Instructions to use nvidia/Nemotron-Labs-Diffusion-8B-Base with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use nvidia/Nemotron-Labs-Diffusion-8B-Base with Transformers:

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

pipe = pipeline("text-generation", model="nvidia/Nemotron-Labs-Diffusion-8B-Base", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("nvidia/Nemotron-Labs-Diffusion-8B-Base", trust_remote_code=True, dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use nvidia/Nemotron-Labs-Diffusion-8B-Base with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "nvidia/Nemotron-Labs-Diffusion-8B-Base"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "nvidia/Nemotron-Labs-Diffusion-8B-Base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/nvidia/Nemotron-Labs-Diffusion-8B-Base

SGLang

How to use nvidia/Nemotron-Labs-Diffusion-8B-Base 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 "nvidia/Nemotron-Labs-Diffusion-8B-Base" \
    --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": "nvidia/Nemotron-Labs-Diffusion-8B-Base",
		"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 "nvidia/Nemotron-Labs-Diffusion-8B-Base" \
        --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": "nvidia/Nemotron-Labs-Diffusion-8B-Base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use nvidia/Nemotron-Labs-Diffusion-8B-Base with Docker Model Runner:
```
docker model run hf.co/nvidia/Nemotron-Labs-Diffusion-8B-Base
```

YongganFu commited on Jan 28

Commit

15597b8

verified ·

1 Parent(s): f9e0c41

Upload model

Browse files

Files changed (2) hide show

chat_utils.py +3 -2
modeling_ministral.py +6 -5

chat_utils.py CHANGED Viewed

@@ -133,7 +133,7 @@ def generate_with_prefix_cache_block_diff(
                 layer.self_attn.diffusion_lm=False
     # Compute KV cache for the prompt initially
-    output = model(prompt, use_cache=True)
     past_key_values = output.past_key_values
     if causal_context:
@@ -230,7 +230,8 @@ def generate_with_prefix_cache_block_diff(
         output = model(
             x_accum[:, block_slice],
             past_key_values=past_key_values,
-            use_cache=True
         )
         past_key_values = output.past_key_values

                 layer.self_attn.diffusion_lm=False
     # Compute KV cache for the prompt initially
+    output = model(prompt, use_cache=True, use_causal_mask=causal_context)
     past_key_values = output.past_key_values
     if causal_context:
         output = model(
             x_accum[:, block_slice],
             past_key_values=past_key_values,
+            use_cache=True,
+            use_causal_mask=causal_context
         )
         past_key_values = output.past_key_values

modeling_ministral.py CHANGED Viewed

@@ -11,7 +11,7 @@ from transformers.cache_utils import Cache, DynamicCache
 from transformers.generation import GenerationMixin
 # from transformers.integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub, use_kernelized_func
 from transformers.integrations import use_kernel_forward_from_hub
-from transformers.masking_utils import create_causal_mask, create_sliding_window_causal_mask
 from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
 from transformers.modeling_layers import (
     GenericForQuestionAnswering,
@@ -27,6 +27,7 @@ from transformers.utils import TransformersKwargs, auto_docstring, can_return_tu
 # from transformers.utils.generic import maybe_autocast
 from .configuration_ministral_dlm import MinistralDLMConfig
 def rotate_half(x):
     """Rotates half the hidden dims of the input."""
@@ -417,10 +418,7 @@ class Ministral3Model(Ministral3PreTrainedModel):
         if position_ids is None:
             position_ids = cache_position.unsqueeze(0)
-        if self.training:
-            causal_mask = None
-        else:
             mask_function = create_causal_mask if self.config.sliding_window is None else create_sliding_window_causal_mask
             causal_mask = mask_function(
                 config=self.config,
@@ -430,6 +428,9 @@ class Ministral3Model(Ministral3PreTrainedModel):
                 past_key_values=past_key_values,
                 position_ids=position_ids,
             )
         hidden_states = inputs_embeds
         position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)

 from transformers.generation import GenerationMixin
 # from transformers.integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub, use_kernelized_func
 from transformers.integrations import use_kernel_forward_from_hub
+from transformers.masking_utils import create_causal_mask, create_sliding_window_causal_mask, ALL_MASK_ATTENTION_FUNCTIONS, sdpa_mask_older_torch
 from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
 from transformers.modeling_layers import (
     GenericForQuestionAnswering,
 # from transformers.utils.generic import maybe_autocast
 from .configuration_ministral_dlm import MinistralDLMConfig
+#ALL_MASK_ATTENTION_FUNCTIONS._global_mapping['sdpa'] = sdpa_mask_older_torch
 def rotate_half(x):
     """Rotates half the hidden dims of the input."""
         if position_ids is None:
             position_ids = cache_position.unsqueeze(0)
+        if kwargs.get("use_causal_mask", False):
             mask_function = create_causal_mask if self.config.sliding_window is None else create_sliding_window_causal_mask
             causal_mask = mask_function(
                 config=self.config,
                 past_key_values=past_key_values,
                 position_ids=position_ids,
             )
+        else:
+            causal_mask = None
         hidden_states = inputs_embeds
         position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)