Deploy multiple reasoning model sizes efficiently by embedding nested submodels within a single parent—use end-to-end trained routers to select submodels at inference, achieving 360× cost reduction vs training families separately.
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Deploy multiple reasoning model sizes efficiently by embedding nested submodels within a single parent—use end-to-end trained routers to select submodels at inference, achieving 360× cost reduction vs training families separately.
Deploy Efficient Reasoning Models via Nested Submodel Elasticity
Training multiple model sizes (12B, 9B, 6B) separately requires training three independent models—expensive and wasteful. Nemotron Elastic embeds multiple nested submodels within a single parent model, sharing weights hierarchically. An end-to-end trained router selects which submodel to extract at inference based on computational budget. This enables deploying various model sizes with constant memory overhead while training only once.
The approach achieves 360× cost reduction vs. training family members separately, enabling efficient deployment of reasoning models across diverse hardware constraints.
Core Concept
Model families (like GPT-3 with 7B, 13B, 175B variants) typically train separately, multiplying training cost. Nemotron Elastic inverts this: train once, extract multiple sizes.
The key insight is weight sharing with hierarchical structure:
Parent Model: Largest model (12B); contains full capability
Nested Submodels: Extractable smaller models (9B, 6B, 3B) using prefix of parent's weights
Router: Learnable component selecting which submodel to use based on task/budget
Elastification Techniques: Domain-specific tricks for Mamba (SSM), MLPs, and layer selection
This enables:
Single training pass (using curriculum on budget diversity)
Multiple deployable sizes
Constant memory (all sizes fit in parent's memory)
360× cost reduction vs. training separate models
Architecture Overview
Nested Model Structure: Submodel i uses first k_i layers, sharing weights with parent; i.e., all weights are reused
Router-Based Selection: Learned module predicting optimal submodel for task/input; can be based on task type, input complexity, or deployment budget
Elastification Techniques: Specialized methods for Mamba (group-aware SSM elastification), MLPs (heterogeneous elastification), and layer selection (normalized MSE-based importance)
Two-Stage Training Curriculum: Stage 1: train parent + routers; Stage 2: distillation from parent to submodels
Knowledge Distillation: Multi-budget optimization ensuring each submodel learns from parent
classLayerImportanceEstimator(nn.Module):
"""
Estimate which layers are most important for each submodel depth.
Use normalized MSE-based layer importance.
"""def__init__(self, num_layers):
super().__init__()
# Learn importance scores for each layerself.layer_importance = nn.Parameter(torch.ones(num_layers))
defcompute_importance(self):
"""Normalize to probability distribution."""return torch.softmax(self.layer_importance, dim=0)
defselect_important_layers(self, num_select):
"""Select top-k important layers."""
importance = self.compute_importance()
_, indices = torch.topk(importance, num_select)
returnsorted(indices.tolist())
defelastify_depth(parent_model, submodel_depth, parent_depth):
"""
Elastification: create submodel using important layers.
For simplicity, use prefix of layers (can extend to importance-based selection).
"""classElasticSubmodel(nn.Module):
def__init__(self, parent, depth):
super().__init__()
self.embedding = parent.embedding
self.layers = parent.layers[:depth]
self.output_proj = parent.output_proj
defforward(self, x):
hidden = self.embedding(x)
for layer inself.layers:
hidden = layer(hidden)
returnself.output_proj(hidden)
return ElasticSubmodel(parent_model, submodel_depth)
classMambaElastification(nn.Module):
"""
Elastification for Mamba SSM models: group-aware SSM elastification.
"""def__init__(self, hidden_dim, num_groups=4):
super().__init__()
self.num_groups = num_groups
self.group_projections = nn.ModuleList([
nn.Linear(hidden_dim, hidden_dim // num_groups) for _ inrange(num_groups)
])
defelastify_ssm(self, ssm_layer, target_hidden_dim):
"""
Reduce SSM hidden dimension while maintaining group structure.
"""classElasticSSM(nn.Module):
def__init__(self, parent_ssm, target_dim, group_proj):
super().__init__()
self.parent_ssm = parent_ssm
self.target_dim = target_dim
self.group_proj = group_proj
defforward(self, x):
# Reduce via learned projection
x_reduced = self.group_proj(x)
# Forward through parent SSM (handles partial dims)# Note: real implementation would adapt SSM internals
output = self.parent_ssm(x_reduced)
return output
return ElasticSSM(ssm_layer, target_hidden_dim, self.group_projections[0])
Step 4: Two-Stage Training Curriculum.
deftrain_elastic_model(
model, train_loader, num_epochs=100,
submodel_depths=[24, 18, 12], lr=5e-5):
"""
Two-stage training: Stage 1 trains parent + routers; Stage 2 distillation.
"""
optimizer = torch.optim.AdamW(model.parameters(), lr=lr)
# Stage 1: Train parent + router (50 epochs)print("Stage 1: Training parent model and router...")
for epoch inrange(num_epochs // 2):
total_loss = 0for batch in train_loader:
input_ids = batch['input_ids']
target_ids = batch['target_ids']
budget_level = batch.get('budget_level', torch.zeros(input_ids.shape[0])) # Optional# Forward: get parent logits and router logits
parent_logits, router_logits = model(input_ids, output_router_logits=True)
# Loss 1: Parent model loss
parent_loss = torch.nn.functional.cross_entropy(
parent_logits.reshape(-1, parent_logits.shape[-1]),
target_ids.reshape(-1)
)
# Loss 2: Router loss (encourage selecting appropriate submodel)# Router logits should align with complexity of input
router_target = estimate_task_difficulty(input_ids)
router_loss = torch.nn.functional.cross_entropy(router_logits, router_target)
total_loss = parent_loss + 0.1 * router_loss
optimizer.zero_grad()
total_loss.backward()
optimizer.step()
print(f"Epoch {epoch} (Stage 1): loss={total_loss.item():.4f}")
# Stage 2: Knowledge distillation from parent to submodels (50 epochs)print("\nStage 2: Knowledge distillation to submodels...")
for epoch inrange(num_epochs // 2, num_epochs):
total_loss = 0for batch in train_loader:
input_ids = batch['input_ids']
target_ids = batch['target_ids']
# Parent outputs (frozen or with slow learning)with torch.no_grad():
parent_logits, _ = model(input_ids, output_router_logits=True)
# Submodel distillation
distill_loss = 0for submodel_idx, depth inenumerate(submodel_depths):
submodel = model.extract_submodel(submodel_idx)
submodel_logits = submodel(input_ids)
# KL divergence between parent and submodel
kl_loss = torch.nn.functional.kl_div(
torch.log_softmax(submodel_logits, dim=-1),
torch.softmax(parent_logits, dim=-1),
reduction='batchmean'
)
distill_loss += kl_loss / len(submodel_depths)
optimizer.zero_grad()
distill_loss.backward()
optimizer.step()
total_loss += distill_loss
print(f"Epoch {epoch} (Stage 2): distill_loss={distill_loss.item():.4f}")
return model
defestimate_task_difficulty(input_ids):
"""
Heuristic: input complexity determines which submodel.
Can be replaced with learned difficulty estimator.
"""# Length-based: longer inputs need larger models
lengths = (input_ids > 0).sum(dim=-1)
difficulty = torch.clamp(lengths // 256, 0, 2) # 0-2 for 3 submodelsreturn difficulty
Practical Guidance
When to Use: Deploying reasoning models across diverse hardware (GPUs, CPUs, edge devices) where model size flexibility is crucial. Cost reduction from single training pass is massive advantage.
Architecture Decisions:
Submodel depths: geometric progression (e.g., 36→24→12 layers) works well; adjust based on task requirements
Router mechanism: simple MLPs sufficient; can add task-specific gating for specialized routing
Elastification: layer prefix works for Transformers; Mamba requires group-aware techniques
Training schedule: 50/50 split between parent training and distillation; adjust ratio based on dataset size
Pitfalls:
Router overconfidence: Router may converge to always selecting one submodel; use entropy regularization
Distillation divergence: Submodels diverging from parent; use stronger KL weight, periodic teacher updates
Submodel collapse: Smaller submodels may learn trivial solutions; use curriculum learning (start large, gradually shrink)
Depth selection: Equal-spaced layer selection may be suboptimal; profile on validation set
When NOT to Use: Single-model deployment where multi-size support unnecessary; models with highly specialized architecture (e.g., custom expert selection).
Integration: Works with Transformers and Mamba; minimal changes to base architecture.
