File size: 3,333 Bytes
05b0ab9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 | """
HSAQ — Hybrid Sensitivity-Aware Quantization
=============================================
Novel mixed-precision quantization pipeline that combines:
1. Per-layer sensitivity profiling (output drift measurement)
2. Memory-budget-aware tier assignment (critical / normal / tolerant)
3. Mixed-precision HQQ/AWQ/GPTQ quantization (3/4-bit) per sensitivity tier
4. Optional structured attention-head pruning for tolerant layers
5. Optional 2-bit quantization for tolerant layers (quality cliff risk)
6. LoRA domain adapter training for quality recovery
7. 5-stage model hunter: discover → filter → score → profile → emit
Target: Fit 13-20B models on 12 GB consumer GPUs with all layers on GPU.
No CPU offload — PCIe shuffle per token tanks inference 5-10x.
Usage:
from quantization.hsaq import HSAQPipeline, ModelHunterPipeline
# Single-model quantization
pipeline = HSAQPipeline(
model_id="Qwen/Qwen2.5-14B-Instruct",
gpu_budget_gb=11.2,
calibration_dataset="wikitext",
)
pipeline.run() # profiles → classifies → quantizes → adapts
# Multi-model hunter
hunter = ModelHunterPipeline(HunterConfig())
results = hunter.run() # discover → filter → score → profile → emit
"""
from quantization.hsaq.adapter import LoRAAdapterTrainer
from quantization.hsaq.assignment import (
Assignment,
AssignmentResult,
BudgetInfeasibleError,
LayerCandidate,
LayerOption,
assign_bit_widths,
pareto_frontier,
)
from quantization.hsaq.budget import MemoryBudgetCalculator
from quantization.hsaq.candidate import (
DiscoveryStage,
EmitStage,
FilterConfig,
FilterStage,
HunterConfig,
ModelHunterPipeline,
ScoreStage,
compute_model_hash,
extract_arch_from_config,
kv_bytes_per_token,
predict_vram_mixed_34bit,
)
from quantization.hsaq.candidate_record import (
ArchType,
CandidateRecord,
EligibilityTier,
predict_kv_gb,
predict_weights_gb,
)
from quantization.hsaq.config import (
HSAQBudget,
HSAQConfig,
LayerSensitivity,
LayerTier,
SensitivityResult,
TierBudget,
)
from quantization.hsaq.pipeline import HSAQPipeline
from quantization.hsaq.pruner import AttentionHeadPruner
from quantization.hsaq.sensitivity import (
PIPELINE_VERSION,
SensitivityCacheDB,
SensitivityProfiler,
)
__all__ = [
"PIPELINE_VERSION",
"ArchType",
"Assignment",
"AssignmentResult",
"AttentionHeadPruner",
"BudgetInfeasibleError",
"CandidateRecord",
"DiscoveryStage",
"EligibilityTier",
"EmitStage",
"FilterConfig",
"FilterStage",
"HSAQBudget",
"HSAQConfig",
"HSAQPipeline",
"HunterConfig",
"LayerCandidate",
"LayerOption",
"LayerSensitivity",
"LayerTier",
"LoRAAdapterTrainer",
"MemoryBudgetCalculator",
"ModelHunterPipeline",
"ScoreStage",
"SensitivityCacheDB",
"SensitivityProfiler",
"SensitivityResult",
"TierBudget",
"assign_bit_widths",
"compute_model_hash",
"extract_arch_from_config",
"kv_bytes_per_token",
"pareto_frontier",
"predict_kv_gb",
"predict_vram_mixed_34bit",
"predict_weights_gb",
]
|