index_put_ # img_set_tensor: a list of tensors, each tensor has shape (1, N_tokens, C) assert all([_img_set_tensor.shape[0] == 1 for _img_set_tensor in img_set_tensor]), 'img_set_tensor should have shape (1, N_tokens, C)' # Shape: (merged_N_tokens, C) merged_img_set_tensor = torch.cat(img_set_tensor, dim=1).squeeze(0) merged_img_set_tensor = merged_img_set_tensor.to(hidden_states.dtype).to(hidden_states.device) # Temporarily disable autocast to avoid issue on bf16 tensors # Ref: https://github.com/pytorch/pytorch/issues/132715 with torch.autocast(device_type=hidden_states.device.type, enabled=False): new_hidden_states = hidden_states.index_put( indices=positions_tuple, values=merged_img_set_tensor, accumulate=False ) hidden_states = new_hidden_states else: raise NotImplementedError if fake_image_forward and self.training: hidden_states = hidden_states + (0 * img_set_tensor[0].to(hidden_states.dtype).to(hidden_states.device)).sum() if self.drop is not None: hidden_states = self.drop(hidden_states) return hidden_states class Phi4MMAudioEmbedding(nn.Module): """Audio embedding.""" def __init__(self, config: PretrainedConfig, **kwargs) -> None: super().__init__() self.config = config # n_embed or hidden_size for text LM hidden_size = config.n_embd if hasattr(config, 'n_embd') else config.hidden_size if hasattr(config, 'embd_pdrop') or hasattr(config, 'embed_pdrop'): embd_drop = config.embd_pdrop if hasattr(config, 'embd_pdrop') else config.embed_pdrop self.drop = nn.Dropout(embd_drop) else: self.drop = None audio_dim_out = None # Set this variable according to the actual audio processor logger.info(f"create audio processor {config.audio_processor}") self.layer_idx = -2 if isinstance(config.audio_processor, dict) and config.audio_processor.get('name', None) == "cascades": encoder_config = config.audio_processor.get("config", None) assert encoder_config is not None self.encoder = ConformerEncoder(**encoder_config) # fake initialization, create encoder_embedding layer only so that # in decoding, all parameters can be loaded in from_pretrained_function # in training, we do post init after from_pretrained function to make sure the correct initialization self.encoder.post_init({}) audio_dim_out = encoder_config["attention_dim"] n_mels = encoder_config["input_size"] else: raise NotImplementedError assert audio_dim_out is not None, "Remember to set values for audio_dim_out" self.audio_dim_out = audio_dim_out self.audio_dim_in = n_mels self.freeze_audio_processor = kwargs.get('freeze_audio_processor', False) logger.info(f'freeze_audio_processor = {self.freeze_audio_processor}') self.downsample_rate = kwargs.get('downsample_rate', 1) enable_gradient_checkpointing = kwargs.get('enable_gradient_checkpointing', False) if enable_gradient_checkpointing: self.encoder.gradient_checkpointing_enable() logger.info(f'gradient checkpointing enabled for audio processor') projection_cls = kwargs.get('projection_cls', 'linear') if projection_cls == 'linear': self.audio_projection = nn.Linear(audio_dim_out, hidden_size) elif projection_cls == 'mlp': # follow llava-v1.5's implementation # (do not use image_projection and image_proj_norm) dim_projection = hidden_size depth = 2 self.linear_downsample_rate = self.downsample_rate layers_for_speech = [nn.Linear(audio_dim_out * self.linear_downsample_rate, dim_projection)] for _ in range(1, depth): layers_for_speech.extend([nn.GELU(), nn.Linear(dim_projection, dim_projection)]) audio_projection_for_speech = nn.Sequential(*layers_for_speech) layers_for_vision = [nn.Linear(audio_dim_out * self.linear_downsample_rate, dim_projection)] for _ in range(1, depth): layers_for_vision.extend([nn.GELU(), nn.Linear(dim_projection, dim_projection)]) audio_projection_for_vision = nn.Sequential(*layers_for_vision) self.audio_projection = nn.ModuleDict({ 'speech': audio_projection_for_speech, 'vision': audio_projection_for_vision }) else: raise NotImplementedError(f'projection_cls = {projection_cls}, not implemented') self.vocab_size = config.vocab_size self.input_embeds = None self.audio_embed_sizes = None def post_init(self, audio_config): # execute after the from_pretrained() initialization of the phi4mm model if audio_config.get('name', None) == "cascades": init_model_config = audio_config.get("init_model", {}) self.encoder.post_init(init_model_config) # remove the init model in config so it is not saved in the config. # This might affect the model loading in resuming training and decoding. if "init_model" in audio_config: audio_config.pop("init_model") def set_audio_embeds(self, input_embeds: torch.FloatTensor) -> None: self.input_embeds = input_embeds def set_audio_embed_sizes(self, audio_embed_sizes: torch.LongTensor) -> None: self.audio_embed_sizes = audio_embed_sizes def get_audio_features(self, input_embeds: torch.FloatTensor, audio_attention_mask: torch.Tensor, audio_projection_mode: str='speech'): if self.freeze_audio_processor: with torch.no_grad(): audio_features, masks = self.encoder(input_embeds, audio_attention_mask) else: audio_features, masks = self.encoder(input_embeds, audio_attention_mask) if isinstance(self.audio_projection, nn.Sequential): audio_set_tensor = self.audio_projection(audio_features) elif isinstance(self.audio_projection, nn.ModuleDict): audio_set_tensor = self.audio_projection[audio_projection_mode](audio_features) else: raise NotImplementedError return audio_set_tensor def forward(self, input_ids: torch.LongTensor, input_embeds: torch.FloatTensor, audio_embed_sizes=None, audio_attention_mask=None, audio_projection_mode='speech', **kwargs) -> torch.FloatTensor: ''' arguments: input_ids: input text ids (B, U) input_embeds: audio features (B, T, D) B: num audios in a sequence ''' if self.input_embeds is not None: input_embeds = self.input_embeds.clone() if self.audio_embed_sizes is not None: audio_embed_sizes = self.audio_embed_sizes.clone() input_shape = input_ids.size() input_ids = input_ids.view(-1, input_shape[-1]) MAX_INPUT_ID = int(1e9) with torch.no_grad(): positions = torch.nonzero(input_ids == _AUDIO_SPECIAL_TOKEN_ID, as_tuple=False) positions_tuple = torch.nonzero(input_ids == _AUDIO_SPECIAL_TOKEN_ID, as_tuple=True) if isinstance(self.audio_projection, nn.Sequential): target_device = self.audio_projection[0].bias.device target_dtype = self.audio_projection[0].bias.dtype elif isinstance(self.audio_projection, nn.ModuleDict): target_device = self.audio_projection[audio_projection_mode][0].bias.device target_dtype = self.audio_projection[audio_projection_mode][0].bias.dtype else: # It's a single nn.Linear layer target_device = self.audio_projection.bias.device target_dtype = self.audio_projection.bias.dtype if input_embeds is not None: input_embeds = input_embeds.to(target_device).to(target_dtype) if len(positions.tolist()) > 0: audio_set_tensor = self.get_audio_features(input_embeds, audio_attention_mask, audio_projection_mode) else: # # create an audio tensor # To do: not sure if this is required for text only input if self.training: audio_embeds = torch.zeros(1, 500, self.audio_dim_in).to(target_device).to(target_dtype) audio_attention_mask = audio_embeds.new_ones(audio_embeds.size()[:2]).long() audio_set_tensor = self.get_audio_features(audio_embeds, audio_attention_mask, audio_projection_mode) hidden_states = kwargs['wte'](input_ids) if len(positions.tolist()) > 0: assert audio_embed_sizes.sum().item() == len(positions), \ f"please ensure the encoder outputs have the same length as defined in input_ids! \n audio_embed_sizes.sum().item(): {audio_embed_sizes.sum().item()} \n len(positions): {len(positions)} \n audio_embed_sizes: {audio_embed_sizes} \n positions: {positions} \n input_ids.shape \n {input_ids.shape}" # new implementation without in-place operation # Ref: https://huggingface.co/microsoft/Phi-3.5-vision-instruct/blob/4a0d683eba9f1d0cbfb6151705d1ee73c25a80ca/modeling_phi3_v.py#L233 # Ref: https://pytorch.org/docs/stable/generated/torch.Tensor.index_put.html # Ref: https://pytorch.org/docs/stable/generated/torch.Tensor.index_put_.html#torch.Tensor.index_put_ # audio_set_tensor: shape (N_audios, N_padded_tokens, C) # Shape: (merged_N_tokens, C) merged_audio_set_tensor = torch.cat([ audio_set_tensor[i, :audio_embed_sizes[i], :] for i in range(len(audio_embed_sizes)) ], dim=0) merged_audio_set_tensor = merged_audio_set_tensor.to(hidden_states.dtype).to(hidden_states.device) # Temporarily disable autocast to avoid issue on bf16 tensors # Ref: https://github.com/pytorch/pytorch/issues/132715 with torch.autocast(device_type=hidden_states.device.type, enabled=False): new_hidden_states = hidden_states.index_put( indices=positions_tuple, values=merged_audio_set_tensor, accumulate=False ) hidden_states = new_hidden_states else: if self.training: hidden_states = hidden_states + (0 * audio_set_tensor[:,0].to(hidden_states.dtype).to(hidden_states.device)).sum() if self.drop is not None: hidden_states = self.drop(hidden_states) return hidden_states class Phi4MMImageAudioEmbedding(nn.Module): """Image-audio embedding.""" def __init__(self, config: PretrainedConfig, **kwargs) -> None: super().__init__() self.vocab_size = config.vocab_size self.image_input_id = kwargs.get('image_input_id', -1) self.audio_input_id = kwargs.get('audio_input_id', -10000) assert self.image_input_id != self.audio_input_id, 'image_input_id and audio_input_id should be different' self.image_embd_layer_kwargs = kwargs['image_embd_layer'] self.image_embed = Phi4MMImageEmbedding(config, **self.image_embd_layer_kwargs) self.audio_embd_layer_kwargs = kwargs['audio_embd_layer'] self.audio_embed = Phi4MMAudioEmbedding(config, **self.audio_embd_layer_kwargs) self.input_image_embeds = None self.image_sizes = None self.image_attention_mask = None self.input_audio_embeds = None self.audio_embed_sizes = None def post_init(self, audio_config): # post init for audio embedding # ref: model.model.embed_tokens_extend.post_init(audio_config) in phyagi/getters/model.py self.audio_embed.post_init(audio_config) def set_input_image_embeds(self, input_image_embeds: torch.FloatTensor) -> None: self.input_image_embeds = input_image_embeds def set_image_sizes(self, image_sizes: torch.LongTensor) -> None: self.image_sizes = image_sizes def set_img_attn_mask(self, image_attention_mask: torch.FloatTensor) -> None: self.image_attention_mask = image_attention_mask def set_input_audio_embeds(self, input_audio_embeds: torch.FloatTensor) -> None: self.input_audio_embeds = input_audio_embeds def set_audio_embed_sizes(self, audio_embed_sizes: torch.LongTensor) -> None: self.audio_embed_sizes = audio_embed_sizes def forward( self, input_ids: torch.LongTensor, input_embeds, input_image_embeds: Optional[torch.FloatTensor]=None, input_audio_embeds: Optional[torch.FloatTensor]=None, image_sizes=None, image_attention_mask=None, audio_embed_sizes=None, audio_attention_mask=None, audio_projection_mode='speech', wte=None, ) -> torch.FloatTensor: MAX_INPUT_ID = int(1e9) assert -MAX_INPUT_ID < self.audio_input_id < self.image_input_id # override image and audio embeddings and sizes from object itself # this is for inference # ref: phyagi/eval/utils/text_generation_vision_audio_pipeline.py if self.input_image_embeds is not None: assert input_image_embeds is None input_image_embeds = self.input_image_embeds.clone() # NOTE weijian: set input_image_embeds to None after first call in for eval stage # during evaluation, it will call model's forward() multiple times # the first time input_ids contains the prompt (including <|image_{}|>) and input_embeds exists # from the second time, the input_ids will only contain the generated text # thus, the input_image_embeds is no longer needed self.input_image_embeds = None if self.image_sizes is not None: assert image_sizes is None image_sizes = self.image_sizes if self.input_audio_embeds is not None: assert input_audio_embeds is None input_audio_embeds = self.input_audio_embeds.clone() self.input_audio_embeds = None if self.audio_embed_sizes is not None: assert audio_embed_sizes is None audio_embed_sizes = self.audio_embed_sizes.clone() if self.image_attention_mask is not None: assert image_attention_mask is None image_attention_mask = self.image_attention_mask.clone() self.image_attention_mask = None input_shape = input_ids.size() input_ids = input_ids.view(-1, input_shape[-1]) # backward compatibility with torch.no_grad(): new_input_ids = input_ids.clone() new_input_ids[(input_ids >= _COMPATIBLE_IMAGE_SPECIAL_TOKEN_ID_RANGE[0]) & (input_ids <= _COMPATIBLE_IMAGE_SPECIAL_TOKEN_ID_RANGE[1])] = _IMAGE_SPECIAL_TOKEN_ID new_input_ids[(input_ids >= _COMPATIBLE_AUDIO_SPECIAL_TOKEN_ID_RANGE[0]) & (input_ids <= _COMPATIBLE_AUDIO_SPECIAL_TOKEN_ID_RANGE[1])] = _AUDIO_SPECIAL_TOKEN_ID input_ids = new_input_ids with torch.no_grad(): image_position_mask = input_ids == _IMAGE_SPECIAL_TOKEN_ID non_image_position_mask = ~image_position_mask assert input_embeds is None if self.training: assert input_image_embeds is not None or input_audio_embeds is not None if input_image_embeds is not None: image_hidden_states = self.image_embed( input_ids=input_ids, input_embeds=input_image_embeds, image_sizes=image_sizes, wte=wte, image_attention_mask=image_attention_mask ) if input_audio_embeds is not None: audio_hidden_states = self.audio_embed( input_ids=input_ids, input_embeds=input_audio_embeds, audio_embed_sizes=audio_embed_sizes, audio_attention_mask=audio_attention_mask, wte=wte, audio_projection_mode=audio_projection_mode, ) # merge image and audio hidden states # NOTE weijian: for non-image-audio tokens, here we use audio hidden states # actually, in the debug code above, the non-image-audio tokens from image_hidden_states and audio_hidden_states should be the same if input_image_embeds is not None and input_audio_embeds is not None: dtype = image_hidden_states.dtype hidden_states = image_hidden_states * image_position_mask.to(dtype).unsqueeze(-1) + audio_hidden_states * non_image_position_mask.to(dtype).unsqueeze(-1) elif input_image_embeds is not None: hidden_states = image_hidden_states elif input_audio_embeds is not None: hidden_states = audio_hidden_states else: assert wte is not None hidden_states = wte(input_ids) return hidden_states # Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Phi3 class Phi4MMRMSNorm(nn.Module): def __init__(self, hidden_size, eps=1e-6): """ Phi4MMRMSNorm is equivalent to T5LayerNorm """ super().__init__() self.weight = nn.Parameter(torch.ones(hidden_size)) self.variance_epsilon = eps def forward(self, hidden_states): input_dtype = hidden_states.dtype hidden_states = hidden_states.to(torch.float32) variance = hidden_states.pow(2).mean(-1, keepdim=True) hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) return self.weight * hidden_states.to(input_dtype) def extra_repr(self): return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}" # Copied from transformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding with gemma->phi3, Gemma->Phi3 class Phi4MMRotaryEmbedding(nn.Module): def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None): super().__init__() self.dim = dim self.max_position_embeddings = max_position_embeddings self.base = base inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float() / self.dim)) self.register_buffer("inv_freq", tensor=inv_freq, persistent=False) @torch.no_grad() def forward(self, x, position_ids, seq_len=None): # x: [bs, num_attention_heads, seq_len, head_size] self.inv_freq.to(x.device) inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1) position_ids_expanded = position_ids[:, None, :].float() # Force float32 since bfloat16 loses precision on long contexts # See https://github.com/huggingface/transformers/pull/29285 device_type = x.device.type device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu" with torch.autocast(device_type=device_type, enabled=False): freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) emb = torch.cat((freqs, freqs), dim=-1) cos = emb.cos() sin = emb.sin() return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) class Phi4MMSuScaledRotaryEmbedding(Phi4MMRotaryEmbedding): def __init__(self, dim, config, device=None): warnings.warn( "The class Phi4MMSuScaledRotaryEmbedding is deprecated and will be removed in version 5 of Transformers. Please" " use Phi4MMLongRoPEScaledRotaryEmbedding instead.", FutureWarning, ) super().__init__(dim, config.max_position_embeddings, config.rope_theta, device) self.short_factor = config.rope_scaling["short_factor"] self.long_factor = config.rope_scaling["long_factor"] self.original_max_position_embeddings = config.original_max_position_embeddings @torch.no_grad() def forward(self, x, position_ids, seq_len=None): seq_len = torch.max(position_ids) + 1 if seq_len > self.original_max_position_embeddings: ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device) else: ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device) inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape) inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1) position_ids_expanded = position_ids[:, None, :].float() # Force float32 since bfloat16 loses precision on long contexts # See https://github.com/huggingface/transformers/pull/29285 device_type = x.device.type device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu" with torch.autocast(device_type=device_type, enabled=False): freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) emb = torch.cat((freqs, freqs), dim=-1) scale = self.max_position_embeddings / self.original_max_position_embeddings if scale <= 1.0: scaling_factor = 1.0 else: scaling_factor = math.sqrt(1 + math.log(scale) / math.log(self.original_max_position_embeddings)) cos = emb.cos() * scaling_factor sin = emb.sin() * scaling_factor return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) class Phi4MMYarnScaledRotaryEmbedding(Phi4MMRotaryEmbedding): def __init__(self, dim, config, device=None): warnings.warn( "The class Phi4MMYarnScaledRotaryEmbedding is deprecated and will be removed in version 5 of Transformers", FutureWarning, ) super().__init__(dim, config.max_position_embeddings, config.rope_theta, device) self.short_factor = config.rope_scaling["short_factor"] self.long_factor = config.rope_scaling["long_factor"] self.original_max_position_embeddings = config.original_max_position_embeddings @torch.no_grad() def forward(self, x, position_ids, seq_len=None): seq_len = torch.max(position_ids) + 1 if seq_len > self.original_max_position_embeddings: ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device) else: ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device) inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape) inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1) position_ids_expanded = position_ids[:, None, :].float() # Force float32 since bfloat16 loses precision on long contexts # See https://github.com/huggingface/transformers/pull/29285 device_type = x.device.type device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu" with torch.autocast(device_type=device_type, enabled=False): freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) emb = torch.cat((freqs, freqs), dim=-1) scale = self.max_position_embeddings / self.original_max_position_embeddings if scale <= 1.0: scaling_factor = 1.0 else: scaling_factor = 0.1 * math.log(scale) + 1.0 cos = emb.cos() * scaling_factor sin = emb.sin() * scaling_factor return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) class Phi4MMLongRoPEScaledRotaryEmbedding(Phi4MMRotaryEmbedding): def __init__(self, dim, config, device=None): super().__init__(dim, config.max_position_embeddings, config.rope_theta, device) self.short_factor = config.rope_scaling["short_factor"] self.long_factor = config.rope_scaling["long_factor"] self.original_max_position_embeddings = config.original_max_position_embeddings @torch.no_grad() def forward(self, x, position_ids, seq_len=None): seq_len = seq_len or torch.max(position_ids) + 1 if seq_len > self.original_max_position_embeddings: ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device) else: ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device) inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape) inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1) position_ids_expanded = position_ids[:, None, :].float() # Force float32 since bfloat16 loses precision on long contexts # See https://github.com/huggingface/transformers/pull/29285 device_type = x.device.type device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu" with torch.autocast(device_type=device_type, enabled=False): freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) emb = torch.cat((freqs, freqs), dim=-1) scale = self.max_position_embeddings / self.original_max_position_embeddings if scale <= 1.0: scaling_factor = 1.0 else: scaling_factor = math.sqrt(1 + math.log(scale) / math.log(self.original_max_position_embeddings)) cos = emb.cos() * scaling_factor sin = emb.sin() * scaling_factor return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) # Copied from transformers.models.llama.modeling_llama.rotate_half def rotate_half(x): """Rotates half the hidden dims of the input.""" x1 = x[..., : x.shape[-1] // 2] x2 = x[..., x.shape[-1] // 2 :] return torch.cat((-x2, x1), dim=-1) def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1): """Applies Rotary Position Embedding to the query and key tensors. Args: q (`torch.Tensor`): The query tensor. k (`torch.Tensor`): The key tensor. cos (`torch.Tensor`): The cosine part of the rotary embedding. sin (`torch.Tensor`): The sine part of the rotary embedding. position_ids (`torch.Tensor`, *optional*): Deprecated and unused. unsqueeze_dim (`int`, *optional*, defaults to 1): The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. Returns: `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. """ cos = cos.unsqueeze(unsqueeze_dim) sin = sin.unsqueeze(unsqueeze_dim) rotary_dim = cos.shape[-1] q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:] k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:] q_embed = torch.cat([(q_rot * cos) + (rotate_half(q_rot) * sin), q_pass], dim=-1) k_embed = torch.cat([(k_rot * cos) + (rotate_half(k_rot) * sin), k_pass], dim=-1) return q_embed, k_embed class Phi4MMMLP(nn.Module): def __init__(self, config): super().__init__() self.config = config self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False) self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False) self.activation_fn = ACT2FN[config.hidden_act] def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor: up_states = self.gate_up_proj(hidden_states) gate, up_states = up_states.chunk(2, dim=-1) up_states = up_states * self.activation_fn(gate) return self.down_proj(up_states) # Copied from transformers.models.llama.modeling_llama.repeat_kv with llama->phi def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor: """ This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) """ batch, num_key_value_heads, slen, head_dim = hidden_states.shape if n_rep == 1: return hidden_states hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim) return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim) class Phi4MMAttention(nn.Module): """Multi-headed attention from 'Attention Is All You Need' paper""" def __init__(self, config: Phi4MMConfig, layer_idx: Optional[int] = None): super().__init__() self.config = config self.layer_idx = layer_idx if layer_idx is None: logger.warning_once( f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will " "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` " "when creating this class." ) self.attention_dropout = config.attention_dropout self.hidden_size = config.hidden_size self.num_heads = config.num_attention_heads self.head_dim = self.hidden_size // self.num_heads self.num_key_value_heads = config.num_key_value_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.max_position_embeddings = config.max_position_embeddings self.original_max_position_embeddings = config.original_max_position_embeddings self.rope_theta = config.rope_theta self.rope_scaling = config.rope_scaling self.rotary_ndims = int(self.head_dim * config.partial_rotary_factor) self.is_causal = True if (self.head_dim * self.num_heads) != self.hidden_size: raise ValueError( f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}" f" and `num_heads`: {self.num_heads})." ) op_size = self.num_heads * self.head_dim + 2 * (self.num_key_value_heads * self.head_dim) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) self.qkv_proj = nn.Linear(self.hidden_size, op_size, bias=False) self._init_rope() def _init_rope(self): if self.rope_scaling is None: self.rotary_emb = Phi4MMRotaryEmbedding( self.rotary_ndims, max_position_embeddings=self.max_position_embeddings, base=self.rope_theta, ) else: scaling_type = self.config.rope_scaling["type"] if scaling_type == "longrope": self.rotary_emb = Phi4MMLongRoPEScaledRotaryEmbedding(self.rotary_ndims, self.config) else: raise ValueError(f"Unknown RoPE scaling type {scaling_type}") def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: logger.warning_once("You are not running the flash-attention implementation, expect numerical differences.") bsz, q_len, _ = hidden_states.size() qkv = self.qkv_proj(hidden_states) query_pos = self.num_heads * self.head_dim query_states = qkv[..., :query_pos] key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim] value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :] query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) kv_seq_len = key_states.shape[-2] if past_key_value is not None: if self.layer_idx is None: raise ValueError( f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} " "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class " "with a layer index." ) kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx) cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len) query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads if n_kv_heads < n_heads key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim) if attention_mask is not None: causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] attn_weights += causal_mask # upcast attention to fp32 attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_states.dtype) attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training) attn_output = torch.matmul(attn_weights, value_states) if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim): raise ValueError( f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is" f" {attn_output.size()}" ) attn_output = attn_output.transpose(1, 2).contiguous() attn_output = attn_output.reshape(bsz, q_len, self.hidden_size) attn_output = self.o_proj(attn_output) if not output_attentions: attn_weights = None return attn_output, attn_weights, past_key_value class Phi4MMFlashAttention2(Phi4MMAttention): """ Phi-4-MM flash attention module. This module inherits from `Phi4MMAttention` as the weights of the module stays untouched. The only required change would be on the forward pass where it needs to correctly call the public API of flash attention and deal with padding tokens in case the input contains any of them. """ # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__ def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1. # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0. # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left). self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10() def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.LongTensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: # Phi4MMFlashAttention2 attention does not support output_attentions output_attentions = False bsz, q_len, _ = hidden_states.size() qkv = self.qkv_proj(hidden_states) query_pos = self.num_heads * self.head_dim query_states = qkv[..., :query_pos] key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim] value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :] # Flash attention requires the input to have the shape # batch_size x seq_length x head_dim x hidden_dim # therefore we just need to keep the original shape query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) kv_seq_len = key_states.shape[-2] if past_key_value is not None: if self.layer_idx is None: raise ValueError( f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} " "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class " "with a layer index." ) kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx) # Because the input can be padded, the absolute sequence length depends on the max position id. rotary_seq_len = ( max(kv_seq_len, position_ids[:, -1].max().item() + 1) if position_ids is not None else kv_seq_len ) cos, sin = self.rotary_emb(value_states, seq_len=rotary_seq_len, position_ids=position_ids) query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads if n_kv_heads < n_heads key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) attn_dropout = self.attention_dropout if self.training else 0.0 # In PEFT, usually we cast the layer norms in float32 for training stability reasons # therefore the input hidden states gets silently casted in float32. Hence, we need # cast them back in the correct dtype just to be sure everything works as expected. # This might slowdown training & inference so it is recommended to not cast the LayerNorms # in fp32. if query_states.dtype == torch.float32: if torch.is_autocast_enabled(): target_dtype = torch.get_autocast_gpu_dtype() # Handle the case where the model is quantized elif hasattr(self.config, "_pre_quantization_dtype"): target_dtype = self.config._pre_quantization_dtype else: target_dtype = self.qkv_proj.weight.dtype logger.warning_once( f"The input hidden states seems to be silently casted in float32, this might be related to" f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in" f" {target_dtype}." ) query_states = query_states.to(target_dtype) key_states = key_states.to(target_dtype) value_states = value_states.to(target_dtype) # Reashape to the expected shape for Flash Attention query_states = query_states.transpose(1, 2) key_states = key_states.transpose(1, 2) value_states = value_states.transpose(1, 2) attn_output = _flash_attention_forward( query_states, key_states, value_states, attention_mask, q_len, position_ids=position_ids, dropout=attn_dropout, sliding_window=getattr(self.config, "sliding_window", None), use_top_left_mask=self._flash_attn_uses_top_left_mask, is_causal=self.is_causal, ) attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous() attn_output = self.o_proj(attn_output) if not output_attentions: attn_weights = None return attn_output, attn_weights, past_key_value # copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Phi # TODO @Arthur no longer copied from LLama after static cache class Phi4MMSdpaAttention(Phi4MMAttention): """ Phi4MM attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from `Phi4MMAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to SDPA API. """ # Adapted from Phi4MMAttention.forward def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: if output_attentions: # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. logger.warning_once( "Phi4MMModel is using Phi4MMSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, " 'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.' ) return super().forward( hidden_states=hidden_states, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, ) bsz, q_len, _ = hidden_states.size() qkv = self.qkv_proj(hidden_states) query_pos = self.num_heads * self.head_dim query_states = qkv[..., :query_pos] key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim] value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :] query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) kv_seq_len = key_states.shape[-2] if past_key_value is not None: kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx) cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len) query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) causal_mask = attention_mask if attention_mask is not None: causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, # Reference: https://github.com/pytorch/pytorch/issues/112577. if query_states.device.type == "cuda" and attention_mask is not None: query_states = query_states.contiguous() key_states = key_states.contiguous() value_states = value_states.contiguous() # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling. # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1. is_causal = True if causal_mask is None and q_len > 1 else False attn_output = torch.nn.functional.scaled_dot_product_attention( query_states, key_states, value_states, attn_mask=causal_mask, dropout_p=self.attention_dropout if self.training else 0.0, is_causal=is_causal, ) attn_output = attn_output.transpose(1, 2).contiguous() attn_output = attn_output.view(bsz, q_len, self.hidden_size) attn_output = self.o_proj(attn_output) return attn_output, None, past_key_value PHI4MM_ATTENTION_CLASSES = { "eager": Phi4MMAttention, "flash_attention_2": Phi4MMFlashAttention2, "sdpa": Phi4MMSdpaAttention, } class Phi4MMDecoderLayer(nn.Module): def __init__(self, config: Phi4MMConfig, layer_idx: int): super().__init__() self.config = config self.self_attn = PHI4MM_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx) self.mlp = Phi4MMMLP(config) self.input_layernorm = Phi4MMRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.resid_attn_dropout = nn.Dropout(config.resid_pdrop) self.resid_mlp_dropout = nn.Dropout(config.resid_pdrop) self.post_attention_layernorm = Phi4MMRMSNorm(config.hidden_size, eps=config.rms_norm_eps) def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Tuple[torch.Tensor]] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, cache_position: Optional[torch.LongTensor] = None, **kwargs, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: """ Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`, *optional*): attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. position_ids (`torch.LongTensor` of shape `({0})`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids) output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence kwargs (`dict`, *optional*): Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code into the model """ residual = hidden_states hidden_states = self.input_layernorm(hidden_states) # Self Attention attn_outputs, self_attn_weights, present_key_value = self.self_attn( hidden_states=hidden_states, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, ) hidden_states = residual + self.resid_attn_dropout(attn_outputs) residual = hidden_states hidden_states = self.post_attention_layernorm(hidden_states) hidden_states = self.mlp(hidden_states) hidden_states = residual + self.resid_mlp_dropout(hidden_states) outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights,) if use_cache: outputs += (present_key_value,) return outputs PHI4MM_START_DOCSTRING = r""" This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.) This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior. Parameters: config ([`Phi4MMConfig`]): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights. """ @add_start_docstrings( "The bare Phi-4-MM model outputting raw hidden-states without any specific head on top.", PHI4MM_START_DOCSTRING, ) class Phi4MMPreTrainedModel(PreTrainedModel): config_class = Phi4MMConfig base_model_prefix = "model" supports_gradient_checkpointing = True _no_split_modules = ["Phi4MMDecoderLayer"] _skip_keys_device_placement = "past_key_values" _supports_flash_attn_2 = True _supports_sdpa = True _supports_cache_class = True _version = "0.0.5" def _init_weights(self, module): std = self.config.initializer_range if isinstance(module, nn.Linear): module.weight.data.normal_(mean=0.0, std=std) if module.bias is not None: module.bias.data.zero_() elif isinstance(module, nn.Embedding): module.weight.data.normal_(mean=0.0, std=std) if module.padding_idx is not None: module.weight.data[module.padding_idx].zero_() PHI4MM_INPUTS_DOCSTRING = r""" Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. If `past_key_values` is used, optionally only the last `input_ids` have to be input (see `past_key_values`). If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy. - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids) past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*): Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values` returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`. Two formats are allowed: - a [`~cache_utils.Cache`] instance, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache); - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy cache format. The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the legacy cache format will be returned. If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, this tensor is not affected by padding. It is used to update the cache in the correct position and to infer the complete sequence length. """ @add_start_docstrings( "The bare Phi-4-MM model outputting raw hidden-states without any specific head on top.", PHI4MM_START_DOCSTRING, ) class Phi4MMModel(Phi4MMPreTrainedModel): """ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Phi4MMDecoderLayer`] Args: config: Phi4MMConfig """ def __init__(self, config: Phi4MMConfig): super().__init__(config) self.padding_idx = config.pad_token_id self.vocab_size = config.vocab_size self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx) self.embed_dropout = nn.Dropout(config.embd_pdrop) self.embed_tokens_extend = None if isinstance(config.embd_layer, dict): embedding_config = { 'embedding_cls': config.embd_layer['embedding_cls'], **config.embd_layer } self.embed_tokens_extend = Phi4MMImageAudioEmbedding(config, **embedding_config) self.layers = nn.ModuleList( [Phi4MMDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] ) self._attn_implementation = config._attn_implementation self.norm = Phi4MMRMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.gradient_checkpointing = False # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self): return self.embed_tokens def set_input_embeddings(self, value): self.embed_tokens = value @add_start_docstrings_to_model_forward(PHI4MM_INPUTS_DOCSTRING) def forward( self, input_ids: torch.LongTensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[List[torch.FloatTensor]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, input_image_embeds: Optional[torch.FloatTensor] = None, image_sizes: Optional[torch.LongTensor] = None, image_attention_mask=None, input_audio_embeds: Optional[torch.FloatTensor] = None, audio_embed_sizes=None, audio_attention_mask=None, audio_projection_mode=None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, **kwargs, ) -> Union[Tuple, BaseModelOutputWithPast]: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) use_cache = use_cache if use_cache is not None else self.config.use_cache return_dict = return_dict if return_dict is not None else self.config.use_return_dict if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError("You must specify exactly one of input_ids or inputs_embeds") if self.gradient_checkpointing and self.training: if use_cache: logger.warning_once( "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." ) use_cache = False # kept for BC (non `Cache` `past_key_values` inputs) return_legacy_cache = False if use_cache and not isinstance(past_key_values, Cache): return_legacy_cache = True if past_key_values is None: past_key_values = DynamicCache() else: past_key_values = DynamicCache.from_legacy_cache(past_key_values) logger.warning_once( "We detected that you are passing `past_key_values` as a tuple of tuples. This is deprecated and " "will be removed in v4.47. Please convert your cache or use an appropriate `Cache` class " "(https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)" ) if inputs_embeds is None: inputs_embeds = self.embed_tokens_extend( input_ids=input_ids, input_embeds=inputs_embeds, input_image_embeds=input_image_embeds, input_audio_embeds=input_audio_embeds, image_sizes=image_sizes, image_attention_mask=image_attention_mask, audio_embed_sizes=audio_embed_sizes, audio_attention_mask=audio_attention_mask, audio_projection_mode=audio_projection_mode, wte=self.embed_tokens, ) if cache_position is None: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 cache_position = torch.arange( past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device ) if position_ids is None: position_ids = cache_position.unsqueeze(0) causal_mask = self._update_causal_mask( attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions ) hidden_states = inputs_embeds # decoder layers all_hidden_states = () if output_hidden_states else None all_self_attns = () if output_attentions else None next_decoder_cache = None for decoder_layer in self.layers: if output_hidden_states: all_hidden_states += (hidden_states,) if self.gradient_checkpointing and self.training: layer_outputs = self._gradient_checkpointing_func( decoder_layer.__call__, hidden_states, causal_mask, position_ids, past_key_values, output_attentions, use_cache, cache_position, ) else: layer_outputs = decoder_layer( hidden_states, attention_mask=causal_mask, position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, ) hidden_states = layer_outputs[0] if use_cache: next_decoder_cache = layer_outputs[2 if output_attentions else 1] if output_attentions: all_self_attns += (layer_outputs[1],) hidden_states = self.norm(hidden_states) # add hidden states from the last decoder layer if output_hidden_states: all_hidden_states += (hidden_states,) next_cache = next_decoder_cache if use_cache else None if return_legacy_cache: next_cache = next_cache.to_legacy_cache() if not return_dict: return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None) return BaseModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=next_cache, hidden_states=all_hidden_states, attentions=all_self_attns, ) def _update_causal_mask( self, attention_mask: torch.Tensor, input_tensor: torch.Tensor, cache_position: torch.Tensor, past_key_values: Cache, output_attentions: bool, ): if self.config._attn_implementation == "flash_attention_2": if attention_mask is not None and 0.0 in attention_mask: return attention_mask return None # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail # to infer the attention mask. past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 using_static_cache = isinstance(past_key_values, StaticCache) using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache) # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward if ( self.config._attn_implementation == "sdpa" and not (using_static_cache or using_sliding_window_cache) and not output_attentions ): if AttentionMaskConverter._ignore_causal_mask_sdpa( attention_mask, inputs_embeds=input_tensor, past_key_values_length=past_seen_tokens, sliding_window=self.config.sliding_window, is_training=self.training, ): return None dtype, device = input_tensor.dtype, input_tensor.device min_dtype = torch.finfo(dtype).min sequence_length = input_tensor.shape[1] # SlidingWindowCache or StaticCache if using_sliding_window_cache or using_static_cache: target_length = past_key_values.get_max_cache_shape() # DynamicCache or no cache else: target_length = ( attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else past_seen_tokens + sequence_length + 1 ) # 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], config=self.config, past_key_values=past_key_values, ) if ( self.config._attn_implementation == "sdpa" and attention_mask is not None and attention_mask.device.type == "cuda" and not output_attentions ): # 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 causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) return causal_mask @staticmethod # Copied from transformers.models.mistral.modeling_mistral.MistralModel._prepare_4d_causal_attention_mask_with_cache_position with Mistral->Phi3 def _prepare_4d_causal_attention_mask_with_cache_position( attention_mask: torch.Tensor, sequence_length: int, target_length: int, dtype: torch.dtype, device: torch.device, cache_position: torch.Tensor, batch_size: int, config: Phi4MMConfig, past_key_values: Cache, ): """ Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing. Args: attention_mask (`torch.Tensor`): A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`. sequence_length (`int`): The sequence length being processed. target_length (`int`): The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet. dtype (`torch.dtype`): The dtype to use for the 4D attention mask. device (`torch.device`): The device to plcae the 4D attention mask on. cache_position (`torch.Tensor`): Indices depicting the position of the input sequence tokens in the sequence. batch_size (`torch.Tensor`): Batch size. config (`Phi4MMConfig`): The model's configuration class past_key_values (`Cache`): The cache class that is being used currently to generate """ 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: min_dtype = torch.finfo(dtype).min causal_mask = torch.full( (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device ) diagonal_attend_mask = torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) if config.sliding_window is not None: # if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also # the check is needed to verify is current checkpoint was trained with sliding window or not if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length: sliding_attend_mask = torch.arange(target_length, device=device) <= ( cache_position.reshape(-1, 1) - config.sliding_window ) diagonal_attend_mask.bitwise_or_(sliding_attend_mask) causal_mask *= diagonal_attend_mask causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1) if attention_mask is not None: causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit if attention_mask.shape[-1] > target_length: attention_mask = attention_mask[:, :target_length] 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 ) return causal_mask class Phi4MMForCausalLM(Phi4MMPreTrainedModel, GenerationMixin): _tied_weights_keys = ["lm_head.weight"] # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.__init__ with Llama->Phi def __init__(self, config): super().__init__(config) self.model = Phi4MMModel(config) self.vocab_size = config.vocab_size self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) # Initialize weights and apply final processing self.post_init() # LoRA related settings assert getattr(config, "vision_lora", None) is not None from peft import LoraConfig, get_peft_model vision_lora_config = LoraConfig( r=config.vision_lora['r'], lora_alpha=config.vision_lora['lora_alpha'], target_modules=config.vision_lora['layer'], lora_dropout=config.vision_lora['dp'], task_type="CAUSAL_LM", ) peft_model = get_peft_model(self.model, vision_lora_config, adapter_name="vision") self.config.vision_lora['r'] = config.vision_lora['r'] self.config.vision_lora['lora_alpha'] = config.vision_lora['lora_alpha'] self.config.vision_lora['layer'] = config.vision_lora['layer'] self.config.vision_lora['dp'] = config.vision_lora['dp'] assert getattr(config, "speech_lora", None) is not None speech_lora_config = LoraConfig( r=config.speech_lora['r'], lora_alpha=config.speech_lora['lora_alpha'], target_modules=config.speech_lora['layer'], lora_dropout=config.speech_lora['dp'], task_type="CAUSAL_LM", ) peft_model.base_model.active_adapter.append("speech") peft_model.add_adapter("speech", speech_lora_config) self.config.speech_lora['r'] = config.speech_lora['r'] self.config.speech_lora['lora_alpha'] = config.speech_lora['lora_alpha'] self.config.speech_lora['layer'] = config.speech_lora['layer'] self.config.speech_lora['dp'] = config.speech_lora['dp'] def set_lora_adapter(self, adapter_name) -> None: from peft.tuners.lora.layer import LoraLayer for module in self.modules(): if isinstance(module, LoraLayer): if module.merged: warnings.warn("Adapter cannot be set when the model is merged. Unmerging the model first.") module.unmerge() module.set_adapter(adapter_name) module._disable_adapters = False def unset_lora_adapter(self) -> None: # Ref: peft/tuners/tuners_utils.py - enable_adapters() # Ref: peft/tuners/lora/layer.py from peft.tuners.lora.layer import LoraLayer for module in self.modules(): if isinstance(module, LoraLayer): # disable grads on all adapter layers # TODO weijian: may use enable_adapters() instead for layer_name in module.adapter_layer_names: layer = getattr(module, layer_name) layer.requires_grad_(False) module._disable_adapters = True # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_input_embeddings def get_input_embeddings(self): return self.model.embed_tokens # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_input_embeddings def set_input_embeddings(self, value): self.model.embed_tokens = value # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_output_embeddings def get_output_embeddings(self): return self.lm_head # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_output_embeddings def set_output_embeddings(self, new_embeddings): self.lm_head = new_embeddings # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_decoder def set_decoder(self, decoder): self.model = decoder # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_decoder def get_decoder(self): return self.model # Ignore copy @add_start_docstrings_to_model_forward(PHI4MM_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC) def forward( self, input_ids: torch.LongTensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[List[torch.FloatTensor]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, input_image_embeds: Optional[torch.FloatTensor] = None, image_sizes: Optional[torch.LongTensor] = None, image_attention_mask=None, input_audio_embeds: Optional[torch.FloatTensor] = None, audio_embed_sizes=None, audio_attention_mask=None, input_mode=None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, num_logits_to_keep: int = 0, ) -> Union[Tuple, CausalLMOutputWithPast]: r""" Args: labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. num_logits_to_keep (`int`, *optional*): Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that token can save memory, which becomes pretty significant for long sequences or large vocabulary size. Returns: Example: ```python >>> from transformers import AutoTokenizer, Phi4MMForCausalLM >>> model = Phi4MMForCausalLM.from_pretrained("TBA") >>> tokenizer = AutoTokenizer.from_pretrained("TBA") >>> prompt = "This is an example script ." >>> inputs = tokenizer(prompt, return_tensors="pt") >>> # Generate >>> generate_ids = model.generate(inputs.input_ids, max_length=30) >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] 'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum' ```""" if ( use_cache and self.config.rope_scaling and cache_position is not None and cache_position[0] == self.config.original_max_position_embeddings ): logger.warning( f"If you are not using the generate method, you may encounter nonsensical outputs after the {self.config.original_max_position_embeddings}th token, as the KV cache needs to be recomputed." ) output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict if isinstance(input_mode, torch.Tensor): assert len(input_mode) == 1 input_mode = input_mode[0].item() input_mode = InputMode(input_mode) if input_mode in [InputMode.VISION_SPEECH, InputMode.VISION]: self.set_lora_adapter('vision') audio_projection_mode = 'vision' elif input_mode == InputMode.SPEECH: self.set_lora_adapter('speech') audio_projection_mode = 'speech' elif input_mode == InputMode.LANGUAGE: self.unset_lora_adapter() audio_projection_mode = 'speech' else: raise ValueError(f"Invalid input_mode: {input_mode}") # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) outputs = self.model( input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, input_image_embeds=input_image_embeds, image_sizes=image_sizes, image_attention_mask=image_attention_mask, input_audio_embeds=input_audio_embeds, audio_embed_sizes=audio_embed_sizes, audio_attention_mask=audio_attention_mask, audio_projection_mode=audio_projection_mode, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) hidden_states = outputs[0] # Only compute necessary logits, and do not upcast them to float if we are not computing the loss logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :]) loss = None if labels is not None: loss = self.loss_function(logits, labels, self.vocab_size) if not return_dict: output = (logits,) + outputs[1:] return (loss,) + output if loss is not None else output return CausalLMOutputWithPast( loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) def prepare_inputs_for_generation( self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, input_image_embeds=None, image_sizes=None, image_attention_mask=None, input_audio_embeds=None, audio_embed_sizes=None, audio_attention_mask=None, input_mode=None, cache_position=None, position_ids=None, use_cache=True, num_logits_to_keep=None, **kwargs ): # Overwritten -- this model may need to switch between short and long rope, invalidating the cache in the # process # When the first time input length reached long and short factor switching point, enforce re-compute cache # It will cause downside of slower at this single token position, however, better than current failure. if ( past_key_values and self.config.rope_scaling and input_ids.shape[1] >= self.config.original_max_position_embeddings + 1 ): past_length = cache_position[0] if past_length <= self.config.original_max_position_embeddings: past_key_values = None model_inputs = super().prepare_inputs_for_generation( input_ids=input_ids, past_key_values=past_key_values, attention_mask=attention_mask, inputs_embeds=inputs_embeds, input_image_embeds=input_image_embeds, image_sizes=image_sizes, image_attention_mask=image_attention_mask, input_audio_embeds=input_audio_embeds, audio_embed_sizes=audio_embed_sizes, audio_attention_mask=audio_attention_mask, input_mode=input_mode, cache_position=cache_position, position_ids=position_ids, use_cache=use_cache, num_logits_to_keep=num_logits_to_keep, **kwargs, ) return model_inputs @add_start_docstrings( """ The [`Phi4MMModel`] with a sequence classification head on top (linear layer). [`Phi4MMForSequenceClassification`] uses the last token in order to do the classification, as other causal models (e.g. GPT-2) do. Since it does classification on the last token, it requires to know the position of the last token. If a `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in each row of the batch). """, PHI4MM_START_DOCSTRING, ) # Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->Phi, LLAMA->PHI, self.transformer->self.model, transformer_outputs->model_outputs class Phi4MMForSequenceClassification(Phi4MMPreTrainedModel): def __init__(self, config): super().__init__(config) self.num_labels = config.num_labels self.model = Phi4MMModel(config) self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False) # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self): return self.model.embed_tokens def set_input_embeddings(self, value): self.model.embed_tokens = value @add_start_docstrings_to_model_forward(PHI4MM_INPUTS_DOCSTRING) def forward( self, input_ids: torch.LongTensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[List[torch.FloatTensor]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, SequenceClassifierOutputWithPast]: r""" labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). """ return_dict = return_dict if return_dict is not None else self.config.use_return_dict model_outputs = self.model( input_ids, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) hidden_states = model_outputs[0] logits = self.score(hidden_states) if input_ids is not None: batch_size = input_ids.shape[0] else: batch_size = inputs_embeds.shape[0] if self.config.pad_token_id is None and batch_size != 1: raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.") if self.config.pad_token_id is None: sequence_lengths = -1 else: if input_ids is not None: # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1 sequence_lengths = sequence_lengths % input_ids.shape[-1] sequence_lengths = sequence_lengths.to(logits.device) else: sequence_lengths = -1 pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths] loss = None if labels is not None: loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config) if not return_dict: output = (pooled_logits,) + model_outputs[1:] return ((loss,) + output) if loss is not None else output return SequenceClassifierOutputWithPast( loss=loss, logits=pooled_logits, past_key_values=model_outputs.past_key_values, hidden_states=model_outputs.hidden_states, attentions=model_outputs.attentions, ) @add_start_docstrings( """ [`Phi4MMModel`] with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """, PHI4MM_START_DOCSTRING, ) # Copied from transformers.models.mpt.modeling_mpt.MptForTokenClassification with Mpt->Phi,MPT->PHI,self.transformer->self.model,transformer_outputs->model_outputs class Phi4MMForTokenClassification(Phi4MMPreTrainedModel): def __init__(self, config: Phi4MMConfig): super().__init__(config) self.num_labels = config.num_labels self.model = Phi4MMModel(config) if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None: classifier_dropout = config.classifier_dropout elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None: classifier_dropout = config.hidden_dropout else: classifier_dropout = 0.1 self.dropout = nn.Dropout(classifier_dropout) self.classifier = nn.Linear(config.hidden_size, config.num_labels) # Initialize weights and apply final processing self.post_init() @add_start_docstrings_to_model_forward(PHI4MM_INPUTS_DOCSTRING) @add_code_sample_docstrings( checkpoint=_CHECKPOINT_FOR_DOC, output_type=TokenClassifierOutput, config_class=_CONFIG_FOR_DOC, ) def forward( self, input_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None, attention_mask: Optional[torch.Tensor] = None, inputs_embeds: Optional[torch.Tensor] = None, labels: Optional[torch.Tensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, **deprecated_arguments, ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]: r""" labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). """ return_dict = return_dict if return_dict is not None else self.config.use_return_dict model_outputs = self.model( input_ids, past_key_values=past_key_values, attention_mask=attention_mask, inputs_embeds=inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) hidden_states = model_outputs[0] hidden_states = self.dropout(hidden_states) logits = self.classifier(hidden_states) loss = None if labels is not None: # move labels to correct device to enable model parallelism labels = labels.to(logits.device) batch_size, seq_length = labels.shape loss_fct = CrossEntropyLoss() loss = loss_fct( logits.view(batch_size * seq_length, self.num_labels), labels.view(batch_size * seq_length) ) if not return_dict: output = (logits,) + model_outputs[2:] return ((loss,) + output) if loss is not None else output return TokenClassifierOutput( loss=loss, logits=logits, hidden_states=model_outputs.hidden_states, attentions=model_outputs.attentions, ) AutoConfig.register("phi4mm", Phi4MMConfig) AutoModelForCausalLM.register(Phi4MMConfig, Phi4MMForCausalLM) Phi4MMConfig.register_for_auto_class() Phi4MMForCausalLM.register_for_auto_class("AutoModelForCausalLM")