from __future__ import annotations # Model Constants IMAGE_TOKEN = "" IMG_START_TOKEN = "" IMG_END_TOKEN = "" IGNORE_INDEX = -100 PAD_FOR_EOS = -300 import torch import torch.nn.functional as F from PIL import Image import torch def mask_token_segment( start_id: int, end_id: int, input_ids: torch.Tensor, fill_value: int = -100): """ Replace *every* token from each `start_id` **through** its matching `end_id` (boundaries included) with `fill_value`. Any spans that start with some other token are left untouched. Works on CUDA, TorchScript, batched via vmap, etc.—no Python loops. """ if input_ids.dim() != 1: raise ValueError("`input_ids` must be 1-D") device = input_ids.device n = input_ids.size(0) # where the *target* start-tokens and end-tokens sit start_pos = (input_ids == start_id).nonzero(as_tuple=True)[0] # ascending end_pos = (input_ids == end_id).nonzero(as_tuple=True)[0] # ascending if start_pos.numel() == 0: return input_ids.clone() # ── pair every start with the first end that comes *after* it ──────────────── # searchsorted gives the insertion index into the (sorted) end positions idx_in_end = torch.searchsorted(end_pos, start_pos, right=False) have_match = idx_in_end < end_pos.size(0) # safety: drop unmatched start_pos = start_pos[have_match] end_pos = end_pos[idx_in_end[have_match]] # (rare) guard against pathological orderings keep = end_pos > start_pos start_pos, end_pos = start_pos[keep], end_pos[keep] if start_pos.numel() == 0: return input_ids # ── differential “scan-line” trick to build the span mask in O(N) ─────────── # +1 at each start index, -1 at the element *after* each end delta = torch.zeros(n + 1, dtype=torch.int8, device=device) delta[start_pos] += 1 delta[end_pos + 1] -= 1 # +1 is safe because delta is length n+1 inside = torch.cumsum(delta[:-1], dim=0) > 0 # boolean mask, incl. boundaries # ── apply ──────────────────────────────────────────────────────────────────── out = input_ids.clone() out[inside] = fill_value return out def maybe_zero_3(param, ignore_status=False, name=None): from deepspeed import zero from deepspeed.runtime.zero.partition_parameters import ZeroParamStatus if hasattr(param, "ds_id"): if param.ds_status == ZeroParamStatus.NOT_AVAILABLE: if not ignore_status: print(name, 'no ignore status') with zero.GatheredParameters([param]): param = param.data.detach().cpu().clone() else: param = param.detach().cpu().clone() return param # Borrowed from peft.util.get_peft_model_state_dict def get_peft_state_maybe_zero_3(named_params, bias): if bias == "none": to_return = {k: t for k, t in named_params if "lora_" in k} elif bias == "all": to_return = {k: t for k, t in named_params if "lora_" in k or "bias" in k} elif bias == "lora_only": to_return = {} maybe_lora_bias = {} lora_bias_names = set() for k, t in named_params: if "lora_" in k: to_return[k] = t bias_name = k.split("lora_")[0] + "bias" lora_bias_names.add(bias_name) elif "bias" in k: maybe_lora_bias[k] = t for k, t in maybe_lora_bias: if bias_name in lora_bias_names: to_return[bias_name] = t else: raise NotImplementedError to_return = {k: maybe_zero_3(v, ignore_status=True) for k, v in to_return.items()} return to_return def get_peft_state_non_lora_maybe_zero_3(named_params, require_grad_only=True): to_return = {k: t for k, t in named_params if "lora_" not in k} if require_grad_only: to_return = {k: t for k, t in to_return.items() if t.requires_grad} to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()} return to_return def find_all_linear_names(modules): lora_module_names = set() for name, module in modules(): if isinstance(module, torch.nn.Linear): names = name.split('.') lora_module_names.add(names[0] if len(names) == 1 else names[-1]) if 'lm_head' in lora_module_names: # needed for 16-bit lora_module_names.remove('lm_head') return list(lora_module_names) def expand2square(pil_img, background_color): width, height = pil_img.size if width == height: return pil_img elif width > height: result = Image.new(pil_img.mode, (width, width), background_color) result.paste(pil_img, (0, (width - height) // 2)) return result else: result = Image.new(pil_img.mode, (height, height), background_color) result.paste(pil_img, ((height - width) // 2, 0)) return result def pad_and_stack(img_list, pad_value=0.0): """ img_list : list[Tensor] each (C, H, W) already *normalised* pad_value: float or tuple/list of 3 floats (one per channel) Use 0.0 if your processor has already centred to mean 0. Returns ------- batch : Tensor (B, C, H_max, W_max) """ # 1. target square size --------------------------------------------------- h_max = max(t.shape[1] for t in img_list) w_max = max(t.shape[2] for t in img_list) H, W = max(h_max, w_max), max(h_max, w_max) # 2. create padded copies ------------------------------------------------- padded = [] for img in img_list: c, h, w = img.shape canvas = img.new_full((c, H, W), pad_value) # filled with mean/zeros canvas[:, :h, :w] = img # top-left corner padded.append(canvas) return torch.stack(padded, 0) # (B,C,H,W) # ------------------------------------------------------------------------------------------ # Copyright (c) 2024 Baifeng Shi. # All rights reserved. # # Licensed under the MIT License (MIT). See LICENSE in the repo root for license information. # ------------------------------------------------------------------------------------------ import torch def split_chessboard(x, num_split): """ x: b * c * h * w Deividing x into num_split**2 sub-squares, and concatenate all the sub-squares on the batch dimension """ B, C, H, W = x.shape assert H % num_split == 0 and W % num_split == 0 h, w = H // num_split, W // num_split x_split = torch.cat([x[:, :, i*h:(i+1)*h, j*w:(j+1)*w] for i in range(num_split) for j in range(num_split)], dim=0) return x_split def merge_chessboard(x, num_split): """ x: b * c * h * w Assuming x contains num_split**2 sub-squares concatenated along batch dimension, merge the sub-squares back to the original whole square. (inverse of split_chessboard) """ B, C, H, W = x.shape assert B % (num_split**2) == 0 b = B // (num_split**2) x_merge = torch.cat([torch.cat([x[(i*num_split + j)*b:(i*num_split + j + 1)*b] for j in range(num_split)], dim=-1) for i in range(num_split)], dim=-2) return x_merge def batched_forward(model, x, batch_size=-1): if batch_size == -1: return model(x) else: x_batched = x.split(batch_size) outs = [model(x) for x in x_batched] return torch.cat(outs, dim=0) # ------------------------------------------------------------------------------------------ # Copyright (c) 2024 Baifeng Shi. # All rights reserved. # # Licensed under the MIT License (MIT). See LICENSE in the repo root for license information. # ------------------------------------------------------------------------------------------ import math import torch import torch.nn.functional as F from einops import rearrange def multiscale_forward(model, input, scales=None, img_sizes=None, max_split_size=None, resize_output_to_idx=0, num_prefix_token=0, output_shape='bnc', split_forward=False): # print(f"Input shape: {input.shape}") assert input.dim() == 4, "Input image must be in the shape of BxCxHxW." assert input.shape[2] == input.shape[3], "Currently only square images are supported." assert output_shape in ['bnc', 'bchw'], "Output shape should be either BxNxC (e.g., ViT) or BxCxHxW (e.g., ConvNet)." assert output_shape == 'bnc' or num_prefix_token == 0, "For ConvNet there shouldn't be any prefix token." b, c, input_size, _ = input.shape # image size for each scale assert scales is not None or img_sizes is not None, "Please assign either scales or img_sizes." img_sizes = img_sizes or [int(input_size * scale) for scale in scales] # prepare multiscale inputs max_split_size = max_split_size or input_size # The maximum size of each split of image. Set as the input size by default num_splits = [math.ceil(size / max_split_size) for size in img_sizes] # number of splits each scale input_multiscale = [] for size, num_split in zip(img_sizes, num_splits): x = F.interpolate(input.to(torch.float32), size=size, mode='bicubic').to(input.dtype) x = split_chessboard(x, num_split=num_split) input_multiscale.append(x) # run feedforward on each scale outs_multiscale = [batched_forward(model, x, b) if split_forward else model(x) for x in input_multiscale] if num_prefix_token > 0: outs_prefix_multiscale = [out[:, :num_prefix_token] for out in outs_multiscale] outs_multiscale = [out[:, num_prefix_token:] for out in outs_multiscale] if output_shape == 'bnc': outs_multiscale = [rearrange(out, 'b (h w) c -> b c h w', h=int(out.shape[1] ** 0.5), w=int(out.shape[1] ** 0.5)) for out in outs_multiscale] # merge outputs of different splits for each scale separately outs_multiscale = [merge_chessboard(out, num_split=num_split) for num_split, out in zip(num_splits, outs_multiscale)] # interpolate outputs from different scales and concat together output_size = outs_multiscale[resize_output_to_idx].shape[-2] out = torch.cat([F.interpolate(outs_multiscale[i].to(torch.float32), size=output_size, mode='area').to(outs_multiscale[i].dtype) for i in range(len(outs_multiscale))], dim=1) if output_shape == 'bnc': out = rearrange(out, 'b c h w -> b (h w) c') if num_prefix_token > 0: # take the mean of prefix tokens from different splits for each scale outs_prefix_multiscale = [torch.stack(out.split(b, dim=0), dim=0).mean(dim=0) for out in outs_prefix_multiscale] out_prefix_multiscale = torch.cat(outs_prefix_multiscale, dim=-1) out = torch.cat([out_prefix_multiscale, out], dim=1) return out import torch import torch.nn as nn class MLPAdapter(nn.Module): def __init__(self, input_dim, hidden_dim, output_dim, num_layers=2, activation='gelu', checkpoint_path=None, device=None, **kwargs): """ Initialize the MLPAdapter with the given dimensions and activation function. Args: input_dim (int): Input dimension. hidden_dim (int): Hidden dimension. output_dim (int): Output dimension. layers (int): Number of layers in the MLP. activation (str): Activation function to use ('gelu' or 'relu'). """ super().__init__() self.num_layers = num_layers self.activation = activation self.output_dim = output_dim # Define the first layer layers_list = [nn.Linear(input_dim, hidden_dim, device=device)] if activation == 'gelu': layers_list.append(nn.GELU()) elif activation == 'relu': layers_list.append(nn.ReLU()) else: raise ValueError("Unsupported activation function. Use 'gelu' or 'relu'.") # Define the subsequent layers for _ in range(1, num_layers): layers_list.append(nn.Linear(hidden_dim, hidden_dim, device=device)) if activation == 'gelu': layers_list.append(nn.GELU()) elif activation == 'relu': layers_list.append(nn.ReLU()) # Define the final output layer layers_list.append(nn.Linear(hidden_dim, output_dim, device=device)) self.mlp = nn.Sequential(*layers_list) # Load checkpoint if provided if checkpoint_path: self.load_state_dict(torch.load(checkpoint_path, map_location=device), strict=False) print(f"Loaded MLPAdapter from {checkpoint_path}") if device: self.to(device) def forward(self, x): """ Forward pass through the MLPAdapter. Args: x (torch.Tensor): Input tensor. Returns: torch.Tensor: Output tensor after passing through the MLP. """ return self.mlp(x) import torch import torch.nn as nn import torch.nn.functional as F import PIL.Image from typing import List from transformers import AutoModel, AutoImageProcessor class FastVitVisionTower(nn.Module): def __init__(self, pretrained_model_name_or_path, model_params={}, pad_to_square=True, **kwargs): super().__init__() self.is_loaded = False self.pretrained_model_name_or_path = pretrained_model_name_or_path self.model_params = model_params self.pad_to_square = pad_to_square self.load_model() @property def output_dim(self): return self.vision_tower.config.embed_dim if self.vision_tower else None def load_model(self): if self.is_loaded: return self.image_processor = AutoImageProcessor.from_pretrained(self.pretrained_model_name_or_path) self.image_processor.crop_size = self.image_processor.size self.vision_tower = AutoModel.from_pretrained( self.pretrained_model_name_or_path, **self.model_params, ) self.vision_tower.requires_grad_(False) self.is_loaded = True def preprocess_images(self, imgs: List[PIL.Image.Image], pad_and_stack_tensors=True) -> torch.Tensor: img_mean = tuple(int(x * 255) for x in self.image_processor.image_mean) if self.pad_to_square: imgs = [expand2square(img, img_mean) for img in imgs] imgs = [self.image_processor(img, do_resize=True, do_center_crop=False, return_tensors="pt")['pixel_values'][0] for img in imgs] if pad_and_stack_tensors: imgs = pad_and_stack(imgs, pad_value=0.0) imgs = imgs.to(dtype=torch.float32, device=self.device) return imgs def forward(self, images): if type(images) is list: image_features = [] for image in images: image_feature = self.vision_tower( image.to(device=self.device, dtype=self.dtype).unsqueeze(0) ) image_features.append(image_feature) else: image_features = self.vision_tower( images.to(device=self.device, dtype=self.dtype), ) return image_features @property def dummy_feature(self): return torch.zeros(1, self.embed_dim, device=self.device, dtype=self.dtype) @property def dtype(self): return self.vision_tower.dtype @property def device(self): return self.vision_tower.device @property def config(self): if self.is_loaded: return self.vision_tower.config else: return self.cfg_only @property def hidden_size(self): return self.config.embed_dim @property def num_patches(self): return (self.config.image_size // self.config.patch_size) ** 2 class FastVitVisionTowerS2(FastVitVisionTower): def __init__(self, pretrained_model_name_or_path, s2_scales, model_params={}, **kwargs): self.s2_scales = list(map(int, s2_scales.split(','))) self.s2_scales.sort() self.s2_split_size = self.s2_scales[0] self.s2_image_size = self.s2_scales[-1] super().__init__(pretrained_model_name_or_path, model_params) self.multiscale_forward = multiscale_forward @property def output_dim(self): return (2*self.vision_tower.config.embed_dim) if self.vision_tower else None def load_model(self): if self.is_loaded: return super().load_model() self.image_processor.size = self.image_processor.crop_size = { "height": self.s2_image_size, "width": self.s2_image_size } def forward_feature(self, images): image_size = self.vision_tower.config.image_size if images.shape[2] != image_size or images.shape[3] != image_size: images = F.interpolate( images, size=(image_size, image_size), mode="bilinear", align_corners=False, antialias=True ) return self.vision_tower( images.to(device=self.device, dtype=self.dtype), ) def forward(self, images): if type(images) is list: image_features = [] for image in images: image_feature = self.multiscale_forward( self.forward_feature, image.unsqueeze(0), img_sizes=self.s2_scales, max_split_size=self.s2_split_size ) image_features.append(image_feature) else: image_features = self.multiscale_forward( self.forward_feature, images, img_sizes=self.s2_scales, max_split_size=self.s2_split_size ) return image_features @property def hidden_size(self): return self.config.embed_dim * len(self.s2_scales) import torch import torch.nn as nn import PIL.Image from typing import List from transformers import SiglipVisionModel, SiglipImageProcessor, SiglipVisionConfig class SiglipVisionTower(nn.Module): def __init__(self, pretrained_model_name_or_path, model_params={}, pad_to_square=True, **kwargs): super().__init__() self.is_loaded = False self.pretrained_model_name_or_path = pretrained_model_name_or_path self.model_params = model_params self.pad_to_square = pad_to_square self.select_layer = -2 self.load_model() @property def output_dim(self): return self.vision_tower.config.hidden_size if self.vision_tower else None def load_model(self): if self.is_loaded: return self.image_processor = SiglipImageProcessor.from_pretrained(self.pretrained_model_name_or_path) self.image_processor.crop_size = self.image_processor.size self.vision_tower = SiglipVisionModel.from_pretrained( self.pretrained_model_name_or_path, **self.model_params, ) self.vision_tower.requires_grad_(False) self.is_loaded = True def preprocess_images(self, imgs: List[PIL.Image.Image], pad_and_stack_tensors=True) -> torch.Tensor: img_mean = tuple(int(x * 255) for x in self.image_processor.image_mean) if self.pad_to_square: imgs = [expand2square(img, img_mean) for img in imgs] imgs = [self.image_processor(img, return_tensors="pt")['pixel_values'][0] for img in imgs] if pad_and_stack_tensors: imgs = pad_and_stack(imgs, pad_value=0.0) imgs = imgs.to(dtype=torch.float32, device=self.device) return imgs def feature_select(self, image_forward_outs): image_features = image_forward_outs.hidden_states[self.select_layer] return image_features def forward(self, images): if type(images) is list: image_features = [] for image in images: image_forward_out = self.vision_tower(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True) image_feature = self.feature_select(image_forward_out).to(image.dtype) image_features.append(image_feature) else: image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True) image_features = self.feature_select(image_forward_outs).to(images.dtype) return image_features @property def dummy_feature(self): return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype) @property def dtype(self): return self.vision_tower.dtype @property def device(self): return self.vision_tower.device @property def config(self): if self.is_loaded: return self.vision_tower.config else: return self.cfg_only @property def hidden_size(self): return self.config.hidden_size @property def num_patches(self): return (self.config.image_size // self.config.patch_size) ** 2 class SiglipVisionTowerS2(SiglipVisionTower): def __init__(self, pretrained_model_name_or_path, s2_scales, model_params={}, **kwargs): self.s2_scales = list(map(int, s2_scales.split(','))) self.s2_scales.sort() self.s2_split_size = self.s2_scales[0] self.s2_image_size = self.s2_scales[-1] super().__init__(pretrained_model_name_or_path, model_params) self.multiscale_forward = multiscale_forward self.image_processor.size['height'] = self.image_processor.size['width'] = self.s2_image_size self.image_processor.crop_size['height'] = self.image_processor.crop_size['width'] = self.s2_image_size @property def output_dim(self): return (2*self.vision_tower.config.hidden_size) if self.vision_tower else None def load_model(self): if self.is_loaded: return super().load_model() self.image_processor.size['height'] = self.image_processor.size['width'] = self.s2_image_size self.image_processor.crop_size['height'] = self.image_processor.crop_size['width'] = self.s2_image_size def forward_feature(self, images): image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True) image_features = self.feature_select(image_forward_outs).to(images.dtype) return image_features def forward(self, images): if type(images) is list: image_features = [] for image in images: image_feature = self.multiscale_forward( self.forward_feature, image.unsqueeze(0), img_sizes=self.s2_scales, max_split_size=self.s2_split_size ) image_features.append(image_feature) else: image_features = self.multiscale_forward( self.forward_feature, images, img_sizes=self.s2_scales, max_split_size=self.s2_split_size ) return image_features @property def hidden_size(self): return self.config.hidden_size * len(self.s2_scales) import torch import torch.nn as nn import torch.nn.functional as F from torchvision import transforms from typing import List, Tuple, Optional, Union import PIL from transformers import AutoTokenizer, AutoConfig from transformers.modeling_outputs import CausalLMOutputWithPast from .configuration_phi3 import Phi3Config from .modeling_phi3 import Phi3Model, Phi3ForCausalLM DEFAULT_CFG_SPECIAL_TOKENS = { "image_token_id": 200029, "image_start_token_id": 200030, "image_end_token_id": 200031, } DEFAULT_CFG_VISION_TOWER = { "pretrained_model_name_or_path": "kevin510/fast-vit-hd", "type": "fastvit", "s2_scales": "512,1024", "use_s2": True, "pad_to_square": True, "freeze": False, "model_params": { "trust_remote_code": True } } DEFAULT_CFG_VISION_ADAPTER = { "input_dim": 6144, "hidden_dim": 3072, "output_dim": 3072, "layers": 2, "activation": "gelu", "freeze": False, } class FridayConfig(Phi3Config): model_type = "friday" def __init__(self, base_model_name_or_path: str | None = "microsoft/Phi-4-mini-reasoning", delay_load=False, tokenizer_model_max_length=None, **kwargs ): base_kwargs = {} if base_model_name_or_path is not None: base_cfg = Phi3Config.from_pretrained( base_model_name_or_path, trust_remote_code=True, # Phi‑4 uses custom code in the repo ) base_kwargs = base_cfg.to_dict() merged = {**base_kwargs, **kwargs} self.delay_load = delay_load self.tokenizer_model_max_length = tokenizer_model_max_length self._cfg_vision_tower = DEFAULT_CFG_VISION_TOWER.copy() if "cfg_vision_tower" in kwargs: self._cfg_vision_tower.update(kwargs["cfg_vision_tower"]) self._cfg_vision_adapter = DEFAULT_CFG_VISION_ADAPTER.copy() if "cfg_vision_adapter" in kwargs: self._cfg_vision_adapter.update(kwargs["cfg_vision_adapter"]) self._cfg_special_tokens = DEFAULT_CFG_SPECIAL_TOKENS.copy() if "cfg_special_tokens" in kwargs: self._cfg_special_tokens.update(kwargs["cfg_special_tokens"]) super().__init__(**merged) @property def cfg_vision_tower(self): return self._cfg_vision_tower @cfg_vision_tower.setter def cfg_vision_tower(self, value): if not value: raise ValueError("Name cannot be empty") self._cfg_vision_tower.update(value) @property def cfg_vision_adapter(self): return self._cfg_vision_adapter @cfg_vision_adapter.setter def cfg_vision_adapter(self, value): if not value: raise ValueError("Name cannot be empty") self._cfg_vision_adapter.update(value) @property def cfg_special_tokens(self): return self._cfg_special_tokens @cfg_special_tokens.setter def cfg_special_tokens(self, value): if not value: raise ValueError("Name cannot be empty") self._cfg_special_tokens.update(value) class FridayModel(Phi3Model): config_class = FridayConfig def __init__(self, config: FridayConfig): super().__init__(config) self.cfg_vision_adapter = config.cfg_vision_adapter self.cfg_vision_tower = config.cfg_vision_tower self.vision_tower = None self.mm_projector = None if not config.delay_load: self.initialize_vision_modules() def get_vision_tower(self): return self.vision_tower def initialize_vision_modules(self): if self.vision_tower is not None: return if self.cfg_vision_tower.get("type", "siglip").lower() == "siglip": if self.cfg_vision_tower.get("use_s2", True): self.vision_tower = SiglipVisionTowerS2(**self.cfg_vision_tower) else: self.vision_tower = SiglipVisionTower(**self.cfg_vision_tower) elif self.cfg_vision_tower.get("type", "siglip").lower() == "fastvit": if self.cfg_vision_tower.get("use_s2", True): self.vision_tower = FastVitVisionTowerS2(**self.cfg_vision_tower) else: self.vision_tower = FastVitVisionTower(**self.cfg_vision_tower) else: raise ValueError(f"Unsupported vision tower type: {self.cfg_vision_tower.get('type', 'siglip')}. Supported types are 'siglip' and 'fastvit'.") self.vision_tower.load_model() self.mm_projector = MLPAdapter(**self.cfg_vision_adapter) if self.cfg_vision_tower.get("freeze", False): self.set_vision_tower_requires_grad(False) if self.cfg_vision_adapter.get("freeze", False): self.set_vision_adapter_requires_grad(False) def compute_image_features(self, imgs: torch.Tensor) -> torch.Tensor: features = self.vision_tower(imgs) if isinstance(features, list): features = torch.stack(features, dim=1) return self.mm_projector(features) def set_vision_tower_requires_grad(self, requires_grad: bool): if self.vision_tower is not None: for param in self.vision_tower.parameters(): param.requires_grad = requires_grad else: raise ValueError("Vision tower is not initialized. Please call initialize_vision_modules() first.") def set_vision_adapter_requires_grad(self, requires_grad: bool): if self.mm_projector is not None: for param in self.mm_projector.parameters(): param.requires_grad = requires_grad else: raise ValueError("Vision adapter is not initialized. Please call initialize_vision_modules() first.") def set_vision_tower_dtype(self, dtype: torch.dtype): if self.vision_tower is not None: for p in self.vision_tower.parameters(): p.data = p.data.to(dtype) else: raise ValueError("Vision tower is not initialized. Please call initialize_vision_modules() first.") def set_vision_adapter_dtype(self, dtype: torch.dtype): if self.mm_projector is not None: for p in self.mm_projector.parameters(): p.data = p.data.to(dtype) else: raise ValueError("Vision adapter is not initialized. Please call initialize_vision_modules() first.") def is_vision_tower_frozen(self): if self.vision_tower is not None: return all(not p.requires_grad for p in self.vision_tower.parameters()) else: raise ValueError("Vision tower is not initialized. Please call initialize_vision_modules() first.") def is_vision_adapter_frozen(self): if self.mm_projector is not None: return all(not p.requires_grad for p in self.mm_projector.parameters()) else: raise ValueError("Vision adapter is not initialized. Please call initialize_vision_modules() first.") class FridayForCausalLM(Phi3ForCausalLM): config_class = FridayConfig def __init__(self, config: FridayConfig): super().__init__(config) self.config = config self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) self.image_token_id = config.cfg_special_tokens["image_token_id"] self.image_start_id = config.cfg_special_tokens["image_start_token_id"] self.image_end_id = config.cfg_special_tokens["image_end_token_id"] self.model = FridayModel(config) self.post_init() def get_model(self) -> FridayModel: return self.model def get_vision_tower(self) -> SiglipVisionTower: return self.model.get_vision_tower() def get_vision_adapter(self) -> MLPAdapter: return self.model.mm_projector def get_llm_parameters(self, exclude_lora: bool = False): return [ p for n, p in self.named_parameters() if "vision_tower" not in n and "mm_projector" not in n and (not exclude_lora or ("lora_" not in n)) ] def get_llm_named_modules(self): return {n: m for n, m in self.named_modules() if "vision_tower" not in n and "mm_projector" not in n} def set_llm_requires_grad(self, requires_grad: bool, exclude_lora: bool = True): for n, p in self.named_parameters(): if exclude_lora and ("lora_A" in n or "lora_B" in n): continue if "vision_tower" in n or "mm_projector" in n: continue p.requires_grad = requires_grad def set_vision_tower_requires_grad(self, requires_grad: bool): self.model.set_vision_tower_requires_grad(requires_grad) def set_vision_adapter_requires_grad(self, requires_grad: bool): self.model.set_vision_adapter_requires_grad(requires_grad) def set_llm_dtype(self, dtype: torch.dtype): for p in self.get_llm_parameters(): p.data = p.data.to(dtype) def set_vision_tower_dtype(self, dtype: torch.dtype): self.model.set_vision_tower_dtype(dtype) def set_vision_adapter_dtype(self, dtype: torch.dtype): self.model.set_vision_adapter_dtype(dtype) def is_llm_frozen(self): return all(not p.requires_grad for p in self.get_llm_parameters()) def is_vision_tower_frozen(self): return self.model.is_vision_tower_frozen() def is_vision_adapter_frozen(self): return self.model.is_vision_adapter_frozen() def initialize_vision_modules(self): self.model.initialize_vision_modules() def get_multimodal_input_embeddings(self, input_ids, image_features, return_labels=True) -> torch.Tensor: emb_start_image_id = self.model.embed_tokens(torch.tensor([self.image_start_id], device=self.device)) emb_end_image_id = self.model.embed_tokens(torch.tensor([self.image_end_id], device=self.device)) id_ignore = torch.tensor([IGNORE_INDEX], device=self.device) # repetition‑penalty safety ???? # input_ids[input_ids == self.image_token_id] = 0 # Iterate over each batch item embeds_list, labels_list = [], [] for batch_id, item_ids in enumerate(input_ids): image_token_positions = (item_ids == self.image_token_id).nonzero(as_tuple=True)[0] if len(image_token_positions) != image_features[batch_id].shape[0]: raise ValueError( f"Mismatch between number of image tokens ({len(image_token_positions)}) and number of image features ({image_features[batch_id].shape[0]})" ) cursor = 0 emb_parts, lbl_parts = [], [] for indx_image, image_token_pos in enumerate(image_token_positions): if image_token_pos > cursor: span = item_ids[cursor:image_token_pos] emb_parts.append(self.model.embed_tokens(span)) lbl_parts.append(span) # emb_parts.append(emb_start_image_id) lbl_parts.append(id_ignore) # vision embeddings image_tokens = image_features[batch_id][indx_image] if image_tokens.shape[0] == 1 and image_tokens.ndim == 3: image_tokens = image_tokens.squeeze(0) emb_parts.append(image_tokens) lbl_parts.append(id_ignore.repeat(image_tokens.shape[0])) # emb_parts.append(emb_end_image_id) lbl_parts.append(id_ignore) cursor = image_token_pos + 1 # tail text if cursor < item_ids.shape[0]: tail = item_ids[cursor:] emb_parts.append(self.model.embed_tokens(tail)) lbl_parts.append(tail) embeds_list.append(torch.cat(emb_parts, dim=0)) labels_list.append(torch.cat(lbl_parts, dim=0)) return (embeds_list, labels_list) if return_labels else embeds_list def prepare_inputs_for_multimodal( self, input_ids: torch.LongTensor, images: List[List[PIL.Image.Image]], # B x N position_ids: Optional[torch.LongTensor], attention_mask: Optional[torch.Tensor], past_key_values: Optional[List[torch.FloatTensor]], labels: Optional[torch.LongTensor], ) -> Tuple[Optional[torch.Tensor], Optional[torch.LongTensor], Optional[torch.Tensor], Optional[List[torch.FloatTensor]], torch.Tensor, Optional[torch.Tensor]]: # ─────────────────── early return (no image / streaming step) ─────────────────── # if we have already processed images and are in a streaming step we can skip the multimodal processing # but we need to ensure the attention mask and position ids are correct if past_key_values is not None and attention_mask is not None and input_ids.shape[1] == 1: tgt = past_key_values[-1][-1].shape[-2] + 1 attention_mask = torch.cat( [attention_mask, torch.ones((attention_mask.size(0), tgt - attention_mask.size(1)), dtype=attention_mask.dtype, device=attention_mask.device)], dim=1, ) position_ids = (attention_mask.sum(dim=1, keepdim=True) - 1).long() return input_ids, position_ids, attention_mask, past_key_values, None, labels # ─────────────────────────── images: (B, N) ─────────────────────────── if isinstance(images, list) and isinstance(images[0], list): # images is a list of lists, each containing multiple images, B x N # e.g. [[img1, img2], [img3, img4]] assert len(images) == input_ids.shape[0], f"Batch size mismatch: {len(images)} vs {input_ids.shape[0]}" image_features = [] for sublst_images in images: if len(sublst_images) == 0: image_features.append(torch.zeros((0, self.get_model().mm_projector.output_dim), device=self.device)) else: if isinstance(sublst_images[0], PIL.Image.Image): image_features.append( self.model.compute_image_features( self.model.vision_tower.preprocess_images(sublst_images, pad_and_stack_tensors=True) ) ) elif isinstance(sublst_images[0], torch.Tensor): # This should be a list of tensors of pre-processed images, [(N X 3 X W x H), ...] image_features.append( self.model.compute_image_features(sublst_images) ) elif isinstance(images, list) and isinstance(images[0], PIL.Image.Image): # images is a list of images for a single batch item, 1 x N # e.g. [img1, img2, img3] assert input_ids.shape[0] == 1, f"Batch size mismatch: {len(images)} vs {input_ids.shape[0]}" image_features = [ self.model.compute_image_features( self.model.vision_tower.preprocess_images(images, pad_and_stack_tensors=True) ) ] elif isinstance(images, list) and isinstance(images[0], torch.Tensor): # This should be a list of tensors of pre-processed images, [(N X 3 X W x H), ...] # The list length should match the batch size assert input_ids.shape[0] == len(images), f"Batch size mismatch: {len(images)} vs {input_ids.shape[0]}" image_features = [ self.model.compute_image_features(imgs) for imgs in images ] elif isinstance(images, PIL.Image.Image): # images is a single image, 1 x 1 # e.g. img1 assert input_ids.shape[0] == 1, f"Batch size mismatch: {len(images)} vs {input_ids.shape[0]}" image_features = [ self.model.compute_image_features( self.model.vision_tower.preprocess_images([images]) ) ] else: raise ValueError(f"Unsupported images format: {type(images)}. Expected list of PIL images, a single PIL image or a Tensor of pre-processed images") # ─────────────────────────── image_features: (B x N x D) ─────────────────────────── if isinstance(image_features, list): assert input_ids.shape[0] == len(image_features), f"Incorrectly formatted image_features: list length should match batch size" assert isinstance(image_features[0], torch.Tensor), f"Incorrectly formatted image_features: list items should be tensors" elif isinstance(image_features, torch.Tensor): assert input_ids.shape[0] == image_features.shape[0], f"Incorrectly formatted image_features: tensor should match batch size" # ───────────────────────────── pad handling prelims ────────────────────────────── if attention_mask is None: attention_mask = torch.ones_like(input_ids, dtype=torch.bool) else: attention_mask = attention_mask.bool() if position_ids is None: position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device) input_ids_nopad = [ids[mask] for ids, mask in zip(input_ids, attention_mask)] embeds_list, labels_list = self.get_multimodal_input_embeddings( input_ids_nopad, image_features, return_labels=True ) # ───────────────────── truncate then pad back to rectangle ────────────────────── new_input_embeds = torch.nn.utils.rnn.pad_sequence( embeds_list, batch_first=True, padding_value=0.0 ).to(dtype=self.dtype) new_labels = torch.nn.utils.rnn.pad_sequence( labels_list, batch_first=True, padding_value=IGNORE_INDEX ).long() if self.config.tokenizer_model_max_length is not None: new_input_embeds = new_input_embeds[:, :self.config.tokenizer_model_max_length] new_labels = new_labels[:, :self.config.tokenizer_model_max_length] # ────────────────────────────── attention mask and position ids ──────────────── attention_mask = ( torch.arange(new_input_embeds.size(1), device=input_ids.device) .unsqueeze(0) < torch.tensor([e.size(0) for e in embeds_list], device=input_ids.device).unsqueeze(1) ) raw_pos = attention_mask.cumsum(dim=1) - 1 position_ids = raw_pos.masked_fill(~attention_mask, 0).long() if not self.training: new_labels = None return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels # ------------------------------------------------------------------ def forward( self, input_ids: torch.LongTensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Union[Cache, 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, cache_position: Optional[torch.LongTensor] = None, logits_to_keep: Union[int, torch.Tensor] = 0, images: Optional[PIL.Image.Image] = None, **kwargs: Unpack[KwargsForCausalLM], ) -> Union[Tuple, CausalLMOutputWithPast]: is_multi_modal = images is not None and not ( ( isinstance(images, list) and (len(images) == 0 or all(i == [] for i in images)) ) ) if inputs_embeds is None and is_multi_modal: ( input_ids, position_ids, attention_mask, past_key_values, inputs_embeds, labels ) = self.prepare_inputs_for_multimodal( input_ids=input_ids, images=images, position_ids=position_ids, attention_mask=attention_mask, past_key_values=past_key_values, labels=labels, ) if cache_position is not None and inputs_embeds is not None and cache_position.shape[0] != inputs_embeds.shape[1]: cache_position = torch.arange(inputs_embeds.shape[1], device=self.device) return Phi3ForCausalLM.forward( self, input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, labels=labels, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, cache_position=cache_position, logits_to_keep=logits_to_keep, **kwargs ) def print_device_configuration(self): print("*************Device Configuration*********") if len(self.get_llm_parameters()) > 0: llm_device = set({str(p.device) for p in self.get_llm_parameters()}) llm_dtype = set({p.dtype for p in self.get_llm_parameters()}) print(f"LLM Parameters:\t\t\tdevice: {llm_device}\tdtype: {llm_dtype}\tfrozen: {self.is_llm_frozen()}") else: print("LLM parameters have not been initialized") if self.get_model().vision_tower is not None: vt_device = set({str(p.device) for p in self.get_model().vision_tower.parameters()}) vt_dtype = set({p.dtype for p in self.get_model().vision_tower.parameters()}) print(f"Vision Tower Parameters:\tdevice: {vt_device}\tdtype: {vt_dtype}\tfrozen: {self.is_vision_tower_frozen()}") else: print("Vision tower parameters have not been initialized") if self.get_model().mm_projector is not None: mm_device = set({str(p.device) for p in self.get_model().mm_projector.parameters()}) mm_dtype = set({p.dtype for p in self.get_model().mm_projector.parameters()}) print(f"MM Projector Parameters:\tdevice: {mm_device}\tdtype: {mm_dtype}\tfrozen: {self.is_vision_adapter_frozen()}") else: print("MM Projector parameters have not been initialized") print("******************************************") def build_tokenizer(base_model_id: str) -> Tuple[AutoTokenizer, dict]: tok = AutoTokenizer.from_pretrained(base_model_id, padding_side="right") specials = {t: tok.convert_tokens_to_ids(t) for t in [IMAGE_TOKEN, IMG_START_TOKEN, IMG_END_TOKEN] if t in tok.vocab} if len(specials) < 3: n = tok.add_tokens([IMAGE_TOKEN, IMG_START_TOKEN, IMG_END_TOKEN], special_tokens=True) tok.pad_token = tok.eos_token specials = { "image": tok.convert_tokens_to_ids(IMAGE_TOKEN), "start": tok.convert_tokens_to_ids(IMG_START_TOKEN), "end": tok.convert_tokens_to_ids(IMG_END_TOKEN), } return tok, specials