Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,11 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/figures/data.png filter=lfs diff=lfs merge=lfs -text
+assets/figures/framework.png filter=lfs diff=lfs merge=lfs -text
+assets/figures/mtpt_scaling.png filter=lfs diff=lfs merge=lfs -text
+assets/figures/performance.png filter=lfs diff=lfs merge=lfs -text
+assets/figures/performance_rl.png filter=lfs diff=lfs merge=lfs -text
+assets/figures/performance_table_2b.png filter=lfs diff=lfs merge=lfs -text
+assets/figures/performance_table_8b.png filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,187 @@
----
-license: apache-2.0
----
+---
+license: apache-2.0
+---
+
+# SAIL-VL2
+
+<div align="center">
+  <img src="assets/logo/logo_with_name.jpeg" width="80%" alt="SAIL-VL2 Logo">
+</div>
+
+<font size=3><div align='center' >  
+[[📖 Technique Report]()] 
+[[🤗 SAIL-VL2-2B](https://huggingface.co/BytedanceDouyinContent/SAIL-VL2-2B)]
+[[🤗 SAIL-VL2-8B](https://huggingface.co/BytedanceDouyinContent/SAIL-VL2-8B)]
+[[🤗 SAIL-VL2-2B-Thinking](https://huggingface.co/BytedanceDouyinContent/SAIL-VL2-2B-Thinking)]
+[[🤗 SAIL-VL2-8B-Thinking](https://huggingface.co/BytedanceDouyinContent/SAIL-VL2-8B-Thinking)]
+[[💻 Github](https://github.com/BytedanceDouyinContent/SAIL-VL2)]
+</div></font>
+
+We are very excited to introduce **SAIL-VL2** 🚀, a state-of-the-art visual language model that significantly outperforms existing models in various visual language tasks.
+
+## 🔥 Updates
+
+- **`2025.09.08`** 🌟 **SAIL-VL2 Technical Report** is now available at [arxiv]().
+
+
+## 🌟 Highlights
+- SAIL-VL2 is powerful, efficient, and achieves top results under 2B parameters.
+- SAIL-VL2-Thinking boosts complex reasoning, matching larger models.
+- SAIL-VL2 excels in fine-grained visual tasks beyond similar-scale models.
+
+<div align="center">
+  <img src="assets/figures/performance.png" width="100%" alt="SAIL-VL2 Performance">
+</div>
+
+## Model Architecture:
+
+| Architecture | ViT | LLM | Adapter | Token Merge | Resolution |
+| --- | --- | --- | --- | --- | --- |
+| [🤗SAIL-VL2-2B](https://huggingface.co/BytedanceDouyinContent/SAIL-VL2-2B) | [🤗SAILViT-Huge](https://huggingface.co/BytedanceDouyinContent/SAILViT-Huge-600M-448px) | [🤗Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B) | 2-layer MLP | 2x2 | 448x448xN |
+| [🤗SAIL-VL2-8B](https://huggingface.co/BytedanceDouyinContent/SAIL-VL2-8B) | [🤗SAILViT-Huge](https://huggingface.co/BytedanceDouyinContent/SAILViT-Huge-600M-448px) | [🤗Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) | 2-layer MLP | 2x2 | 448x448xN |
+| [🤗SAIL-VL2-2B-Thinking](https://huggingface.co/BytedanceDouyinContent/SAIL-VL2-2B-Thinking) | [🤗SAILViT-Huge](https://huggingface.co/BytedanceDouyinContent/SAILViT-Huge-600M-448px) | [🤗Qwen3-1.7B](https://huggingface.co/Qwen/Qwen3-1.7B) | 2-layer MLP | 2x2 | 448x448xN |
+| [🤗SAIL-VL2-8B-Thinking](https://huggingface.co/BytedanceDouyinContent/SAIL-VL2-8B-Thinking) | [🤗SAILViT-Huge](https://huggingface.co/BytedanceDouyinContent/SAILViT-Huge-600M-448px) | [🤗Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) | 2-layer MLP | 2x2 | 448x448xN |
+
+
+## 🎬 Quick Start
+
+
+```python
+import torch
+from transformers import AutoTokenizer, AutoModel, AutoProcessor
+from PIL import Image
+
+
+model_path = "your model path"
+
+tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
+processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
+device = torch.cuda.current_device()
+model = AutoModel.from_pretrained(model_path, trust_remote_code=True, torch_dtype=torch.bfloat16,).to(device)
+
+print("##### with images")
+cot_prompt = r"You FIRST think about the reasoning process as an internal monologue and then provide the final answer. The reasoning process MUST BE enclosed within <think> </think> tags. The final answer MUST BE put in \boxed{}."
+messages = [
+    {"role": "user", "content": [{"type": "image", "image": 'image_path'}, 
+    {"type": "text", "text": "describe the image" + cot_prompt}]}
+]
+text = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
+
+image_path = 'your image path'
+image = Image.open(image_path)
+inputs = processor(images=image, text=text, return_tensors="pt", padding=True, truncation=True).to(model.device).to(torch.bfloat16)
+
+generated_ids = model.generate(**inputs, max_new_tokens=512)
+response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+response = response.split('<|im_end|>')[0].strip()
+print(response)
+
+
+print("##### without images")
+cot_prompt = r"You FIRST think about the reasoning process as an internal monologue and then provide the final answer. The reasoning process MUST BE enclosed within <think> </think> tags. The final answer MUST BE put in \boxed{}."
+messages = [
+    {
+        "role": "user",
+        "content": [{"type": "text", "text": "中国的首都是哪里？" + cot_prompt}]
+    }
+]
+text = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
+inputs = processor(images=None, text=text, return_tensors="pt", padding=True, truncation=True).to(model.device).to(torch.bfloat16)
+generated_ids = model.generate(**inputs, max_new_tokens=512)
+response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+response = response.split('<|im_end|>')[0].strip()
+print(response)
+
+```
+
+## 👀 Introduction
+- **SAIL-VL2 is powerful yet efficient:** With training on 776B tokens, SAIL-VL2 has verified its effectiveness across 106 datasets, achieving state-of-the-art results on a broad spectrum of influential benchmarks under the 2B-parameter scale. Remarkably, even without specialized prompting, the base SAIL-VL2 model delivers highly competitive performance on challenging reasoning benchmarks such as MMMU and MathVista, demonstrating strong out-of-the-box capabilities.
+
+- **SAIL-VL2 as a deep thinker:** Many real-world tasks demand sophisticated reasoning and multi-step thought processes, which remain challenging for standard LVMs. To address this, we develop SAIL-VL2-Thinking, a specialized variant trained with advanced Chain-of-Thought (CoT) and reinforcement learning (RL) strategies. This design substantially improves performance on complex reasoning benchmarks, often matching or even surpassing models with far larger parameter scales, thereby setting a new standard for efficient architectures in high-level reasoning.
+
+- **SAIL-VL2 perceives with clarity:** Fine-grained visual understanding is a critical challenge for multimodal models. SAIL-VL2 delivers high-fidelity perception in tasks such as OCR, high-resolution document layout analysis, and complex chart interpretation, achieving detailed visual grounding beyond models of similar scale.
+
+
+<div align="center">
+  <img src="assets/figures/framework.png" width="100%" alt="SAIL-VL2 Framework">
+  <i> Overview of the SAIL-VL2 framework. The architecture is composed of a vision encoder that aligns visual inputs into the representation space of the LLM. A lightweight adapter further transforms visual embeddings into tokenized representations, which are jointly processed with linguistic embeddings for multimodal reasoning and prediction. SAIL-VL2 accommodates multiple LLM backbones, ensuring flexibility and scalability across model configurations.</i>
+</div>
+
+
+## 📚 Training Strategy
+### 🌟 Data construction
+
+<div align="center">
+  <img src="assets/figures/data.png" width="100%" alt="SAIL-VL2 Data">
+  <i> Data construction pipeline for SAIL-VL2 training. High-quality multimodal corpora are construted by curating and filtering open-source datasets, and generating synthetic data, with both components systematically organized to meet the requirements of different training stages.</i>
+</div>
+
+### 🌟 Pre-Train
+
+- **Basic Multimodal Pre-Training:** develops SAIL-VL2’s multimodal alignment via SAIL-ViT, LLM and a random MLP adapter, using 64M samples, AdaLRS and 2048 batch size.
+
+- **Multi-task Pre-Training:** strengthens SAIL-VL2’s visual and instruction-following abilities, unfreezes all params, adds instruction-tuning data, uses 180M samples, and skips AdaLRS.
+
+<div align="center">
+  <img src="assets/figures/mtpt_scaling.png" width="100%" alt="SAIL-VL2 MTPT Scaling">
+  <i> Scaling curves of SAIL-VL2-2B during the multi-task pre-training stage. Results are reported on overall benchmarks, natural-scene VQA datasets, and OCR VQA tasks. ’BMK Score’ denotes the average benchmark score.</i>
+</div>
+
+
+### 🌟 Post-Train
+
+- **Basic Supervised Fine-Tuning:** 4 phases; Model Soup merges homogeneous models.
+
+- **LongCoT Supervised Fine-tuning:** enhances the model’s step-by-step reasoning capabilities for complex problems.
+
+- **RL with Verifiable Rewards:** refines the model by optimizing it against a reward system focused on two primary objectives: the correctness of the final answer and adherence to the specified output format
+
+- **Think-Fusion Supervised Fine-tuning:** enhances the model’s reasoning capabilities while maintaining its broad general understanding.
+
+- **RL with a Mixed Reward System:** enhances the model’s reasoning capabilities through a RL stage
+
+## 📈 Experimental Results
+### 🌟 Performance of 2B series
+<div align="center">
+  <img src="assets/figures/performance_table_2b.png" width="100%" alt="SAIL-VL2 Performance">
+  <i> Overall comparison of the SAIL-VL2 series and existing open-source MLLMs (<4B).</i>
+</div>
+
+### 🌟 Performance of 8B series
+<div align="center">
+  <img src="assets/figures/performance_table_8b.png" width="100%" alt="SAIL-VL2 Performance">
+  <i> Overall comparison of the SAIL-VL2 series with existing open-source 8B MLLMs and closed-source models.</i>
+</div>
+
+### 🌟 Performance of Thinking-mode models
+<div align="center">
+  <img src="assets/figures/performance_rl.png" width="100%" alt="SAIL-VL2 Performance">
+  <i> Evaluation results on OpenCompass multimodal reasoning benchmarks.</i>
+</div>
+
+## 🙏 Acknowledge
+
+Our model is built upon numerous outstanding open-source projects, and we are grateful for their contributions. We extend special thanks to the InternVL team, Qwen team, and Apple team for their great base models, and to the BAAI team (Infinity-MM), MAmmoTH-VL team(MAmmoTH-VL-Insturction-12M) for their generous release of data, and to the OpenCompass team for their valuable benchmarks.
+
+## ✒️ Citation
+
+If you find our work helpful for your research, please consider citing our work.   
+
+```
+@article{dong2025scalable,
+  title={Scalable vision language model training via high quality data curation},
+  author={Dong, Hongyuan and Kang, Zijian and Yin, Weijie and Liang, Xiao and Feng, Chao and Ran, Jiao},
+  journal={arXiv preprint arXiv:2501.05952},
+  year={2025}
+}
+```
+
+
+## 📜 License
+
+This project is licensed under [Apache License 2.0](LICENSE).
+
+## 📧Contact
+
+If you have any question, please feel free to contact us: [email protected]
+

added_tokens.json

@@ -0,0 +1,37 @@
+{
+  "</box>": 151677,
+  "</img>": 151670,
+  "</quad>": 151673,
+  "</ref>": 151675,
+  "</think>": 151668,
+  "</tool_call>": 151658,
+  "</tool_response>": 151666,
+  "<IMG_CONTEXT>": 151671,
+  "<box>": 151676,
+  "<img>": 151669,
+  "<quad>": 151672,
+  "<ref>": 151674,
+  "<think>": 151667,
+  "<tool_call>": 151657,
+  "<tool_response>": 151665,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

chat_template.jinja

@@ -0,0 +1,30 @@
+{%- for message in messages -%}
+  {%- if loop.first and messages[0]['role'] != 'system' -%}
+    {{'<|im_start|>system\n你是由抖音内容理解组开发的多模态大模型，英文名叫UniVL, 是一个有用无害的人工智能助手。<|im_end|>'}}
+  {%- endif -%}
+  {%- if message['role'] == 'system' -%}
+    {{'<|im_system|>'}}
+  {%- endif -%}
+  {%- if message['role'] == 'user' -%}
+    {{'<|im_start|>user\n'}}
+  {%- endif -%}
+  {%- if message['role'] == 'assistant' -%}
+    {{'<|im_assistant|>'}}
+  {%- endif -%}
+
+  {%- if message['content'] is string -%}
+    {{- message['content'] + '<|im_end|>' -}}
+  {%- else -%}
+    {%- for content in message['content'] -%}
+      {%- if content['type'] == 'image' or 'image' in content or 'image_url' in content -%}
+        {{'<image>\n'}}
+      {%- else -%}
+        {{content['text']}}
+      {%- endif -%}
+    {%- endfor -%}
+    {{'<|im_end|>'}}
+  {%- endif -%}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+  {{'<|im_start|>assistant\n'}}
+{%- endif -%}

config.json

@@ -0,0 +1,193 @@
+{
+  "_commit_hash": null,
+  "architectures": [
+    "SAILVLModel"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_sailvl.SAILVLConfig",
+    "AutoModel": "modeling_sailvl.SAILVLModel",
+    "AutoModelForCausalLM": "modeling_sailvl.SAILVLModel"
+  },
+  "downsample_ratio": 0.5,
+  "dynamic_image_size": true,
+  "force_image_size": 448,
+  "llm_config": {
+    "_name_or_path": "/tmp/huggingface_cache/Qwen3-1.7B-Instruct",
+    "add_cross_attention": false,
+    "architectures": [
+      "Qwen3ForCausalLM"
+    ],
+    "attn_implementation": "flash_attention_2",
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": 151643,
+    "chunk_size_feed_forward": 0,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 151645,
+    "exponential_decay_length_penalty": null,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "head_dim": 128,
+    "hidden_act": "silu",
+    "hidden_size": 2048,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "initializer_range": 0.02,
+    "intermediate_size": 6144,
+    "is_decoder": false,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "max_position_embeddings": 40960,
+    "max_window_layers": 28,
+    "min_length": 0,
+    "model_type": "qwen3",
+    "no_repeat_ngram_size": 0,
+    "num_attention_heads": 16,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_hidden_layers": 28,
+    "num_key_value_heads": 8,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": 151643,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "rms_norm_eps": 1e-06,
+    "rope_scaling": null,
+    "rope_theta": 1000000,
+    "sep_token_id": null,
+    "sliding_window": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": true,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": "bfloat16",
+    "torchscript": false,
+    "transformers_version": "4.51.0",
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "use_cache": true,
+    "use_sliding_window": false,
+    "vocab_size": 151936
+  },
+  "max_dynamic_patch": 12,
+  "min_dynamic_patch": 1,
+  "model_type": "sailvl",
+  "pad2square": false,
+  "ps_version": "v2",
+  "select_layer": -1,
+  "template": "univl-chat",
+  "torch_dtype": "bfloat16",
+  "transformers_version": null,
+  "use_backbone_lora": 0,
+  "use_llm_lora": 0,
+  "use_thumbnail": true,
+  "vision_config": {
+    "add_cross_attention": false,
+    "architectures": [
+      "SAILViTModel"
+    ],
+    "attention_dropout": 0.0,
+    "auto_map": {
+      "AutoConfig": "configuration_sailvit.SAILViTConfig",
+      "AutoModel": "modeling_sailvit.SAILViTModel"
+    },
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": null,
+    "chunk_size_feed_forward": 0,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": null,
+    "exponential_decay_length_penalty": null,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "hidden_size": 1536,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "image_size": 448,
+    "intermediate_size": 4096,
+    "is_decoder": false,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "min_length": 0,
+    "model_type": "sailvit",
+    "no_repeat_ngram_size": 0,
+    "num_attention_heads": 12,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_channels": 3,
+    "num_hidden_layers": 24,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": null,
+    "patch_size": 14,
+    "prefix": null,
+    "problem_type": null,
+    "projection_dropout": 0.0,
+    "pruned_heads": {},
+    "qkv_bias": false,
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "rms_norm_eps": 1e-05,
+    "sep_token_id": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": true,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": "bfloat16",
+    "torchscript": false,
+    "transformers_version": "4.51.0",
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "use_bias": false
+  }
+}

configuration_qwen3.py

@@ -0,0 +1,212 @@
+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Qwen3 model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class Qwen3Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Qwen3Model`]. It is used to instantiate a
+    Qwen3 model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of
+    Qwen3-8B [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B).
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 151936):
+            Vocabulary size of the Qwen3 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Qwen3Model`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 22016):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 32):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
+        head_dim (`int`, *optional*, defaults to 128):
+            The attention head dimension.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 32768):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'llama3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        use_sliding_window (`bool`, *optional*, defaults to `False`):
+            Whether to use sliding window attention.
+        sliding_window (`int`, *optional*, defaults to 4096):
+            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
+        max_window_layers (`int`, *optional*, defaults to 28):
+            The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    ```python
+    >>> from transformers import Qwen3Model, Qwen3Config
+
+    >>> # Initializing a Qwen3 style configuration
+    >>> configuration = Qwen3Config()
+
+    >>> # Initializing a model from the Qwen3-8B style configuration
+    >>> model = Qwen3Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "qwen3"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    # Default tensor parallel plan for base model `Qwen3`
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=4096,
+        intermediate_size=22016,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        head_dim=128,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        use_sliding_window=False,
+        sliding_window=4096,
+        max_window_layers=28,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        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.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window  # we check `use_sliding_window` in the modeling code
+        self.max_window_layers = max_window_layers
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.head_dim = head_dim
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, move it to 'rope_type'.
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+        rope_config_validation(self)
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+__all__ = ["Qwen3Config"]

configuration_sailvit.py

@@ -0,0 +1,61 @@
+# copied from https://huggingface.co/apple/aimv2-huge-patch14-448
+from typing import Any
+
+from transformers.configuration_utils import PretrainedConfig
+
+__all__ = ["SAILViTConfig"]
+
+
+class SAILViTConfig(PretrainedConfig):
+    """This is the configuration class to store the configuration of an [`SAIlViTModel`].
+    Instantiating a configuration with the defaults will yield a similar configuration
+    to that of the [apple/aimv2-large-patch14-224](https://huggingface.co/apple/aimv2-large-patch14-224).
+    Args:
+        hidden_size: Dimension of the hidden representations.
+        intermediate_size: Dimension of the SwiGLU representations.
+        num_hidden_layers: Number of hidden layers in the Transformer.
+        num_attention_heads: Number of attention heads for each attention layer
+            in the Transformer.
+        num_channels: Number of input channels.
+        image_size: Image size.
+        patch_size: Patch size.
+        rms_norm_eps: Epsilon value used for the RMS normalization layer.
+        attention_dropout: Dropout ratio for attention probabilities.
+        projection_dropout: Dropout ratio for the projection layer after the attention.
+        qkv_bias: Whether to add a bias to the queries, keys and values.
+        use_bias: Whether to add a bias in the feed-forward and projection layers.
+        kwargs: Keyword arguments for the [`PretrainedConfig`].
+    """
+
+    model_type: str = "sailvit"
+
+    def __init__(
+        self,
+        hidden_size: int = 1024,
+        intermediate_size: int = 2816,
+        num_hidden_layers: int = 24,
+        num_attention_heads: int = 8,
+        num_channels: int = 3,
+        image_size: int = 224,
+        patch_size: int = 14,
+        rms_norm_eps: float = 1e-5,
+        attention_dropout: float = 0.0,
+        projection_dropout: float = 0.0,
+        qkv_bias: bool = False,
+        use_bias: bool = False,
+        **kwargs: Any,
+    ):
+        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.image_size = image_size
+        self.attention_dropout = attention_dropout
+        self.rms_norm_eps = rms_norm_eps
+
+        self.projection_dropout = projection_dropout
+        self.qkv_bias = qkv_bias
+        self.use_bias = use_bias

configuration_sailvl.py

@@ -0,0 +1,101 @@
+# --------------------------------------------------------
+# adapted from internvl_chat/internvl/model/internvl_chat/configuration_internvl_chat.py
+# --------------------------------------------------------
+
+import copy
+
+from transformers import LlamaConfig
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+from .configuration_sailvit import SAILViTConfig
+# from .configuration_qwen2 import Qwen2Config
+from .configuration_qwen3 import Qwen3Config
+
+logger = logging.get_logger(__name__)
+
+
+class SAILVLConfig(PretrainedConfig):
+    model_type = 'sailvl'
+    is_composition = True
+
+    def __init__(
+        self,
+        vision_config=None,
+        llm_config=None,
+        use_backbone_lora=0,
+        use_llm_lora=0,
+        pad2square=False,
+        select_layer=-4,
+        force_image_size=None,
+        downsample_ratio=0.5,
+        template=None,
+        dynamic_image_size=False,
+        use_thumbnail=False,
+        ps_version='v1',
+        min_dynamic_patch=1,
+        max_dynamic_patch=6,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info('vision_config is None. Initializing the SAILViTConfig with default values.')
+
+        if llm_config is None:
+            llm_config = {'architectures': ['Qwen2ForCausalLM']}
+            logger.info('llm_config is None. Initializing the LlamaConfig config with default values (`LlamaConfig`).')
+
+        self.vision_config = SAILViTConfig(**vision_config)
+        if llm_config['architectures'][0] == 'LlamaForCausalLM':
+            self.llm_config = LlamaConfig(**llm_config)
+        elif llm_config['architectures'][0] == 'Qwen3ForCausalLM':
+            self.llm_config = Qwen3Config(**llm_config)
+        else:
+            raise ValueError('Unsupported architecture: {}'.format(llm_config['architectures'][0]))
+        
+        self.use_backbone_lora = use_backbone_lora
+        self.use_llm_lora = use_llm_lora
+        self.pad2square = pad2square
+        self.select_layer = select_layer
+        self.force_image_size = force_image_size
+        self.downsample_ratio = downsample_ratio
+        self.template = template
+        self.dynamic_image_size = dynamic_image_size
+        self.use_thumbnail = use_thumbnail
+        self.ps_version = ps_version  # pixel shuffle version
+        self.min_dynamic_patch = min_dynamic_patch
+        self.max_dynamic_patch = max_dynamic_patch
+
+        logger.info(f'vision_select_layer: {self.select_layer}')
+        logger.info(f'ps_version: {self.ps_version}')
+        logger.info(f'min_dynamic_patch: {self.min_dynamic_patch}')
+        logger.info(f'max_dynamic_patch: {self.max_dynamic_patch}')
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+
+        Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+        
+        output['vision_config'] = self.vision_config.to_dict()
+        output['llm_config'] = self.llm_config.to_dict()
+        output['model_type'] = self.__class__.model_type
+        output['use_backbone_lora'] = self.use_backbone_lora
+        output['use_llm_lora'] = self.use_llm_lora
+        output['pad2square'] = self.pad2square
+        output['select_layer'] = self.select_layer
+        output['force_image_size'] = self.force_image_size
+        output['downsample_ratio'] = self.downsample_ratio
+        output['template'] = self.template
+        output['dynamic_image_size'] = self.dynamic_image_size
+        output['use_thumbnail'] = self.use_thumbnail
+        output['ps_version'] = self.ps_version
+        output['min_dynamic_patch'] = self.min_dynamic_patch
+        output['max_dynamic_patch'] = self.max_dynamic_patch
+
+        return output

conversation.py

@@ -0,0 +1,424 @@
+# ------------------------------------------------------------
+# adapted from InternVL/internvl_chat/internvl/conversation.py
+# ------------------------------------------------------------
+
+import dataclasses
+from enum import IntEnum, auto
+from typing import Any, Dict, List, Tuple, Union
+
+
+class SeparatorStyle(IntEnum):
+    """Separator styles."""
+
+    ADD_COLON_SINGLE = auto()
+    ADD_COLON_TWO = auto()
+    ADD_COLON_SPACE_SINGLE = auto()
+    NO_COLON_SINGLE = auto()
+    NO_COLON_TWO = auto()
+    ADD_NEW_LINE_SINGLE = auto()
+    LLAMA2 = auto()
+    CHATGLM = auto()
+    CHATML = auto()
+    CHATINTERN = auto()
+    DOLLY = auto()
+    RWKV = auto()
+    PHOENIX = auto()
+    ROBIN = auto()
+    FALCON_CHAT = auto()
+    CHATGLM3 = auto()
+    INTERNVL_ZH = auto()
+    MPT = auto()
+
+
+@dataclasses.dataclass
+class Conversation:
+    """A class that manages prompt templates and keeps all conversation history."""
+
+    # The name of this template
+    name: str
+    # The template of the system prompt
+    system_template: str = '{system_message}'
+    # The system message
+    system_message: str = ''
+    # The names of two roles
+    roles: Tuple[str] = ('USER', 'ASSISTANT')
+    # All messages. Each item is (role, message).
+    messages: List[List[str]] = ()
+    # The number of few shot examples
+    offset: int = 0
+    # The separator style and configurations
+    sep_style: SeparatorStyle = SeparatorStyle.ADD_COLON_SINGLE
+    sep: str = '\n'
+    sep2: str = None
+    # Stop criteria (the default one is EOS token)
+    stop_str: Union[str, List[str]] = None
+    # Stops generation if meeting any token in this list
+    stop_token_ids: List[int] = None
+
+    def get_prompt(self) -> str:
+        """Get the prompt for generation."""
+        system_prompt = self.system_template.format(system_message=self.system_message)
+        if self.sep_style == SeparatorStyle.ADD_COLON_SINGLE:
+            ret = system_prompt + self.sep
+            for role, message in self.messages:
+                if message:
+                    ret += role + ': ' + message + self.sep
+                else:
+                    ret += role + ':'
+            return ret
+        elif self.sep_style == SeparatorStyle.ADD_COLON_TWO:
+            seps = [self.sep, self.sep2]
+            ret = system_prompt + seps[0]
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    ret += role + ': ' + message + seps[i % 2]
+                else:
+                    ret += role + ':'
+            return ret
+        elif self.sep_style == SeparatorStyle.ADD_COLON_SPACE_SINGLE:
+            ret = system_prompt + self.sep
+            for role, message in self.messages:
+                if message:
+                    ret += role + ': ' + message + self.sep
+                else:
+                    ret += role + ': '  # must be end with a space
+            return ret
+        elif self.sep_style == SeparatorStyle.ADD_NEW_LINE_SINGLE:
+            ret = '' if system_prompt == '' else system_prompt + self.sep
+            for role, message in self.messages:
+                if message:
+                    ret += role + '\n' + message + self.sep
+                else:
+                    ret += role + '\n'
+            return ret
+        elif self.sep_style == SeparatorStyle.NO_COLON_SINGLE:
+            ret = system_prompt
+            for role, message in self.messages:
+                if message:
+                    ret += role + message + self.sep
+                else:
+                    ret += role
+            return ret
+        elif self.sep_style == SeparatorStyle.NO_COLON_TWO:
+            seps = [self.sep, self.sep2]
+            ret = system_prompt
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    ret += role + message + seps[i % 2]
+                else:
+                    ret += role
+            return ret
+        elif self.sep_style == SeparatorStyle.RWKV:
+            ret = system_prompt
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    ret += (
+                        role
+                        + ': '
+                        + message.replace('\r\n', '\n').replace('\n\n', '\n')
+                    )
+                    ret += '\n\n'
+                else:
+                    ret += role + ':'
+            return ret
+        elif self.sep_style == SeparatorStyle.LLAMA2:
+            seps = [self.sep, self.sep2]
+            if self.system_message:
+                ret = system_prompt
+            else:
+                ret = '[INST] '
+            for i, (role, message) in enumerate(self.messages):
+                tag = self.roles[i % 2]
+                if message:
+                    if i == 0:
+                        ret += message + ' '
+                    else:
+                        ret += tag + ' ' + message + seps[i % 2]
+                else:
+                    ret += tag
+            return ret
+        elif self.sep_style == SeparatorStyle.CHATGLM:
+            # source: https://huggingface.co/THUDM/chatglm-6b/blob/1d240ba371910e9282298d4592532d7f0f3e9f3e/modeling_chatglm.py#L1302-L1308
+            # source2: https://huggingface.co/THUDM/chatglm2-6b/blob/e186c891cf64310ac66ef10a87e6635fa6c2a579/modeling_chatglm.py#L926
+            round_add_n = 1 if self.name == 'chatglm2' else 0
+            if system_prompt:
+                ret = system_prompt + self.sep
+            else:
+                ret = ''
+
+            for i, (role, message) in enumerate(self.messages):
+                if i % 2 == 0:
+                    ret += f'[Round {i//2 + round_add_n}]{self.sep}'
+
+                if message:
+                    ret += f'{role}：{message}{self.sep}'
+                else:
+                    ret += f'{role}：'
+            return ret
+        elif self.sep_style == SeparatorStyle.CHATML:
+            ret = '' if system_prompt == '' else system_prompt + self.sep + '\n'
+            for role, message in self.messages:
+                if message:
+                    ret += role + '\n' + message + self.sep + '\n'
+                else:
+                    ret += role + '\n'
+            return ret
+        elif self.sep_style == SeparatorStyle.CHATGLM3:
+            ret = ''
+            if self.system_message:
+                ret += system_prompt
+            for role, message in self.messages:
+                if message:
+                    ret += role + '\n' + ' ' + message
+                else:
+                    ret += role
+            return ret
+        elif self.sep_style == SeparatorStyle.CHATINTERN:
+            # source: https://huggingface.co/internlm/internlm-chat-7b-8k/blob/bd546fa984b4b0b86958f56bf37f94aa75ab8831/modeling_internlm.py#L771
+            seps = [self.sep, self.sep2]
+            ret = system_prompt
+            for i, (role, message) in enumerate(self.messages):
+                # if i % 2 == 0:
+                #     ret += "<s>"
+                if message:
+                    ret += role + ':' + message + seps[i % 2] + '\n'
+                else:
+                    ret += role + ':'
+            return ret
+        elif self.sep_style == SeparatorStyle.DOLLY:
+            seps = [self.sep, self.sep2]
+            ret = system_prompt
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    ret += role + ':\n' + message + seps[i % 2]
+                    if i % 2 == 1:
+                        ret += '\n\n'
+                else:
+                    ret += role + ':\n'
+            return ret
+        elif self.sep_style == SeparatorStyle.PHOENIX:
+            ret = system_prompt
+            for role, message in self.messages:
+                if message:
+                    ret += role + ': ' + '<s>' + message + '</s>'
+                else:
+                    ret += role + ': ' + '<s>'
+            return ret
+        elif self.sep_style == SeparatorStyle.ROBIN:
+            ret = system_prompt + self.sep
+            for role, message in self.messages:
+                if message:
+                    ret += role + ':\n' + message + self.sep
+                else:
+                    ret += role + ':\n'
+            return ret
+        elif self.sep_style == SeparatorStyle.FALCON_CHAT:
+            ret = ''
+            if self.system_message:
+                ret += system_prompt + self.sep
+            for role, message in self.messages:
+                if message:
+                    ret += role + ': ' + message + self.sep
+                else:
+                    ret += role + ':'
+
+            return ret
+        elif self.sep_style == SeparatorStyle.INTERNVL_ZH:
+            seps = [self.sep, self.sep2]
+            ret = self.system_message + seps[0]
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    ret += role + ': ' + message + seps[i % 2]
+                else:
+                    ret += role + ':'
+            return ret
+        elif self.sep_style == SeparatorStyle.MPT:
+            ret = system_prompt + self.sep if system_prompt else ''
+            for role, message in self.messages:
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    ret += role + message + self.sep
+                else:
+                    ret += role
+            return ret
+        else:
+            raise ValueError(f'Invalid style: {self.sep_style}')
+
+    def set_system_message(self, system_message: str):
+        """Set the system message."""
+        self.system_message = system_message
+
+    def append_message(self, role: str, message: str):
+        """Append a new message."""
+        self.messages.append([role, message])
+
+    def update_last_message(self, message: str):
+        """Update the last output.
+
+        The last message is typically set to be None when constructing the prompt,
+        so we need to update it in-place after getting the response from a model.
+        """
+        self.messages[-1][1] = message
+
+    def to_gradio_chatbot(self):
+        """Convert the conversation to gradio chatbot format."""
+        ret = []
+        for i, (role, msg) in enumerate(self.messages[self.offset :]):
+            if i % 2 == 0:
+                ret.append([msg, None])
+            else:
+                ret[-1][-1] = msg
+        return ret
+
+    def to_openai_api_messages(self):
+        """Convert the conversation to OpenAI chat completion format."""
+        ret = [{'role': 'system', 'content': self.system_message}]
+
+        for i, (_, msg) in enumerate(self.messages[self.offset :]):
+            if i % 2 == 0:
+                ret.append({'role': 'user', 'content': msg})
+            else:
+                if msg is not None:
+                    ret.append({'role': 'assistant', 'content': msg})
+        return ret
+
+    def copy(self):
+        return Conversation(
+            name=self.name,
+            system_template=self.system_template,
+            system_message=self.system_message,
+            roles=self.roles,
+            messages=[[x, y] for x, y in self.messages],
+            offset=self.offset,
+            sep_style=self.sep_style,
+            sep=self.sep,
+            sep2=self.sep2,
+            stop_str=self.stop_str,
+            stop_token_ids=self.stop_token_ids,
+        )
+
+    def dict(self):
+        return {
+            'template_name': self.name,
+            'system_message': self.system_message,
+            'roles': self.roles,
+            'messages': self.messages,
+            'offset': self.offset,
+        }
+
+
+# A global registry for all conversation templates
+conv_templates: Dict[str, Conversation] = {}
+
+
+def register_conv_template(template: Conversation, override: bool = False):
+    """Register a new conversation template."""
+    if not override:
+        assert (
+            template.name not in conv_templates
+        ), f'{template.name} has been registered.'
+
+    conv_templates[template.name] = template
+
+
+def get_conv_template(name: str) -> Conversation:
+    """Get a conversation template."""
+    return conv_templates[name].copy()
+
+
+# Both Hermes-2 and internlm2-chat are chatml-format conversation templates. The difference
+# is that during training, the preprocessing function for the Hermes-2 template doesn't add
+# <s> at the beginning of the tokenized sequence, while the internlm2-chat template does.
+# Therefore, they are completely equivalent during inference.
+register_conv_template(
+    Conversation(
+        name='Hermes-2',
+        system_template='<|im_start|>system\n{system_message}',
+        # note: The new system prompt was not used here to avoid changes in benchmark performance.
+        # system_message='我是书生·万象，英文名是InternVL，是由上海人工智能实验室及多家合作单位联合开发的多模态大语言模型。',
+        system_message='你是由上海人工智能实验室联合商汤科技开发的书生多模态大模型，英文名叫InternVL, 是一个有用无害的人工智能助手。',
+        roles=('<|im_start|>user\n', '<|im_start|>assistant\n'),
+        sep_style=SeparatorStyle.MPT,
+        sep='<|im_end|>',
+        stop_token_ids=[
+            2,
+            6,
+            7,
+            8,
+        ],
+        stop_str='<|endoftext|>',
+    )
+)
+
+
+register_conv_template(
+    Conversation(
+        name='univl-chat',
+        system_template='<|im_start|>system\n{system_message}',
+        # note: The new system prompt was not used here to avoid changes in benchmark performance.
+        system_message='你是由抖音内容理解组开发的多模态大模型，英文名叫UniVL, 是一个有用无害的人工智能助手。',
+        roles=('<|im_start|>user\n', '<|im_start|>assistant\n'),
+        sep_style=SeparatorStyle.MPT,
+        sep='<|im_end|>',
+        stop_token_ids=[
+            2,
+            92543,
+            92542
+        ]
+    )
+)
+
+
+register_conv_template(
+    Conversation(
+        name='internlm2-chat',
+        system_template='<|im_start|>system\n{system_message}',
+        # note: The new system prompt was not used here to avoid changes in benchmark performance.
+        # system_message='我是书生·万象，英文名是InternVL，是由上海人工智能实验室及多家合作单位联合开发的多模态大语言模型。',
+        system_message='你是由上海人工智能实验室联合商汤科技开发的书生多模态大模型，英文名叫InternVL, 是一个有用无害的人工智能助手。',
+        roles=('<|im_start|>user\n', '<|im_start|>assistant\n'),
+        sep_style=SeparatorStyle.MPT,
+        sep='<|im_end|>',
+        stop_token_ids=[
+            2,
+            92543,
+            92542
+        ]
+    )
+)
+
+register_conv_template(
+    Conversation(
+        name='internlm2-plain',
+        system_template='',
+        system_message='',
+        roles=('<|im_start|>user\n', '<|im_start|>assistant\n'),
+        sep_style=SeparatorStyle.MPT,
+        sep='<|im_end|>',
+        stop_token_ids=[
+            2,
+            92543,
+            92542
+        ]
+    )
+)
+
+
+register_conv_template(
+    Conversation(
+        name='phi3-chat',
+        system_template='<|system|>\n{system_message}',
+        # note: The new system prompt was not used here to avoid changes in benchmark performance.
+        # system_message='我是书生·万象，英文名是InternVL，是由上海人工智能实验室及多家合作单位联合开发的多模态大语言模型。',
+        system_message='你是由上海人工智能实验室联合商汤科技开发的书生多模态大模型，英文名叫InternVL, 是一个有用无害的人工智能助手。',
+        roles=('<|user|>\n', '<|assistant|>\n'),
+        sep_style=SeparatorStyle.MPT,
+        sep='<|end|>',
+        stop_token_ids=[
+            2,
+            32000,
+            32007
+        ]
+    )
+)

generation_config.json

@@ -0,0 +1,4 @@
+{
+  "_from_model_config": true,
+  "transformers_version": "4.51.0"
+}

image_processing_sailvl.py

@@ -0,0 +1,262 @@
+"""Image processor class for KimiVL."""
+
+import math
+import numpy as np
+from PIL import Image
+from typing import Optional, Union
+
+import torch
+
+import torchvision.transforms as T
+from torchvision.transforms.functional import InterpolationMode
+from transformers.image_utils import ImageInput, make_list_of_images, valid_images
+from transformers.image_processing_utils import BaseImageProcessor, BatchFeature
+from transformers.utils import TensorType
+
+IMAGENET_MEAN = (0.485, 0.456, 0.406)
+IMAGENET_STD = (0.229, 0.224, 0.225)
+
+def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
+    best_ratio_diff = float('inf')
+    best_ratio = (1, 1)
+    area = width * height
+    for ratio in target_ratios:
+        target_aspect_ratio = ratio[0] / ratio[1]
+        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
+        if ratio_diff < best_ratio_diff:
+            best_ratio_diff = ratio_diff
+            best_ratio = ratio
+        elif ratio_diff == best_ratio_diff:
+            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
+                best_ratio = ratio
+    return best_ratio
+
+
+def dynamic_preprocess(image, min_num=1, max_num=6, image_size=448, use_thumbnail=False):
+    orig_width, orig_height = image.size
+    aspect_ratio = orig_width / orig_height
+
+    # calculate the existing image aspect ratio
+    target_ratios = set(
+        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
+        i * j <= max_num and i * j >= min_num)
+    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+    # find the closest aspect ratio to the target
+    target_aspect_ratio = find_closest_aspect_ratio(
+        aspect_ratio, target_ratios, orig_width, orig_height, image_size)
+
+    # calculate the target width and height
+    target_width = image_size * target_aspect_ratio[0]
+    target_height = image_size * target_aspect_ratio[1]
+    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
+
+    # resize the image
+    resized_img = image.resize((target_width, target_height))
+    processed_images = []
+    for i in range(blocks):
+        box = (
+            (i % (target_width // image_size)) * image_size,
+            (i // (target_width // image_size)) * image_size,
+            ((i % (target_width // image_size)) + 1) * image_size,
+            ((i // (target_width // image_size)) + 1) * image_size
+        )
+        # split the image
+        split_img = resized_img.crop(box)
+        processed_images.append(split_img)
+    assert len(processed_images) == blocks
+    if use_thumbnail and len(processed_images) != 1:
+        thumbnail_img = image.resize((image_size, image_size))
+        processed_images.append(thumbnail_img)
+    return processed_images
+
+def dynamic_preprocess_msac1(image, min_num=1, max_num=6, image_size=448, use_thumbnail=False):
+    orig_width, orig_height = image.size
+    aspect_ratio = orig_width / orig_height
+
+    # calculate the existing image aspect ratio
+    target_ratios = set(
+        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
+        i * j <= max_num and i * j >= min_num)
+    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+    # find the closest aspect ratio to the target
+    target_aspect_ratio = find_closest_aspect_ratio(
+        aspect_ratio, target_ratios, orig_width, orig_height, image_size)
+
+    # calculate the target width and height
+    target_width = image_size * target_aspect_ratio[0]
+    target_height = image_size * target_aspect_ratio[1]
+    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
+
+    # resize the image
+    resized_img = image.resize((target_width, target_height))
+    processed_images = []
+    for i in range(blocks):
+        box = (
+            (i % (target_width // image_size)) * image_size,
+            (i // (target_width // image_size)) * image_size,
+            ((i % (target_width // image_size)) + 1) * image_size,
+            ((i // (target_width // image_size)) + 1) * image_size
+        )
+        # split the image
+        split_img = resized_img.crop(box)
+        processed_images.append(split_img)
+    assert len(processed_images) == blocks
+    if use_thumbnail and len(processed_images) != 1:
+        thumbnail_img = image.resize((image_size, image_size))
+        processed_images.append(thumbnail_img)
+    return processed_images, target_aspect_ratio
+
+def dynamic_preprocess_msac2(image, min_num=1, max_num=6, image_size=448, use_thumbnail=False, prior_aspect_ratio=None):
+    orig_width, orig_height = image.size
+    aspect_ratio = orig_width / orig_height
+
+    # calculate the existing image aspect ratio
+    target_ratios = set(
+        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
+        i * j <= max_num and i * j >= min_num)
+    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
+
+    new_target_ratios = []
+    if prior_aspect_ratio is not None:
+        for i in target_ratios:
+            if prior_aspect_ratio[0]%i[0] != 0 or prior_aspect_ratio[1]%i[1] != 0:
+                new_target_ratios.append(i)
+            else:
+                continue
+
+    # find the closest aspect ratio to the target
+    target_aspect_ratio = find_closest_aspect_ratio(
+        aspect_ratio, new_target_ratios, orig_width, orig_height, image_size)
+
+    # calculate the target width and height
+    target_width = image_size * target_aspect_ratio[0]
+    target_height = image_size * target_aspect_ratio[1]
+    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
+
+    # resize the image
+    resized_img = image.resize((target_width, target_height))
+    processed_images = []
+    for i in range(blocks):
+        box = (
+            (i % (target_width // image_size)) * image_size,
+            (i // (target_width // image_size)) * image_size,
+            ((i % (target_width // image_size)) + 1) * image_size,
+            ((i // (target_width // image_size)) + 1) * image_size
+        )
+        # split the image
+        split_img = resized_img.crop(box)
+        processed_images.append(split_img)
+    assert len(processed_images) == blocks
+    if use_thumbnail and len(processed_images) != 1:
+        thumbnail_img = image.resize((image_size, image_size))
+        processed_images.append(thumbnail_img)
+    return processed_images
+
+
+class SAILVLImageProcessor(BaseImageProcessor):
+    model_type = "sailvl"
+
+    def __init__(
+        self,
+        patch_size: int = 14,
+        image_mean: tuple[float, float, float] = IMAGENET_MEAN,
+        image_std: tuple[float, float, float] = IMAGENET_STD,
+        max_dynamic_patch: int = 10,
+        image_size: int = 448,
+        use_msac: bool = False,
+
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.patch_size = patch_size
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.max_dynamic_patch = max_dynamic_patch
+        self.image_size = image_size
+        self.use_msac = use_msac
+    
+    def build_transform(self, input_size):
+        MEAN, STD = self.image_mean, self.image_std
+        transform = T.Compose([
+            T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
+            T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
+            T.ToTensor(),
+            T.Normalize(mean=MEAN, std=STD)
+        ])
+        return transform
+
+    def load_image(self, image, input_size=448, max_num=6, upscale=False):
+        # image = Image.open(image_file).convert('RGB')
+        if upscale:
+            image = image.resize((image.width * 2, image.height * 2), Image.BILINEAR)
+        transform = self.build_transform(input_size=input_size)
+        images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
+        pixel_values = [transform(image) for image in images]
+        pixel_values = torch.stack(pixel_values)
+        return pixel_values
+    
+    def load_image_msac(self, image, input_size=448, max_num=6, upscale=False):
+        # image = Image.open(image_file).convert('RGB')
+        if upscale:
+            image = image.resize((image.width * 2, image.height * 2), Image.BILINEAR)
+        transform = self.build_transform(input_size=input_size)
+        images,target_aspect_ratio = dynamic_preprocess_msac1(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
+        images = images[:-1] + dynamic_preprocess_msac2(image,max_num=max_num,image_size=input_size,use_thumbnail=False,prior_aspect_ratio=target_aspect_ratio) + images[-1:]
+
+        pixel_values = [transform(image) for image in images]
+        pixel_values = torch.stack(pixel_values)
+        return pixel_values
+
+    def preprocess(
+        self,
+        images: ImageInput,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+    ) -> BatchFeature:
+        images = make_list_of_images(images)
+
+        if not valid_images(images):
+            raise ValueError(
+                "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
+                "torch.Tensor, tf.Tensor or jax.ndarray."
+            )
+        # print('图片个数：',len(images))
+        image_num = len(images)
+        if image_num > 1:
+            # image_path = [x['value'] for x in message if x['type'] == 'image']
+            num_patches_list = []
+            pixel_values_list = []
+            for image_idx, image_pil in enumerate(images):
+                upscale_flag = False
+                curr_pixel_values = self.load_image(
+                    image_pil, max_num=self.max_dynamic_patch, upscale=upscale_flag, input_size=self.image_size).cuda().to(torch.bfloat16)
+                num_patches_list.append(curr_pixel_values.size(0))
+                pixel_values_list.append(curr_pixel_values)
+            pixel_values = torch.cat(pixel_values_list, dim=0)
+            
+        elif image_num == 1:
+            # image_path = [x['value'] for x in message if x['type'] == 'image'][0]
+            image_pil = images[0]
+            upscale_flag = False
+            if self.use_msac:
+                pixel_values = self.load_image_msac(
+                image_pil, max_num=self.max_dynamic_patch, upscale=upscale_flag, input_size=self.image_size).cuda().to(torch.bfloat16)
+            else:
+                pixel_values = self.load_image(
+                    image_pil, max_num=self.max_dynamic_patch, upscale=upscale_flag, input_size=self.image_size).cuda().to(torch.bfloat16)
+            num_patches_list = [pixel_values.size(0)]
+        else:
+            pixel_values = None
+            num_patches_list = None
+
+        # pixel_values, image_grid_hws = [], []
+        # for image in images:
+        #     patches, image_grid_hw = self._preprocess(image)
+        #     pixel_values.append(patches)
+        #     image_grid_hws.append(image_grid_hw)
+        # pixel_values = torch.concat(pixel_values, dim=0)
+        # image_grid_hws = np.array(image_grid_hws)
+        data = {"pixel_values": pixel_values, "num_patches_list": num_patches_list}
+
+        return BatchFeature(data=data, tensor_type=return_tensors)

modeling_qwen3.py

@@ -0,0 +1,1247 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/qwen3/modular_qwen3.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_qwen3.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2025 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from functools import partial
+from typing import Callable, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache
+from transformers.generation import GenerationMixin
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.utils import (
+    LossKwargs,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    can_return_tuple,
+    logging,
+    replace_return_docstrings,
+)
+from transformers.utils.deprecation import deprecate_kwarg
+from .configuration_qwen3 import Qwen3Config
+# from .spec_sdpa_attention import spec_sdpa_attention_forward
+
+from typing import Optional, Tuple
+
+import torch
+try:
+    import torch_npu
+except:
+    print('Using N* GPU...')
+import math
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def spec_sdpa_attention_forward(
+    module: torch.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: Optional[bool] = None,
+    **kwargs,
+) -> Tuple[torch.Tensor, None]:
+    if hasattr(module, "num_key_value_groups"):
+        key = repeat_kv(key, module.num_key_value_groups)
+        value = repeat_kv(value, module.num_key_value_groups)
+
+    causal_mask = attention_mask
+    if attention_mask is not None and causal_mask.ndim == 4:
+        causal_mask = causal_mask[:, :, :, : key.shape[-2]]
+
+    # SDPA with memory-efficient backend is bugged with non-contiguous inputs and custom attn_mask for some torch versions
+    # Reference: https://github.com/pytorch/pytorch/issues/112577.
+    query = query.contiguous()
+    key = key.contiguous()
+    value = value.contiguous()
+
+    # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+    # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
+    # Note that it is important to check first for the shape, otherwise compile will fail with `argument 'is_causal' must be bool, not SymBool`
+    if is_causal is None:
+        is_causal = query.shape[2] > 1 and causal_mask is None
+
+    # Shapes (e.g. query.shape[2]) are tensors during jit tracing, resulting in `is_causal` being a tensor.
+    # We convert it to a bool for the SDPA kernel that only accepts bools.
+
+    if attention_mask is None:
+        atten_mask_npu = torch.triu(torch.ones([query.size(-2), 
+                                                query.size(-2)]), diagonal=1).bool().to(query.device)
+    elif attention_mask.dtype == torch.bool:
+        atten_mask_npu = torch.logical_not(attention_mask.bool()).to(attention_mask.device) # atten_mask需要取反
+    else:
+        atten_mask_npu = attention_mask.bool().to(attention_mask.device)
+
+    if torch.jit.is_tracing() and isinstance(is_causal, torch.Tensor):
+        is_causal = is_causal.item()
+
+    # attn_output = torch.nn.functional.scaled_dot_product_attention(
+    #     query,
+    #     key,
+    #     value,
+    #     attn_mask=causal_mask,
+    #     dropout_p=dropout,
+    #     scale=scaling,
+    #     is_causal=is_causal,
+    # )
+
+    head_num = query.shape[1]
+    attn_output = torch_npu.npu_fusion_attention(
+                    query, key, value, head_num, input_layout="BNSD", 
+                    pse=None,
+                    atten_mask=atten_mask_npu,
+                    scale=1.0 / math.sqrt(query.shape[-1]),
+                    pre_tockens=2147483647,
+                    next_tockens=2147483647,
+                    keep_prob=1
+                )[0]
+
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, None
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "Qwen/Qwen3-8B"
+_CONFIG_FOR_DOC = "Qwen3Config"
+
+
+class Qwen3RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Qwen3RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+class Qwen3MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+class Qwen3Attention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: Qwen3Config, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+
+        self.q_proj = nn.Linear(
+            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.k_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.v_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.o_proj = nn.Linear(
+            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
+        )
+        self.q_norm = Qwen3RMSNorm(self.head_dim, eps=config.rms_norm_eps)  # unlike olmo, only on the head dim!
+        self.k_norm = Qwen3RMSNorm(self.head_dim, eps=config.rms_norm_eps)  # thus post q_norm does not need reshape
+        self.sliding_window = config.sliding_window
+        if not (
+            self.config.use_sliding_window
+            and getattr(self.config, "sliding_window", None) is not None
+            and self.layer_idx >= self.config.max_window_layers
+        ):
+            self.sliding_window = None
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+        key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
+                logger.warning_once(
+                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
+                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+                )
+            else:
+                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            sliding_window=self.sliding_window,  # diff with Llama
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class Qwen3DecoderLayer(nn.Module):
+    def __init__(self, config: Qwen3Config, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = Qwen3Attention(config=config, layer_idx=layer_idx)
+        self.mlp = Qwen3MLP(config)
+        self.input_layernorm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        if (
+            config.sliding_window and config._attn_implementation != "flash_attention_2"
+        ):  # diff with Llama is this warning
+            logger.warning_once(
+                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
+                "unexpected results may be encountered."
+            )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        return outputs
+
+
+class Qwen3RotaryEmbedding(nn.Module):
+    def __init__(self, config: Qwen3Config, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    @torch.no_grad()
+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+QWEN3_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`Qwen3Config`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare Qwen3 Model outputting raw hidden-states without any specific head on top.",
+    QWEN3_START_DOCSTRING,
+)
+class Qwen3PreTrainedModel(PreTrainedModel):
+    config_class = Qwen3Config
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Qwen3DecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = True
+    _supports_attention_backend = True
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+QWEN3_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+            the complete sequence length.
+"""
+
+
+@add_start_docstrings(
+    "The bare Qwen3 Model outputting raw hidden-states without any specific head on top.",
+    QWEN3_START_DOCSTRING,
+)
+class Qwen3Model(Qwen3PreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Qwen3DecoderLayer`]
+
+    Args:
+        config: Qwen3Config
+    """
+
+    def __init__(self, config: Qwen3Config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [Qwen3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = Qwen3RotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @can_return_tuple
+    @add_start_docstrings_to_model_forward(QWEN3_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
+    ) -> BaseModelOutputWithPast:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        # TODO (joao): remove this exception in v4.56 -- it exists for users that try to pass a legacy cache
+        if not isinstance(past_key_values, (type(None), Cache)):
+            raise ValueError("The `past_key_values` should be either a `Cache` object or `None`.")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
+
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                    position_embeddings,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings,
+                    **flash_attn_kwargs,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool = 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"]
+            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
+
+
+class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...
+
+
+class Qwen3ForCausalLM(Qwen3PreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Qwen3Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        if config._attn_implementation == 'sdpa':
+            ALL_ATTENTION_FUNCTIONS['sdpa'] = spec_sdpa_attention_forward
+            print("use specific sdpa attention")
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @can_return_tuple
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    @add_start_docstrings_to_model_forward(QWEN3_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[KwargsForCausalLM],
+    ) -> CausalLMOutputWithPast:
+        r"""
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+            logits_to_keep (`int` or `torch.Tensor`, *optional*):
+                If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+                If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+                This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, Qwen3ForCausalLM
+
+        >>> model = Qwen3ForCausalLM.from_pretrained("Qwen/Qwen3-8B")
+        >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs: BaseModelOutputWithPast = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        hidden_states = outputs.last_hidden_state
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    The Qwen3 Model transformer with a sequence classification head on top (linear layer).
+
+    [`Qwen3ForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+    (e.g. GPT-2) do.
+
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
+    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
+    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
+    each row of the batch).
+    """,
+    QWEN3_START_DOCSTRING,
+)
+class Qwen3ForSequenceClassification(Qwen3PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = Qwen3Model(config)
+        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @can_return_tuple
+    @add_start_docstrings_to_model_forward(QWEN3_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> SequenceClassifierOutputWithPast:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+
+        transformer_outputs: BaseModelOutputWithPast = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+        hidden_states = transformer_outputs.last_hidden_state
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size = input_ids.shape[0]
+        else:
+            batch_size = inputs_embeds.shape[0]
+
+        if self.config.pad_token_id is None and batch_size != 1:
+            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
+        if self.config.pad_token_id is None:
+            last_non_pad_token = -1
+        elif input_ids is not None:
+            # To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id
+            non_pad_mask = (input_ids != self.config.pad_token_id).to(logits.device, torch.int32)
+            token_indices = torch.arange(input_ids.shape[-1], device=logits.device, dtype=torch.int32)
+            last_non_pad_token = (token_indices * non_pad_mask).argmax(-1)
+        else:
+            last_non_pad_token = -1
+            logger.warning_once(
+                f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+            )
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), last_non_pad_token]
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config)
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    The Qwen3 Model transformer with a token classification head on top (a linear layer on top of the hidden-states
+    output) e.g. for Named-Entity-Recognition (NER) tasks.
+    """,
+    QWEN3_START_DOCSTRING,
+)
+class Qwen3ForTokenClassification(Qwen3PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = Qwen3Model(config)
+        if getattr(config, "classifier_dropout", None) is not None:
+            classifier_dropout = config.classifier_dropout
+        elif getattr(config, "hidden_dropout", None) is not None:
+            classifier_dropout = config.hidden_dropout
+        else:
+            classifier_dropout = 0.1
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.score = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @can_return_tuple
+    @add_start_docstrings_to_model_forward(QWEN3_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> TokenClassifierOutput:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+
+        outputs: BaseModelOutputWithPast = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+        sequence_output = outputs.last_hidden_state
+        sequence_output = self.dropout(sequence_output)
+        logits = self.score(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.config)
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+The Qwen3 Model transformer with a span classification head on top for extractive question-answering tasks like
+SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    QWEN3_START_DOCSTRING,
+)
+class Qwen3ForQuestionAnswering(Qwen3PreTrainedModel):
+    base_model_prefix = "transformer"
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.transformer = Qwen3Model(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, 2)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.transformer.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.transformer.embed_tokens = value
+
+    @can_return_tuple
+    @add_start_docstrings_to_model_forward(QWEN3_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        **kwargs,
+    ) -> QuestionAnsweringModelOutput:
+        r"""
+        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
+            are not taken into account for computing the loss.
+        """
+
+        outputs: BaseModelOutputWithPast = self.transformer(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+
+        sequence_output = outputs.last_hidden_state
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1).contiguous()
+        end_logits = end_logits.squeeze(-1).contiguous()
+
+        loss = None
+        if start_positions is not None and end_positions is not None:
+            loss = self.loss_function(start_logits, end_logits, start_positions, end_positions, **kwargs)
+
+        return QuestionAnsweringModelOutput(
+            loss=loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+__all__ = [
+    "Qwen3ForCausalLM",
+    "Qwen3ForQuestionAnswering",
+    "Qwen3Model",
+    "Qwen3PreTrainedModel",
+    "Qwen3ForSequenceClassification",
+    "Qwen3ForTokenClassification",
+]

modeling_sailvit.py

@@ -0,0 +1,198 @@
+# adapted from https://huggingface.co/apple/aimv2-huge-patch14-448 (modification: add gradient checkpoint support)
+from typing import Optional, Tuple, Union
+
+import torch
+from .configuration_sailvit import SAILViTConfig
+from torch import nn
+from torch.nn import functional as F
+from transformers.modeling_outputs import BaseModelOutputWithNoAttention
+from transformers.modeling_utils import PreTrainedModel
+
+__all__ = ["SAILViTModel"]
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(dim))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
+    def extra_repr(self) -> str:
+        return f"{tuple(self.weight.shape)}, eps={self.eps}"
+
+    def _norm(self, x: torch.Tensor) -> torch.Tensor:
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+
+class SAILViTSwiGLUFFN(nn.Module):
+    def __init__(self, config: SAILViTConfig):
+        super().__init__()
+        hidden_features = config.intermediate_size
+        in_features = config.hidden_size
+        bias = config.use_bias
+
+        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias)
+        self.fc2 = nn.Linear(hidden_features, in_features, bias=bias)
+        self.fc3 = nn.Linear(in_features, hidden_features, bias=bias)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = F.silu(self.fc1(x)) * self.fc3(x)
+        x = self.fc2(x)
+        return x
+
+
+class SAILViTPatchEmbed(nn.Module):
+    def __init__(self, config: SAILViTConfig):
+        super().__init__()
+        self.proj = nn.Conv2d(
+            config.num_channels,
+            config.hidden_size,
+            kernel_size=(config.patch_size, config.patch_size),
+            stride=(config.patch_size, config.patch_size),
+        )
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x).flatten(2).transpose(1, 2)
+        x = self.norm(x)
+        return x
+
+
+class SAILViTPreprocessor(nn.Module):
+    def __init__(self, config: SAILViTConfig):
+        super().__init__()
+        num_patches = (config.image_size // config.patch_size) ** 2
+
+        self.patchifier = SAILViTPatchEmbed(config)
+        self.pos_embed = nn.Parameter(torch.zeros((1, num_patches, config.hidden_size)))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        tokens = self.patchifier(x)
+        _, N, _ = tokens.shape
+        pos_embed = self.pos_embed.to(tokens.device)
+        tokens = tokens + pos_embed[:, :N]
+        return tokens
+
+
+class SAILViTAttention(nn.Module):
+    def __init__(self, config: SAILViTConfig):
+        super().__init__()
+        dim = config.hidden_size
+
+        self.num_heads = config.num_attention_heads
+        self.qkv = nn.Linear(dim, dim * 3, bias=config.qkv_bias)
+        self.attn_drop = nn.Dropout(config.attention_dropout)
+        self.proj = nn.Linear(dim, dim, bias=config.use_bias)
+        self.proj_drop = nn.Dropout(config.projection_dropout)
+
+    def forward(
+        self, x: torch.Tensor, mask: Optional[torch.Tensor] = None
+    ) -> torch.Tensor:
+        B, N, C = x.shape
+        qkv = (
+            self.qkv(x)
+            .reshape(B, N, 3, self.num_heads, C // self.num_heads)
+            .permute(2, 0, 3, 1, 4)
+        )
+        q, k, v = qkv.unbind(0)
+
+        x = F.scaled_dot_product_attention(q, k, v, attn_mask=mask)
+        x = x.transpose(1, 2).contiguous().reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class SAILViTBlock(nn.Module):
+    def __init__(self, config: SAILViTConfig):
+        super().__init__()
+        self.attn = SAILViTAttention(config)
+        self.norm_1 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.mlp = SAILViTSwiGLUFFN(config)
+        self.norm_2 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self, x: torch.Tensor, mask: Optional[torch.Tensor] = None
+    ) -> torch.Tensor:
+        x = x + self.attn(self.norm_1(x), mask)
+        x = x + self.mlp(self.norm_2(x))
+        return x
+
+
+class SAILViTTransformer(nn.Module):
+    def __init__(self, config: SAILViTConfig):
+        super().__init__()
+        self.blocks = nn.ModuleList(
+            [SAILViTBlock(config) for _ in range(config.num_hidden_layers)]
+        )
+        self.post_trunk_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        tokens: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        output_hidden_states: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, ...]]]:
+        hidden_states = () if output_hidden_states else None
+        for block in self.blocks:
+            if self.gradient_checkpointing and self.training:
+                tokens = self._gradient_checkpointing_func(block.__call__, tokens, mask)
+            else:
+                tokens = block(tokens, mask)
+            if output_hidden_states:
+                hidden_states += (tokens,)
+        tokens = self.post_trunk_norm(tokens)
+        return tokens, hidden_states
+
+
+class SAILViTPretrainedModel(PreTrainedModel):
+    config_class = SAILViTConfig
+    base_model_prefix = "sailvit"
+    supports_gradient_checkpointing = True
+    main_input_name = "pixel_values"
+    _no_split_modules = ["SAILViTPreprocessor", "SAILViTBlock"]
+    _supports_sdpa = True
+
+
+class SAILViTModel(SAILViTPretrainedModel):
+    def __init__(self, config: SAILViTConfig):
+        super().__init__(config)
+        self.preprocessor = SAILViTPreprocessor(config)
+        self.trunk = SAILViTTransformer(config)
+
+    def forward(
+        self,
+        pixel_values: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[
+        Tuple[torch.Tensor],
+        Tuple[torch.Tensor, Tuple[torch.Tensor, ...]],
+        BaseModelOutputWithNoAttention,
+    ]:
+        if output_hidden_states is None:
+            output_hidden_states = self.config.output_hidden_states
+        if return_dict is None:
+            return_dict = self.config.use_return_dict
+
+        x = self.preprocessor(pixel_values)
+        x, hidden_states = self.trunk(
+            x, mask, output_hidden_states=output_hidden_states
+        )
+
+        if not return_dict:
+            res = (x,)
+            res += (hidden_states,) if output_hidden_states else ()
+            return res
+
+        return BaseModelOutputWithNoAttention(
+            last_hidden_state=x,
+            hidden_states=hidden_states,
+        )
+

modeling_sailvl.py

@@ -0,0 +1,349 @@
+# --------------------------------------------------------
+# SAILVL
+# Copyright (c) 2024 OpenGVLab
+# Licensed under The MIT License [see LICENSE for details]
+# --------------------------------------------------------
+import warnings
+from typing import Any, List, Optional, Tuple, Union
+
+import torch.utils.checkpoint
+import transformers
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from transformers import (AutoModel, GenerationConfig, LlamaForCausalLM,
+                          LlamaTokenizer)
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import ModelOutput, logging
+# from .modeling_qwen2 import Qwen2ForCausalLM
+from .modeling_qwen3 import Qwen3ForCausalLM
+
+from .configuration_sailvl import SAILVLConfig
+from .conversation import get_conv_template
+from .modeling_sailvit import SAILViTModel
+
+logger = logging.get_logger(__name__)
+
+
+def version_cmp(v1, v2, op='eq'):
+    import operator
+
+    from packaging import version
+    op_func = getattr(operator, op)
+    return op_func(version.parse(v1), version.parse(v2))
+
+
+class SAILVLModel(PreTrainedModel):
+    config_class = SAILVLConfig
+    main_input_name = 'pixel_values'
+    _supports_flash_attn_2 = True
+    _no_split_modules = ['InternVisionModel', 'LlamaDecoderLayer', 'InternLM2DecoderLayer']
+
+    def __init__(self, config: SAILVLConfig, vision_model=None, language_model=None):
+        super().__init__(config)
+
+        assert version_cmp(transformers.__version__, '4.36.2', 'ge')
+        image_size = config.force_image_size or config.vision_config.image_size
+        patch_size = config.vision_config.patch_size
+        self.patch_size = patch_size
+        self.select_layer = config.select_layer
+        self.template = config.template
+        self.num_image_token = int((image_size // patch_size) ** 2 * (config.downsample_ratio ** 2))
+        self.downsample_ratio = config.downsample_ratio
+        self.ps_version = config.ps_version
+
+        logger.info(f'num_image_token: {self.num_image_token}')
+        logger.info(f'ps_version: {self.ps_version}')
+        if vision_model is not None:
+            self.vision_model = vision_model
+        else:
+            self.vision_model = SAILViTModel(config.vision_config)
+        if language_model is not None:
+            self.language_model = language_model
+            self.config.llm_config = language_model.config
+        else:
+            if config.llm_config.architectures[0] == 'LlamaForCausalLM':
+                self.language_model = LlamaForCausalLM(config.llm_config)
+            elif config.llm_config.architectures[0] == 'Qwen3ForCausalLM':
+                self.language_model = Qwen3ForCausalLM(config.llm_config)
+            else:
+                raise NotImplementedError(f'{config.llm_config.architectures[0]} is not implemented.')
+
+        vit_hidden_size = config.vision_config.hidden_size
+        llm_hidden_size = config.llm_config.hidden_size
+
+        self.mlp1 = nn.Sequential(
+            nn.LayerNorm(vit_hidden_size * int(1 / self.downsample_ratio) ** 2),
+            nn.Linear(vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_hidden_size),
+            nn.GELU(),
+            nn.Linear(llm_hidden_size, llm_hidden_size)
+        )
+
+        self.img_context_token_id = None
+        self.conv_template = get_conv_template(self.template)
+        self.system_message = self.conv_template.system_message
+
+    def forward(
+            self,
+            pixel_values: torch.FloatTensor,
+            input_ids: torch.LongTensor = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            image_flags: Optional[torch.LongTensor] = None,
+            past_key_values: Optional[List[torch.FloatTensor]] = None,
+            labels: Optional[torch.LongTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        image_flags = image_flags.squeeze(-1)
+        input_embeds = self.language_model.get_input_embeddings()(input_ids)
+
+        vit_embeds = self.extract_feature(pixel_values)
+        vit_embeds = vit_embeds[image_flags == 1]
+        vit_batch_size = pixel_values.shape[0]
+
+        B, N, C = input_embeds.shape
+        input_embeds = input_embeds.reshape(B * N, C)
+
+        if torch.distributed.get_rank() == 0:
+            print(f'dynamic ViT batch size: {vit_batch_size}, images per sample: {vit_batch_size / B}, dynamic token length: {N}')
+
+        input_ids = input_ids.reshape(B * N)
+        selected = (input_ids == self.img_context_token_id)
+        try:
+            input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds.reshape(-1, C)
+        except Exception as e:
+            vit_embeds = vit_embeds.reshape(-1, C)
+            print(f'warning: {e}, input_embeds[selected].shape={input_embeds[selected].shape}, '
+                  f'vit_embeds.shape={vit_embeds.shape}')
+            n_token = selected.sum()
+            input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds[:n_token]
+
+        input_embeds = input_embeds.reshape(B, N, C)
+
+        outputs = self.language_model(
+            inputs_embeds=input_embeds,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        logits = outputs.logits
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.language_model.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def pixel_shuffle(self, x, scale_factor=0.5):
+        n, w, h, c = x.size()
+        # N, W, H, C --> N, W, H * scale, C // scale
+        x = x.reshape(n, w, int(h * scale_factor), int(c / scale_factor))
+        # N, W, H * scale, C // scale --> N, H * scale, W, C // scale
+        x = x.permute(0, 2, 1, 3).contiguous()
+        # N, H * scale, W, C // scale --> N, H * scale, W * scale, C // (scale ** 2)
+        x = x.view(n, int(h * scale_factor), int(w * scale_factor),
+                   int(c / (scale_factor * scale_factor)))
+        if self.ps_version == 'v1':
+            warnings.warn("In ps_version 'v1', the height and width have not been swapped back, "
+                          'which results in a transposed image.')
+        else:
+            x = x.permute(0, 2, 1, 3).contiguous()
+        return x
+
+    def extract_feature(self, pixel_values):
+        if self.select_layer == -1:
+            vit_embeds = self.vision_model(
+                pixel_values=pixel_values,
+                output_hidden_states=False,
+                return_dict=True).last_hidden_state
+        else:
+            vit_embeds = self.vision_model(
+                pixel_values=pixel_values,
+                output_hidden_states=True,
+                return_dict=True).hidden_states[self.select_layer]
+        vit_embeds = vit_embeds
+
+        h = w = int(vit_embeds.shape[1] ** 0.5)
+        vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1)
+        vit_embeds = self.pixel_shuffle(vit_embeds, scale_factor=self.downsample_ratio)
+        vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], -1, vit_embeds.shape[-1])
+        vit_embeds = self.mlp1(vit_embeds)
+        return vit_embeds
+
+    def batch_chat(self, tokenizer, pixel_values, questions, generation_config, num_patches_list=None,
+                   history=None, return_history=False, IMG_START_TOKEN='<img>', IMG_END_TOKEN='</img>',
+                   IMG_CONTEXT_TOKEN='<IMG_CONTEXT>', verbose=False, image_counts=None):
+        if history is not None or return_history:
+            print('Now multi-turn chat is not supported in batch_chat.')
+            raise NotImplementedError
+
+        if image_counts is not None:
+            num_patches_list = image_counts
+            print('Warning: `image_counts` is deprecated. Please use `num_patches_list` instead.')
+
+        img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
+        self.img_context_token_id = img_context_token_id
+
+        if verbose and pixel_values is not None:
+            image_bs = pixel_values.shape[0]
+            print(f'dynamic ViT batch size: {image_bs}')
+
+        queries = []
+        for idx, num_patches in enumerate(num_patches_list):
+            question = questions[idx]
+            if pixel_values is not None and '<image>' not in question:
+                question = '<image>\n' + question
+            template = get_conv_template(self.template)
+            template.append_message(template.roles[0], question)
+            template.append_message(template.roles[1], None)
+            query = template.get_prompt()
+
+            image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN
+            query = query.replace('<image>', image_tokens, 1)
+            queries.append(query)
+
+        tokenizer.padding_side = 'left'
+        model_inputs = tokenizer(queries, return_tensors='pt', padding=True)
+        input_ids = model_inputs['input_ids'].cuda()
+        attention_mask = model_inputs['attention_mask'].cuda()
+        eos_token_id = tokenizer.convert_tokens_to_ids(template.sep)
+        generation_config['eos_token_id'] = eos_token_id
+        generation_output = self.generate(
+            pixel_values=pixel_values,
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            **generation_config
+        )
+        responses = tokenizer.batch_decode(generation_output, skip_special_tokens=True)
+        responses = [response.split(template.sep)[0].strip() for response in responses]
+        return responses
+
+    def chat(self, tokenizer, pixel_values, question, generation_config, history=None, return_history=False,
+             num_patches_list=None, IMG_START_TOKEN='<img>', IMG_END_TOKEN='</img>', IMG_CONTEXT_TOKEN='<IMG_CONTEXT>',
+             verbose=False):
+
+        if history is None and pixel_values is not None and '<image>' not in question:
+            question = '<image>\n' + question
+
+        if num_patches_list is None:
+            num_patches_list = [pixel_values.shape[0]] if pixel_values is not None else []
+        assert pixel_values is None or len(pixel_values) == sum(num_patches_list)
+
+        img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
+        self.img_context_token_id = img_context_token_id
+
+        template = get_conv_template(self.template)
+        template.system_message = self.system_message
+        eos_token_id = tokenizer.convert_tokens_to_ids(template.sep)
+
+        history = [] if history is None else history
+        for (old_question, old_answer) in history:
+            template.append_message(template.roles[0], old_question)
+            template.append_message(template.roles[1], old_answer)
+        template.append_message(template.roles[0], question)
+        template.append_message(template.roles[1], None)
+        query = template.get_prompt()
+
+        if verbose and pixel_values is not None:
+            image_bs = pixel_values.shape[0]
+            print(f'dynamic ViT batch size: {image_bs}')
+
+        for num_patches in num_patches_list:
+            image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN
+            query = query.replace('<image>', image_tokens, 1)
+
+        model_inputs = tokenizer(query, return_tensors='pt')
+        input_ids = model_inputs['input_ids'].cuda()
+        attention_mask = model_inputs['attention_mask'].cuda()
+        generation_config['eos_token_id'] = eos_token_id
+        generation_output = self.generate(
+            pixel_values=pixel_values,
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            **generation_config
+        )
+        response = tokenizer.batch_decode(generation_output, skip_special_tokens=True)[0]
+        response = response.split(template.sep)[0].strip()
+        history.append((question, response))
+        if return_history:
+            return response, history
+        else:
+            query_to_print = query.replace(IMG_CONTEXT_TOKEN, '')
+            query_to_print = query_to_print.replace(f'{IMG_START_TOKEN}{IMG_END_TOKEN}', '<image>')
+            if verbose:
+                print(query_to_print, response)
+            return response
+
+    @torch.no_grad()
+    def generate(
+            self,
+            pixel_values: Optional[torch.FloatTensor] = None,
+            input_ids: Optional[torch.FloatTensor] = None,
+            attention_mask: Optional[torch.LongTensor] = None,
+            visual_features: Optional[torch.FloatTensor] = None,
+            generation_config: Optional[GenerationConfig] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+            **generate_kwargs,
+    ) -> torch.LongTensor:
+        # img_context_token_id = tokenizer.convert_tokens_to_ids('<IMG_CONTEXT>')
+        # self.img_context_token_id = img_context_token_id
+        self.img_context_token_id = 151671
+
+        assert self.img_context_token_id is not None
+        if pixel_values is not None:
+            if visual_features is not None:
+                vit_embeds = visual_features
+            else:
+                vit_embeds = self.extract_feature(pixel_values)
+            input_embeds = self.language_model.get_input_embeddings()(input_ids)
+            B, N, C = input_embeds.shape
+            input_embeds = input_embeds.reshape(B * N, C)
+
+            input_ids = input_ids.reshape(B * N)
+            selected = (input_ids == self.img_context_token_id)
+            assert selected.sum() != 0
+            input_embeds[selected] = vit_embeds.reshape(-1, C).to(input_embeds.device)
+
+            input_embeds = input_embeds.reshape(B, N, C)
+        else:
+            input_embeds = self.language_model.get_input_embeddings()(input_ids)
+
+        outputs = self.language_model.generate(
+            inputs_embeds=input_embeds,
+            attention_mask=attention_mask,
+            generation_config=generation_config,
+            output_hidden_states=output_hidden_states,
+            use_cache=True,
+            **generate_kwargs,
+        )
+
+        return outputs

preprocessor_config.json

@@ -0,0 +1,21 @@
+{
+    "auto_map": {
+        "AutoImageProcessor": "image_processing_sailvl.SAILVLImageProcessor",
+        "AutoProcessor": "processing_sailvl.SAILVLProcessor"
+    },
+    "max_dynamic_patch": 10,
+    "patch_size": 14,
+    "image_size": 448,
+    "use_msac": false,
+    "image_mean": [
+        0.485,
+        0.456,
+        0.406
+    ],
+    "image_std": [
+        0.229,
+        0.224,
+        0.225
+    ],
+    "pad_input": true
+}

processing_sailvl.py

@@ -0,0 +1,168 @@
+# coding=utf-8
+# Copyright 2025 The Moonshot Team and HuggingFace Inc. team. All rights reserved.
+#
+# The code is based on the Qwen2VL processor (qwen2_vl/processing_qwen2_vl.py), but modified for KimiVL.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Processor class for KimiVL.
+"""
+
+from typing import List, Union
+
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.image_utils import ImageInput
+from transformers.processing_utils import ProcessingKwargs, ProcessorMixin, Unpack, _validate_images_text_input_order
+from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
+from transformers.utils import logging
+import torch
+
+
+logger = logging.get_logger(__name__)
+
+
+class SAILVLProcessorKwargs(ProcessingKwargs, total=False):
+    _defaults = {
+        "text_kwargs": {
+            "padding": False,
+        },
+        "images_kwargs": {},
+    }
+
+
+class SAILVLProcessor(ProcessorMixin):
+    # r"""
+    # Constructs a KimiVL processor which wraps a KimiVL image processor and a tokenizer into a single processor.
+    # [`KimiVLProcessor`] offers all the functionalities of [`KimiVLImageProcessor`] and [`TikTokenTokenizer`]. See the
+    # [`~KimiVLProcessor.__call__`] and [`~KimiVLProcessor.decode`] for more information.
+    # Args:
+    #     image_processor ([`KimiVLImageProcessor`], *optional*):
+    #         The image processor is a required input.
+    #     tokenizer ([`TikTokenTokenizer`], *optional*):
+    #         The tokenizer is a required input.
+    #     chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
+    #         in a chat into a tokenizable string.
+    # """
+
+    attributes = ["image_processor", "tokenizer"]
+    valid_kwargs = [ "chat_template"]
+    image_processor_class = "AutoImageProcessor"
+    tokenizer_class = "AutoTokenizer"
+
+    def __init__(
+        self,
+        image_processor=None,
+        tokenizer=None,
+        chat_template=None,
+        **kwargs,
+    ):
+        self.image_token = "<image>"
+        self.num_image_token = 256
+        super().__init__(image_processor, tokenizer, chat_template=chat_template)
+
+    def __call__(
+        self,
+        images: ImageInput = None,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
+        **kwargs: Unpack[SAILVLProcessorKwargs],
+    ) -> BatchFeature:
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to TikTokenTokenizer's [`~TikTokenTokenizer.__call__`] if `text` is not `None` to encode
+        the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
+        CLIPImageProcessor's [`~CLIPImageProcessor.__call__`] if `images` is not `None`. Please refer to the docstring
+        of the above two methods for more information.
+        Args:
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+            text (`str`, `List[str]`, `List[List[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+        """
+        if images is None and text is None:
+            raise ValueError("You have to specify at least one of `images` or `text`.")
+
+        # check if images and text inputs are reversed for BC
+        images, text = _validate_images_text_input_order(images, text)
+
+        output_kwargs = self._merge_kwargs(
+            SAILVLProcessorKwargs,
+            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
+            **kwargs,
+        )
+        # print(output_kwargs)
+        if images is not None:
+            image_inputs = self.image_processor(images, **output_kwargs["images_kwargs"])
+
+        else:
+            image_inputs = {}
+
+        if isinstance(text, str):
+            text = [text]
+        elif not isinstance(text, list) and not isinstance(text[0], str):
+            raise ValueError("Invalid input text. Please provide a string, or a list of strings")
+
+        query = text[0]
+        if images is not None:
+            for num_patches in image_inputs['num_patches_list']:
+                image_tokens = '<img>' + '<IMG_CONTEXT>' * self.num_image_token * num_patches + '</img>'
+                query = query.replace('<image>', image_tokens, 1)
+            image_inputs.pop('num_patches_list')
+
+        model_inputs = self.tokenizer(query, return_tensors='pt')
+        input_ids = model_inputs['input_ids'].cuda()
+        attention_mask = model_inputs['attention_mask'].cuda()
+        stop_word = '<|im_end|>'
+        eos_token_id = self.tokenizer.convert_tokens_to_ids(stop_word)
+
+
+        text_inputs = {'input_ids': input_ids, 'attention_mask': attention_mask, 'eos_token_id': eos_token_id}
+        return BatchFeature(data={**text_inputs, **image_inputs})
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
+
+
+__all__ = ["SAILVLProcessorKwargs"]

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81bf64fd18d98deef69ea2152dc76af3ea92c09caf85966dda8630966db43eeb
+size 5461342238

spec_sdpa_attention.py

@@ -0,0 +1,91 @@
+from typing import Optional, Tuple
+
+import torch
+try:
+    import torch_npu
+except:
+    print('Using N* GPU...')
+import math
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def spec_sdpa_attention_forward(
+    module: torch.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: Optional[bool] = None,
+    **kwargs,
+) -> Tuple[torch.Tensor, None]:
+    if hasattr(module, "num_key_value_groups"):
+        key = repeat_kv(key, module.num_key_value_groups)
+        value = repeat_kv(value, module.num_key_value_groups)
+
+    causal_mask = attention_mask
+    if attention_mask is not None and causal_mask.ndim == 4:
+        causal_mask = causal_mask[:, :, :, : key.shape[-2]]
+
+    # SDPA with memory-efficient backend is bugged with non-contiguous inputs and custom attn_mask for some torch versions
+    # Reference: https://github.com/pytorch/pytorch/issues/112577.
+    query = query.contiguous()
+    key = key.contiguous()
+    value = value.contiguous()
+
+    # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+    # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
+    # Note that it is important to check first for the shape, otherwise compile will fail with `argument 'is_causal' must be bool, not SymBool`
+    if is_causal is None:
+        is_causal = query.shape[2] > 1 and causal_mask is None
+
+    # Shapes (e.g. query.shape[2]) are tensors during jit tracing, resulting in `is_causal` being a tensor.
+    # We convert it to a bool for the SDPA kernel that only accepts bools.
+
+    if attention_mask is None:
+        atten_mask_npu = torch.triu(torch.ones([query.size(-2), 
+                                                query.size(-2)]), diagonal=1).bool().to(query.device)
+    elif attention_mask.dtype == torch.bool:
+        atten_mask_npu = torch.logical_not(attention_mask.bool()).to(attention_mask.device) # atten_mask需要取反
+    else:
+        atten_mask_npu = attention_mask.bool().to(attention_mask.device)
+
+    if torch.jit.is_tracing() and isinstance(is_causal, torch.Tensor):
+        is_causal = is_causal.item()
+
+    # attn_output = torch.nn.functional.scaled_dot_product_attention(
+    #     query,
+    #     key,
+    #     value,
+    #     attn_mask=causal_mask,
+    #     dropout_p=dropout,
+    #     scale=scaling,
+    #     is_causal=is_causal,
+    # )
+
+    head_num = query.shape[1]
+    attn_output = torch_npu.npu_fusion_attention(
+                    query, key, value, head_num, input_layout="BNSD", 
+                    pse=None,
+                    atten_mask=atten_mask_npu,
+                    scale=1.0 / math.sqrt(query.shape[-1]),
+                    pre_tockens=2147483647,
+                    next_tockens=2147483647,
+                    keep_prob=1
+                )[0]
+
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, None

special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc7a4004b65007afc1409bed6f57ca23287da9c8889b79974ba6ccb077d30eb9
+size 11424306

tokenizer_config.json

@@ -0,0 +1,311 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151665": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151666": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151667": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151668": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151669": {
+      "content": "<img>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151670": {
+      "content": "</img>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151671": {
+      "content": "<IMG_CONTEXT>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151672": {
+      "content": "<quad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151673": {
+      "content": "</quad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151674": {
+      "content": "<ref>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151675": {
+      "content": "</ref>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151676": {
+      "content": "<box>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151677": {
+      "content": "</box>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "chat_template": "{%- for message in messages -%}{%- if loop.first and messages[0]['role'] != 'system' -%}{{'<|im_system|>system<|im_middle|>You are a helpful assistant<|im_end|>'}}{%- endif -%}{%- if message['role'] == 'system' -%}{{'<|im_system|>'}}{%- endif -%}{%- if message['role'] == 'user' -%}{{'<|im_user|>'}}{%- endif -%}{%- if message['role'] == 'assistant' -%}{{'<|im_assistant|>'}}{%- endif -%}{{- message['role'] -}}{{'<|im_middle|>'}}{%- if message['content'] is string -%}{{- message['content'] + '<|im_end|>' -}}{%- else -%}{%- for content in message['content'] -%}{%- if content['type'] == 'image' or 'image' in content or 'image_url' in content -%}{{'<|media_start|>image<|media_content|><|media_pad|><|media_end|>'}}{%- else -%}{{content['text']}}{%- endif -%}{%- endfor -%}{{'<|im_end|>'}}{%- endif -%}{%- endfor -%}{%- if add_generation_prompt -%}{{'<|im_assistant|>assistant<|im_middle|>'}}{%- endif -%}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}