diff --git "a/modular_isaac.py" "b/modular_isaac.py" --- "a/modular_isaac.py" +++ "b/modular_isaac.py" @@ -1,172 +1,107 @@ -# Copyright (c) 2024 Perceptron, Inc. All rights reserved. -# Perceptron, Inc. Non-Production License (2024-01-01) - - -### 1. Scope and acceptance - -# **1.1. Scope of the Agreement.** -# This Agreement applies to any use, modification, or Distribution of any Perceptron Model by You, regardless of the source You obtained a copy of such Perceptron Model. -# -# **1.2. Acceptance.** By accessing, using, modifying, Distributing a Perceptron Model, or by creating, using or distributing a Derivative of the Perceptron Model, You agree to be bound by this Agreement. -# -# **1.3. Acceptance on behalf of a third-party.** If You accept this Agreement on behalf of Your employer or another person or entity, You warrant and represent that You have the authority to act and accept this Agreement on their behalf. In such a case, the word “You” in this Agreement will refer to Your employer or such other person or entity. -# -# ## 2. License -# **2.1. Grant of rights.** Subject to Section 3 below, Perceptron, Inc. hereby grants You a non-exclusive, royalty-free, worldwide, non-sublicensable, non-transferable, limited license to use, copy, modify, and Distribute under the conditions provided in Section 2.2 below, the Perceptron Model and any Derivatives made by or for Perceptron, Inc. and to create Derivatives of the Perceptron Model. -# -# **2.2. Distribution of Perceptron Model and Derivatives made by or for Perceptron, Inc..** Subject to Section 3 below, You may Distribute copies of the Perceptron Model and/or Derivatives made by or for Perceptron, Inc., under the following conditions: -# - You must make available a copy of this Agreement to third-party recipients of the Perceptron Models and/or Derivatives made by or for Perceptron, Inc. you Distribute, it being specified that any rights to use the Perceptron Models and/or Derivatives made by or for Perceptron, Inc. shall be directly granted by Perceptron, Inc. to said third-party recipients pursuant to the Perceptron, Inc. Non-Production License agreement executed between these parties; -# - You must retain in all copies of the Perceptron Models the following attribution notice within a “Notice” text file distributed as part of such copies: “Licensed by Perceptron, Inc. under the Perceptron, Inc. Non-Production License”. -# -# **2.3. Distribution of Derivatives made by or for You.** Subject to Section 3 below, You may Distribute any Derivatives made by or for You under additional or different terms and conditions, provided that: -# - In any event, the use and modification of Perceptron Model and/or Derivatives made by or for Perceptron, Inc. shall remain governed by the terms and conditions of this Agreement; -# - You include in any such Derivatives made by or for You prominent notices stating that You modified the concerned Perceptron Model; and -# - Any terms and conditions You impose on any third-party recipients relating to Derivatives made by or for You shall neither limit such third-party recipients’ use of the Perceptron Model or any Derivatives made by or for Perceptron, Inc. in accordance with the Perceptron, Inc. Non-Production License nor conflict with any of its terms and conditions. -# -# ## 3. Limitations -# **3.1. Misrepresentation.** You must not misrepresent or imply, through any means, that the Derivatives made by or for You and/or any modified version of the Perceptron Model You Distribute under your name and responsibility is an official product of Perceptron, Inc. or has been endorsed, approved or validated by Perceptron, Inc., unless You are authorized by Us to do so in writing. -# -# **3.2. Usage Limitation** -# - You shall only use the Perceptron Models and Derivatives (whether or not created by Perceptron, Inc.) for testing, research, Personal, or evaluation purposes in Non-Production Environments; -# - Subject to the foregoing, You shall not supply the Perceptron Models or Derivatives in the course of a commercial activity, whether in return for payment or free of charge, in any medium or form, including but not limited to through a hosted or managed service (e.g. SaaS, cloud instances, etc.), or behind a software layer. -# -# **3.3. Usage not permitted under this Agreement.** If You want to use a Perceptron Model or a Derivative for any purpose that is not expressly authorized under this Agreement, You must request a license from Perceptron, Inc., which Perceptron, Inc. may grant to You in Perceptron, Inc.’s sole discretion. Please contact Perceptron, Inc. at the following e-mail address if You want to discuss such a license: sales@perceptron.inc -# -# ## 4. Intellectual Property -# **4.1. Trademarks.** No trademark licenses are granted under this Agreement, and in connection with the Perceptron Models, You may not use any name or mark owned by or associated with Perceptron, Inc. or any of its affiliates, except (i) as required for reasonable and customary use in describing and Distributing the Perceptron Models and Derivatives made by or for Perceptron, Inc. and (ii) for attribution purposes as required by this Agreement. -# -# **4.2. Outputs.** We claim no ownership rights in and to the Outputs. You are solely responsible for the Outputs You generate and their subsequent uses in accordance with this Agreement. -# -# **4.3. Derivatives.** By entering into this Agreement, You accept that any Derivatives that You may create or that may be created for You shall be subject to the restrictions set out in Section 3 of this Agreement. -# -# # 5. Liability -# **5.1. Limitation of liability.** In no event, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall Perceptron, Inc. be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this Agreement or out of the use or inability to use the Perceptron Models and Derivatives (including but not limited to damages for loss of data, loss of goodwill, loss of expected profit or savings, work stoppage, computer failure or malfunction, or any damage caused by malware or security breaches), even if Perceptron, Inc. has been advised of the possibility of such damages. -# -# **5.2. Indemnification.** You agree to indemnify and hold harmless Perceptron, Inc. from and against any claims, damages, or losses arising out of or related to Your use or Distribution of the Perceptron Models and Derivatives. +# coding=utf-8 +# Copyright 2025 Perceptron, Inc and The HuggingFace Team. All rights reserved. # -# ## 6. Warranty -# **6.1. Disclaimer.** Unless required by applicable law or agreed to in writing, Perceptron, Inc. provides the Perceptron Models and Derivatives on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. Perceptron, Inc. does not represent nor warrant that the Perceptron Models and Derivatives will be error-free, meet Your or any third party’s requirements, be secure or will allow You or any third party to achieve any kind of result or generate any kind of content. You are solely responsible for determining the appropriateness of using or Distributing the Perceptron Models and Derivatives and assume any risks associated with Your exercise of rights under this Agreement. +# 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 # -# # 7. Termination -# **7.1. Term.** This Agreement is effective as of the date of your acceptance of this Agreement or access to the concerned Perceptron Models or Derivatives and will continue until terminated in accordance with the following terms. +# http://www.apache.org/licenses/LICENSE-2.0 # -# **7.2. Termination.** Perceptron, Inc. may terminate this Agreement at any time if You are in breach of this Agreement. Upon termination of this Agreement, You must cease to use all Perceptron Models and Derivatives and shall permanently delete any copy thereof. Sections 5, 6, 7 and 8 shall survive the termination of this Agreement. -# -# **7.3. Litigation.** If You initiate any legal action or proceedings against Us or any other entity (including a cross-claim or counterclaim in a lawsuit), alleging that the Model or a Derivative, or any part thereof, infringe upon intellectual property or other rights owned or licensable by You, then any licenses granted to You under this Agreement will immediately terminate as of the date such legal action or claim is filed or initiated. -# -# # 8. General provisions -# 8.1. Governing Law. This Agreement will be governed by and construed in accordance with the laws of the State of Washington, without regard to its conflict of law principles. -# -# 8.2. Jurisdiction. The state and federal courts located in King County, Washington shall have exclusive jurisdiction over any dispute arising out of or relating to this Agreement, and You and We consent to personal jurisdiction and venue in such courts. -# -# **8.3. Severability.** If any provision of this Agreement is held to be invalid, illegal or unenforceable, the remaining provisions shall be unaffected thereby and remain valid as if such provision had not been set forth herein. -# -# # 9. Definitions -# **“Agreement”**: means this Perceptron, Inc. Non-Production License agreement governing the access, use, and Distribution of the Perceptron Models and Derivatives. -# -# **“Derivative”**: means any (i) modified version of the Perceptron Model (including but not limited to any customized or fine-tuned version thereof), (ii) work based on the Perceptron Model, or (iii) any other derivative work thereof. For the avoidance of doubt, Outputs are not considered as Derivatives under this Agreement. -# -# **“Distribution”**, **“Distributing”**, **“Distribute”** or **“Distributed”**: means providing or making available, by any means, a copy of the Perceptron Models and/or the Derivatives as the case may be, subject to Section 3 of this Agreement. -# -# **“Perceptron, Inc.”**, **“We”** or **“Us”**: means Perceptron, Inc., a Delaware corporation with its principal place of business at 10900 NE 8th St Suite 613, Bellevue, WA 98004. -# -# **“Perceptron Model”**: means the foundational large language model(s), and its elements which include algorithms, software, instructed checkpoints, parameters, source code (inference code, evaluation code and, if applicable, fine-tuning code) and any other elements associated thereto made available by Perceptron, Inc. under this Agreement, including, if any, the technical documentation, manuals and instructions for the use and operation thereof. -# -# **“Non-Production Environment”**: means any setting, use case, or application of the Perceptron Models or Derivatives that expressly excludes live, real-world conditions, commercial operations, revenue-generating activities, or direct interactions with or impacts on end users (such as, for instance, Your employees or customers). Non-Production Environment may include, but is not limited to, any setting, use case, or application for research, development, testing, quality assurance, training, internal evaluation (other than any internal usage by employees in the context of the company’s business activities), and demonstration purposes. -# -# **“Outputs”**: means any content generated by the operation of the Perceptron Models or the Derivatives from a prompt (i.e., text instructions) provided by users. For the avoidance of doubt, Outputs do not include any components of a Perceptron Models, such as any fine-tuned versions of the Perceptron Models, the weights, or parameters. -# -# **“Personal”**: means any use of a Perceptron Model or a Derivative that is (i) solely for personal, non-profit and non-commercial purposes and (ii) not directly or indirectly connected to any commercial activities, business operations, or employment responsibilities. For illustration purposes, Personal use of a Model or a Derivative does not include any usage by individuals employed in companies in the context of their daily tasks, any activity that is intended to generate revenue, or that is performed on behalf of a commercial entity. -# -# **“You”**: means the individual or entity entering into this Agreement with Perceptron, Inc.. +# 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. from __future__ import annotations import copy import math -import re -from collections import defaultdict -from typing import Any, Callable, Optional, Sequence, Union - -import PIL.Image -import torch -import torch.nn as nn -import torch.nn.functional as F -from transformers import ( - AutoImageProcessor, - AutoModel, - AutoTokenizer, - BatchFeature, - Cache, - Qwen3Config, - Qwen3ForCausalLM, - Qwen3PreTrainedModel, -) -from transformers.cache_utils import SlidingWindowCache, StaticCache +from collections.abc import Callable, Sequence +from enum import IntEnum +from typing import Any, Optional, Union + +from transformers.cache_utils import DynamicCache +from transformers.configuration_utils import PretrainedConfig, layer_type_validation +from transformers.feature_extraction_utils import BatchFeature from transformers.generation.utils import GenerationMixin from transformers.image_processing_utils_fast import ( BaseImageProcessorFast, - DefaultFastImageProcessorKwargs, + ImagesKwargs, SizeDict, group_images_by_shape, reorder_images, ) from transformers.image_utils import ( - ChannelDimension, PILImageResampling, ) -from transformers.modeling_attn_mask_utils import AttentionMaskConverter -from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast +from transformers.masking_utils import ( + ALL_MASK_ATTENTION_FUNCTIONS, + create_masks_for_generate, + packed_sequence_mask_function, +) +from transformers.modeling_outputs import ( + BaseModelOutputWithPast, + CausalLMOutputWithPast, +) from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS -from transformers.models.qwen2.tokenization_qwen2 import Qwen2Tokenizer -from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLRotaryEmbedding +from transformers.models.qwen3.configuration_qwen3 import Qwen3Config +from transformers.models.qwen3.modeling_qwen3 import ( + Qwen3ForCausalLM, + Qwen3Model, + Qwen3PreTrainedModel, +) +from transformers.processing_utils import ProcessorMixin, Unpack +from transformers.utils import TensorType, auto_docstring +from transformers.utils.constants import IMAGENET_STANDARD_MEAN as VISION_MEAN +from transformers.utils.constants import IMAGENET_STANDARD_STD as VISION_STD +from transformers.utils.generic import ( + OutputRecorder, + TransformersKwargs, + can_return_tuple, + check_model_inputs, +) +from transformers.utils.import_utils import ( + is_torch_available, + is_torchdynamo_compiling, + is_torchvision_available, + is_vision_available, +) +from transformers.models.qwen2_5_vl import modeling_qwen2_5_vl as qwen2_5_vl_modeling from transformers.models.siglip2.configuration_siglip2 import Siglip2VisionConfig from transformers.models.siglip2.modeling_siglip2 import ( Siglip2Attention, - Siglip2Encoder as HFSiglip2Encoder, - Siglip2EncoderLayer as HFSiglip2EncoderLayer, - Siglip2VisionEmbeddings as HFSiglip2VisionEmbeddings, + Siglip2Encoder, + Siglip2EncoderLayer, + Siglip2VisionEmbeddings, ) -from transformers.processing_utils import ProcessorMixin, Unpack -from transformers.tokenization_utils import TensorType -from transformers.utils import auto_docstring -from transformers.utils.generic import can_return_tuple -# Vision preprocessing constants -from transformers.utils.constants import IMAGENET_STANDARD_MEAN as VISION_MEAN -from transformers.utils.constants import IMAGENET_STANDARD_STD as VISION_STD -from transformers.utils.import_utils import is_torchdynamo_compiling - -try: - from perceptron.tensorstream.tensorstream import ( - Event, - Stream, - TensorStream, - TextType, - VisionType, - create_stream, - group_streams, - ) - from perceptron.tensorstream.ops import ( - compute_mrope_pos_tensor, - modality_mask, - reconstruct_tensor_stream_from_compact_dict, - tensor_stream_token_view, - ) - from perceptron.tensorstream.ops import ( - slice as ts_slice, + +if is_torch_available(): + import torch + import torch.nn as nn + import torch.nn.functional as F +if is_vision_available(): + from PIL.Image import Image +else: + Image = None +if is_torchvision_available(): + from transformers.models.pix2struct.image_processing_pix2struct_fast import ( + torch_extract_patches, ) -except ModuleNotFoundError as exc: # pragma: no cover - import guard - raise ModuleNotFoundError( - "perceptron.tensorstream is required for the Isaac HuggingFace integration. " - "Ensure the TensorStream package is installed and on PYTHONPATH." - ) from exc -_ORIGINAL_ATTENTION_FUNCTIONS: dict[str, Callable[..., tuple[torch.Tensor, Optional[torch.Tensor]]]] = {} -for _attn_name in ("flash_attention_2", "sdpa", "eager"): - if _attn_name in ALL_ATTENTION_FUNCTIONS: - _ORIGINAL_ATTENTION_FUNCTIONS[_attn_name] = ALL_ATTENTION_FUNCTIONS[_attn_name] +class ModalityType(IntEnum): + """ + Modality identifiers for events. + + Members: + image: Vision tokens (e.g., patches). + text: Textual tokens. + """ + + image = 0 + text = 1 class IsaacVisionConfig(Siglip2VisionConfig): @@ -183,76 +118,79 @@ class IsaacVisionConfig(Siglip2VisionConfig): model_type = "isaac_vision" base_config_key = "vision_config" - _attn_implementation: str | None = None def __init__( self, - pixel_shuffle_scale_factor: int = 1, - num_patches: int = 256, + hidden_size=768, + intermediate_size=3072, + num_hidden_layers=12, + num_attention_heads=12, + num_channels=3, + num_patches=256, + patch_size=16, + hidden_act="gelu_pytorch_tanh", + layer_norm_eps=1e-6, + attention_dropout=0.0, + pixel_shuffle_scale_factor=1, **kwargs, ): super().__init__(**kwargs) + self.hidden_size = hidden_size + self.intermediate_size = intermediate_size + self.num_hidden_layers = num_hidden_layers + self.num_attention_heads = num_attention_heads + self.num_channels = num_channels + self.patch_size = patch_size + self.attention_dropout = attention_dropout + self.layer_norm_eps = layer_norm_eps + self.hidden_act = hidden_act + self.num_patches = num_patches + # Add our custom fields self.pixel_shuffle_scale_factor = pixel_shuffle_scale_factor - self.num_patches = num_patches - if self._attn_implementation is None: - self._attn_implementation = "flash_attention_2" + # Ensure a sensible default attention backend + if getattr(self, "_attn_implementation", None) is None: + self._attn_implementation = "sdpa" -class IsaacImageProcessorKwargs(DefaultFastImageProcessorKwargs, total=False): - patch_size: int | None - max_num_patches: int | None - min_num_patches: int | None - pixel_shuffle_scale: int | None +class IsaacImageProcessorFastKwargs(ImagesKwargs, total=False): + patch_size: Optional[int] + max_num_patches: Optional[int] + min_num_patches: Optional[int] + pixel_shuffle_scale: Optional[int] @auto_docstring class IsaacImageProcessorFast(BaseImageProcessorFast): MAX_PIXELS = 60_000_000 # 60‑megapixel ceiling ≈ 8200 × 7300 px - r"""Fast torch-based image processor for Isaac vision inputs.""" resample = PILImageResampling.BILINEAR model_input_names = ["patches", "token_grids"] - valid_kwargs = IsaacImageProcessorKwargs - unused_kwargs = ["size", "do_center_crop", "crop_size"] + valid_kwargs = IsaacImageProcessorFastKwargs + unused_kwargs = ["size", "do_center_crop", "crop_size", "pad_size", "do_pad"] do_resize = True - size: SizeDict | None = None - default_to_square: bool | None = None do_center_crop = False - crop_size: SizeDict | None = None - patch_size: int | None = 16 - max_num_patches: int | None = 256 - min_num_patches: int | None = None - pixel_shuffle_scale: int | None = 1 + patch_size: Optional[int] = 16 + max_num_patches: Optional[int] = 256 + min_num_patches: Optional[int] = None + pixel_shuffle_scale: Optional[int] = 1 do_pad = False - pad_size: SizeDict | None = None do_rescale = True - rescale_factor = 1 / 255 do_normalize = True image_mean = list(VISION_MEAN) image_std = list(VISION_STD) do_convert_rgb = True - return_tensors = None - data_format = ChannelDimension.FIRST - input_data_format = None - device = None disable_grouping = False - size_divisor: int | None = None def __init__( self, - **kwargs: Unpack[IsaacImageProcessorKwargs], + **kwargs: Unpack[IsaacImageProcessorFastKwargs], ) -> None: super().__init__(**kwargs) - pixel_shuffle_scale = 1 if self.pixel_shuffle_scale is None else int(self.pixel_shuffle_scale) - if pixel_shuffle_scale < 1: - raise ValueError("`pixel_shuffle_scale` must be >= 1") - self.pixel_shuffle_scale = pixel_shuffle_scale - def _validate_preprocess_kwargs(self, **kwargs): # Allow callers to omit resize-related placeholders that BaseImageProcessorFast checks for. kwargs.pop("do_resize", None) @@ -264,31 +202,12 @@ class IsaacImageProcessorFast(BaseImageProcessorFast): def resize( self, - image: "torch.Tensor", + image: torch.Tensor, size: SizeDict, - interpolation: Optional[Any] = None, - antialias: bool = True, **kwargs, ) -> torch.Tensor: - if size.height is None or size.width is None: - raise ValueError("IsaacImageProcessorFast requires explicit `height` and `width` when resizing.") - - resize_mode: Any = interpolation - if hasattr(resize_mode, "value"): - resize_mode = resize_mode.value - elif hasattr(resize_mode, "name"): - resize_mode = resize_mode.name.lower() - elif resize_mode is None: - resize_mode = "bilinear" - - if isinstance(resize_mode, str): - mode_key = resize_mode.lower() - else: - mode_key = resize_mode - - resize_kwargs: dict[str, Any] = {} - if mode_key in {"linear", "bilinear", "bicubic", "trilinear"}: - resize_kwargs["align_corners"] = False + resize_kwargs: dict[str, Any] = {"align_corners": False} + resize_mode = "bilinear" return F.interpolate( image, @@ -299,12 +218,9 @@ class IsaacImageProcessorFast(BaseImageProcessorFast): def _preprocess( self, - images: list["torch.Tensor"], + images: list[torch.Tensor], do_resize: bool, - size: Optional[SizeDict], interpolation: Optional[Any], - do_center_crop: bool, - crop_size: Optional[SizeDict], do_rescale: Optional[bool], rescale_factor: Optional[float], do_normalize: Optional[bool], @@ -312,38 +228,24 @@ class IsaacImageProcessorFast(BaseImageProcessorFast): image_std: Optional[Union[float, Sequence[float]]], disable_grouping: Optional[bool] = None, return_tensors: Optional[Union[str, TensorType]] = None, - do_pad: Optional[bool] = None, - pad_size: Optional[SizeDict] = None, *, - patch_size: int | None = None, - max_num_patches: int | None = None, - min_num_patches: int | None = None, - pixel_shuffle_scale: int | None = None, + patch_size: Optional[int] = None, + max_num_patches: Optional[int] = None, + min_num_patches: Optional[int] = None, + pixel_shuffle_scale: Optional[int] = None, **kwargs, ) -> BatchFeature: - if do_center_crop: - raise ValueError("`do_center_crop` is not supported by IsaacImageProcessorFast.") - if do_pad: - raise ValueError("`do_pad` is not supported by IsaacImageProcessorFast.") + grouped_images, grouped_images_index = group_images_by_shape( + images, disable_grouping=disable_grouping + ) - grouped_images, grouped_images_index = group_images_by_shape(images, disable_grouping=disable_grouping) - processed_patches_grouped: dict[tuple[int, ...], torch.Tensor] = {} - token_grids_grouped: dict[tuple[int, ...], torch.Tensor] = {} - virtual_dims_grouped: dict[tuple[int, ...], torch.Tensor] = {} - real_dims_grouped: dict[tuple[int, ...], torch.Tensor] = {} + grouped_outputs = {} for shape, stacked_images in grouped_images.items(): - if stacked_images.ndim != 4: - raise ValueError("Expected batched channel-first image tensors.") - batch_size, channels, original_height, original_width = stacked_images.shape if bool(self.do_convert_rgb) and channels == 1: stacked_images = stacked_images.repeat(1, 3, 1, 1) - channels = 3 - - if original_height * original_width > self.MAX_PIXELS: - raise ValueError(f"Image (w={original_width}, h={original_height}) > MAX=`{self.MAX_PIXELS}`") target_height, target_width = get_image_size_for_max_num_patches( original_height, @@ -353,44 +255,39 @@ class IsaacImageProcessorFast(BaseImageProcessorFast): min_num_patches=min_num_patches, pixel_shuffle_scale=pixel_shuffle_scale, ) - if do_resize: - resize_size = SizeDict(height=target_height, width=target_width) image_batch = self.resize( - image=stacked_images, - size=resize_size, + stacked_images, + SizeDict(height=target_height, width=target_width), interpolation=interpolation, ) else: - if ((original_height % patch_size) != 0) or ((original_width % patch_size) != 0): - raise ValueError("Image dimensions must be divisible by patch_size when resize is disabled.") - image_batch = stacked_images - target_height, target_width = original_height, original_width - - if do_rescale: - image_batch = self.rescale_and_normalize( - image_batch, - do_rescale=do_rescale, - rescale_factor=rescale_factor, - do_normalize=do_normalize, - image_mean=image_mean, - image_std=image_std, + if (original_height % patch_size) or (original_width % patch_size): + raise ValueError( + f"Image dimensions (h={original_height}, w={original_width}) must be divisible by patch_size={patch_size} when resize is disabled; enable resizing or adjust the input resolution." + ) + image_batch, target_height, target_width = ( + stacked_images, + original_height, + original_width, ) - nhwc_images = image_batch.permute(0, 2, 3, 1) - nhwc_images = _compute_residual_p_frames(nhwc_images, is_p_frame=[False] * batch_size) + image_batch = self.rescale_and_normalize( + image_batch, + do_rescale=do_rescale, + rescale_factor=rescale_factor, + do_normalize=do_normalize, + image_mean=image_mean, + image_std=image_std, + ) - patches = patchify_vision(nhwc_images, patch_size=patch_size) + patches = torch_extract_patches(image_batch, patch_size, patch_size) _, height_tokens, width_tokens, _ = patches.shape token_grid = ( - torch.tensor( - [height_tokens, width_tokens], - dtype=torch.long, - device=patches.device, - ) - .unsqueeze(0) - .repeat(batch_size, 1) + torch.tensor([height_tokens, width_tokens], device=patches.device) + .long() + .expand(batch_size, 2) ) real_dim = ( @@ -403,16 +300,14 @@ class IsaacImageProcessorFast(BaseImageProcessorFast): .repeat(batch_size, 1) ) - if pixel_shuffle_scale > 1: - if (height_tokens % pixel_shuffle_scale) or (width_tokens % pixel_shuffle_scale): - raise ValueError( - "Spatial dimensions must be divisible by pixel_shuffle_scale when pixel shuffle is enabled." - ) - virtual_height = height_tokens // pixel_shuffle_scale - virtual_width = width_tokens // pixel_shuffle_scale - else: - virtual_height = height_tokens - virtual_width = width_tokens + if (height_tokens % pixel_shuffle_scale) or ( + width_tokens % pixel_shuffle_scale + ): + raise ValueError( + f"Token grid (h={height_tokens}, w={width_tokens}) must be divisible by pixel_shuffle_scale={pixel_shuffle_scale}; adjust resize/patch parameters or disable pixel shuffle." + ) + virtual_height = height_tokens // pixel_shuffle_scale + virtual_width = width_tokens // pixel_shuffle_scale virtual_dim = ( torch.tensor( @@ -423,617 +318,302 @@ class IsaacImageProcessorFast(BaseImageProcessorFast): .unsqueeze(0) .repeat(batch_size, 1) ) + grouped_outputs[shape] = (patches, token_grid, virtual_dim, real_dim) + + def _reorder_grouped_item( # reorder an item of tuple payloads using the same grouped_images_index + grouped: dict[ + tuple[int, ...], + tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor], + ], + grouped_index: dict[tuple[int, ...], list[int]], + item_idx: int, + ) -> list[torch.Tensor]: + return reorder_images( + {k: v[item_idx] for k, v in grouped.items()}, grouped_index + ) - processed_patches_grouped[shape] = patches - token_grids_grouped[shape] = token_grid - virtual_dims_grouped[shape] = virtual_dim - real_dims_grouped[shape] = real_dim - - patches_slices = reorder_images(processed_patches_grouped, grouped_images_index) - token_grid_slices = reorder_images(token_grids_grouped, grouped_images_index) - virtual_dim_slices = reorder_images(virtual_dims_grouped, grouped_images_index) - real_dim_slices = reorder_images(real_dims_grouped, grouped_images_index) - - patches_tensor = torch.stack(patches_slices, dim=0) - token_grids_tensor = torch.stack(token_grid_slices, dim=0) - virtual_dims_tensor = torch.stack(virtual_dim_slices, dim=0) - real_dims_tensor = torch.stack(real_dim_slices, dim=0) - - return BatchFeature( - data={ - "patches": patches_tensor, - "token_grids": token_grids_tensor, - "virtual_pixel_size": virtual_dims_tensor, - "real_pixel_size": real_dims_tensor, - }, - tensor_type=return_tensors, - ) - + keys = ("patches", "token_grids", "virtual_pixel_size", "real_pixel_size") + tensors: dict[str, torch.Tensor] = {} -def _max_from_cu(cu: torch.Tensor | None, fallback: int) -> int: - """Helper to compute max sequence length from cumulative sequence lengths.""" - if cu is None or len(cu) < 2: - return fallback - return int((cu[1:] - cu[:-1]).max().item()) + for i, key in enumerate(keys): + slices = _reorder_grouped_item(grouped_outputs, grouped_images_index, i) + tensors[key] = torch.stack(slices, dim=0) + return BatchFeature(data=tensors, tensor_type=return_tensors) -def build_document_attention_mask( - cu_seqlens: torch.Tensor | None, - total_tokens: int, - dtype: torch.dtype, - device: torch.device, -) -> torch.Tensor | None: - """Creates an additive attention mask that blocks cross-document attention.""" - if cu_seqlens is None: - return None +def create_document_attention_mask( + config: PretrainedConfig, + input_embeds: torch.Tensor, + cu_seqlens: Optional[torch.Tensor], +) -> Optional[Union[torch.Tensor, Any]]: + """ + Materialize a backend-specific block-diagonal attention mask from packed cu_seqlens. - if cu_seqlens.numel() < 2: - return None + Returns None if cu_seqlens is missing/degenerate. + """ + if cu_seqlens is None or cu_seqlens.numel() < 2: + return None # Degenerate input: nothing to mask seq_sizes = (cu_seqlens[1:] - cu_seqlens[:-1]).long() - if seq_sizes.numel() == 0: - return None + if seq_sizes.numel() == 0 or int(seq_sizes.sum()) == 0: + return None # All-empty segments produce no attention blocks - seg_ids = torch.repeat_interleave(torch.arange(seq_sizes.numel(), device=device), seq_sizes) - block_mask = seg_ids[:, None] != seg_ids[None, :] - additive_mask = torch.zeros((total_tokens, total_tokens), dtype=dtype, device=device) - additive_mask.masked_fill_(block_mask, float("-inf")) - return additive_mask.view(1, 1, total_tokens, total_tokens) - - -def ensure_document_attention_mask( - attention_mask: Optional[torch.Tensor], - cu_seqlens: Optional[torch.Tensor], - total_tokens: int, - dtype: torch.dtype, - device: torch.device, -) -> Optional[torch.Tensor]: - if attention_mask is not None or cu_seqlens is None: - return attention_mask - - return build_document_attention_mask( - cu_seqlens=cu_seqlens, - total_tokens=total_tokens, - dtype=dtype, - device=device, + seg_ids = torch.repeat_interleave( + torch.arange(seq_sizes.numel(), device=cu_seqlens.device), + seq_sizes, ) - - -def flash_attention_document_mask_forward( - module: torch.nn.Module, - q_lhd: torch.Tensor, # (L, H, D) - k_lhd: torch.Tensor, # (L, H, D) - v_lhd: torch.Tensor, # (L, H, D) - attention_mask: torch.Tensor | None = None, # unused for FA path - dropout: float = 0.0, - scaling: float | None = None, - cum_seq_q: torch.Tensor | None = None, - cum_seq_k: torch.Tensor | None = None, - max_seqlen: int | None = None, - is_causal: bool = False, - **kwargs, -) -> tuple[torch.Tensor, None]: - """FlashAttention that consumes (L, H, D) directly to avoid layout churn.""" - L, H, D = q_lhd.shape - - # Compute max block length once (honor caller when provided) - if max_seqlen is not None: - max_q = max_k = int(max_seqlen) - else: - max_q = _max_from_cu(cum_seq_q, L) - max_k = _max_from_cu(cum_seq_k, L) - - # Ensure contiguity only if needed - if not q_lhd.is_contiguous(): - q_lhd = q_lhd.contiguous() - if not k_lhd.is_contiguous(): - k_lhd = k_lhd.contiguous() - if not v_lhd.is_contiguous(): - v_lhd = v_lhd.contiguous() - - out_lhd, *_ = torch.ops.aten._flash_attention_forward( - query=q_lhd, # (L, H, D) - key=k_lhd, # (L, H, D) - value=v_lhd, # (L, H, D) - cum_seq_q=cum_seq_q, - cum_seq_k=cum_seq_k, - max_q=max_q, - max_k=max_k, - dropout_p=dropout, - is_causal=is_causal, - return_debug_mask=False, - scale=scaling, - window_size_left=-1, - window_size_right=-1, - alibi_slopes=None, + mask_function = packed_sequence_mask_function(seg_ids.view(1, -1)) + + seq_len = input_embeds.shape[1] + cache_position = torch.arange(seq_len, device=input_embeds.device, dtype=torch.long) + + mask_interface = ALL_MASK_ATTENTION_FUNCTIONS[config._attn_implementation] + return mask_interface( + batch_size=input_embeds.shape[0], + cache_position=cache_position, + kv_length=seq_len, + kv_offset=0, + mask_function=mask_function, + attention_mask=None, + allow_is_causal_skip=False, + allow_is_bidirectional_skip=False, + dtype=input_embeds.dtype, + config=config, + use_vmap=False, ) - return out_lhd, None # (L, H, D) - - -def sdpa_document_mask_forward( - q_lhd: torch.Tensor, # (L, H, D) - k_lhd: torch.Tensor, # (L, H, D) - v_lhd: torch.Tensor, # (L, H, D) - dropout: float, - scaling: float | None, - attention_mask: torch.Tensor | None = None, - cu_seqlens: torch.Tensor | None = None, -) -> torch.Tensor: - """SDPA with block-diagonal masking for variable-length sequences.""" - L, H, D = q_lhd.shape - - # Transpose to (1, H, L, D) format for SDPA - Q = q_lhd.permute(1, 0, 2).unsqueeze(0) - K = k_lhd.permute(1, 0, 2).unsqueeze(0) - V = v_lhd.permute(1, 0, 2).unsqueeze(0) - - # Build block-diagonal mask for variable-length sequences - attn_mask = attention_mask - if attn_mask is None: - attn_mask = build_document_attention_mask( - cu_seqlens=cu_seqlens, - total_tokens=L, - dtype=q_lhd.dtype, - device=q_lhd.device, - ) - - if attn_mask is not None and attn_mask.dtype != Q.dtype: - attn_mask = attn_mask.to(Q.dtype) - Y = F.scaled_dot_product_attention(Q, K, V, attn_mask=attn_mask, dropout_p=dropout, scale=scaling) - return Y.squeeze(0).permute(1, 0, 2) # Back to (L, H, D) +class IsaacVisionEmbeddings(Siglip2VisionEmbeddings): + """Adapter around SigLIP2 vision embeddings that consumes packed patch sequences. -class IsaacVisionEmbeddings(HFSiglip2VisionEmbeddings): - """Adapter around SigLIP2 vision embeddings that consumes packed patch sequences.""" + Isaac accepts variable-resolution vision inputs as a single packed sequence with per-image + `token_grids`; packing/unpacking here reconstructs per-image shapes so we can resize positional + embeddings and build `cu_seqlens` for variable-length attention (not generic generation packing). + """ def __init__(self, config: IsaacVisionConfig): super().__init__(config) + self.config = config + self.embed_dim = config.hidden_size + self.patch_size = config.patch_size + + self.patch_embedding = nn.Linear( + in_features=config.num_channels * self.patch_size * self.patch_size, + out_features=self.embed_dim, + ) - def forward(self, seq_patches: torch.Tensor, spatial_shapes: torch.Tensor) -> torch.Tensor: - packed_pixel_values, seq_lengths = self._pack_to_batch(seq_patches, spatial_shapes) + self.num_patches = config.num_patches + self.position_embedding_size = int(self.num_patches**0.5) + self.position_embedding = nn.Embedding(self.num_patches, self.embed_dim) + + @check_model_inputs + def forward( + self, seq_patches: torch.Tensor, spatial_shapes: torch.Tensor + ) -> torch.Tensor: + # Rebatch packed variable-resolution patches to resize per-image position embeddings + # and track lengths for varlen attention metadata. + packed_pixel_values, seq_lengths = self._pack_to_batch( + seq_patches, spatial_shapes + ) if packed_pixel_values is None: return seq_patches.new_zeros((0, self.embed_dim)) - embeddings = super().forward(packed_pixel_values, spatial_shapes) + target_dtype = self.patch_embedding.weight.dtype + patch_embeds = self.patch_embedding(packed_pixel_values.to(dtype=target_dtype)) + + positional_embeddings = self.position_embedding.weight.reshape( + self.position_embedding_size, + self.position_embedding_size, + -1, + ) + resized_positional_embeddings = self.resize_positional_embeddings( + positional_embeddings, + spatial_shapes, + max_length=packed_pixel_values.shape[1], + ) + + embeddings = patch_embeds + resized_positional_embeddings return self._unpack_from_batch(embeddings, seq_lengths) def _pack_to_batch( self, seq_patches: torch.Tensor, spatial_shapes: torch.Tensor, - ) -> tuple[torch.Tensor | None, torch.Tensor]: - if seq_patches.ndim != 2: - raise ValueError("`seq_patches` is expected to be 2D (total_patches, patch_dim).") - if spatial_shapes.ndim != 2 or spatial_shapes.size(-1) != 2: - raise ValueError("`spatial_shapes` must have shape (num_images, 2) with (height_tokens, width_tokens).") - - seq_lengths = spatial_shapes.long().prod(dim=-1) - total_patches = int(seq_lengths.sum().item()) - if total_patches != seq_patches.size(0): - raise ValueError( - "Mismatch between packed patches and spatial shapes: got " - f"{seq_patches.size(0)} patches but spatial shapes imply {total_patches}." - ) + ) -> tuple[Optional[torch.Tensor], torch.Tensor]: + """Rebatch a packed patch sequence using per-image grids to align embeddings. + + Args: + seq_patches: Packed patches of shape (total_patches, patch_dim). + spatial_shapes: Per-image patch grids of shape (num_images, 2) as (H_tokens, W_tokens). - batch_size = spatial_shapes.size(0) + Returns: + (packed_pixel_values, seq_lengths) where: + - packed_pixel_values: (batch, max_len, patch_dim) padded with zeros, or None if batch_size == 0 + - seq_lengths: (batch,) lengths for each image + """ + seq_lengths = spatial_shapes.long().prod(dim=-1) # (B,) + batch_size = int(seq_lengths.numel()) if batch_size == 0: return None, seq_lengths - max_length = int(seq_lengths.max().item()) - patch_dim = seq_patches.size(-1) - device = seq_patches.device - - packed_pixel_values = seq_patches.new_zeros((batch_size, max_length, patch_dim), device=device) - - start = 0 - for batch_idx, length in enumerate(seq_lengths.tolist()): - if length == 0: - continue - end = start + length - packed_pixel_values[batch_idx, :length] = seq_patches[start:end] - start = end - + # Split the packed sequence into per-image chunks, then pad to a batch + lengths_list = seq_lengths.tolist() + chunks = seq_patches.split(lengths_list, dim=0) + packed_pixel_values = nn.utils.rnn.pad_sequence( + chunks, batch_first=True + ) # zero-padded by default return packed_pixel_values, seq_lengths - def _unpack_from_batch(self, embeddings: torch.Tensor, seq_lengths: torch.Tensor) -> torch.Tensor: - output_chunks: list[torch.Tensor] = [] - for batch_idx, length in enumerate(seq_lengths.tolist()): - if length == 0: - continue - output_chunks.append(embeddings[batch_idx, :length]) - - if not output_chunks: - return embeddings.new_zeros((0, embeddings.size(-1))) - - return torch.cat(output_chunks, dim=0) + def _unpack_from_batch( + self, embeddings: torch.Tensor, seq_lengths: torch.Tensor + ) -> torch.Tensor: + """Flatten a padded batch back to packed sequence order using `seq_lengths`.""" + lengths = seq_lengths.to(device=embeddings.device).tolist() + chunks = [embeddings[i, :l] for i, l in enumerate(lengths) if l > 0] + return torch.cat(chunks, dim=0) class IsaacVisionAttention(Siglip2Attention): """Custom attention that supports variable-length sequences with flash attention.""" - ATTENTION_KEY_MAP: dict[str, str] = { - "flash_attention_2": "isaac_flash_attention_2", - "flash_attention_3": "isaac_flash_attention_3", - "isaac_flash_attention_2": "isaac_flash_attention_2", - "isaac_flash_attention_3": "isaac_flash_attention_3", - "sdpa": "isaac_sdpa", - "isaac_sdpa": "isaac_sdpa", - "eager": "isaac_eager", - "isaac_eager": "isaac_eager", - } - - def __init__(self, vision_config): - super().__init__(vision_config) - self.vision_config = vision_config - self._variable_length_metadata = None - - def _variable_length_context(self, *, cu_seqlens=None, max_seqlen=None): - """Store packed-sequence metadata for the next forward call.""" - self._variable_length_metadata = (cu_seqlens, max_seqlen) - - def _consume_variable_length_metadata(self): - if self._variable_length_metadata is None: - return None, None - cu_seqlens, max_seqlen = self._variable_length_metadata - self._variable_length_metadata = None - return cu_seqlens, max_seqlen - - def forward(self, hidden_states, attention_mask=None, **kwargs): - cu_seqlens = kwargs.pop("cu_seqlens", None) - max_seqlen = kwargs.pop("max_seqlen", None) - kwargs.pop("output_attentions", None) - if kwargs: - unexpected = ", ".join(sorted(kwargs)) - raise TypeError(f"Unexpected kwargs for IsaacVisionAttention.forward: {unexpected}") - cached_cu, cached_max = self._consume_variable_length_metadata() - if cu_seqlens is None: - cu_seqlens = cached_cu - if max_seqlen is None: - max_seqlen = cached_max - - # Expect packed sequences with batch_size == 1 - batch_size, L, _ = hidden_states.shape - if batch_size != 1: - raise ValueError("packed variable-length attention expects batch_size=1") - x = hidden_states[0] # (L, E) - - H = self.num_heads - D = self.head_dim - p_drop = self.dropout if self.training else 0.0 - - # Project and reshape to (L, H, D) - q = self.q_proj(x).view(L, H, D) - k = self.k_proj(x).view(L, H, D) - v = self.v_proj(x).view(L, H, D) - - attn_impl = getattr(self.vision_config, "_attn_implementation", "flash_attention_3") - - attn_mask = ensure_document_attention_mask( - attention_mask, - cu_seqlens, - L, - q.dtype, - q.device, - ) - - resolved_key = self.ATTENTION_KEY_MAP.get(attn_impl) - attention_fn = ALL_ATTENTION_FUNCTIONS.get(resolved_key) if resolved_key is not None else None - if attention_fn is None: - raise ValueError(f"Attention implementation {attn_impl} not found.") - - query_states = q.transpose(0, 1).unsqueeze(0) - key_states = k.transpose(0, 1).unsqueeze(0) - value_states = v.transpose(0, 1).unsqueeze(0) - - attention_kwargs: dict[str, Any] = { - "dropout": p_drop, - "scaling": self.scale, - "is_causal": False, - } - if cu_seqlens is not None: - attention_kwargs["cu_seq_lens_q"] = cu_seqlens - attention_kwargs["cu_seq_lens_k"] = cu_seqlens - if max_seqlen is not None: - attention_kwargs["max_length_q"] = max_seqlen - attention_kwargs["max_length_k"] = max_seqlen - - attn_output, _ = attention_fn( - self, - query_states, - key_states, - value_states, - attn_mask, - **attention_kwargs, - ) - - y_lhd = attn_output.squeeze(0).permute(1, 0, 2).contiguous() - - # Merge heads and project - y = self.out_proj(y_lhd.reshape(L, self.embed_dim)) - return y.unsqueeze(0), None # (1, L, E) - - -class IsaacVisionEncoderLayer(HFSiglip2EncoderLayer): - """Isaac vision encoder layer with variable-length attention.""" - - def __init__(self, vision_config: IsaacVisionConfig): - super().__init__(vision_config) - self.self_attn = IsaacVisionAttention(vision_config) - def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, + output_attentions: bool = False, cu_seqlens: Optional[torch.Tensor] = None, max_seqlen: Optional[int] = None, - output_attentions: bool = False, **kwargs, ): - if cu_seqlens is not None or max_seqlen is not None: - self.self_attn._variable_length_context( - cu_seqlens=cu_seqlens, - max_seqlen=max_seqlen, + kwargs.pop("output_hidden_states", None) + kwargs.pop("return_dict", None) + + batch_size, seq_length, embed_dim = hidden_states.shape + queries = self.q_proj(hidden_states) + keys = self.k_proj(hidden_states) + values = self.v_proj(hidden_states) + + queries = queries.view( + batch_size, seq_length, self.num_heads, self.head_dim + ).transpose(1, 2) + keys = keys.view( + batch_size, seq_length, self.num_heads, self.head_dim + ).transpose(1, 2) + values = values.view( + batch_size, seq_length, self.num_heads, self.head_dim + ).transpose(1, 2) + + attn_impl = self.config._attn_implementation + attention_interface: Callable = ALL_ATTENTION_FUNCTIONS["sdpa"] + if attn_impl != "sdpa": + attention_interface = ALL_ATTENTION_FUNCTIONS[attn_impl] + + attention_kwargs: dict[str, Any] = { + "is_causal": False, + "scaling": self.scale, + } + + supports_varlen = cu_seqlens is not None and attn_impl in { + "flash_attention_2", + "flash_attention_3", + "flex_attention", + "paged|flash_attention_2", + "paged|flash_attention_3", + } + if supports_varlen: + if max_seqlen is not None: + max_q = max_k = int(max_seqlen) + elif cu_seqlens.numel() >= 2: + lengths = cu_seqlens[1:] - cu_seqlens[:-1] + max_q = max_k = lengths.max() if lengths.numel() > 0 else seq_length + else: + max_q = max_k = seq_length + + attention_kwargs.update( + { + "cu_seq_lens_q": cu_seqlens, + "cu_seq_lens_k": cu_seqlens, + "max_length_q": max_q, + "max_length_k": max_k, + } ) - attention_mask = ensure_document_attention_mask( + attn_output, attn_weights = attention_interface( + self, + queries, + keys, + values, attention_mask, - cu_seqlens, - hidden_states.size(1), - hidden_states.dtype, - hidden_states.device, + **attention_kwargs, ) + attn_output = attn_output.reshape( + batch_size, seq_length, embed_dim + ).contiguous() + attn_output = self.out_proj(attn_output) - return super().forward( - hidden_states, - attention_mask=attention_mask, - output_attentions=output_attentions, - **kwargs, - ) + return attn_output, attn_weights -class IsaacVisionEncoder(HFSiglip2Encoder): - """Encoder using Isaac encoder layers with variable-length attention support.""" +class IsaacVisionEncoderLayer(Siglip2EncoderLayer): + """Isaac vision encoder layer with variable-length attention.""" def __init__(self, config: IsaacVisionConfig): super().__init__(config) - self.layers = nn.ModuleList([IsaacVisionEncoderLayer(config) for _ in range(config.num_hidden_layers)]) - - def __variable_length_context(self, cu_seqlens, max_seqlen) -> None: - if cu_seqlens is None and max_seqlen is None: - return - - for layer in self.layers: - if isinstance(layer, IsaacVisionEncoderLayer): - layer.self_attn._variable_length_context( - cu_seqlens=cu_seqlens, - max_seqlen=max_seqlen, - ) + self.self_attn = IsaacVisionAttention(config) - @can_return_tuple def forward( self, - inputs_embeds, + hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, cu_seqlens: Optional[torch.Tensor] = None, max_seqlen: Optional[int] = None, - output_attentions: Optional[bool] = None, - output_hidden_states: Optional[bool] = None, - return_dict: Optional[bool] = None, + output_attentions: bool = False, + **kwargs: Unpack[TransformersKwargs], ): - self.__variable_length_context(cu_seqlens, max_seqlen) - - attention_mask = ensure_document_attention_mask( - attention_mask, - cu_seqlens, - inputs_embeds.size(1), - inputs_embeds.dtype, - inputs_embeds.device, - ) - - return super().forward( - inputs_embeds, + r""" + cu_seqlens (`torch.Tensor`, *optional*): + Prefix-sum tensor whose length equals the number of documents + 1. The difference between successive + entries gives each document's token count and enables block-diagonal attention masking for packed batches. + max_seqlen (`int`, *optional*): + Maximum document length referenced by `cu_seqlens`. Passed to FlashAttention so it can size temporary + buffers for packed variable-length attention. + """ + # Run attention directly so variable-length metadata reaches FlashAttention. + residual = hidden_states + hidden_states = self.layer_norm1(hidden_states) + attn_output, _ = self.self_attn( + hidden_states, attention_mask=attention_mask, - output_attentions=output_attentions, - ) - - -def _isaac_flash_attention_forward( - module: nn.Module, - query: torch.Tensor, - key: torch.Tensor, - value: torch.Tensor, - attention_mask: Optional[torch.Tensor], - dropout: float = 0.0, - scaling: Optional[float] = None, - is_causal: bool = False, - **kwargs, -) -> tuple[torch.Tensor, Optional[torch.Tensor]]: - base_fn = _ORIGINAL_ATTENTION_FUNCTIONS.get("flash_attention_2") - if not isinstance(module, IsaacVisionAttention) or base_fn is None: - if base_fn is None: - raise ValueError("Base flash attention function unavailable for fallback.") - return base_fn( - module, - query, - key, - value, - attention_mask, - dropout=dropout, - scaling=scaling, - is_causal=is_causal, + cu_seqlens=cu_seqlens, + max_seqlen=max_seqlen, **kwargs, ) + hidden_states = residual + attn_output - if query.dim() != 4 or query.size(0) != 1: - raise ValueError("IsaacVisionAttention expects packed sequences with batch size 1 when using packed attention.") - - _, num_heads, seq_len, head_dim = query.shape - q_lhd = query.transpose(1, 2).reshape(seq_len, num_heads, head_dim) - k_lhd = key.transpose(1, 2).reshape(seq_len, num_heads, head_dim) - v_lhd = value.transpose(1, 2).reshape(seq_len, num_heads, head_dim) - - cum_seq_q = kwargs.get("cu_seq_lens_q") - cum_seq_k = kwargs.get("cu_seq_lens_k", cum_seq_q) - max_seqlen = kwargs.get("max_length_q") - - effective_dropout = dropout if dropout is not None else (module.dropout if module.training else 0.0) - effective_scaling = module.scale if scaling is None else scaling - - attn_mask = attention_mask - if attn_mask is None: - attn_mask = build_document_attention_mask( - cu_seqlens=cum_seq_q, - total_tokens=seq_len, - dtype=q_lhd.dtype, - device=q_lhd.device, - ) - - attn_output_lhd, attn_weights = flash_attention_document_mask_forward( - module, - q_lhd, - k_lhd, - v_lhd, - attention_mask=attn_mask, - dropout=effective_dropout, - scaling=effective_scaling, - cum_seq_q=cum_seq_q, - cum_seq_k=cum_seq_k, - max_seqlen=max_seqlen, - is_causal=is_causal, - ) + residual = hidden_states + hidden_states = self.layer_norm2(hidden_states) + hidden_states = self.mlp(hidden_states) + hidden_states = residual + hidden_states - attn_output = attn_output_lhd.permute(1, 0, 2).unsqueeze(0) - return attn_output, attn_weights - - -def _isaac_sdpa_forward( - module: nn.Module, - query: torch.Tensor, - key: torch.Tensor, - value: torch.Tensor, - attention_mask: Optional[torch.Tensor], - dropout: float = 0.0, - scaling: Optional[float] = None, - is_causal: bool = False, - **kwargs, -) -> tuple[torch.Tensor, Optional[torch.Tensor]]: - base_fn = _ORIGINAL_ATTENTION_FUNCTIONS.get("sdpa") - if not isinstance(module, IsaacVisionAttention) or base_fn is None: - if base_fn is None: - raise ValueError("Base SDPA function unavailable for fallback.") - return base_fn( - module, - query, - key, - value, - attention_mask, - dropout=dropout, - scaling=scaling, - is_causal=is_causal, - **kwargs, - ) + return hidden_states - if query.dim() != 4 or query.size(0) != 1: - raise ValueError("IsaacVisionAttention expects packed sequences with batch size 1 when using packed attention.") - - _, num_heads, seq_len, head_dim = query.shape - q_lhd = query.transpose(1, 2).reshape(seq_len, num_heads, head_dim) - k_lhd = key.transpose(1, 2).reshape(seq_len, num_heads, head_dim) - v_lhd = value.transpose(1, 2).reshape(seq_len, num_heads, head_dim) - - cum_seq = kwargs.get("cu_seq_lens_q") - effective_dropout = dropout if dropout is not None else (module.dropout if module.training else 0.0) - effective_scaling = module.scale if scaling is None else scaling - - attn_mask = attention_mask - if attn_mask is None: - attn_mask = build_document_attention_mask( - cu_seqlens=cum_seq, - total_tokens=seq_len, - dtype=q_lhd.dtype, - device=q_lhd.device, - ) - attn_output_lhd = sdpa_document_mask_forward( - q_lhd, - k_lhd, - v_lhd, - dropout=effective_dropout, - scaling=effective_scaling, - attention_mask=attn_mask, - cu_seqlens=cum_seq, - ) +class IsaacVisionEncoder(Siglip2Encoder): + """Encoder using Isaac encoder layers with variable-length attention support.""" - attn_output = attn_output_lhd.permute(1, 0, 2).unsqueeze(0) - return attn_output, None - - -def _isaac_eager_forward( - module: nn.Module, - query: torch.Tensor, - key: torch.Tensor, - value: torch.Tensor, - attention_mask: Optional[torch.Tensor], - dropout: float = 0.0, - scaling: Optional[float] = None, - is_causal: bool = False, - **kwargs, -) -> tuple[torch.Tensor, Optional[torch.Tensor]]: - base_fn = _ORIGINAL_ATTENTION_FUNCTIONS.get("eager") - if not isinstance(module, IsaacVisionAttention) or base_fn is None: - if base_fn is None: - raise ValueError("Base eager attention function unavailable for fallback.") - return base_fn( - module, - query, - key, - value, - attention_mask, - dropout=dropout, - scaling=scaling, - is_causal=is_causal, - **kwargs, + def __init__(self, config: IsaacVisionConfig): + super().__init__(config) + self.layers = nn.ModuleList( + [IsaacVisionEncoderLayer(config) for _ in range(config.num_hidden_layers)] ) - if query.dim() != 4 or query.size(0) != 1: - raise ValueError("IsaacVisionAttention expects packed sequences with batch size 1 when using packed attention.") - - _, num_heads, seq_len, head_dim = query.shape - q_lhd = query.transpose(1, 2).reshape(seq_len, num_heads, head_dim) - k_lhd = key.transpose(1, 2).reshape(seq_len, num_heads, head_dim) - v_lhd = value.transpose(1, 2).reshape(seq_len, num_heads, head_dim) - - effective_scaling = module.scale if scaling is None else scaling - attn_weights = torch.matmul(q_lhd, k_lhd.transpose(1, 2)) * effective_scaling - - if attention_mask is not None: - mask = attention_mask - if mask.dim() == 4: - mask = mask.squeeze(0).squeeze(0) - attn_weights = attn_weights + mask - - attn_weights = torch.softmax(attn_weights, dim=-1) - if dropout and module.training: - attn_weights = F.dropout(attn_weights, p=dropout, training=True) - - attn_output_lhd = torch.matmul(attn_weights, v_lhd) - attn_output = attn_output_lhd.permute(1, 0, 2).unsqueeze(0) - return attn_output, attn_weights - - -ALL_ATTENTION_FUNCTIONS.register("isaac_flash_attention_2", _isaac_flash_attention_forward) -ALL_ATTENTION_FUNCTIONS.register("isaac_flash_attention_3", _isaac_flash_attention_forward) -ALL_ATTENTION_FUNCTIONS.register("isaac_sdpa", _isaac_sdpa_forward) -ALL_ATTENTION_FUNCTIONS.register("isaac_eager", _isaac_eager_forward) - def create_pixel_shuffle_index_map( seq_sizes: torch.Tensor, token_grids: torch.Tensor, scale_factor: int = 1, - device: torch.device | None = None, + device: Optional[torch.device] = None, ) -> torch.Tensor: """ Build a gather-index map that tells us, for every *output* token after @@ -1053,47 +633,32 @@ def create_pixel_shuffle_index_map( packed sequence for the j-th sub-patch that forms the i-th output token. """ - if device is None: - device = seq_sizes.device - - scale_factor = int(scale_factor) - if scale_factor < 2: - raise ValueError("`scale_factor` must be ≥ 2") - - # Safety: all spatial dims must be divisible by the scale factor - # Cannot run under torch compile fullgraph mode hence if not is_torchdynamo_compiling(): - if not ((token_grids[:, 0] % scale_factor == 0).all() and (token_grids[:, 1] % scale_factor == 0).all()): + if (token_grids % scale_factor).any(): raise AssertionError( - "Every (H,W) in `token_grids` must be divisible by " - f"scale_factor={scale_factor}, got {token_grids.tolist()}" + f"Every (H,W) in token_grids must be divisible by scale_factor={scale_factor}, got {token_grids.tolist()}" ) gather_chunks: list[torch.Tensor] = [] tok_offset = 0 + for seq_len, (h, w) in zip(seq_sizes.tolist(), token_grids.tolist()): + # Flat indices for this image's packed segment + grid = ( + torch.arange(seq_len, device=device, dtype=torch.int64).view(h, w) + + tok_offset + ) - for seq_len, (h, w) in zip(seq_sizes.tolist(), token_grids.tolist(), strict=False): - # Build the (H, W) grid of flat indices for this image - grid = torch.arange(seq_len, device=device, dtype=torch.int64) + tok_offset - grid = grid.view(h, w) # (H, W) - - # -------- identical ordering to your fixed-res routine -------- - # Step 1: split width into blocks of scale_factor - grid = grid.view(h, w // scale_factor, scale_factor) # (H, W/scale_factor, scale_factor) - # Step 2: now split height into blocks of scale_factor - grid = grid.view(h // scale_factor, scale_factor, w // scale_factor, scale_factor) - # (H/scale_factor, scale_factor, W/scale_factor, scale_factor) - # Step 3: final permutation to (H/scale_factor, W/scale_factor, scale_factor, scale_factor) - grid = grid.permute(0, 2, 1, 3).contiguous() # (H/scale_factor, W/scale_factor, scale_factor, scale_factor) - # Step 4: each (scale_factor, scale_factor) block forms one output token - gather_chunks.append(grid.reshape(-1, scale_factor * scale_factor)) - # (H*W / scale_factor**2, scale_factor**2) + # Block into (H/s, W/s) groups; each group contributes s*s indices + grid = ( + grid.view(h // scale_factor, scale_factor, w // scale_factor, scale_factor) + .permute(0, 2, 1, 3) + .contiguous() + ) + gather_chunks.append(grid.view(-1, scale_factor * scale_factor)) tok_offset += seq_len - # Concatenate over all images in the packed batch - gather_idx = torch.cat(gather_chunks, dim=0) # (Σ_i HᵢWᵢ/scale_factor**2, scale_factor**2) - return gather_idx + return torch.cat(gather_chunks, dim=0) def pixel_shuffle_varlen( @@ -1122,47 +687,67 @@ def pixel_shuffle_varlen( Raises: ValueError: If more than one batch item is provided. """ - keep_batch_dim = x.dim() == 3 - if keep_batch_dim: + return_with_batch_dim = x.dim() == 3 + if return_with_batch_dim: if x.size(0) != 1: - raise AssertionError("Packed sequence is expected to have batch_size == 1") - x_ = x.squeeze(0) # (seq, embed) + raise ValueError( + f"Packed vision sequences expect a singleton batch dimension; received batch_size={x.size(0)}." + ) + embeddings = x.squeeze(0) # (seq, embed) else: - x_ = x # (seq, embed) + embeddings = x # (seq, embed) - embed_dim = x_.size(-1) + embed_dim = embeddings.size(-1) scale_factor = int(scale_factor) # Calculate seq_sizes from token_grids seq_sizes = torch.prod(token_grids, dim=-1) - # Build index map and gather in one go + # Build a single gather index so pixel shuffle works on the packed stream + # without unpacking per-image grids. gather_idx = create_pixel_shuffle_index_map( seq_sizes=seq_sizes, token_grids=token_grids, scale_factor=scale_factor, - device=x_.device, + device=embeddings.device, ) # (new_seq, scale_factor**2) # Gather → (new_seq, scale_factor**2, embed_dim) - gathered = x_[gather_idx] # fancy indexing keeps gradient + gathered = embeddings[gather_idx] # fancy indexing keeps gradient # Merge the scale_factor**2 group dimension into channels to finish the shuffle out = gathered.reshape(gathered.size(0), embed_dim * scale_factor * scale_factor) # Restore batch dimension if needed - if keep_batch_dim: + if return_with_batch_dim: out = out.unsqueeze(0) return out class IsaacVisionTransformer(nn.Module): + """Vision tower that packs variable-resolution patches, applies varlen attention, and pixel-shuffles outputs. + + Args: + config (IsaacVisionConfig): Vision configuration with pixel-shuffle and patching parameters. + + Inputs: + packed_seq_patches (Tuple[Tensor, Tensor]): ``(patches, token_grids)`` where ``patches`` is a packed + patch sequence and ``token_grids`` holds per-image (H_tokens, W_tokens). + + Returns: + torch.Tensor: Vision embeddings after encoder + pixel shuffle, shaped ``(seq_len, hidden_size * s^2)``. + """ + + _supports_sdpa = True + def __init__(self, config: IsaacVisionConfig): super().__init__() self.config = config self.embeddings = IsaacVisionEmbeddings(config) self.encoder = IsaacVisionEncoder(config) - self.post_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) + self.post_layernorm = nn.LayerNorm( + config.hidden_size, eps=config.layer_norm_eps + ) self.pixel_shuffle_scale_factor = config.pixel_shuffle_scale_factor def forward(self, packed_seq_patches: tuple[torch.Tensor, torch.Tensor]): @@ -1172,31 +757,33 @@ class IsaacVisionTransformer(nn.Module): # Get embeddings from packed sequence hidden_states = self.embeddings(seq_patches, token_grids) - # Add a pseudo batch dimension for the encoder + # Add a pseudo batch dimension so we can reuse the batch-first encoder stack + # while still driving per-image cu_seqlens through the varlen attention path. hidden_states = hidden_states.unsqueeze(0) # Generate cumulative sequence lengths for variable-length attention - cu_seqlens = torch.zeros(seq_sizes.size(0) + 1, dtype=torch.int32, device=hidden_states.device) - cu_seqlens[1:] = seq_sizes.cumsum(0) - max_seqlen = int(seq_sizes.max().item()) if seq_sizes.numel() > 0 else 0 + cu_seqlens = F.pad(seq_sizes.cumsum(0).to(torch.int32), (1, 0)) + + attention_mask = create_document_attention_mask( + self.config, hidden_states, cu_seqlens + ) # Pass through encoder with variable-length attention parameters encoder_outputs = self.encoder( inputs_embeds=hidden_states, + attention_mask=attention_mask, cu_seqlens=cu_seqlens, - max_seqlen=max_seqlen, ) hidden_states = encoder_outputs.last_hidden_state # Apply final layer normalization hidden_states = self.post_layernorm(hidden_states) - if self.pixel_shuffle_scale_factor > 1: - hidden_states = pixel_shuffle_varlen( - x=hidden_states, - token_grids=token_grids, - scale_factor=self.pixel_shuffle_scale_factor, - ) + hidden_states = pixel_shuffle_varlen( + x=hidden_states, + token_grids=token_grids, + scale_factor=self.pixel_shuffle_scale_factor, + ) # Remove the pseudo batch dimension we added earlier hidden_states = hidden_states.squeeze(0) @@ -1204,6 +791,21 @@ class IsaacVisionTransformer(nn.Module): return hidden_states +class IsaacVisionEmbedding(nn.Module): + _supports_sdpa = True + + def __init__(self, config: IsaacConfig): + super().__init__() + vision_cfg = config.vision_config + + self.vision_tower = IsaacVisionTransformer(vision_cfg) + self.multimodal_projector = IsaacMultiModalProjector(config) + + def forward(self, vision_tokens: tuple[torch.Tensor, torch.Tensor]) -> torch.Tensor: + hidden_states = self.vision_tower(vision_tokens) + return self.multimodal_projector(hidden_states) + + def get_scaled_image_size( scale: float, original_size: int, @@ -1222,7 +824,7 @@ def get_image_size_for_max_num_patches( image_width: int, patch_size: int, max_num_patches: int, - min_num_patches: int | None = None, + min_num_patches: Optional[int] = None, eps: float = 1e-5, pixel_shuffle_scale: int = 1, ) -> tuple[int, int]: @@ -1259,20 +861,29 @@ def get_image_size_for_max_num_patches( num_patches = (adjusted_height / patch_size) * (adjusted_width / patch_size) if min_num_patches is not None and num_patches < min_num_patches: - # Scale up + # Scale up via binary search to satisfy the minimum patch budget while + # preserving divisibility by patch_size * pixel_shuffle_scale. scale_min, scale_max = 1.0, 100.0 while (scale_max - scale_min) >= eps: scale = (scale_min + scale_max) / 2 - target_height = get_scaled_image_size(scale, image_height, patch_size, pixel_shuffle_scale) - target_width = get_scaled_image_size(scale, image_width, patch_size, pixel_shuffle_scale) + target_height = get_scaled_image_size( + scale, image_height, patch_size, pixel_shuffle_scale + ) + target_width = get_scaled_image_size( + scale, image_width, patch_size, pixel_shuffle_scale + ) num_patches = (target_height / patch_size) * (target_width / patch_size) if num_patches >= min_num_patches: scale_max = scale else: scale_min = scale scale = scale_max - target_height = get_scaled_image_size(scale, image_height, patch_size, pixel_shuffle_scale) - target_width = get_scaled_image_size(scale, image_width, patch_size, pixel_shuffle_scale) + target_height = get_scaled_image_size( + scale, image_height, patch_size, pixel_shuffle_scale + ) + target_width = get_scaled_image_size( + scale, image_width, patch_size, pixel_shuffle_scale + ) return target_height, target_width elif num_patches <= max_num_patches: return adjusted_height, adjusted_width @@ -1281,46 +892,33 @@ def get_image_size_for_max_num_patches( scale_min, scale_max = eps / 10, 1.0 while (scale_max - scale_min) >= eps: scale = (scale_min + scale_max) / 2 - target_height = get_scaled_image_size(scale, image_height, patch_size, pixel_shuffle_scale) - target_width = get_scaled_image_size(scale, image_width, patch_size, pixel_shuffle_scale) + target_height = get_scaled_image_size( + scale, image_height, patch_size, pixel_shuffle_scale + ) + target_width = get_scaled_image_size( + scale, image_width, patch_size, pixel_shuffle_scale + ) num_patches = (target_height / patch_size) * (target_width / patch_size) if num_patches <= max_num_patches: scale_min = scale else: scale_max = scale scale = scale_min - target_height = get_scaled_image_size(scale, image_height, patch_size, pixel_shuffle_scale) - target_width = get_scaled_image_size(scale, image_width, patch_size, pixel_shuffle_scale) + target_height = get_scaled_image_size( + scale, image_height, patch_size, pixel_shuffle_scale + ) + target_width = get_scaled_image_size( + scale, image_width, patch_size, pixel_shuffle_scale + ) return target_height, target_width -def patchify_vision(image: torch.Tensor, patch_size: int) -> torch.Tensor: - r"""Convert normalized images into flattened ViT-style patches. - - Args: - image (`torch.Tensor`): - Tensor of shape `(num_images, height, width, channels)`. - patch_size (`int`): - Edge length of the square patches - - Returns: - `torch.Tensor`: - Patch tensor where each position stores the flattened pixels belonging to that patch. +class IsaacConfig(PretrainedConfig): + """Configuration class for Isaac multimodal model. - Raises: - ValueError: If `height` or `width` is not divisible by `patch_size`. + This configuration corresponds to checkpoints such as + [Perceptron/isaac-base](https://huggingface.co/Perceptron/isaac-base). """ - num_images, height, width, channels = image.shape - if height % patch_size or width % patch_size: - raise ValueError(f"Dimensions of images {image.shape} are not divisible by patch_size={patch_size}.") - patches = image.reshape(num_images, height // patch_size, patch_size, width // patch_size, patch_size, channels) - patches = patches.permute(0, 1, 3, 2, 4, 5) - patches = patches.reshape(num_images, height // patch_size, width // patch_size, channels * patch_size * patch_size) - return patches - - -class IsaacConfig(Qwen3Config): - """Configuration class for Isaac multimodal model.""" model_type = "isaac" sub_configs = {"vision_config": IsaacVisionConfig, "text_config": Qwen3Config} @@ -1328,41 +926,65 @@ class IsaacConfig(Qwen3Config): def __init__( self, - vision_config: IsaacVisionConfig | None = None, - text_config: Qwen3Config | dict | None = None, + vision_config: Optional[IsaacVisionConfig] = None, + text_config: Optional[Union[Qwen3Config, dict]] = None, vision_rescale_factor: float = 1 / 255, max_sequence_length: int = 16384, vision_token: str = "", **kwargs, ): - self._rope_scaling: dict[str, Any] | None = None - resolved_text_config = kwargs.pop("text_config", text_config) - if isinstance(resolved_text_config, Qwen3Config): - text_config_kwargs = copy.deepcopy(resolved_text_config.to_dict()) - elif isinstance(resolved_text_config, dict): - text_config_kwargs = copy.deepcopy(resolved_text_config) - elif resolved_text_config is None: - text_config_kwargs = {} - else: - raise TypeError("`text_config` must be a mapping or `Qwen3Config` instance when provided.") + attn_implementation = kwargs.get("attn_implementation") - text_config_kwargs.update(kwargs) + if isinstance(text_config, dict): + self.text_config = self.sub_configs["text_config"](**text_config) + elif isinstance(text_config, Qwen3Config): + self.text_config = text_config + elif text_config is None: + self.text_config = self.sub_configs["text_config"]() - super().__init__(**text_config_kwargs) - self.text_config = Qwen3Config(**text_config_kwargs) - if self._rope_scaling is None: - self._rope_scaling = getattr(self.text_config, "rope_scaling", None) - else: - self.text_config.rope_scaling = self._rope_scaling + # Seed RoPE parameters before base init so the shared mixin can standardize/validate them. + self.rope_parameters = getattr(self.text_config, "rope_parameters", None) + self.layer_types = getattr(self.text_config, "layer_types", None) - # Handle vision config - either dict or IsaacVisionConfig instance - if isinstance(vision_config, dict): + super().__init__(**kwargs) + + # Keep rope parameters aligned between the composite and text sub-configs. + self.text_config.rope_parameters = self.rope_parameters + + # Mirror frequently accessed Qwen3 attributes at the composite config level + self.vocab_size = self.text_config.vocab_size + self.hidden_size = self.text_config.hidden_size + self.num_hidden_layers = self.text_config.num_hidden_layers + self.num_attention_heads = self.text_config.num_attention_heads + self.head_dim = self.text_config.head_dim + self.hidden_act = self.text_config.hidden_act + self.use_cache = self.text_config.use_cache + self.rope_theta = self.rope_parameters["rope_theta"] + + self.layer_types = getattr(self.text_config, "layer_types", None) + layer_type_validation(self.layer_types, self.num_hidden_layers) + + # Handle vision config - either dict or IsaacVisionConfig instance + if isinstance(vision_config, dict): self.vision_config = self.sub_configs["vision_config"](**vision_config) elif isinstance(vision_config, IsaacVisionConfig): self.vision_config = vision_config elif vision_config is None: self.vision_config = self.sub_configs["vision_config"]() + # Propagate user-requested attention backend to the vision sub-config when provided. + if attn_implementation is not None: + if isinstance(attn_implementation, dict): + vision_attn = attn_implementation.get( + "vision_config", attn_implementation.get("", None) + ) + else: + vision_attn = attn_implementation + if vision_attn is not None: + self.vision_config._attn_implementation = vision_attn + + if getattr(self, "_attn_implementation", None) is None: + self._attn_implementation = "sdpa" # Vision normalization parameters self.vision_rescale_factor = float(vision_rescale_factor) @@ -1370,293 +992,447 @@ class IsaacConfig(Qwen3Config): self.max_sequence_length = max_sequence_length self.vision_token = vision_token - def get_text_config(self, *_, **kwargs) -> Qwen3Config: - # Accept optional decoder/encoder flags to align with HF composite configs - kwargs.pop("decoder", None) - kwargs.pop("encoder", None) - return self.text_config - - @property - def rope_scaling(self): + def to_dict(self): + output = super().to_dict() + # Ensure nested configs round-trip through dict serialization if hasattr(self, "text_config") and self.text_config is not None: - return getattr(self.text_config, "rope_scaling", None) - return self._rope_scaling - - @rope_scaling.setter - def rope_scaling(self, value): - self._rope_scaling = value - if hasattr(self, "text_config") and self.text_config is not None: - self.text_config.rope_scaling = value - - @property - def vision_attn_implementation(self) -> str | None: - value = getattr(self.vision_config, "_attn_implementation", None) - if value is None: - value = getattr(self.vision_config, "attn_implementation", None) - return value - - @vision_attn_implementation.setter - def vision_attn_implementation(self, value: str | None) -> None: - self.vision_config._attn_implementation = value - if value is not None: - self.vision_config.attn_implementation = value - elif hasattr(self.vision_config, "attn_implementation"): - delattr(self.vision_config, "attn_implementation") - - -# ============================================================================ -# Processor Components -# ============================================================================ + output["text_config"] = self.text_config.to_dict() + if hasattr(self, "vision_config") and self.vision_config is not None: + output["vision_config"] = self.vision_config.to_dict() + return output -def create_text_event(tokenizer: AutoTokenizer, text: str, time: float = 0.0) -> Event: - r"""Wrap a text into an `Event` compatible with the multimodal TensorStream. +class IsaacProcessor(ProcessorMixin): + """Processor that pairs the Isaac image processor with the Qwen2 tokenizer. Args: - tokenizer (`AutoTokenizer`): - Tokenizer used to convert text into model vocabulary ids. - text (`str`): - Plain-text fragment to encode. - time (`float`, *optional*, defaults to 0.0): - Timeline coordinate associated with the event. Both start and end times use the same value because text - segments are instantaneous in the scheduler. + image_processor: Vision preprocessor (fast) used for patch extraction. + tokenizer: Qwen2 tokenizer instance. + vision_token (str, optional): Placeholder token marking image locations. Defaults to "". + max_sequence_length (int, optional): Maximum combined text+vision tokens kept. Defaults to 16384. + rescale_factor (float, optional): Image rescale factor; defaults to 1/255. + config (IsaacConfig | dict, optional): If provided, overrides processor defaults from the model config. Returns: - `Event`: Event carrying a `(num_tokens, 1)` tensor of token ids with matching - metadata so that downstream processors can compute modality-specific embeddings. + BatchFeature: Contains ``input_ids`` and ``packed_inputs`` (patch tensors, grids, offsets, lengths, modality, positions). """ - tokens = tokenizer.encode(text, add_special_tokens=False, return_tensors="pt").squeeze(0) - - # Calculate dimensions for the event - num_tokens = len(tokens) - dims_virtual = [num_tokens, 1] # [sequence_length, 1] - dims_real = dims_virtual.copy() - - # Ensure tokens has the right shape for tensor_stream_token_view - # It expects a 2D tensor where sum(dim=-1) gives the token IDs - if tokens.dim() == 1: - tokens = tokens.unsqueeze(-1) - - return Event( - data=tokens, - type=TextType.text, - time=(time, time), - dims_virtual=dims_virtual, - dims_real=dims_real, - idx_range=(0, num_tokens), - ) - - -# ============================================================================ -# Processor -# ============================================================================ - -class IsaacProcessor(ProcessorMixin): attributes = ["image_processor", "tokenizer"] image_processor_class = ("IsaacImageProcessorFast",) - tokenizer_class = ("Qwen2Tokenizer", "Qwen2TokenizerFast") + tokenizer_class = ("Qwen2Tokenizer",) + pad_token_id = 151643 def __init__( self, - image_processor: IsaacImageProcessorFast | None = None, - tokenizer: Qwen2Tokenizer | None = None, + image_processor, + tokenizer, *, vision_token: str = "", max_sequence_length: int = 16384, - rescale_factor: float | None = None, - config: IsaacConfig | dict | None = None, + rescale_factor: Optional[float] = None, + config: Optional[Union[IsaacConfig, dict]] = None, ) -> None: - if tokenizer is None: - raise ValueError("`tokenizer` must be provided to initialize IsaacProcessor.") - if isinstance(config, dict): config = IsaacConfig(**config) if config is not None: - max_sequence_length = config.max_sequence_length vision_token = config.vision_token + max_sequence_length = config.max_sequence_length rescale_factor = config.vision_rescale_factor - resolved_rescale_factor = float(rescale_factor) if rescale_factor is not None else float(1 / 255) - + resolved_rescale_factor = ( + float(rescale_factor) if rescale_factor is not None else float(1 / 255) + ) if config is not None: config.vision_rescale_factor = resolved_rescale_factor self.image_processor = image_processor - super().__init__(image_processor, tokenizer) + + text_pad_token_id = getattr(self.tokenizer, "pad_token_id", None) + image_pad_token_id = self.tokenizer.convert_tokens_to_ids("<|image_pad|>") + + self.text_pad_token_id = int(text_pad_token_id) + self.image_pad_token_id = int(image_pad_token_id) + self.pad_token_id = self.text_pad_token_id + self.current_processor = self.image_processor self.config = config - - # Mirror tokenizer chat template so ProcessorMixin.apply_chat_template works. self.chat_template = getattr(self.tokenizer, "chat_template", None) - self.vision_token = vision_token self.max_sequence_length = max_sequence_length - def build_event_stream_simple( - self, - text: str, - images: list[PIL.Image.Image] | None = None, - ) -> Stream: - events = [] - # Process text and images - # Find all occurrences of vision token - - pattern = re.escape(self.vision_token) - parts = re.split(f"({pattern})", text) # Keep the delimiter in the result - - image_idx = 0 - for current_time, part in enumerate(parts): - if part == self.vision_token: - # Replace vision token with image event - if images is None or image_idx >= len(images): - raise ValueError("Encountered vision token without a corresponding image.") - - features = self.image_processor( - images=images[image_idx], - return_tensors=TensorType.PYTORCH, - ) - - patches = features["patches"][0] # (H_tokens, W_tokens, embed) - virtual_dims = features["virtual_pixel_size"][0].tolist() - real_dims = features["real_pixel_size"][0].tolist() - - vision_event = Event( - data=patches.reshape(-1, patches.shape[-1]), - type=VisionType.image, - time=(current_time, current_time), - dims_virtual=virtual_dims, - dims_real=real_dims, - idx_range=(0, math.prod(virtual_dims)), - ) - events.append(vision_event) - image_idx += 1 - elif part: # Non-empty text part - # tokens = self.text_processor.tokenize(part, add_special_tokens=False) - text_event = create_text_event(self.tokenizer, part, time=current_time) - events.append(text_event) - - # Create stream without scheduling (events already in order) - return create_stream(events, priority=[TextType.text, VisionType.image], schedule=True) + def _pack_batch( + self, texts: list[str], images_list: Optional[list[Optional[list[Image]]]] + ) -> dict[str, Optional[torch.Tensor]]: + if images_list is None: + pairs = ((t, None) for t in texts) + else: + pairs = zip(texts, images_list, strict=True) + + per_sample: list[dict[str, Any]] = [] + for txt, imgs in pairs: + if imgs is not None and isinstance(imgs, Image): + imgs = [imgs] + per_sample.append(self._pack_single(txt, imgs)) + + lengths = [int(p["input_ids"].shape[1]) for p in per_sample] + max_len = max(lengths, default=0) + batch = len(per_sample) + + # Use first device with data as anchor + base_device = torch.device("cpu") + for p in per_sample: + if p["input_ids"].numel() > 0: + base_device = p["input_ids"].device + break + + pad_id = self.text_pad_token_id + padded_input_ids = torch.full( + (batch, max_len), pad_id, device=base_device, dtype=torch.long + ) + padded_modality = torch.full( + (batch, max_len), + ModalityType.text.value, + device=base_device, + dtype=torch.long, + ) + padded_position_ids = torch.zeros( + (batch, max_len, 3), device=base_device, dtype=torch.long + ) - def __call__( - self, - text: str | list[str], - images: PIL.Image.Image | list[PIL.Image.Image] | None = None, - return_tensors: str | TensorType | None = TensorType.PYTORCH, - **kwargs, - ) -> BatchFeature: - """ - Process text and images into TensorStream format. - Args: - text: Input text or list of texts with vision tokens - images: PIL image or list of images (optional) - return_tensors: Format for output tensors + for i, (sample, l) in enumerate(zip(per_sample, lengths)): + if l: + padded_input_ids[i, -l:] = sample["input_ids"][0] + padded_modality[i, -l:] = sample["modality_tensor"][0] + padded_position_ids[i, -l:] = sample["position_ids"][0] + + # Vision-side aggregation (batch-first, padded) + max_images = max((s["vision_image_count"] for s in per_sample), default=0) + max_vision_len = max((s["vision_patches_total"] for s in per_sample), default=0) + + if max_images > 0 and max_vision_len > 0: + # Determine patch dimension/dtype from first available patch tensor + patch_dim = None + patch_dtype = None + for s in per_sample: + vps = s["vision_patches"] + if vps: + patch_dim = int(vps[0].shape[-1]) + patch_dtype = vps[0].dtype + break + assert patch_dim is not None and patch_dtype is not None + + vision_patches = torch.zeros( + (batch, max_vision_len, patch_dim), + device=base_device, + dtype=patch_dtype, + ) + vision_patches_len = torch.zeros( + (batch,), device=base_device, dtype=torch.long + ) + vision_token_grids = torch.zeros( + (batch, max_images, 2), device=base_device, dtype=torch.long + ) + vision_token_offsets = torch.zeros( + (batch, max_images), device=base_device, dtype=torch.long + ) + vision_token_lengths = torch.zeros( + (batch, max_images), device=base_device, dtype=torch.long + ) + vision_image_count = torch.zeros( + (batch,), device=base_device, dtype=torch.long + ) - Returns: - BatchFeature with input_ids and tensor_stream - """ - # Normalize inputs to lists - if isinstance(text, str): - texts = [text] - else: - texts = text + for i, sample in enumerate(per_sample): + vps: list[torch.Tensor] = sample["vision_patches"] + grids: list[tuple[int, int]] = sample["vision_token_grids"] + offs: list[int] = sample["vision_token_offsets"] + lens: list[int] = sample["vision_token_lengths"] + if not vps: + continue + vision_image_count[i] = len(vps) + cursor = 0 + for img_idx, (vp, grid, off, ln) in enumerate( + zip(vps, grids, offs, lens) + ): + plen = int(vp.shape[0]) + if plen <= 0 or img_idx >= max_images: + continue + vision_patches[i, cursor : cursor + plen] = vp.to(base_device) + cursor += plen + vision_token_grids[i, img_idx] = torch.tensor( + grid, device=base_device, dtype=torch.long + ) + vision_token_offsets[i, img_idx] = int(off) + vision_token_lengths[i, img_idx] = int(ln) + vision_patches_len[i] = cursor - if images is not None: - if isinstance(images, PIL.Image.Image): - images_list = [images] - else: - images_list = images else: - images_list = None - - if len(texts) != 1: - raise ValueError("IsaacProcessor currently supports batch_size=1") - if images_list is not None: - # Count vision tokens in text to validate image count - vision_token_count = texts[0].count(self.vision_token) - if vision_token_count != len(images_list): - raise ValueError( - f"Number of {self.vision_token} tokens in text ({vision_token_count}) " - f"must match number of images ({len(images_list)})" - ) - - # Build event stream - stream = self.build_event_stream_simple( - text=texts[0], - images=images_list, - ) + vision_patches = vision_patches_len = vision_token_grids = None + vision_token_offsets = vision_token_lengths = vision_image_count = None + + return { + "input_ids": padded_input_ids, + "vision_patches": vision_patches, + "vision_patches_len": vision_patches_len, + "vision_token_grids": vision_token_grids, + "vision_token_offsets": vision_token_offsets, + "vision_token_lengths": vision_token_lengths, + "vision_image_count": vision_image_count, + "modality_tensor": padded_modality, + "position_ids": padded_position_ids, + } - # Create TensorStream - tensor_stream = TensorStream([stream]) + def _pack_single(self, text: str, images: Optional[list[Image]]) -> dict[str, Any]: + segments = text.split( + self.vision_token + ) # Parse by vision_token; interleave text segments and image segments. + num_images = len(segments) - 1 + items: list[dict[str, Any]] = [] + total = 0 + num_provided_images = len(images) if images is not None else 0 + if not num_images == num_provided_images: + raise ValueError( + f"IsaacProcessor expects one image per image token, got {num_images} tokens and {num_provided_images} images in sample with text {text} " + ) - # Slice to max length if needed - _, T = tensor_stream.shape - if T > self.max_sequence_length: - tensor_stream = ts_slice(tensor_stream, start=T - self.max_sequence_length, end=T) + for index, segment in enumerate(segments): + if segment: + tok = ( + self.tokenizer.encode( + segment, add_special_tokens=False, return_tensors="pt" + ) + .squeeze(0) + .to(torch.long) + ) + segment_length = int(tok.numel()) + items.append( + {"type": "text", "segment_length": segment_length, "tok": tok} + ) + total += segment_length - # Get token view - tokens = tensor_stream_token_view(tensor_stream) - if return_tensors in (TensorType.PYTORCH, "pt"): - input_ids = torch.as_tensor(tokens, dtype=torch.long) - else: - input_ids = tokens + if index < num_images: + feat = self.image_processor( + images=images[index], return_tensors=TensorType.PYTORCH + ) + patches = feat["patches"][0].reshape(-1, feat["patches"].shape[-1]) - data = { - "input_ids": input_ids, - "tensor_stream": tensor_stream, - } + virtual_pixel_size = ( + feat["virtual_pixel_size"][0].to(torch.long).tolist() + ) + real_pixel_size = feat["real_pixel_size"][0].to(torch.long).tolist() + dims = tuple( + (virtual_pixel_size + [1, 1, 1])[:3] + ) # (T,H,W) in virtual space + segment_length = int(dims[0] * dims[1] * dims[2]) + + items.append( + { + "type": "image", + "segment_length": segment_length, + "dims": dims, + "patches": patches, + "grid": (int(real_pixel_size[1]), int(real_pixel_size[2])), + } + ) + total += segment_length + + # Tail crop window. + start = max(0, total - self.max_sequence_length) + end = total + + image_pad_value = self.image_pad_token_id + base_device: Optional[torch.device] = None + position_ids, modality, input_ids = [], [], [] + vpatches: list[torch.Tensor] = [] + grids: list[tuple[int, int]] = [] + vision_token_offsets: list[int] = [] + vision_token_lengths: list[int] = [] + + global_offset = 0 + position_offset = 0 + + for item in items: + segment_length = int(item["segment_length"]) + current_window_start = max(start, global_offset) + current_window_end = min(end, global_offset + segment_length) + has_overlap = current_window_end > current_window_start + + if has_overlap and base_device is None: + base_device = ( + item["patches"].device + if item["type"] == "image" + else item["tok"].device + ) - return BatchFeature(data=data) + if has_overlap: + segment_local_start = int(current_window_start - global_offset) + segment_local_end = int(current_window_end - global_offset) + segment_local_indices = torch.arange( + segment_local_start, + segment_local_end, + device=base_device, + dtype=torch.long, + ) + segment_kept_length = segment_local_end - segment_local_start + if item["type"] == "text": + slice_index = segment_local_indices + position_offset + zero_axis_pad = torch.zeros_like(slice_index) + position_ids.append( + torch.stack((slice_index, zero_axis_pad, zero_axis_pad), -1) + ) + modality.append( + torch.full( + (segment_kept_length,), + ModalityType.text.value, + device=base_device, + dtype=torch.long, + ) + ) + input_ids.append( + item["tok"].to(base_device)[ + segment_local_start:segment_local_end + ] + ) + position_offset += segment_length + else: + num_pos_slices, grid_height_tokens, grid_width_tokens = item["dims"] + hw = grid_height_tokens * grid_width_tokens + slice_index = (segment_local_indices // hw) + position_offset + rem = segment_local_indices % hw + row_index = rem // grid_width_tokens + col_index = rem % grid_width_tokens + position_ids.append( + torch.stack((slice_index, row_index, col_index), -1) + ) + modality.append( + torch.full( + (segment_kept_length,), + ModalityType.image.value, + device=base_device, + dtype=torch.long, + ) + ) + input_ids.append( + torch.full( + (segment_kept_length,), + image_pad_value, + device=base_device, + dtype=torch.long, + ) + ) -# ============================================================================ -# Model -# ============================================================================ + vpatches.append(item["patches"].to(base_device)) + grids.append(item["grid"]) + vision_token_offsets.append(segment_local_start) + vision_token_lengths.append(segment_kept_length) + position_offset += int(num_pos_slices) -def compute_position_ids_input_ids(input_ids: torch.Tensor) -> torch.Tensor: - r"""Create 3D positional indices for token input. + else: + position_offset += ( + segment_length if item["type"] == "text" else int(item["dims"][0]) + ) - Args: - input_ids (`torch.Tensor`): - Tensor of shape `(batch_size, seq_len)` containing token ids. + global_offset += segment_length - Returns: - `torch.Tensor`: Positional indices with shape `(batch_size, seq_len, 3)` where each channel duplicates the - 1D position so it can be consumed by the 3-axis MRoPE rotary embedding. - """ - batch_size, seq_length = input_ids.shape - position_ids = torch.arange(seq_length, device=input_ids.device) - position_ids = position_ids.view(1, -1).expand(batch_size, -1) - position_ids = position_ids.unsqueeze(2).expand(-1, -1, 3) # Add 3D for MRoPE - return position_ids + if base_device is None: + base_device = torch.device("cpu") + modality_tensor = ( + torch.cat(modality, 0).unsqueeze(0) + if modality + else torch.zeros((1, 0), device=base_device, dtype=torch.long) + ) + position_ids = ( + torch.cat(position_ids, 0).unsqueeze(0) + if position_ids + else torch.zeros((1, 0, 3), device=base_device, dtype=torch.long) + ) + input_ids = ( + torch.cat(input_ids, 0).unsqueeze(0) + if input_ids + else torch.zeros((1, 0), device=base_device, dtype=torch.long) + ) -class IsaacRotaryEmbedding(nn.Module): - EXTRA_ROPE_KEYS = {"mrope_section", "mrope_interleaved"} + return { + "input_ids": input_ids, + "modality_tensor": modality_tensor, + "position_ids": position_ids, + "vision_patches": vpatches, + "vision_token_grids": grids, + "vision_token_offsets": vision_token_offsets, + "vision_token_lengths": vision_token_lengths, + "vision_patches_total": sum(int(v.shape[0]) for v in vpatches), + "vision_image_count": len(vpatches), + } + def __call__( + self, + text: Union[str, list[str]], + images: Optional[Union[Image, list[Image]]] = None, + return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH, + **kwargs, + ) -> BatchFeature: + texts = [text] if isinstance(text, str) else text + images_list: Optional[list[Optional[list[Image]]]] = None + if images is not None: + if isinstance(images, list) and len(images) == len(texts): + if not images: + images_list = [] + elif isinstance(images[0], list): + images_list = images # already per-sample + else: + images_list = [ + [img] for img in images + ] # list of images, one per sample + else: + images_list = [] + for t in texts: + n_tok = t.count(self.vision_token) + if n_tok == 0: + images_list.append(None) + else: + if isinstance(images, list): + images_list.append(images) + else: + images_list.append([images]) + + packed = self._pack_batch(texts, images_list) + input_ids = packed.pop("input_ids") + return BatchFeature(data={"input_ids": input_ids, "packed_inputs": packed}) + + +class IsaacRotaryEmbedding(qwen2_5_vl_modeling.Qwen2_5_VLRotaryEmbedding): def __init__(self, config: IsaacConfig, device=None): - super().__init__() - - rope_source_cfg = config.get_text_config() if hasattr(config, "get_text_config") else config + rope_source_cfg = ( + config.get_text_config() if hasattr(config, "get_text_config") else config + ) rope_scaling = getattr(rope_source_cfg, "rope_scaling", None) or {} - - sanitized_scaling = {k: v for k, v in rope_scaling.items() if k not in self.EXTRA_ROPE_KEYS} config_for_rope = copy.copy(rope_source_cfg) - config_for_rope.rope_scaling = sanitized_scaling if sanitized_scaling else None + config_for_rope.rope_scaling = rope_scaling - init_device = device if device is not None and getattr(device, "type", None) != "meta" else None - self._qwen_rotary = Qwen2_5_VLRotaryEmbedding(config_for_rope, device=init_device) + init_device = ( + device + if device is not None and getattr(device, "type", None) != "meta" + else None + ) + super().__init__(config_for_rope, device=init_device) - rotary_half_dim = self._qwen_rotary.inv_freq.shape[0] - self.mrope_section = self._resolve_mrope_section(rope_scaling.get("mrope_section"), rotary_half_dim) - self.hidden_size = getattr(rope_source_cfg, "hidden_size", None) or config.hidden_size + rotary_half_dim = self.inv_freq.shape[0] + self.mrope_section = self._resolve_mrope_section( + rope_scaling.get("mrope_section"), rotary_half_dim + ) + self.hidden_size = ( + getattr(rope_source_cfg, "hidden_size", None) or config.hidden_size + ) @staticmethod - def _resolve_mrope_section(section: list[int] | None, rotary_half_dim: int) -> list[int]: + def _resolve_mrope_section( + section: Optional[list[int]], rotary_half_dim: int + ) -> list[int]: if section is None: weights = (2, 1, 1) base = [rotary_half_dim * w // sum(weights) for w in weights] @@ -1664,12 +1440,6 @@ class IsaacRotaryEmbedding(nn.Module): return base section = [int(v) for v in section] - if len(section) != 3: - raise ValueError("`mrope_section` must contain exactly three elements (temporal, height, width)") - if sum(section) != rotary_half_dim: - raise ValueError( - f"`mrope_section` must sum to the rotary half-dimension ({rotary_half_dim}). Received {section}." - ) return section def _combine_axes(self, tensor: torch.Tensor) -> torch.Tensor: @@ -1677,21 +1447,12 @@ class IsaacRotaryEmbedding(nn.Module): chunks = tensor.split(split_sections, dim=-1) return torch.cat([chunk[i % 3] for i, chunk in enumerate(chunks)], dim=-1) - @property - def inv_freq(self) -> torch.Tensor: - return self._qwen_rotary.inv_freq - def forward( self, position_ids: torch.Tensor, modality_tensor: torch.Tensor, - hidden_states: torch.Tensor | None = None, + hidden_states: Optional[torch.Tensor] = None, ) -> tuple[torch.Tensor, torch.Tensor]: - if position_ids.ndim != 3 or position_ids.size(-1) != 3: - raise ValueError("`position_ids` must have shape (batch, seq_len, 3) for MRoPE") - if modality_tensor.shape != position_ids.shape[:2]: - raise ValueError("`modality_tensor` must align with the first two dims of `position_ids`") - if hidden_states is None: batch, seq_len, _ = position_ids.shape hidden_states = torch.zeros( @@ -1704,43 +1465,49 @@ class IsaacRotaryEmbedding(nn.Module): with torch.no_grad(): pos = position_ids.clone() - not_spatial = modality_tensor != VisionType.image.value - if not_spatial.any(): - data_1d = pos[not_spatial][..., 0].unsqueeze(-1) - pos[not_spatial] = data_1d.expand(-1, pos.shape[-1]) - + not_spatial = modality_tensor != ModalityType.image.value + data_1d = pos[not_spatial][..., 0].unsqueeze( + -1 + ) # Collapse non-vision modalities to 1D positions + pos[not_spatial] = data_1d.expand(-1, pos.shape[-1]) pos_axes = pos.permute(2, 0, 1).contiguous() - cos_axes, sin_axes = self._qwen_rotary(hidden_states, pos_axes) - - cos_axes = cos_axes.to(hidden_states.dtype) - sin_axes = sin_axes.to(hidden_states.dtype) - - cos_combined = self._combine_axes(cos_axes) - sin_combined = self._combine_axes(sin_axes) + cos_axes, sin_axes = super().forward(hidden_states, pos_axes) + cos_axes, sin_axes = ( + cos_axes.to(hidden_states.dtype), + sin_axes.to(hidden_states.dtype), + ) + cos_combined, sin_combined = ( + self._combine_axes(cos_axes), + self._combine_axes(sin_axes), + ) return cos_combined, sin_combined +@auto_docstring class IsaacModel(Qwen3PreTrainedModel): supports_gradient_checkpointing = True + _can_compile_fullgraph = False + _supports_flex_attn = False + _can_record_outputs = {"attentions": OutputRecorder(IsaacVisionAttention, index=1)} + all_tied_weights_keys: dict[str, str] = {} def __init__(self, config: IsaacConfig): Qwen3PreTrainedModel.__init__(self, config) - text_cfg_source = getattr(config, "get_text_config", lambda: config)() + text_cfg_source = config.text_config text_cfg = copy.deepcopy(text_cfg_source) - text_cfg._attn_implementation = config._attn_implementation - self.text_model = AutoModel.from_config(text_cfg) - # Ensure downstream callers observe the composed config - self.text_model.config = config + self.text_model = Qwen3Model._from_config(text_cfg) + self.text_model.config = ( + config # Ensure downstream callers observe the composed config + ) self.rotary_emb = IsaacRotaryEmbedding(config, device=self.device) - if config.vision_config is None: - raise ValueError("IsaacConfig should always have vision_config") - - hidden_dim = config.vision_config.hidden_size * (config.vision_config.pixel_shuffle_scale_factor**2) + hidden_dim = config.vision_config.hidden_size * ( + config.vision_config.pixel_shuffle_scale_factor**2 + ) self.vision_embedding = nn.Sequential( IsaacVisionTransformer(config.vision_config), nn.Linear( @@ -1751,18 +1518,29 @@ class IsaacModel(Qwen3PreTrainedModel): nn.SiLU(), nn.Linear(4 * hidden_dim, config.hidden_size, bias=False), ) + self.vision_embedding._supports_sdpa = True + self.max_sequence_length = config.max_sequence_length + self.vision_rescale_factor = config.vision_rescale_factor + self.vision_token = config.vision_token + self.rope_deltas = None - # Dispatch table for TensorStream balanced embedding (text + vision) - self.embed_fns = { - TextType: self.embed_text_tokens, - VisionType: self.embed_vision, - } + self.post_init() + + # Respect config-specified gradient checkpointing + if getattr(config, "gradient_checkpointing", False): + self.gradient_checkpointing_enable() def get_input_embeddings(self) -> nn.Module: return self.text_model.get_input_embeddings() def set_input_embeddings(self, value: nn.Module) -> None: self.text_model.set_input_embeddings(value) + vocab_size = getattr(value, "num_embeddings", None) + if vocab_size is not None: + self.config.vocab_size = vocab_size + if hasattr(self.config, "text_config"): + self.config.text_config.vocab_size = vocab_size + self.text_model.config.vocab_size = vocab_size @property def embed_tokens(self) -> nn.Module: @@ -1773,454 +1551,441 @@ class IsaacModel(Qwen3PreTrainedModel): self.text_model.embed_tokens = value @property - def layers(self) -> nn.ModuleList: - return self.text_model.layers + def vision_model(self) -> nn.Module: + return self.vision_embedding.vision_tower - @property - def norm(self) -> nn.Module: - return self.text_model.norm + def embed_packed_inputs( + self, input_ids: torch.Tensor, packed_inputs: dict[str, Optional[torch.Tensor]] + ) -> tuple[torch.Tensor, torch.Tensor]: + """ + Batch-first vision packing for DataParallel safety. + Expects packed_inputs containing: + - modality_tensor: (B, S) + - position_ids: (B, S, 3) (optional, used elsewhere) + - vision_patches: (B, max_vision_len, patch_dim) padded + - vision_patches_len: (B,) lengths of real vision patches per sample + - vision_token_grids: (B, max_images, 2) padded + - vision_token_offsets: (B, max_images) padded (virtual offsets) + - vision_token_lengths: (B, max_images) padded (virtual lengths) + - vision_image_count: (B,) number of images per sample + """ - def _set_gradient_checkpointing(self, enable: bool = True, gradient_checkpointing_func=None): - self.text_model._set_gradient_checkpointing( - enable=enable, gradient_checkpointing_func=gradient_checkpointing_func + modality = packed_inputs["modality_tensor"].to( + device=input_ids.device, dtype=torch.long ) + embeds = self.text_model.embed_tokens(input_ids) + + vision_patches = packed_inputs.get("vision_patches") + if vision_patches is None: + return embeds, modality - def embed_text_tokens(self, token_ids: torch.Tensor) -> torch.Tensor: - """Embed text tokens, squeezing singleton dimensions.""" - # Text events are shaped as (..., 1); squeeze the singleton index dim - h = self.text_model.embed_tokens(token_ids) - if h.dim() >= 2 and h.size(-2) == 1: - h = h[..., 0, :] - return h + vision_patches_len = packed_inputs["vision_patches_len"].to( + device=vision_patches.device, dtype=torch.long + ) + vision_token_grids = packed_inputs["vision_token_grids"].to( + device=vision_patches.device, dtype=torch.long + ) + vision_token_offsets = packed_inputs["vision_token_offsets"].to( + device=vision_patches.device, dtype=torch.long + ) + vision_token_lengths = packed_inputs["vision_token_lengths"].to( + device=vision_patches.device, dtype=torch.long + ) + vision_image_count = packed_inputs["vision_image_count"].to( + device=vision_patches.device, dtype=torch.long + ) - def embed_vision(self, vision_tokens: tuple[torch.Tensor, torch.Tensor]) -> torch.Tensor: - """Embed vision tokens using the vision encoder.""" - # vision tokens is (seq_patches, token_grids) - return self.vision_embedding(vision_tokens) + flat_patches: list[torch.Tensor] = [] + flat_grids: list[torch.Tensor] = [] + flat_offsets: list[torch.Tensor] = [] + flat_lengths: list[torch.Tensor] = [] + flat_batch_idx: list[int] = [] - def embed_stream(self, tensor_stream: TensorStream) -> torch.Tensor: - """ - Embed each modality stream independently, preserving the original TensorStream - structure. - """ - flat_stream = tensor_stream.flat_stream() - per_modality_stream = group_streams(flat_stream, group_fn=lambda ev: ev.type, schedule=False) - per_modality_compact_stream = {k: v.compact() for k, v in per_modality_stream.items()} - - # Collect per-event grids for vision tokens (H, W like dims sans time) - token_grids = defaultdict(list) - for stream in tensor_stream.streams: - for event in stream: - token_grids[event.type].append(event.dims(virtual=False)) - - embedded_compact = {} - for stream_type, modality_payload_tensor in per_modality_compact_stream.items(): - if stream_type.modality == VisionType: - # Build a (N_events, 2) grid tensor with spatial dims only - grids = token_grids.get(stream_type, []) - if len(grids) == 0: - input_tensor = modality_payload_tensor - else: - token_grids_tensor = torch.tensor(grids, dtype=torch.long, device=tensor_stream.device)[:, 1:] - input_tensor = (modality_payload_tensor, token_grids_tensor) - embedded_compact[stream_type] = self.embed_fns[stream_type.modality](input_tensor) - else: - embedded_compact[stream_type] = self.embed_fns[stream_type.modality](modality_payload_tensor) + batch_size = vision_patches.shape[0] + for b in range(batch_size): + total_len = int(vision_patches_len[b].item()) + if total_len <= 0: + continue + vp = vision_patches[b, :total_len] + img_cnt = int(vision_image_count[b].item()) + if img_cnt <= 0: + continue + grids_b = vision_token_grids[b, :img_cnt] + offs_b = vision_token_offsets[b, :img_cnt] + lens_b = vision_token_lengths[b, :img_cnt] + sizes = grids_b.prod(-1).tolist() + cursor = 0 + for grid, off, ln, size in zip(grids_b, offs_b, lens_b, sizes): + size_int = int(size) + if size_int <= 0: + continue + chunk = vp[cursor : cursor + size_int] + cursor += size_int + flat_patches.append(chunk) + flat_grids.append(grid) + flat_offsets.append(off) + flat_lengths.append(ln) + flat_batch_idx.append(b) + assert cursor == total_len, ( + f"Vision patches cursor mismatch: cursor={cursor}, total={total_len}" + ) + + if flat_patches: + flat_patches_t = torch.cat(flat_patches, dim=0) + flat_grids_t = torch.stack(flat_grids, dim=0) + flat_offsets_t = torch.stack(flat_offsets, dim=0) + flat_lengths_t = torch.stack(flat_lengths, dim=0) + flat_batch_idx_t = torch.tensor( + flat_batch_idx, device=vision_patches.device, dtype=torch.long + ) + else: + flat_patches_t = vision_patches.new_zeros((0, vision_patches.shape[-1])) + flat_grids_t = vision_patches.new_zeros((0, 2), dtype=torch.long) + flat_offsets_t = vision_patches.new_zeros((0,), dtype=torch.long) + flat_lengths_t = vision_patches.new_zeros((0,), dtype=torch.long) + flat_batch_idx_t = vision_patches.new_zeros((0,), dtype=torch.long) + + grid_sum = ( + int(flat_grids_t.prod(-1).sum().item()) if flat_grids_t.numel() else 0 + ) + assert flat_patches_t.shape[0] == grid_sum, ( + "Packed vision mismatch after flatten: " + f"patches={flat_patches_t.shape[0]}, grid_sum={grid_sum}" + ) + + if flat_patches_t.numel() == 0: + return embeds, modality - # Reconstruct a TensorStream with embedded payloads and compact - embedded_ts = reconstruct_tensor_stream_from_compact_dict(tensor_stream, embedded_compact) - h = embedded_ts.compact() # (B, T, D) - return h + vision = self.vision_embedding((flat_patches_t, flat_grids_t)) + # per-image token counts AFTER pixel-shuffle + vision_reduction_factor = int( + self.config.vision_config.pixel_shuffle_scale_factor + ) + sizes = ( + flat_grids_t.prod(-1) + .div( + vision_reduction_factor * vision_reduction_factor, rounding_mode="floor" + ) + .tolist() + ) + + chunks = vision.split(sizes, dim=0) + vision_chunks: list[torch.Tensor] = [] + vision_batch_idx: list[int] = [] + for chunk, size, offset, length, batch_index in zip( + chunks, + sizes, + flat_offsets_t.tolist(), + flat_lengths_t.tolist(), + flat_batch_idx_t.tolist(), + ): + size_int = int(size) + if size_int <= 0: + continue + offset_int = max(0, min(int(offset), size_int)) + length_int = max(0, min(int(length), size_int - offset_int)) + if length_int: + vision_chunks.append(chunk[offset_int : offset_int + length_int]) + vision_batch_idx.append(int(batch_index)) + + if vision_chunks: + vision_flat = torch.cat(vision_chunks, 0) + else: + vision_flat = vision.new_zeros((0, vision.size(-1))) + + embeds = embeds.clone() + num_batches = modality.shape[0] + image_positions = [ + (modality[b] == ModalityType.image.value) + .nonzero(as_tuple=False) + .squeeze(-1) + for b in range(num_batches) + ] + cursors = [0 for _ in range(num_batches)] + + for chunk, batch_index in zip(vision_chunks, vision_batch_idx): + if chunk.numel() == 0: + continue + positions = image_positions[batch_index] + start = cursors[batch_index] + end = start + chunk.shape[0] + embeds[batch_index, positions[start:end]] = chunk.to( + device=embeds.device, dtype=embeds.dtype + ) + cursors[batch_index] = end + + return embeds, modality + + def get_rope_index( + self, + *, + position_ids: Optional[torch.Tensor] = None, + attention_mask: torch.Tensor, + inputs_embeds: torch.Tensor, + cache_position: torch.Tensor, + ) -> tuple[torch.Tensor, torch.Tensor]: + """Build 3D position ids and per-batch RoPE deltas.""" + + device = inputs_embeds.device + batch_size, seq_len = inputs_embeds.shape[:2] + + if position_ids is None: + cp = cache_position.to(device=device, dtype=torch.long) + if cp.ndim == 1: + cp = cp.view(1, -1).expand(batch_size or 1, -1) + + base_delta = torch.as_tensor( + 0 if self.rope_deltas is None else self.rope_deltas, + device=device, + dtype=torch.long, + ).reshape(-1, 1) + base_delta = torch.broadcast_to(base_delta, (batch_size, 1)) + + mask_delta = attention_mask.to(device=device, dtype=torch.long).sum( + 1, keepdim=True + ) - attention_mask.size(1) + rope_position = cp + base_delta + mask_delta + pos_3d = rope_position.unsqueeze(-1).expand(-1, -1, 3) + return pos_3d, base_delta + + position_ids = position_ids.to(device=device) + if position_ids.ndim == 2: + position_ids = position_ids.unsqueeze(-1).expand(-1, -1, 3) + + if position_ids.shape[1] != seq_len: + start_positions = position_ids[:, :1, 0] + position_ids = ( + torch.arange(seq_len, device=position_ids.device).view(1, -1) + + start_positions + ) + position_ids = position_ids.unsqueeze(-1).expand(-1, -1, 3) + + attn = attention_mask.to(device=device, dtype=torch.long) + m_per_batch = position_ids.amax(dim=(1, 2)) + seq_lens = attn.eq(1).sum(dim=-1).to(dtype=m_per_batch.dtype, device=device) + rope_deltas = ( + (m_per_batch + 1 - seq_lens).to(dtype=position_ids.dtype).unsqueeze(1) + ) + return position_ids, rope_deltas + + @auto_docstring + @check_model_inputs def forward( self, - input_ids: torch.LongTensor | None = None, - tensor_stream: TensorStream | None = None, - attention_mask: torch.Tensor | None = None, - position_ids: torch.LongTensor | None = None, - modality_tensor: torch.LongTensor | None = None, - past_key_values: list[torch.FloatTensor] | None = None, - inputs_embeds: torch.FloatTensor | None = None, - use_cache: bool | None = None, - output_hidden_states: bool | None = None, - return_dict: bool | None = None, - cache_position: torch.LongTensor | None = None, - **kwargs, + input_ids: Optional[torch.LongTensor] = None, + packed_inputs: Optional[dict[str, torch.Tensor]] = 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, + use_cache: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, + **kwargs: Unpack[TransformersKwargs], ) -> tuple | BaseModelOutputWithPast: """ Forward pass with MRoPE position embeddings. Computes position embeddings once and passes them through all layers. + + Args: + packed_inputs (`dict`, *optional*): + Plain tensor payloads. When provided, requires `input_ids` for text tokens (or `text_token_ids` so `input_ids` can be rebuilt). + modality_tensor (`torch.LongTensor`, *optional*): + Modality identifiers aligned with the embedded sequence, shaped `(batch_size, seq_len)` and containing + values from `ModalityType`. Automatically built from `packed_inputs` or treated as text-only when omitted. """ - 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 - - # Get inputs - if tensor_stream is not None and inputs_embeds is not None: - raise ValueError("You cannot specify both tensor_stream and inputs_embeds") - elif tensor_stream is not None: - # Embed TensorStream directly - inputs_embeds = self.embed_stream(tensor_stream) - # Create modality tensor if not provided - if modality_tensor is None: - modality_tensor = modality_mask(tensor_stream) - elif input_ids is not None and inputs_embeds is not None: - raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") + + output_attentions = kwargs.pop("output_attentions", None) + + modality_tensor: Optional[torch.Tensor] = None + + if packed_inputs is not None: + inputs_embeds, modality_tensor = self.embed_packed_inputs( + input_ids, packed_inputs + ) elif input_ids is not None: inputs_embeds = self.text_model.embed_tokens(input_ids) - # Create text modality tensor if not provided - if modality_tensor is None: - batch_size, seq_length = input_ids.shape - modality_tensor = torch.full( - (batch_size, seq_length), TextType.text.value, device=input_ids.device, dtype=torch.long - ) - elif inputs_embeds is None: - raise ValueError("You have to specify either tensor_stream, input_ids or inputs_embeds") - # Create default position_ids if not provided - if position_ids is None: - if tensor_stream is not None: - position_ids = compute_mrope_pos_tensor(tensor_stream) # (B,L,3) - else: - position_ids = compute_position_ids_input_ids(input_ids) + device = inputs_embeds.device + batch_size, seq_len = inputs_embeds.shape[:2] + + if use_cache and past_key_values is None: + past_key_values = DynamicCache(config=self.config.get_text_config()) + + 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 + seq_len, device=device + ) + + if attention_mask is None: + attention_mask = torch.ones( + inputs_embeds.shape[:2], device=inputs_embeds.device, dtype=torch.long + ) + + if ( + position_ids is None + and packed_inputs is not None + and packed_inputs.get("position_ids") is not None + ): + position_ids = packed_inputs.get("position_ids").to(device=device) + + position_ids, rope_deltas = self.get_rope_index( + position_ids=position_ids, + attention_mask=attention_mask, + inputs_embeds=inputs_embeds, + cache_position=cache_position, + ) + self.rope_deltas = rope_deltas + + if modality_tensor is None: + modality_tensor = torch.full( + (batch_size, seq_len), + ModalityType.text.value, + device=device, + dtype=torch.long, + ) - # Compute MRoPE position embeddings if we have custom rotary_emb cos, sin = self.rotary_emb( - position_ids, - modality_tensor, - hidden_states=inputs_embeds, + position_ids, modality_tensor, hidden_states=inputs_embeds ) - cos = cos.to(inputs_embeds.dtype) - sin = sin.to(inputs_embeds.dtype) - # Prepare attention mask - if attention_mask is not None: - attention_mask = self._update_causal_mask( - attention_mask, inputs_embeds, cache_position, past_key_values, False + decoder_position_ids = ( + position_ids[..., 0] if position_ids.ndim == 3 else position_ids + ) + + if not isinstance(attention_mask, dict): + attention_mask = create_masks_for_generate( + config=self.config, + input_embeds=inputs_embeds, + attention_mask=attention_mask, + cache_position=cache_position, + past_key_values=past_key_values, + position_ids=decoder_position_ids, ) - # Initialize hidden states + is_mask_dict = isinstance(attention_mask, dict) hidden_states = inputs_embeds + all_attentions = [] if output_attentions else None - for decoder_layer in self.text_model.layers: - layer_outputs = decoder_layer( + for layer in self.text_model.layers: + layer_mask = ( + attention_mask[layer.attention_type] if is_mask_dict else attention_mask + ) + layer_outputs = layer( hidden_states, - attention_mask=attention_mask, - position_ids=position_ids, - past_key_value=past_key_values, + attention_mask=layer_mask, + position_ids=decoder_position_ids, + past_key_values=past_key_values, use_cache=use_cache, cache_position=cache_position, position_embeddings=(cos, sin), + output_attentions=output_attentions, **kwargs, ) - hidden_states = layer_outputs[0] if isinstance(layer_outputs, tuple) else layer_outputs + layer_outputs_is_tuple = isinstance(layer_outputs, tuple) + hidden_states = ( + layer_outputs[0] if layer_outputs_is_tuple else layer_outputs + ) + if output_attentions and layer_outputs_is_tuple: + all_attentions.append(layer_outputs[1]) - # Final layer norm hidden_states = self.text_model.norm(hidden_states) return BaseModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=past_key_values, + hidden_states=(hidden_states,), + attentions=tuple(all_attentions) if output_attentions else None, ) - def _update_causal_mask( - self, - attention_mask: torch.Tensor, - input_tensor: torch.Tensor, - cache_position: torch.Tensor, - past_key_values: Cache, - output_attentions: bool = False, - ): - if self.config._attn_implementation == "flash_attention_2": - if attention_mask is not None and past_key_values is not None: - is_padding_right = attention_mask[:, -1].sum().item() != input_tensor.size()[0] - if is_padding_right: - raise ValueError( - "You are attempting to perform batched generation with padding_side='right'" - " this may lead to unexpected behaviour for Flash Attention version of Qwen3. Make sure to " - " call `tokenizer.padding_side = 'left'` before tokenizing the input. " - ) - 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 in ["cuda", "xpu", "npu"] - 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 - 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: Qwen3Config, - 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 place 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 (`Qwen3Config`): - 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, :].to( - causal_mask.device - ) - padding_mask = padding_mask == 0 - causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( - padding_mask, min_dtype - ) - return causal_mask - +@auto_docstring class IsaacForConditionalGeneration(Qwen3ForCausalLM, GenerationMixin): - """Isaac multimodal model for conditional generation.""" - config_class = IsaacConfig + _can_compile_fullgraph = False + _tied_weights_keys = {"lm_head.weight": "model.text_model.embed_tokens.weight"} + all_tied_weights_keys: dict[str, str] = { + "lm_head.weight": "model.text_model.embed_tokens.weight" + } def __init__(self, config: IsaacConfig): super().__init__(config) - self.model = IsaacModel(config) # Use our custom model + self.model = IsaacModel(config) self.vocab_size = config.vocab_size self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) - # Tracks rotary position offsets computed during a full forward pass so decode steps can reuse them. - self.rope_deltas = None - - def get_rope_index( - self, - input_ids: torch.Tensor | None, - tensor_stream: TensorStream | None, - attention_mask: torch.Tensor | None, - ) -> tuple[torch.Tensor, torch.Tensor]: - """Compute MRoPE position ids from a TensorStream (or 1D fallback). - - Returns (position_ids, rope_deltas). position_ids is (B,L,3) for MRoPE. - rope_deltas is (B,1) used to advance positions in decode. - """ - # tensor_stream present: compute 3D coords - if tensor_stream is None and input_ids is None: - raise ValueError("`tensor_stream` or `input_ids` must be provided to compute rope indices") - - if tensor_stream is not None: - pos_3d = compute_mrope_pos_tensor(tensor_stream) # (B,L,3) - else: - pos_3d = compute_position_ids_input_ids(input_ids) - B, L, _ = pos_3d.shape - - # Max position per batch across the 3 planes and sequence dimension: (B,) - m_per_batch = pos_3d.amax(dim=(1, 2)) - - # Sequence lengths per batch: (B,) - if attention_mask is None: - seq_lens = torch.full_like(m_per_batch, L) - else: - seq_lens = attention_mask.eq(1).sum(dim=-1).to(dtype=m_per_batch.dtype, device=m_per_batch.device) - - rope_deltas = (m_per_batch + 1 - seq_lens).to(dtype=pos_3d.dtype).unsqueeze(1) - return pos_3d, rope_deltas + @auto_docstring + @can_return_tuple + @check_model_inputs def forward( self, - input_ids: torch.LongTensor | None = None, - tensor_stream: TensorStream | None = None, - attention_mask: torch.Tensor | None = None, - position_ids: torch.LongTensor | None = None, - past_key_values: list[torch.FloatTensor] | None = None, - inputs_embeds: torch.FloatTensor | None = None, - labels: torch.LongTensor | None = None, - use_cache: bool | None = None, - output_hidden_states: bool | None = None, - return_dict: bool | None = None, - cache_position: torch.LongTensor | None = None, - **kwargs, + input_ids: Optional[torch.LongTensor] = None, + packed_inputs: Optional[dict[str, torch.Tensor]] = 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, + cache_position: Optional[torch.LongTensor] = None, + **kwargs: Unpack[TransformersKwargs], ) -> tuple | CausalLMOutputWithPast: - """ - Forward pass for conditional generation supporting both standard inputs and TensorStream. - Uses our embed_stream approach for multimodal inputs. - """ + """Run multimodal CausalLM forward, accepting packed vision/text inputs. - # Don't compute embeddings here - let the model handle it - if tensor_stream is not None: - input_ids = None - if input_ids is None and inputs_embeds is None and tensor_stream is None: - raise ValueError("Either input_ids, inputs_embeds, or tensor_stream must be provided.") - - # Build position ids (MRoPE) if needed and tensor_stream is available - # During decode we reuse `self.rope_deltas` computed on the initial forward pass; `rope_delta` captures how far - # cached rotary phases have progressed so we can advance `position_ids` without rebuilding the TensorStream. - if position_ids is None and tensor_stream is not None: - position_ids, self.rope_deltas = self.get_rope_index(input_ids, tensor_stream, attention_mask) - elif position_ids is None and input_ids is not None: - # For text inputs build position ids and modality tensor - position_ids = compute_position_ids_input_ids(input_ids) - if cache_position is not None and self.rope_deltas is not None: - # Combine the incremental decode step (`cache_position`) with cached offsets so hidden states continue - # rotating in lockstep across generation steps. - rope_delta = (cache_position[0] + self.rope_deltas).to(input_ids.device) - else: - rope_delta = 0 - if cache_position is not None and not isinstance(rope_delta, int): # otherwise `deltas` is an int `0` - batch_size = input_ids.shape[0] - rope_delta = rope_delta.repeat_interleave(batch_size // rope_delta.shape[0], dim=0) - position_ids = position_ids.add(rope_delta) - - if tensor_stream is not None: - modality_tensor = modality_mask(tensor_stream) - else: - batch_size, seq_len = input_ids.shape - modality_tensor = torch.empty(batch_size, seq_len, device=position_ids.device).fill_(TextType.text.value) + Args: + packed_inputs (`dict`, *optional*): + Packed vision/text payload from ``IsaacProcessor`` containing modality ids, MRoPE position ids, and + vision patch tensors/grids (with optional offsets/lengths) used to rebuild embeddings. + + Returns: + CausalLMOutputWithPast: logits, optional loss, caches, hidden states, attentions. + """ + output_attentions = kwargs.pop("output_attentions", None) outputs = self.model( input_ids=input_ids, - tensor_stream=tensor_stream, + packed_inputs=packed_inputs, attention_mask=attention_mask, position_ids=position_ids, - modality_tensor=modality_tensor, past_key_values=past_key_values, inputs_embeds=inputs_embeds, use_cache=use_cache, - output_hidden_states=output_hidden_states, - return_dict=return_dict, + output_attentions=output_attentions, cache_position=cache_position, **kwargs, ) - hidden_states = outputs[0] logits = self.lm_head(hidden_states) - loss = None if labels is not None: - loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size) + loss = self.loss_function( + logits=logits, labels=labels, vocab_size=self.config.vocab_size + ) return CausalLMOutputWithPast( loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, - attentions=None, + attentions=outputs.attentions if output_attentions else None, ) def prepare_inputs_for_generation( self, input_ids: torch.LongTensor, - past_key_values: list[torch.FloatTensor] | None = None, - attention_mask: torch.Tensor | None = None, - inputs_embeds: torch.FloatTensor | None = None, - tensor_stream: TensorStream | None = None, - cache_position: torch.LongTensor | None = None, - position_ids: torch.LongTensor | None = None, - use_cache: bool = True, + past_key_values: Optional[list[torch.FloatTensor]] = None, + attention_mask: Optional[torch.Tensor] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + packed_inputs: Optional[dict[str, torch.Tensor]] = None, + cache_position: Optional[torch.LongTensor] = None, + position_ids: Optional[torch.LongTensor] = None, **kwargs, ) -> dict[str, Any]: - """ - Prepare inputs for generation, handling TensorStream inputs properly. - """ - # Call parent preparation model_inputs = super().prepare_inputs_for_generation( input_ids, past_key_values=past_key_values, @@ -2228,48 +1993,40 @@ class IsaacForConditionalGeneration(Qwen3ForCausalLM, GenerationMixin): inputs_embeds=inputs_embeds, cache_position=cache_position, position_ids=position_ids, - use_cache=use_cache, **kwargs, ) + if packed_inputs is None: + return model_inputs - # Handle TensorStream for first forward pass only - if tensor_stream is not None and (cache_position is None or cache_position[0] == 0): - model_inputs["tensor_stream"] = tensor_stream - # Let forward rebuild position_ids using cached deltas during decode + past_len = ( + past_key_values.get_seq_length() if past_key_values is not None else 0 + ) + first_step = past_len == 0 + model_inputs["packed_inputs"] = packed_inputs if first_step else None model_inputs["position_ids"] = None - # Drop tensor_stream after step 0 - if cache_position is not None and cache_position[0] != 0: - model_inputs["tensor_stream"] = None + return model_inputs - def can_generate(self) -> bool: + @classmethod + def can_generate(cls) -> bool: return True - -AutoImageProcessor.register( - IsaacConfig, - fast_image_processor_class=IsaacImageProcessorFast, - exist_ok=True, -) + def set_input_embeddings(self, value: nn.Module) -> None: + self.model.set_input_embeddings(value) + vocab_size = getattr(value, "num_embeddings", None) + self.config.vocab_size = vocab_size + self.model.config.vocab_size = vocab_size + self.model.text_model.config.vocab_size = vocab_size + if self.lm_head.weight.shape[0] != vocab_size: + self.lm_head = nn.Linear(self.config.hidden_size, vocab_size, bias=False) + self.lm_head.weight = self.model.text_model.embed_tokens.weight __all__ = [ "IsaacConfig", "IsaacModel", + "IsaacPreTrainedModel", # noqa: F822 "IsaacForConditionalGeneration", "IsaacImageProcessorFast", "IsaacProcessor", ] - - -def _compute_residual_p_frames(frames: torch.Tensor, is_p_frame: list[bool]) -> torch.Tensor: - """Compute residuals for P-frames to stay in sync with the training pipeline.""" - if not any(is_p_frame): - return frames - - frame_indices = torch.arange(len(is_p_frame), device=frames.device) - i_frame_mask = torch.tensor([not flag for flag in is_p_frame], device=frames.device) - last_i_indices = torch.cummax((i_frame_mask * (1 + frame_indices)), dim=0).values.long() - 1 - p_indices = frame_indices[torch.tensor(is_p_frame, device=frames.device)] - frames[p_indices] = frames[p_indices] - frames[last_i_indices[p_indices]] - return frames