Text Generation
Transformers
Safetensors
PyTorch
English
nemotron_labs_sigma
feature-extraction
nvidia
nemotron-labs
conversational
custom_code
Instructions to use nvidia/Nemotron-Labs-Sigma-3B-instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use nvidia/Nemotron-Labs-Sigma-3B-instruct with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="nvidia/Nemotron-Labs-Sigma-3B-instruct", trust_remote_code=True) messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("nvidia/Nemotron-Labs-Sigma-3B-instruct", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use nvidia/Nemotron-Labs-Sigma-3B-instruct with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "nvidia/Nemotron-Labs-Sigma-3B-instruct" # 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-Sigma-3B-instruct", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/nvidia/Nemotron-Labs-Sigma-3B-instruct
- SGLang
How to use nvidia/Nemotron-Labs-Sigma-3B-instruct 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-Sigma-3B-instruct" \ --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-Sigma-3B-instruct", "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-Sigma-3B-instruct" \ --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-Sigma-3B-instruct", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use nvidia/Nemotron-Labs-Sigma-3B-instruct with Docker Model Runner:
docker model run hf.co/nvidia/Nemotron-Labs-Sigma-3B-instruct
fix: declare projected_bernstein schedule parameterization
Browse files- config.json +3 -1
- configuration_nemotron_labs_sigma.py +2 -0
- modeling_nemotron_labs_sigma.py +71 -4
config.json
CHANGED
|
@@ -52,5 +52,7 @@
|
|
| 52 |
"pad_token_id": null,
|
| 53 |
"bos_token_id": 1,
|
| 54 |
"eos_token_id": 2,
|
| 55 |
-
"torch_dtype": "bfloat16"
|
|
|
|
|
|
|
| 56 |
}
|
|
|
|
| 52 |
"pad_token_id": null,
|
| 53 |
"bos_token_id": 1,
|
| 54 |
"eos_token_id": 2,
|
| 55 |
+
"torch_dtype": "bfloat16",
|
| 56 |
+
"schedule_parameterization": "projected_bernstein",
|
| 57 |
+
"schedule_degree": 32
|
| 58 |
}
|
configuration_nemotron_labs_sigma.py
CHANGED
|
@@ -78,6 +78,8 @@ class NemotronLabsSigmaConfig(_PretrainedConfig):
|
|
| 78 |
self.rope_parameters = rope_parameters
|
| 79 |
# transformers 4.x reads YaRN settings from `rope_scaling`.
|
| 80 |
self.rope_scaling = dict(rope_parameters)
|
|
|
|
|
|
|
| 81 |
self.embed_dim = embed_dim
|
| 82 |
self.block_size = block_size
|
| 83 |
self.replaid_selfcond = replaid_selfcond
|
|
|
|
| 78 |
self.rope_parameters = rope_parameters
|
| 79 |
# transformers 4.x reads YaRN settings from `rope_scaling`.
|
| 80 |
self.rope_scaling = dict(rope_parameters)
|
| 81 |
+
self.schedule_parameterization = kwargs.pop("schedule_parameterization", "legacy")
|
| 82 |
+
self.schedule_degree = kwargs.pop("schedule_degree", 32)
|
| 83 |
self.embed_dim = embed_dim
|
| 84 |
self.block_size = block_size
|
| 85 |
self.replaid_selfcond = replaid_selfcond
|
modeling_nemotron_labs_sigma.py
CHANGED
|
@@ -238,9 +238,49 @@ class NLCDEncoder(nn.Module):
|
|
| 238 |
# ---------------------------------------------------------------------------
|
| 239 |
|
| 240 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 241 |
class NLCDNoiseSchedule(nn.Module):
|
| 242 |
-
"""
|
| 243 |
-
compatibility; the default edm_sigma sampler does not
|
|
|
|
| 244 |
|
| 245 |
def __init__(self):
|
| 246 |
super().__init__()
|
|
@@ -263,6 +303,8 @@ class NLCDNoiseSchedule(nn.Module):
|
|
| 263 |
|
| 264 |
|
| 265 |
class NLCDGammaBounds(nn.Module):
|
|
|
|
|
|
|
| 266 |
def __init__(self, gamma_0=-3.0, gamma_1=6.0):
|
| 267 |
super().__init__()
|
| 268 |
self.gamma_0 = nn.Parameter(torch.tensor(float(gamma_0)))
|
|
@@ -272,6 +314,22 @@ class NLCDGammaBounds(nn.Module):
|
|
| 272 |
return self.gamma_0.double(), self.gamma_1.double()
|
| 273 |
|
| 274 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 275 |
class NLCDEmbeddingMatrix(nn.Module):
|
| 276 |
"""Vocabulary embedding matrix E in R^{V x embed_dim}, rows on the unit sphere."""
|
| 277 |
|
|
@@ -307,8 +365,17 @@ class NemotronLabsSigmaModel(PreTrainedModel):
|
|
| 307 |
super().__init__(config)
|
| 308 |
self.encoder = NLCDEncoder(config)
|
| 309 |
self.diffusion_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
| 310 |
-
|
| 311 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 312 |
self.replaid_embedding_matrix = NLCDEmbeddingMatrix(config.vocab_size, config.embed_dim)
|
| 313 |
self.replaid_input_linear = nn.Linear(config.embed_dim, config.hidden_size, bias=False)
|
| 314 |
if config.replaid_selfcond:
|
|
|
|
| 238 |
# ---------------------------------------------------------------------------
|
| 239 |
|
| 240 |
|
| 241 |
+
class NLCDBernsteinNoiseSchedule(nn.Module):
|
| 242 |
+
"""Monotone degree-N Bernstein (Bezier) map t -> gamma_tilde in [0, 1].
|
| 243 |
+
|
| 244 |
+
This is the parameterization used by the MS-SUSReg lineage: the free
|
| 245 |
+
parameter is the vector of simplex increments, whose cumulative sum gives
|
| 246 |
+
the Bezier control points. Uniform increments reproduce s(t) = t.
|
| 247 |
+
"""
|
| 248 |
+
|
| 249 |
+
def __init__(self, degree=32):
|
| 250 |
+
super().__init__()
|
| 251 |
+
self.degree = int(degree)
|
| 252 |
+
self.increment_values = nn.Parameter(
|
| 253 |
+
torch.full((self.degree,), 1.0 / self.degree)
|
| 254 |
+
)
|
| 255 |
+
|
| 256 |
+
def increments(self):
|
| 257 |
+
return self.increment_values.double()
|
| 258 |
+
|
| 259 |
+
def control_points(self):
|
| 260 |
+
inc = self.increments()
|
| 261 |
+
return torch.cat((inc.new_zeros(1), inc.cumsum(dim=0)))
|
| 262 |
+
|
| 263 |
+
@staticmethod
|
| 264 |
+
def _de_casteljau(t, control_points):
|
| 265 |
+
t = t.double()
|
| 266 |
+
values = control_points.double().expand(t.numel(), -1)
|
| 267 |
+
flat_t = t.reshape(-1, 1)
|
| 268 |
+
one_minus_t = 1.0 - flat_t
|
| 269 |
+
for remaining in range(control_points.numel() - 1, 0, -1):
|
| 270 |
+
values = (
|
| 271 |
+
one_minus_t * values[:, :remaining]
|
| 272 |
+
+ flat_t * values[:, 1 : remaining + 1]
|
| 273 |
+
)
|
| 274 |
+
return values[:, 0].reshape(t.shape)
|
| 275 |
+
|
| 276 |
+
def forward(self, t):
|
| 277 |
+
return self._de_casteljau(t, self.control_points())
|
| 278 |
+
|
| 279 |
+
|
| 280 |
class NLCDNoiseSchedule(nn.Module):
|
| 281 |
+
"""Legacy learned monotone map t -> gamma_tilde in [0, 1] (kept for
|
| 282 |
+
checkpoint compatibility; the default edm_sigma sampler does not
|
| 283 |
+
evaluate it)."""
|
| 284 |
|
| 285 |
def __init__(self):
|
| 286 |
super().__init__()
|
|
|
|
| 303 |
|
| 304 |
|
| 305 |
class NLCDGammaBounds(nn.Module):
|
| 306 |
+
"""Legacy gamma endpoints stored as two scalar parameters."""
|
| 307 |
+
|
| 308 |
def __init__(self, gamma_0=-3.0, gamma_1=6.0):
|
| 309 |
super().__init__()
|
| 310 |
self.gamma_0 = nn.Parameter(torch.tensor(float(gamma_0)))
|
|
|
|
| 314 |
return self.gamma_0.double(), self.gamma_1.double()
|
| 315 |
|
| 316 |
|
| 317 |
+
class NLCDProjectedGammaBounds(nn.Module):
|
| 318 |
+
"""Gamma endpoints stored as a single 2-vector (MS-SUSReg lineage)."""
|
| 319 |
+
|
| 320 |
+
def __init__(self, gamma_0=-3.0, gamma_1=6.0):
|
| 321 |
+
super().__init__()
|
| 322 |
+
self.endpoint_values = nn.Parameter(
|
| 323 |
+
torch.tensor([float(gamma_0), float(gamma_1)])
|
| 324 |
+
)
|
| 325 |
+
|
| 326 |
+
def forward(self):
|
| 327 |
+
return (
|
| 328 |
+
self.endpoint_values[0].double(),
|
| 329 |
+
self.endpoint_values[1].double(),
|
| 330 |
+
)
|
| 331 |
+
|
| 332 |
+
|
| 333 |
class NLCDEmbeddingMatrix(nn.Module):
|
| 334 |
"""Vocabulary embedding matrix E in R^{V x embed_dim}, rows on the unit sphere."""
|
| 335 |
|
|
|
|
| 365 |
super().__init__(config)
|
| 366 |
self.encoder = NLCDEncoder(config)
|
| 367 |
self.diffusion_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
| 368 |
+
# Two schedule parameterizations exist in the wild. The MS-SUSReg
|
| 369 |
+
# lineage stores Bernstein increments and a gamma endpoint 2-vector;
|
| 370 |
+
# earlier checkpoints store an MLP schedule and two gamma scalars.
|
| 371 |
+
if getattr(config, "schedule_parameterization", "legacy") == "projected_bernstein":
|
| 372 |
+
self.noise_schedule = NLCDBernsteinNoiseSchedule(
|
| 373 |
+
degree=getattr(config, "schedule_degree", 32)
|
| 374 |
+
)
|
| 375 |
+
self.gamma_bounds = NLCDProjectedGammaBounds()
|
| 376 |
+
else:
|
| 377 |
+
self.noise_schedule = NLCDNoiseSchedule()
|
| 378 |
+
self.gamma_bounds = NLCDGammaBounds()
|
| 379 |
self.replaid_embedding_matrix = NLCDEmbeddingMatrix(config.vocab_size, config.embed_dim)
|
| 380 |
self.replaid_input_linear = nn.Linear(config.embed_dim, config.hidden_size, bias=False)
|
| 381 |
if config.replaid_selfcond:
|