# Modifications

Modified by ixim / iximbox: eligible linear weights converted from Qwen-Image-2.1 to bitsandbytes LLM.int8 INT8. Built with Qwen. Non-commercial research/evaluation under the accompanying Qwen Research License.

- Converted eligible transformer and text encoder linear layers to LLM.int8, threshold 6.0.
- Retained sensitive projections, normalization, embeddings, vision model and VAE in floating point.
- Re-serialized component weights, shard indexes and configs; no fine-tuning.
- Weight-file headers and modified JSON files contain modification notices.
- Exact quantized module names and dtype counts are in the component reports.

## Runtime and evaluation correction, 2026-09-21

- Added an instance-local bitsandbytes auxiliary-tensor offload workaround and sequential component loading in scripts/runtime.py.
- Re-ran the paired evaluation in isolated processes with startup, loading and per-image memory records.
- Preserved the initial evaluation and failed 1024px editing outputs; added controlled editing diagnostics and a paired 2048px retest with VAE tiling.
- Model weights, quantization configuration and inference-file fingerprints are unchanged.