Module parallel 

ModelPlan-driven parallel topology and artifact placement contracts.

The automatic loader does not select a family-specific TP/EP recipe. It compiles the canonical operation plan, binds the source tensor census, estimates the checkpoint residency of every legal program, and then selects a registered numeric backend. Family packs remain responsible for semantic tensor/config validation; they do not carry per-layer placement lists.

Structs

GroupedParallelPlan

GroupedTopologyRequest

ModelParallelContract

ParallelPlan

StageGroupRequest

One pipeline stage’s model-specific tensor/expert group.

StageRankGroup

TopologyError

TopologyRequest

Enums

HardwareTarget

Constants

AUTO_PARALLEL_MAX_CARDS

PRODUCT_MAX_CARDS

The execution planner’s supported rank envelope. Hardware qualification and tuned defaults remain model x rig evidence, but the placement/runtime contract must not stop at earlier three-card qualification cells.

STEP37_TRUNK_LAYERS

Functions

validate_fp8_expert_checkpoint

Prove that every routed layer in the structural contract exposes native stacked block-128 E4M3 expert banks. Converted and per-tensor artifacts do not inherit this backend.

validate_nvfp4_expert_checkpoint

Prove that every routed layer in the structural contract exposes native ModelOpt NVFP4 experts: either one stacked bank or the Hugging Face per-expert layout. Both carry packed e2m1 codes, per-16 UE4M3 scales, and finite-positive per-expert macros.

validate_step_fp8_checkpoint

Legacy API retained for existing focused gates.

validate_step_nvfp4_checkpoint

Legacy API retained for existing focused gates.

validate_step_pp_request

Validate the live Step PP request before the loader allocates CUDA state. Checkpoint tensor census is deliberately a separate loader gate: topology legality must remain testable without opening model files, while serving requires both gates.