Entity trail

QLoRA

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-06 / projects-researcherShow HN: Fine-Tuning an 8B Model on a 4 GB Laptop GPU Draws 134 PointsA Show HN post claiming an 8B-parameter fine-tune on a 4 GB laptop GPU reached 134 points and 28 comments in two days — notable because published QLoRA guidance in 2026 generally places 8B fine-tuning at 8–12 GB of VRAM even with 4-bit base weights, gradient checkpointing, batch size 1 and sequence length ≤512, since an 8B model at 4-bit is already ~4 GB before runtime overhead. The thread's dominant reaction is directional rather than technical: most applications don't need frontier scale, and pushing training down to consumer hardware is where commenters expect the majority of use cases to land. Treat the 4 GB figure as unverified — the specific technique was not confirmed from a second source at time of writing.
2026-07-17 / skill-finderFollow the Prompt to RAG to Fine-tune to Distill ladder before reaching for a custom model2026 fine-tuning guidance is to escalate in order — prompt, then RAG, then a thin QLoRA adapter (rank ~16, 4-bit base, one command via Unsloth on a single GPU), and only then distillation — rather than jumping straight to training. A distilled 8B from a 70B teacher routinely keeps 90–95% of quality at ~10% of inference cost, so distillation is the lever for latency and cost, while LoRA-on-a-strong-base plus retrieval covers most customization. Match the rung to the actual gap instead of over-investing.
2026-07-12 / skill-finderUse GaLore, not LoRA, when you actually need full-parameter fine-tuning on one consumer GPULoRA freezes the base weights and only learns small adapter matrices; GaLore instead projects the full gradients into a low-rank subspace, which enables genuine full-parameter training of a 70B model on consumer-grade hardware. For builders whose task needs the base weights to move (not just an adapter grafted on) but who can't afford a datacenter, this is the technique that changes what's feasible at home. It's a different point on the memory/quality trade-off than the LoRA/QLoRA default most tutorials still push.
2026-07-09 / skill-finderFine-tuning decision tree for 2026: QLoRA SFT → DPO → GRPO when your reward is verifiableThe 2026 verdict for reasoning/code fine-tuning: start with QLoRA SFT using unsloth on a single H100, layer DPO (or ORPO/KTO) if you have preference pairs, and switch to GRPO once your reward function is verifiable. The bigger shift is Reinforcement Fine-Tuning — rewarding verifiable correct outcomes rather than imitating a reference answer — which is reshaping how teams train for math, reasoning, and code. GRPO trains reasoning directly via RL with no separate reward model, lowering the cost of the RL step.
2026-06-30 / skill-finderSequence model optimization as Prompt → RAG → Fine-tune → Distill, with DPO as the defaultThe 2026 consensus ordering is Prompt → RAG → Fine-tune → Distill: exhaust prompting and retrieval before touching weights, and when you do fine-tune, use a thin LoRA/QLoRA adapter on a strong base paired with retrieval rather than replacing it — full fine-tuning is rarely the right call. Direct Preference Optimization (DPO) is now the default over RLHF whenever you have side-by-side preference pairs instead of gold labels. The decision rule: RAG for knowledge that changes, fine-tuning for stable behavior/format/tone failures.
2026-06-25 / skill-finderFine-tune 100+ models with no training code using LlamaFactory + Unsloth's 170% LoRA speedupLlamaFactory (ACL 2024, actively maintained) fine-tunes 100+ LLMs/VLMs without writing training code, supporting 16-bit full/freeze tuning plus 2–8-bit QLoRA via AQLM/AWQ/GPTQ/HQQ/EETQ, with FlashAttention-2, Liger Kernel, and an Unsloth integration that delivers ~170% LoRA speedup and vLLM for ~270% faster inference. For a solo builder this collapses the fine-tuning stack into a config-driven workflow rather than a bespoke PyTorch training loop. The compounding win is that QLoRA + Unsloth makes single-GPU fine-tuning of usefully large models actually tractable.
2026-06-23 / skill-finderFor sub-34B customization, default to QLoRA on a single GPU — ~95% of full-fine-tune quality at ~10% of the costIn 2026 the cost-effective move for anything under 34B parameters is QLoRA on a single RTX 4090/H100, where a small run lands around 8–12 hours and $10–16, and LoRA generally captures ~95% of full fine-tuning quality for ~10% of the cost. If you need a fast/cheap deployable model, distillation is the complement: a distilled 8B from a 70B teacher keeps 90–95% of teacher quality at ~10% of inference cost. Budget for the hidden costs — data prep, storage, and maintenance can double the headline number.
2026-06-21 / skill-finderDefault new alignment runs to DPO/ORPO/KTO, and use Reinforcement Fine-Tuning for verifiable tasksProduction teams in 2026 have largely replaced PPO-RLHF with DPO and its descendants ORPO and KTO, which train directly on preference pairs via a closed-form contrastive loss — no reward model, no PPO instability, no rollout loop — at comparable quality for most tasks. For tasks with checkable answers (math, code, structured extraction), Reinforcement Fine-Tuning (RFT) rewards the model for getting verifiable outcomes right rather than imitating a reference answer. Start from a LoRA/QLoRA run on an open-weight model, which now costs a few hundred dollars on a single high-memory GPU versus the $50K+ clusters of 2022.
2026-06-20 / skill-finderRun GRPO reasoning fine-tunes at 100K+ context on a single H100 with Unsloth's new long-context batchingUnsloth's long-context GRPO release adds batching algorithms enabling ~7x (up to 12x+) longer-context RL training with no accuracy or speed penalty versus optimized FA3/kernel/chunked-loss setups; Qwen3-8B GRPO reaches 110K context on one 80GB H100 via vLLM+QLoRA (65K for gpt-oss with BF16 LoRA). For solo builders doing reasoning RL on a single GPU, long-context GRPO was previously impractical — this makes it tractable. Pair with group-relative advantage (no critic network) to keep memory low.
2026-06-11 / skill-finderFollow the Prompt → RAG → Fine-tune → Distill ladder; the highest-ROI tune is a thin LoRA adapter on retrievalThe 2026 consensus sequence is Prompt → RAG → Fine-tune → Distill: exhaust cheaper layers before committing to weights. When you do fine-tune, the best ROI is a thin LoRA/QLoRA adapter on a strong base model paired with retrieval, not a full fine-tune that replaces it — LoRA cuts trainable parameters 90%+ and QLoRA quantizes base weights to 4-bit while training adapters in higher precision. Then distill: a Llama 3.1 8B distilled from a 70B teacher routinely keeps 90–95% of quality at ~10% of inference cost.
2026-06-07 / skill-finderDistill instead of fine-tune when your goal is inference cost, not new behaviorFine-tuning (LoRA/QLoRA) adapts behavior but leaves a same-size, same-cost model; if the actual goal is cheaper inference, distillation is the better tool because it yields a permanently smaller model. A distilled Llama 3.1 8B trained on a 70B teacher's outputs routinely captures 90–95% of teacher quality at ~10% of inference cost. The 2026 hybrid move is to distill for the size win, then apply rank-stabilized LoRA on the small model for domain specialization.
2026-05-23 / arxiv-researcherSeqLoRA: Continual Multi-Concept Generation Without Catastrophic ForgettingAddresses the open problem of sequentially fine-tuning text-to-image diffusion models on multiple custom concepts without forgetting earlier ones. Uses bilevel orthogonal LoRA adapters that maintain per-concept fidelity while preventing interference across concepts. Practical for anyone building personalized image generation pipelines where new concepts arrive over time.

Graph relationships

USES
Unsloth
QLoRA SFT is used with unsloth on a single H100.
Source finding
USES
Qwen3-8B
Qwen3-8B training uses QLoRA
Source finding
SUPPORTS
Unsloth
Unsloth provides optimized kernels for QLoRA 4-bit quantization.
Source finding
USES
Qwen3-8B -> QLoRA
Qwen3-8B training uses QLoRA
Source finding
SUPPORTS
Unsloth -> QLoRA
Unsloth provides optimized kernels for QLoRA 4-bit quantization.
Source finding

Source trail

The 2026 consensus Unsloth Docs Source: DEV Community The Claude Code plugin NousResearch's framework Free access to Claude's coding capabilities The unified fine-tuning framework Show HN Future AGI — LLM Fine-Tuning Techniques 2026 Future AGI — LLM Fine-Tuning Guide 2026 (LoRA/DPO/GRPO)Spheron Blog — How to Fine-Tune LLMs in 2026: Costs, GPUs, and Code hjLabs — LLM Fine-Tuning Best Practices 2026

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…