可以像 qwen3 那样加上 /no_think 关闭深思考模式吗

.gitattributes

@@ -11,7 +11,7 @@
 *.model filter=lfs diff=lfs merge=lfs -text
 *.msgpack filter=lfs diff=lfs merge=lfs -text
 *.onnx filter=lfs diff=lfs merge=lfs -text
- 
+*.ot filter=lfs diff=lfs merge=lfs -text
 *.parquet filter=lfs diff=lfs merge=lfs -text
 *.pb filter=lfs diff=lfs merge=lfs -text
 *.pt filter=lfs diff=lfs merge=lfs -text
@@ -44,6 +44,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.tar filter=lfs diff=lfs merge=lfs -text
 *.wasm filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
-
-notebook.ipynb filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,110 +1,93 @@
----
+---
-license: mit
+license: mit
-language:
+---
-- en
+
-- zh
+# GLM-4.1V-9B-Thinking
-base_model:
+
-- zai-org/GLM-4-9B-0414
+<div align="center">
-pipeline_tag: image-text-to-text
+<img src=https://raw.githubusercontent.com/THUDM/GLM-4.1V-Thinking/99c5eb6563236f0ff43605d91d107544da9863b2/resources/logo.svg width="40%"/>
-library_name: transformers
+</div>
-tags:
+<p align="center">
-- reasoning
+    📖 查看 GLM-4.1V-9B-Thinking <a href="https://arxiv.org/abs/2507.01006" target="_blank">论文</a> 。
----
+    <br>
-
+    💡 立即在线体验 <a href="https://huggingface.co/spaces/THUDM/GLM-4.1V-9B-Thinking-Demo" target="_blank">Hugging Face</a> 或 <a href="https://modelscope.cn/studios/ZhipuAI/GLM-4.1V-9B-Thinking-Demo" target="_blank">ModelScope</a> 上的 GLM-4.1V-9B-Thinking。
-# GLM-4.1V-9B-Thinking
+    <br>
-
+    📍 在 <a href="https://www.bigmodel.cn/dev/api/visual-reasoning-model/GLM-4.1V-Thinking">智谱大模型开放平台</a> 使用 GLM-4.1V-9B-Thinking 的API服务。
-<div align="center">
+</p>
-<img src=https://raw.githubusercontent.com/zai-org/GLM-4.1V-Thinking/99c5eb6563236f0ff43605d91d107544da9863b2/resources/logo.svg width="40%"/>
+
-</div>
+## 模型介绍
-<p align="center">
+
-    📖 View the GLM-4.1V-9B-Thinking <a href="https://arxiv.org/abs/2507.01006" target="_blank">paper</a>.
+视觉语言大模型（VLM）已经成为智能系统的关键基石。随着真实世界的智能任务越来越复杂，VLM模型也亟需在基本的多模态感知之外，
-    <br>
+逐渐增强复杂任务中的推理能力，提升自身的准确性、全面性和智能化程度，使得复杂问题解决、长上下文理解、多模态智能体等智能任务成为可能。
-    📍 Using GLM-4.1V-9B-Thinking API at <a href="https://www.bigmodel.cn/dev/api/visual-reasoning-model/GLM-4.1V-Thinking">Zhipu Foundation Model Open Platform</a>
+
-</p>
+基于 [GLM-4-9B-0414](https://github.com/THUDM/GLM-4) 基座模型，我们推出新版VLM开源模型 **GLM-4.1V-9B-Thinking**
-
+，引入思考范式，通过课程采样强化学习 RLCS（Reinforcement Learning with Curriculum Sampling）全面提升模型能力，
-
+达到 10B 参数级别的视觉语言模型的最强性能，在18个榜单任务中持平甚至超过8倍参数量的 Qwen-2.5-VL-72B。
-## Model Introduction
+我们同步开源基座模型 **GLM-4.1V-9B-Base**，希望能够帮助更多研究者探索视觉语言模型的能力边界。
-
+
-Vision-Language Models (VLMs) have become foundational components of intelligent systems. As real-world AI tasks grow
+![rl](https://raw.githubusercontent.com/THUDM/GLM-4.1V-Thinking/refs/heads/main/resources/rl.jpeg)
-increasingly complex, VLMs must evolve beyond basic multimodal perception to enhance their reasoning capabilities in
+
-complex tasks. This involves improving accuracy, comprehensiveness, and intelligence, enabling applications such as
+与上一代的 CogVLM2 及 GLM-4V 系列模型相比，**GLM-4.1V-Thinking** 有如下改进：
-complex problem solving, long-context understanding, and multimodal agents.
+
-
+1. 系列中首个推理模型，不仅仅停留在数学领域，在多个子领域均达到世界前列的水平。
-Based on the [GLM-4-9B-0414](https://github.com/zai-org/GLM-4) foundation model, we present the new open-source VLM model
+2. 支持 **64k** 上下长度。
-**GLM-4.1V-9B-Thinking**, designed to explore the upper limits of reasoning in vision-language models. By introducing
+3. 支持**任意长宽比**和高达 **4k** 的图像分辨率。
-a "thinking paradigm" and leveraging reinforcement learning, the model significantly enhances its capabilities. It
+4. 提供支持**中英文双语**的开源模型版本。
-achieves state-of-the-art performance among 10B-parameter VLMs, matching or even surpassing the 72B-parameter
+
-Qwen-2.5-VL-72B on 18 benchmark tasks. We are also open-sourcing the base model GLM-4.1V-9B-Base to
+## 榜单信息
-support further research into the boundaries of VLM capabilities.
+
-
+GLM-4.1V-9B-Thinking 通过引入「思维链」（Chain-of-Thought）推理机制，在回答准确性、内容丰富度与可解释性方面，
-![rl](https://raw.githubusercontent.com/zai-org/GLM-4.1V-Thinking/refs/heads/main/resources/rl.jpeg)
+全面超越传统的非推理式视觉模型。在28项评测任务中有23项达到10B级别模型最佳，甚至有18项任务超过8倍参数量的Qwen-2.5-VL-72B。
-
+
-Compared to the previous generation models CogVLM2 and the GLM-4V series, **GLM-4.1V-Thinking** offers the
+![bench](https://raw.githubusercontent.com/THUDM/GLM-4.1V-Thinking/refs/heads/main/resources/bench.jpeg)
-following improvements:
+
-
+## 快速推理
-1. The first reasoning-focused model in the series, achieving world-leading performance not only in mathematics but also
+
-   across various sub-domains.
+这里展现了一个使用`transformers`进行单张图片推理的代码。首先，从源代码安装`transformers`库。
-2. Supports **64k** context length.
+```
-3. Handles **arbitrary aspect ratios** and up to **4K** image resolution.
+pip install git+https://github.com/huggingface/transformers.git
-4. Provides an open-source version supporting both **Chinese and English bilingual** usage.
+```
-
+
-## Benchmark Performance
+接着按照以下代码运行:
-
+
-By incorporating the Chain-of-Thought reasoning paradigm, GLM-4.1V-9B-Thinking significantly improves answer accuracy,
+```python
-richness, and interpretability. It comprehensively surpasses traditional non-reasoning visual models.
+from transformers import AutoProcessor, Glm4vForConditionalGeneration
-Out of 28 benchmark tasks, it achieved the best performance among 10B-level models on 23 tasks,
+import torch
-and even outperformed the 72B-parameter Qwen-2.5-VL-72B on 18 tasks.
+
-
+MODEL_PATH = "THUDM/GLM-4.1V-9B-Thinking"
-![bench](https://raw.githubusercontent.com/zai-org/GLM-4.1V-Thinking/refs/heads/main/resources/bench.jpeg)
+messages = [
-
+    {
-## Quick Inference
+        "role": "user",
-
+        "content": [
-This is a simple example of running single-image inference using the `transformers` library.  
+            {
-First, install the `transformers` library from source:
+                "type": "image",
-
+                "url": "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png"
-```
+            },
-pip install transformers>=4.57.1
+            {
-```
+                "type": "text",
-
+                "text": "describe this image"
-Then, run the following code:
+            }
-
+        ],
-```python
+    }
-from transformers import AutoProcessor, Glm4vForConditionalGeneration
+]
-import torch
+processor = AutoProcessor.from_pretrained(MODEL_PATH, use_fast=True)
-
+model = Glm4vForConditionalGeneration.from_pretrained(
-MODEL_PATH = "zai-org/GLM-4.1V-9B-Thinking"
+    pretrained_model_name_or_path=MODEL_PATH,
-messages = [
+    torch_dtype=torch.bfloat16,
-    {
+    device_map="auto",
-        "role": "user",
+)
-        "content": [
+inputs = processor.apply_chat_template(
-            {
+    messages,
-                "type": "image",
+    tokenize=True,
-                "url": "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png"
+    add_generation_prompt=True,
-            },
+    return_dict=True,
-            {
+    return_tensors="pt"
-                "type": "text",
+).to(model.device)
-                "text": "describe this image"
+generated_ids = model.generate(**inputs, max_new_tokens=8192)
-            }
+output_text = processor.decode(generated_ids[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False)
-        ],
+print(output_text)
-    }
+```
-]
+
-processor = AutoProcessor.from_pretrained(MODEL_PATH, use_fast=True)
+
-model = Glm4vForConditionalGeneration.from_pretrained(
+视频推理，网页端Demo部署等更代码请查看我们的 [github](https://github.com/THUDM/GLM-4.1V-Thinking)。
-    pretrained_model_name_or_path=MODEL_PATH,
+
-    torch_dtype=torch.bfloat16,
-    device_map="auto",
-)
-inputs = processor.apply_chat_template(
-    messages,
-    tokenize=True,
-    add_generation_prompt=True,
-    return_dict=True,
-    return_tensors="pt"
-).to(model.device)
-generated_ids = model.generate(**inputs, max_new_tokens=8192)
-output_text = processor.decode(generated_ids[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False)
-print(output_text)
-```
-
-For video reasoning, web demo deployment, and more code, please check
-our [GitHub](https://github.com/zai-org/GLM-V).

config.json

@@ -3,50 +3,38 @@
     "Glm4vForConditionalGeneration"
   ],
   "model_type": "glm4v",
+  "attention_bias": true,
+  "attention_dropout": 0.0,
+  "pad_token_id": 151329,
+  "eos_token_id": [
+    151329,
+    151336,
+    151338,
+    151348
+  ],
   "image_start_token_id": 151339,
   "image_end_token_id": 151340,
   "video_start_token_id": 151341,
   "video_end_token_id": 151342,
   "image_token_id": 151343,
   "video_token_id": 151344,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 13696,
+  "max_position_embeddings": 32768,
+  "num_attention_heads": 32,
+  "num_hidden_layers": 40,
+  "num_key_value_heads": 2,
+  "rms_norm_eps": 1e-05,
+  "rope_theta": 10000.0,
   "tie_word_embeddings": false,
-  "transformers_version": "4.57.1",
+  "torch_dtype": "bfloat16",
-  "text_config": {
+  "transformers_version": "4.53.0dev",
-    "model_type": "glm4v_text",
+  "use_cache": true,
-    "attention_bias": true,
+  "vocab_size": 151552,
-    "attention_dropout": 0.0,
+  "partial_rotary_factor": 0.5,
-    "pad_token_id": 151329,
-    "eos_token_id": [
-      151329,
-      151336,
-      151338,
-      151348
-    ],
-    "hidden_act": "silu",
-    "hidden_size": 4096,
-    "initializer_range": 0.02,
-    "intermediate_size": 13696,
-    "max_position_embeddings": 65536,
-    "num_attention_heads": 32,
-    "num_hidden_layers": 40,
-    "num_key_value_heads": 2,
-    "rms_norm_eps": 1e-05,
-    "dtype": "bfloat16",
-    "use_cache": true,
-    "vocab_size": 151552,
-    "partial_rotary_factor": 0.5,
-    "rope_theta": 10000,
-    "rope_scaling": {
-      "rope_type": "default",
-      "mrope_section": [
-        8,
-        12,
-        12
-      ]
-    }
-  },
   "vision_config": {
-    "model_type": "glm4v",
     "hidden_size": 1536,
     "depth": 24,
     "num_heads": 12,
@@ -61,5 +49,13 @@
     "rms_norm_eps": 1e-05,
     "spatial_merge_size": 2,
     "temporal_patch_size": 2
+  },
+  "rope_scaling": {
+    "type": "default",
+    "mrope_section": [
+      8,
+      12,
+      12
+    ]
   }
 }

generation_config.json

@@ -1,15 +0,0 @@
-{
-  "_from_model_config": true,
-  "do_sample": true,
-  "eos_token_id": [
-    151329,
-    151336,
-    151338,
-    151348
-  ],
-  "pad_token_id": 151329,
-  "top_p": 0.6,
-  "temperature": 0.8,
-  "top_k": 2,
-  "transformers_version": "4.57.1"
-}

tokenizer_config.json

@@ -134,7 +134,7 @@
       "normalized": false,
       "rstrip": false,
       "single_word": false,
-      "special": false
+      "special": true
     },
      "151346": {
       "content": "</think>",
@@ -142,7 +142,7 @@
       "normalized": false,
       "rstrip": false,
       "single_word": false,
-      "special": false
+      "special": true
     },
      "151347": {
       "content": "<answer>",
@@ -150,7 +150,7 @@
       "normalized": false,
       "rstrip": false,
       "single_word": false,
-      "special": false
+      "special": true
     },
      "151348": {
       "content": "</answer>",
@@ -158,7 +158,7 @@
       "normalized": false,
       "rstrip": false,
       "single_word": false,
-      "special": false
+      "special": true
     },
     "151349": {
       "content": "<|begin_of_box|>",
@@ -166,7 +166,7 @@
       "normalized": false,
       "rstrip": false,
       "single_word": false,
-      "special": false
+      "special": true
     },
     "151350": {
       "content": "<|end_of_box|>",
@@ -174,7 +174,7 @@
       "normalized": false,
       "rstrip": false,
       "single_word": false,
-      "special": false
+      "special": true
     },
     "151351": {
       "content": "<|sep|>",
@@ -182,7 +182,7 @@
       "normalized": false,
       "rstrip": false,
       "single_word": false,
-      "special": false
+      "special": true
     }
   },
   "additional_special_tokens": [
@@ -200,8 +200,15 @@
     "<|end_of_image|>",
     "<|begin_of_video|>",
     "<|end_of_video|>",
+    "<|begin_of_box|>",
+    "<|end_of_box|>",
+    "<|sep|>",
     "<|image|>",
-    "<|video|>"
+    "<|video|>",
+    "<think>",
+    "</think>",
+    "<answer>",
+    "</answer>"
   ],
   "clean_up_tokenization_spaces": false,
   "do_lower_case": false,
@@ -211,7 +218,7 @@
     "input_ids",
     "attention_mask"
   ],
-  "model_max_length": 65536,
+  "model_max_length": 64000,
   "padding_side": "left",
   "remove_space": false,
   "tokenizer_class": "PreTrainedTokenizer"