Module planning 

Greedy Layer Assignment with Fault Tolerance and Adaptive Quantization

This module provides:

• Sequential greedy layer splitting across nodes based on VRAM capacity
• Adaptive quantization trigger (select_quantization_mode)
• Load balancing and fault detection integration

Structs

LayerAssignment

Layer assignment to a specific node

LayerSpec

Layer specification with VRAM requirements

MigrationPlanner

Dynamic migration planner for moving layers between nodes under load.

PlacementPlan

Layer placement plan across nodes

PlanningTuning

Runtime-aware planning hints derived from host auto-detection.

RebalanceTrigger

Trigger mechanism for dynamic rebalancing based on load/health.

Enums

QuantizationMode

Quantization mode enumeration

Constants

DELIVERY_RATIO_INT4_THRESHOLD

DELIVERY_RATIO_INT8_THRESHOLD

Delivery ratio thresholds for adaptive quantization

Functions

assign_layers_sequentially

Assign layers sequentially across nodes based on VRAM capacity.

assign_layers_with_fault_tolerance

Assign layers with fault tolerance and load balancing

assign_layers_with_fault_tolerance_and_runtime

Assign layers with fault tolerance and runtime-aware chunking.

assign_layers_with_runtime_profile

Assign layers using runtime auto-detection to expose worker-parallel chunks.

calculate_cluster_health

Calculate network health across the cluster

chunk_assignments_for_workers

Split large node assignments into smaller contiguous chunks for worker-level parallelism.

rebalance_assignments

Update layer assignments based on node health metrics

select_quantization_mode

Select quantization mode based on cluster health metrics