Instructions to use Maxtimer97/GLM2NSA with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Maxtimer97/GLM2NSA with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Maxtimer97/GLM2NSA", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("Maxtimer97/GLM2NSA", trust_remote_code=True, dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use Maxtimer97/GLM2NSA with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Maxtimer97/GLM2NSA"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Maxtimer97/GLM2NSA",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Maxtimer97/GLM2NSA

SGLang

How to use Maxtimer97/GLM2NSA with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Maxtimer97/GLM2NSA" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Maxtimer97/GLM2NSA",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Maxtimer97/GLM2NSA" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Maxtimer97/GLM2NSA",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Maxtimer97/GLM2NSA with Docker Model Runner:
```
docker model run hf.co/Maxtimer97/GLM2NSA
```

Maxtimer97 commited on Sep 29, 2025

Commit

bb18c9e

verified ·

1 Parent(s): ce81c56

Hardwired nsa attention

Browse files

Files changed (1) hide show

modeling_chatglm.py +3 -1

modeling_chatglm.py CHANGED Viewed

@@ -642,7 +642,7 @@ class SelfAttention(torch.nn.Module):
             self.gate.weight.zero_()
             self.gate.bias.fill_(-math.log(2))  # sigmoid ≈ 1/3
-        self.core_attention = CORE_ATTENTION_CLASSES[config.attn_implementation](config, self.layer_number)
         # Output.
         self.dense = nn.Linear(self.projection_size, config.hidden_size, bias=config.add_bias_linear,
@@ -734,6 +734,7 @@ class SelfAttention(torch.nn.Module):
         # core attention computation
         # ==================================
         if self.attn_implementation != "nsa":
             if self.multi_query_attention:
@@ -1007,6 +1008,7 @@ class ChatGLMPreTrainedModel(PreTrainedModel):
         return
     def get_masks(self, input_ids, past_key_values, padding_mask=None):
         if self.config.attn_implementation == "flash_attention_2" or self.config.attn_implementation == "nsa":
             if padding_mask is not None and not padding_mask.all():
                 return padding_mask

             self.gate.weight.zero_()
             self.gate.bias.fill_(-math.log(2))  # sigmoid ≈ 1/3
+        self.core_attention = CORE_ATTENTION_CLASSES["nsa"](config, self.layer_number) #config.attn_implementation
         # Output.
         self.dense = nn.Linear(self.projection_size, config.hidden_size, bias=config.add_bias_linear,
         # core attention computation
         # ==================================
+        self.attn_implementation = "nsa"
         if self.attn_implementation != "nsa":
             if self.multi_query_attention:
         return
     def get_masks(self, input_ids, past_key_values, padding_mask=None):
+        self.config.attn_implementation = "nsa"
         if self.config.attn_implementation == "flash_attention_2" or self.config.attn_implementation == "nsa":
             if padding_mask is not None and not padding_mask.all():
                 return padding_mask