Portfolio 10 · adapter state

LoRA and QLoRA fine-tuning planner

Derive trainable adapter parameters from selected matrix dimensions, separate every memory component, require activation and throughput calibration, and price the whole rented instance.

Decision this answers: Does the chosen adapter topology fit with reserved HBM, and what runtime and cost follow from a compatible measured training rate?

Scenario

Decision summary

Trainable adapter parameters

11927552exact-derived

Exact sum across selected projections and layers.

[1]

HBM with headroom

15.14 GiBexact-derived

Calibrated fit: true.

[2][4]

Measured runtime

3.91 hmeasured

Rounded steps divided by measured steps/second.

[3]

Whole-instance cost

$13.28exact-derived

Runtime multiplied by complete instance rate.

[2]

Direct-labeled result profile

Trainable adapter parameters HBM with headroom Measured runtime Whole-instance cost

The tables below are the accessible source of truth; bar lengths never carry identity alone.

Fine-tuning strategy ledger

Fine-tuning strategy ledger
Candidate	Base state	Trainable state	Activation evidence	Runtime evidence	Quality claim	Evidence
Full fine-tuning	Full-precision trainable base	All selected weights	Calibration required	Measurement required	None	unknown[4]
LoRA	Frozen base	Exact low-rank matrices	Calibration required	Measurement required	None	exact-derived[1][3]
QLoRA	Quantized frozen base	Exact low-rank matrices	Calibration required	Measurement required	None	exact-derived[2]

Official surface versus this site

LoRA paper

Use the official surface to confirm native prices, counting rules, limits, and product eligibility.

This offline workbench

The papers define low-rank adaptation and quantized fine-tuning. This page expands target matrices into exact parameters and reconciles frozen weights, adapter state, activations, workspace, headroom, steps, and measured runtime.

Sources and methodology

Every result-affecting reference is visible here without JavaScript and is retained in the JSON export.

[1] LoRA paper

Authority: Microsoft Research
Native identifier: arXiv:2106.09685
Unit: rank and matrix dimensions
Evidence: exact-derived
As of: 2026-08-01
Retrieved: 2026-08-01
Freshness: reference
Normalization: Adapter parameters use r(in+out) per targeted matrix without a percentage shortcut.

Open the authoritative source

[2] QLoRA paper

Authority: University of Washington
Native identifier: NeurIPS 2023 QLoRA
Unit: quantized base and metadata
Evidence: exact-derived
As of: 2026-08-01
Retrieved: 2026-08-01
Freshness: reference
Normalization: Quantized base state remains distinct from trainable adapter state and empirical workspace.

Open the authoritative source

[3] Hugging Face PEFT LoRA guide

Authority: Hugging Face
Native identifier: PEFT LoRA
Unit: target modules
Evidence: exact-derived
As of: 2026-08-01
Retrieved: 2026-08-01
Freshness: reference
Normalization: Selected modules are expanded into actual matrix dimensions and layer counts.

Open the authoritative source

[4] PyTorch activation checkpointing

Authority: PyTorch
Native identifier: torch.utils.checkpoint
Unit: activation memory
Evidence: measured
As of: 2026-08-01
Retrieved: 2026-08-01
Freshness: reference
Normalization: Activation memory remains a user calibration and is never manufactured from the model name.

Open the authoritative source

Limits and non-claims

Full, LoRA, and QLoRA feasibility and cost do not imply equal quality or convergence.
Without measured activation memory the fit is incomplete; without measured throughput runtime and cost are unproven.
Framework-specific kernels, allocator fragmentation, and every optimization are outside v1.