Improve model card with paper link, system requirements, and sample usage
Browse filesThis pull request significantly improves the model card for GLM-4.5-Air-Base.
Key improvements include:
- Adding a prominent link to the official Hugging Face paper page ([GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models](https://huggingface.co/papers/2508.06471)) for better discoverability. The existing ArXiv link remains.
- Integrating detailed "System Requirements" from the official GitHub repository, providing crucial information on inference and fine-tuning configurations.
- Enhancing the "Quick Start" section by:
- Including general installation instructions.
- Providing a concrete Python code snippet for text generation using the `transformers` library.
- Incorporating detailed setup commands for `vLLM` and `SGLang` from the GitHub README, making it easier for users to deploy and interact with the model directly from the Hub page.
These updates aim to provide a more comprehensive and user-friendly experience for researchers and developers on Hugging Face.
@@ -1,10 +1,10 @@
|
|
1 |
---
|
2 |
-
license: mit
|
3 |
language:
|
4 |
- en
|
5 |
- zh
|
6 |
-
pipeline_tag: text-generation
|
7 |
library_name: transformers
|
|
|
|
|
8 |
---
|
9 |
|
10 |
# GLM-4.5-Air-Base
|
@@ -21,7 +21,9 @@ library_name: transformers
|
|
21 |
<br>
|
22 |
👉 One click to <a href="https://chat.z.ai">GLM-4.5</a>.
|
23 |
</p>
|
24 |
-
|
|
|
|
|
25 |
## Model Introduction
|
26 |
|
27 |
The **GLM-4.5** series models are foundation models designed for intelligent agents. GLM-4.5 has **355** billion total parameters with **32** billion active parameters, while GLM-4.5-Air adopts a more compact design with **106** billion total parameters and **12** billion active parameters. GLM-4.5 models unify reasoning, coding, and intelligent agent capabilities to meet the complex demands of intelligent agent applications.
|
@@ -30,7 +32,7 @@ Both GLM-4.5 and GLM-4.5-Air are hybrid reasoning models that provide two modes:
|
|
30 |
|
31 |
We have open-sourced the base models, hybrid reasoning models, and FP8 versions of the hybrid reasoning models for both GLM-4.5 and GLM-4.5-Air. They are released under the MIT open-source license and can be used commercially and for secondary development.
|
32 |
|
33 |
-
As demonstrated in our comprehensive evaluation across 12 industry-standard benchmarks, GLM-4.5 achieves exceptional performance with a score of **63.2**, in the **3rd** place among all the proprietary and open-source
|
34 |
|
35 |

|
36 |
|
@@ -39,8 +41,155 @@ our [technical blog](https://z.ai/blog/glm-4.5). The technical report will be re
|
|
39 |
|
40 |
The model code, tool parser and reasoning parser can be found in the implementation of [transformers](https://github.com/huggingface/transformers/tree/main/src/transformers/models/glm4_moe), [vLLM](https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/glm4_moe_mtp.py) and [SGLang](https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/models/glm4_moe.py).
|
41 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
## Quick Start
|
43 |
|
44 |
**Note**: This is a base model, not for chat.
|
45 |
|
46 |
-
Please
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
|
|
2 |
language:
|
3 |
- en
|
4 |
- zh
|
|
|
5 |
library_name: transformers
|
6 |
+
license: mit
|
7 |
+
pipeline_tag: text-generation
|
8 |
---
|
9 |
|
10 |
# GLM-4.5-Air-Base
|
|
|
21 |
<br>
|
22 |
👉 One click to <a href="https://chat.z.ai">GLM-4.5</a>.
|
23 |
</p>
|
24 |
+
|
25 |
+
This repository contains the base model presented in the paper [GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models](https://huggingface.co/papers/2508.06471).
|
26 |
+
|
27 |
## Model Introduction
|
28 |
|
29 |
The **GLM-4.5** series models are foundation models designed for intelligent agents. GLM-4.5 has **355** billion total parameters with **32** billion active parameters, while GLM-4.5-Air adopts a more compact design with **106** billion total parameters and **12** billion active parameters. GLM-4.5 models unify reasoning, coding, and intelligent agent capabilities to meet the complex demands of intelligent agent applications.
|
|
|
32 |
|
33 |
We have open-sourced the base models, hybrid reasoning models, and FP8 versions of the hybrid reasoning models for both GLM-4.5 and GLM-4.5-Air. They are released under the MIT open-source license and can be used commercially and for secondary development.
|
34 |
|
35 |
+
As demonstrated in our comprehensive evaluation across 12 industry-standard benchmarks, GLM-4.5 achieves exceptional performance with a score of **63.2**, in the **3rd** place among all the proprietary and open-source models. Notably, GLM-4.5-Air delivers competitive results at **59.8** while maintaining superior efficiency.
|
36 |
|
37 |

|
38 |
|
|
|
41 |
|
42 |
The model code, tool parser and reasoning parser can be found in the implementation of [transformers](https://github.com/huggingface/transformers/tree/main/src/transformers/models/glm4_moe), [vLLM](https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/glm4_moe_mtp.py) and [SGLang](https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/models/glm4_moe.py).
|
43 |
|
44 |
+
## System Requirements
|
45 |
+
|
46 |
+
### Inference
|
47 |
+
|
48 |
+
We provide minimum and recommended configurations for "full-featured" model inference. The data in the table below is
|
49 |
+
based on the following conditions:
|
50 |
+
|
51 |
+
1. All models use MTP layers and specify
|
52 |
+
`--speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4` to ensure competitive
|
53 |
+
inference speed.
|
54 |
+
2. The `cpu-offload` parameter is not used.
|
55 |
+
3. Inference batch size does not exceed `8`.
|
56 |
+
4. All are executed on devices that natively support FP8 inference, ensuring both weights and cache are in FP8 format.
|
57 |
+
5. Server memory must exceed `1T` to ensure normal model loading and operation.
|
58 |
+
|
59 |
+
The models can run under the configurations in the table below:
|
60 |
+
|
61 |
+
| Model | Precision | GPU Type and Count | Test Framework |
|
62 |
+
|---|---|---|---|
|
63 |
+
| GLM-4.5 | BF16 | H100 x 16 / H200 x 8 | sglang |
|
64 |
+
| GLM-4.5 | FP8 | H100 x 8 / H200 x 4 | sglang |
|
65 |
+
| GLM-4.5-Air | BF16 | H100 x 4 / H200 x 2 | sglang |
|
66 |
+
| GLM-4.5-Air | FP8 | H100 x 2 / H200 x 1 | sglang |
|
67 |
+
|
68 |
+
Under the configurations in the table below, the models can utilize their full 128K context length:
|
69 |
+
|
70 |
+
| Model | Precision | GPU Type and Count | Test Framework |
|
71 |
+
|---|---|---|---|
|
72 |
+
| GLM-4.5 | BF16 | H100 x 32 / H200 x 16 | sglang |
|
73 |
+
| GLM-4.5 | FP8 | H100 x 16 / H200 x 8 | sglang |
|
74 |
+
| GLM-4.5-Air | BF16 | H100 x 8 / H200 x 4 | sglang |
|
75 |
+
| GLM-4.5-Air | FP8 | H100 x 4 / H200 x 2 | sglang |
|
76 |
+
|
77 |
+
### Fine-tuning
|
78 |
+
|
79 |
+
The code can run under the configurations in the table below
|
80 |
+
using [Llama Factory](https://github.com/hiyouga/LLaMA-Factory):
|
81 |
+
|
82 |
+
| Model | GPU Type and Count | Strategy | Batch Size (per GPU) |
|
83 |
+
|---|---|---|---|
|
84 |
+
| GLM-4.5 | H100 x 16 | Lora | 1 |
|
85 |
+
| GLM-4.5-Air | H100 x 4 | Lora | 1 |
|
86 |
+
|
87 |
+
The code can run under the configurations in the table below using [Swift](https://github.com/modelscope/ms-swift):
|
88 |
+
|
89 |
+
| Model | GPU Type and Count | Strategy | Batch Size (per GPU) |
|
90 |
+
|---|---|---|---|
|
91 |
+
| GLM-4.5 | H20 (96GiB) x 16 | Lora | 1 |
|
92 |
+
| GLM-4.5-Air | H20 (96GiB) x 4 | Lora | 1 |
|
93 |
+
| GLM-4.5 | H20 (96GiB) x 128 | SFT | 1 |
|
94 |
+
| GLM-4.5-Air | H20 (96GiB) x 32 | SFT | 1 |
|
95 |
+
| GLM-4.5 | H20 (96GiB) x 128 | RL | 1 |
|
96 |
+
| GLM-4.5-Air | H20 (96GiB) x 32 | RL | 1 |
|
97 |
+
|
98 |
## Quick Start
|
99 |
|
100 |
**Note**: This is a base model, not for chat.
|
101 |
|
102 |
+
Please install the required packages according to `requirements.txt`.
|
103 |
+
|
104 |
+
```bash
|
105 |
+
pip install -r requirements.txt
|
106 |
+
```
|
107 |
+
|
108 |
+
### transformers
|
109 |
+
|
110 |
+
Here's a basic example to use the model with the `transformers` library for text generation:
|
111 |
+
|
112 |
+
```python
|
113 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
114 |
+
import torch
|
115 |
+
|
116 |
+
model_id = "zai-org/GLM-4.5-Air-Base"
|
117 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
118 |
+
# Ensure to load with the correct dtype, e.g., bfloat16 as specified in config.json
|
119 |
+
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16)
|
120 |
+
|
121 |
+
prompt = "Hello, I'm a language model,"
|
122 |
+
inputs = tokenizer(prompt, return_tensors="pt")
|
123 |
+
|
124 |
+
# Generate
|
125 |
+
generate_ids = model.generate(inputs.input_ids, max_new_tokens=100)
|
126 |
+
print(tokenizer.batch_decode(generate_ids, skip_special_tokens=True)[0])
|
127 |
+
```
|
128 |
+
|
129 |
+
The GitHub repository provides more detailed examples, including `trans_infer_cli.py`.
|
130 |
+
|
131 |
+
### vLLM
|
132 |
+
|
133 |
+
Both BF16 and FP8 can be started with the following code:
|
134 |
+
|
135 |
+
```bash
|
136 |
+
vllm serve zai-org/GLM-4.5-Air \
|
137 |
+
--tensor-parallel-size 8 \
|
138 |
+
--tool-call-parser glm45 \
|
139 |
+
--reasoning-parser glm45 \
|
140 |
+
--enable-auto-tool-choice \
|
141 |
+
--served-model-name glm-4.5-air
|
142 |
+
```
|
143 |
+
|
144 |
+
If you're using 8x H100 GPUs and encounter insufficient memory when running the GLM-4.5 model, you'll need
|
145 |
+
`--cpu-offload-gb 16` (only applicable to vLLM).
|
146 |
+
|
147 |
+
If you encounter `flash infer` issues, use `VLLM_ATTENTION_BACKEND=XFORMERS` as a temporary replacement. You can also
|
148 |
+
specify `TORCH_CUDA_ARCH_LIST='9.0+PTX'` to use `flash infer` (different GPUs have different TORCH_CUDA_ARCH_LIST
|
149 |
+
values, please check accordingly).
|
150 |
+
|
151 |
+
### SGLang
|
152 |
+
|
153 |
+
+ BF16
|
154 |
+
|
155 |
+
```bash
|
156 |
+
python3 -m sglang.launch_server \
|
157 |
+
--model-path zai-org/GLM-4.5-Air \
|
158 |
+
--tp-size 8 \
|
159 |
+
--tool-call-parser glm45 \
|
160 |
+
--reasoning-parser glm45 \
|
161 |
+
--speculative-algorithm EAGLE \
|
162 |
+
--speculative-num-steps 3 \
|
163 |
+
--speculative-eagle-topk 1 \
|
164 |
+
--speculative-num-draft-tokens 4 \
|
165 |
+
--mem-fraction-static 0.7 \
|
166 |
+
--served-model-name glm-4.5-air \
|
167 |
+
--host 0.0.0.0 \
|
168 |
+
--port 8000
|
169 |
+
```
|
170 |
+
|
171 |
+
+ FP8
|
172 |
+
|
173 |
+
```bash
|
174 |
+
python3 -m sglang.launch_server \
|
175 |
+
--model-path zai-org/GLM-4.5-Air-FP8 \
|
176 |
+
--tp-size 4 \
|
177 |
+
--tool-call-parser glm45 \
|
178 |
+
--reasoning-parser glm45 \
|
179 |
+
--speculative-algorithm EAGLE \
|
180 |
+
--speculative-num-steps 3 \
|
181 |
+
--speculative-eagle-topk 1 \
|
182 |
+
--speculative-num-draft-tokens 4 \
|
183 |
+
--mem-fraction-static 0.7 \
|
184 |
+
--disable-shared-experts-fusion \
|
185 |
+
--served-model-name glm-4.5-air-fp8 \
|
186 |
+
--host 0.0.0.0 \
|
187 |
+
--port 8000
|
188 |
+
```
|
189 |
+
|
190 |
+
### Request Parameter Instructions
|
191 |
+
|
192 |
+
+ When using `vLLM` and `SGLang`, thinking mode is enabled by default when sending requests. If you want to disable the
|
193 |
+
thinking switch, you need to add the `extra_body={"chat_template_kwargs": {"enable_thinking": False}}` parameter.
|
194 |
+
+ Both support tool calling. Please use OpenAI-style tool description format for calls.
|
195 |
+
+ For specific code, please refer to `api_request.py` in the `inference` folder.
|