Instructions to use nvidia/Nemotron-Labs-Diffusion-3B-Base with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use nvidia/Nemotron-Labs-Diffusion-3B-Base with Transformers:

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

pipe = pipeline("text-generation", model="nvidia/Nemotron-Labs-Diffusion-3B-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-3B-Base", trust_remote_code=True, dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use nvidia/Nemotron-Labs-Diffusion-3B-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-3B-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-3B-Base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/nvidia/Nemotron-Labs-Diffusion-3B-Base

SGLang

How to use nvidia/Nemotron-Labs-Diffusion-3B-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-3B-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-3B-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-3B-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-3B-Base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use nvidia/Nemotron-Labs-Diffusion-3B-Base with Docker Model Runner:
```
docker model run hf.co/nvidia/Nemotron-Labs-Diffusion-3B-Base
```

abhgarg commited on 7 days ago

Commit

96d0421

verified ·

1 Parent(s): 36a1c93

Clean up rope params; ensure transformers 4.55/5.0 compatibility

Browse files

- Remove duplicate top-level rope_scaling block and stray rope_theta from config.json
- Remove duplicate 'type' key from rope_parameters
- For 3B-Base/8B-Base: set max_position_embeddings=4096 and factor=0.25 to match training
- Mirror rope_theta and rope_scaling from rope_parameters in MinistralDLMConfig for v4.55 yarn
- Drop unused sdpa_mask_older_torch import (removed in transformers v5.0)
- Bump transformers_version to 5.0.0

Files changed (3) hide show

config.json +4 -18
configuration_ministral_dlm.py +4 -3
modeling_ministral.py +1 -1

config.json CHANGED Viewed

@@ -31,7 +31,7 @@
   "initializer_range": 0.02,
   "intermediate_size": 9216,
   "mask_token_id": 100,
-  "max_position_embeddings": 262144,
   "mlp_bias": false,
   "model_type": "ministral_dlm",
   "multi_sampling": null,
@@ -47,33 +47,19 @@
   "rope_parameters": {
     "beta_fast": 32.0,
     "beta_slow": 1.0,
-    "factor": 16.0,
     "llama_4_scaling_beta": 0.1,
     "mscale": 1.0,
     "mscale_all_dim": 1.0,
     "original_max_position_embeddings": 16384,
     "rope_theta": 1000000.0,
-    "rope_type": "yarn",
-    "type": "yarn"
   },
-  "rope_scaling": {
-    "beta_fast": 32.0,
-    "beta_slow": 1.0,
-    "factor": 16.0,
-    "llama_4_scaling_beta": 0.1,
-    "mscale": 1.0,
-    "mscale_all_dim": 1.0,
-    "original_max_position_embeddings": 16384,
-    "rope_theta": 1000000.0,
-    "rope_type": "yarn",
-    "type": "yarn"
-  },
-  "rope_theta": 1000000.0,
   "sliding_window": null,
   "tie_word_embeddings": false,
   "tok_mask_half_life_ratio": null,
   "torch_dtype": "bfloat16",
-  "transformers_version": "4.55.4",
   "use_cache": false,
   "vocab_size": 131072
 }

   "initializer_range": 0.02,
   "intermediate_size": 9216,
   "mask_token_id": 100,
+  "max_position_embeddings": 4096,
   "mlp_bias": false,
   "model_type": "ministral_dlm",
   "multi_sampling": null,
   "rope_parameters": {
     "beta_fast": 32.0,
     "beta_slow": 1.0,
+    "factor": 0.25,
     "llama_4_scaling_beta": 0.1,
     "mscale": 1.0,
     "mscale_all_dim": 1.0,
     "original_max_position_embeddings": 16384,
     "rope_theta": 1000000.0,
+    "rope_type": "yarn"
   },
   "sliding_window": null,
   "tie_word_embeddings": false,
   "tok_mask_half_life_ratio": null,
   "torch_dtype": "bfloat16",
+  "transformers_version": "5.0.0",
   "use_cache": false,
   "vocab_size": 131072
 }

configuration_ministral_dlm.py CHANGED Viewed

@@ -153,7 +153,6 @@ class MinistralDLMConfig(PretrainedConfig):
         tie_word_embeddings=False,
         rope_theta=1000000.0,
         rope_parameters=None,
-        rope_scaling=None,
         attention_bias=False,
         attention_dropout=0.0,
         mlp_bias=False,
@@ -200,9 +199,11 @@ class MinistralDLMConfig(PretrainedConfig):
         self.initializer_range = initializer_range
         self.rms_norm_eps = rms_norm_eps
         self.use_cache = use_cache
-        self.rope_theta = rope_theta
         self.rope_parameters = rope_parameters
-        self.rope_scaling = rope_scaling
         self.attention_bias = attention_bias
         self.attention_dropout = attention_dropout
         self.mlp_bias = mlp_bias

         tie_word_embeddings=False,
         rope_theta=1000000.0,
         rope_parameters=None,
         attention_bias=False,
         attention_dropout=0.0,
         mlp_bias=False,
         self.initializer_range = initializer_range
         self.rms_norm_eps = rms_norm_eps
         self.use_cache = use_cache
         self.rope_parameters = rope_parameters
+        # `rope_theta` is read at the top level by transformers v4.55's yarn impl; mirror from rope_parameters when present.
+        self.rope_theta = (rope_parameters or {}).get("rope_theta", rope_theta)
+        # v4.55 reads rope params from `rope_scaling`; in v5.0 `rope_scaling` is a property alias for rope_parameters.
+        self.rope_scaling = rope_parameters
         self.attention_bias = attention_bias
         self.attention_dropout = attention_dropout
         self.mlp_bias = mlp_bias

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, ALL_MASK_ATTENTION_FUNCTIONS, sdpa_mask_older_torch
 from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
 from transformers.modeling_layers import (
     GenericForQuestionAnswering,

 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
 from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
 from transformers.modeling_layers import (
     GenericForQuestionAnswering,