Skip to content

round7-9 (gated): training above 14B on this card — 27B dense QLoRA at the VRAM edge with Unsloth offload, or a recorded rejection #3088

Description

@Xore

Part of #3079 (round 7). Plan §4. Gated: opens only if round7-5 shows a gain worth scaling, and only after round7-8 has scored it.

The envelope

Unsloth's published QLoRA minimums put 27 B dense at ≈ 22 GB — two gigabytes over the RTX 4000 Ada — and 32 B at 26 GB. MoE bases are out for 4-bit QLoRA altogether (Unsloth: not recommended, bitsandbytes gap), and their 16-bit LoRA needs 60 GB+. Compute is local only (operator decision 2026-09-06), so "training bigger" means the VRAM edge, not a bigger card.

What to try, in order

  1. 27 B dense QLoRA at the edge, on the top dense family (Qwen3.8-27B / Qwen3.6-27B, derestricted variant): use_gradient_checkpointing="unsloth" (activations offloaded to system RAM — this is where the two empty DIMM slots finally matter for training, not just for offload inference), unsloth_tiled_mlp=True, batch 1 with gradient accumulation, sequence 2–4 k, LoRA on attention only first, then add MLP if it fits. Record the peak from nvidia-smi per configuration; an OOM is a measurement — write it down with the config that produced it.
  2. Adapter-only export path: if a 27 B adapter trains but the 16-bit merge does not fit in RAM comfortably, merge on the CPU in the llama.cpp image (convert_lora_to_gguf.py exists there) — Ollama's own ADAPTER import covers Llama/Gemma-2 adapters only, so merging stays the rule.
  3. A rented GPU is not in scope for this round; if the edge attempt fails and the local result justifies it, that is a new operator decision with a budget figure, filed as its own issue.

Deliverables

  • a VRAM table per configuration attempted (fits / OOM, peak, tokens/s)
  • if it trains: the artefact through export_to_ollama.sh at the R3 ladder, scored by round7-8 against its base at the same quant
  • if it does not: the recorded rejection with numbers, so the question is not re-raised from the premise that a 20 GB card should train a 27 B

Depends on round7-5, round7-8 (first results). Gated on the operator.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

enhancementNew feature or requestllmLLM analysis workermlML worker and GPU scoringparkedDeliberately waiting on a named gate (not actionable now)

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions