Q1: Qwen-Image-2512 LoRA infrastructure gate
Date: 2026-08-25 UTC
Accepted run: 20260825T230424Z-lora-smoke
Code: 3c56f9fc1eade8f52e7d5b8c2a0b23ae59c1091f
Result: infrastructure pass; no morphology-quality claim
Purpose
Q1 was deliberately limited to eight balanced, hash-checked training images and one epoch. Its purpose was to prove that the pinned Qwen-Image-2512 transformer can produce a loadable rank-32 LoRA checkpoint on one A100 40 GB, and that the checkpoint can generate all four fixed validation cells. Q1 was not sized to improve clinical fidelity.
Failure-driven implementation record
- Attempt 1 loaded and quantised all 720 intended transformer layers, then failed because the offline training processor was missing.
- Attempt 2 fixed the processor and reached the first batch, then measured a 5.06 GiB VAE allocation shortfall with 37.20 GiB already resident.
- Attempt 3 reached the transformer forward pass with CPU offload, then exposed a bitsandbytes NF4/CUDA-stream incompatibility when quantised layers moved between devices.
- Attempt 4 used the publisher's split-training boundary: stage 1 cached fixed text-encoder and VAE features; stage 2 loaded only the NF4 transformer and trained the LoRA from those cached features. This run passed.
Every attempt retained the same eight images, prompts, 1024² maximum pixels, rank, optimiser, learning rate, and one-epoch bound. Failures did not become opportunities to change or cherry-pick the scientific inputs.
Accepted-run results
- Eight unique permitted source images, two per category; all were preassigned to the training split.
- Eight cached feature files produced in stage 1.
- Eight transformer-training steps completed in approximately 18.8 seconds, about 2.35 seconds per step after loading.
- One 472,047,184-byte checkpoint:
epoch-0.safetensors. - Checkpoint SHA-256:
97354a92bb9bd861b0aaf9be00f18a7fe02a29a4b8fd61748960b8a7dfddcd72. - The pinned inference pipeline loaded the checkpoint and patched all 720 LoRA tensors.
- Four fixed 1024×1024 RGB validation PNGs passed decoding, dimensions, manifest, checkpoint-hash linkage, and SHA-256 checks.
- Twenty-three files matched the run's internal
ALL_SHA256SUMSmanifest after retrieval. - Worker runtime was 652 seconds. Total attempt-4 A100 uptime from GCE start/stop timestamps was approximately 853 seconds, about US$0.87 at the previously recorded US$3.673385/hour list rate before disk, credits, and tax.
| Category | Generation seconds | Peak allocated VRAM |
|---|---|---|
| Normal | 113.643 | 19.947 GiB |
| H. pylori with gastritis | 63.021 | 19.949 GiB |
| Intestinal metaplasia | 61.285 | 19.947 GiB |
| Mixed | 61.355 | 19.948 GiB |
The A100 and temporary CPU recovery worker were stopped. Temporary recovery snapshots, disks, and instances were removed after local verification; the stopped original Qwen disk remains the reproducible model/cache source.
Visual audit
The Q0/Q1/held-out comparison sheet has SHA-256 fe132ea0666298b83adc0b27167b2df63e78657020fb6493a22f8a58b4e98945.
Q1 outputs remain visually close to Q0: the normal field is still wood-grain-like, H. pylori objects remain grossly oversized, and the metaplasia/mixed fields do not match held-out morphology. This is expected from eight samples and one epoch. No Q1 output is submission-eligible, and Q1 does not pass the Q2 morphology gate.
Gate decision
Q1 passes its infrastructure exit criterion: a checkpoint was generated, hash-verified, loaded by the pinned Qwen pipeline, and used to produce four mechanically valid cells without a split violation. Advance to a bounded Q2 morphology run using only the accepted HiESD corpus v2. Checkpoint selection in Q2 must be based on aggregate preregistered comparison, not on attractive individual images.