Module precision_plan 

2026-09-28: The declared precision plan: what a checkpoint’s quantization_config declares, per linear layer, for the weight AND the input activations.

This is the one place that reads the precision half of quantization_config (compressed-tensors config_groups, ModelOpt quantized_layers / quant_algo, and the HF fp8 block scheme). Loaders and kernel dispatch ask DeclaredPrecisionPlan::resolve for a module and run what it says by default. A lever may only go BELOW what the plan declares; running above it is not a silent default either.

Owner: config (quantization). Invariants:

• Pure: no environment, no I/O. A malformed block is an error, never a guess.
• A module the plan does not cover (ignored, unmatched, or no quantization at all) resolves to LayerPrecision::UNQUANTIZED: 16-bit weights and activations.

Structs

DeclaredPrecisionPlan

2026-09-28: The per-layer precision a checkpoint declares.

LayerPrecision

2026-09-28: What one linear layer declares. None means that operand is not quantized (16-bit).

Operand

2026-09-28: One quantized operand as declared.

Rule

2026-09-28: One scheme and the modules it targets.

Enums

Granularity

2026-09-28: How many values share one scale.

IgnoreDialect

2026-09-30: Which format’s matching rules an ignore / exclude entry follows.

NumKind

2026-09-28: Element type of a quantized operand.

PlanSource

2026-09-28: Which quantization_config dialect the plan came from.

ScaleTiming

2026-09-28: When an operand’s scales are computed.

Target

2026-09-28: One target or ignore entry.