Qwen.PrecisionPlan.ixx File Reference

The Qwen 3.8-27B per-role weight allocation, as a named plan struct. More...

#include <concepts>
import Dnn.Quantization.Weight.PrecisionPlan;
import Dnn.Quantization.Weight.Policies;

Classes
struct	Mila::Dnn::QwenOraclePrecisionPlan
	The Section 5 oracle: FP4 across the body, everything else as the plan has it. More...
struct	Mila::Dnn::QwenPrecisionPlan
	The section 5 allocation: 25.62 B resident weights at 2.90 bits average, 8.65 GiB. More...
struct	Mila::Dnn::QwenPlanDetail::QwenPrecisionPlanSelector< T, kIsBarePolicy >
struct	Mila::Dnn::QwenPlanDetail::QwenPrecisionPlanSelector< T, true >
struct	Mila::Dnn::QwenUniformPrecisionPlan< TPolicy >
	One policy across every Qwen role, DeltaNet included. More...

Namespaces
namespace	Mila
	Mila main API namespace.

Concepts
concept	Mila::Dnn::DeltaNetPrecisionRoles
	Roles a Gated DeltaNet block instantiates, on top of the feed-forward pair.

Typedefs
template<typename T>
using	Mila::Dnn::QwenPrecisionPlanFor
	The Qwen lift: accepts a plan or a bare policy, and yields a plan WITH the DeltaNet roles.
using	Mila::Dnn::QwenReferencePrecisionPlan = QwenUniformPrecisionPlan<NoWeightQuant>
	The reference-precision plan: everything BF16, nothing quantized.

Detailed Description

The Qwen 3.8-27B per-role weight allocation, as a named plan struct.

This file IS Specifications/Qwen3.8.md section 5, expressed as types. It is the one declaration a reader consults to learn what the model spends its 12 GiB on, which is the property the whole per-role mechanism exists to preserve (section 6).

Typedef Documentation

QwenPrecisionPlanFor

template<typename T>

using Mila::Dnn::QwenPrecisionPlanFor

export

Initial value:

 
typename QwenPlanDetail::QwenPrecisionPlanSelector<T, WeightQuantPolicy<T>>::type

The Qwen lift: accepts a plan or a bare policy, and yields a plan WITH the DeltaNet roles.

The generic PrecisionPlanFor cannot serve the DeltaNet block, because the generic uniform plan deliberately does not name this family's roles – a block kind must not be able to invalidate plans that already work. So the family that adds roles adds the lift that carries them, and QwenDeltaNetBlock<Cuda, BF16, PerGroupFp4<128>> stays a valid spelling alongside <..., QwenPrecisionPlan>.

QwenReferencePrecisionPlan

using Mila::Dnn::QwenReferencePrecisionPlan = QwenUniformPrecisionPlan<NoWeightQuant>

export

The reference-precision plan: everything BF16, nothing quantized.

Phase 4 gates the chassis against the HF reference at BF16 before Phase 5 applies the allocation above, so both must be spellable at a declaration site. Named here so the two arms of that comparison read as siblings rather than as a plan and a bare policy.