The latent cache holds two fp32 posterior tensors per crop/flip variant per
image, pinned on CUDA hosts, so datasets with thousands of images can exhaust
host or pinned memory with no fallback. Estimate the cache size from the first
real encoded latent and fall back to per-step VAE encoding when it exceeds a
4 GiB budget. UNSLOTH_DIFFUSION_FORCE_LATENT_CACHE bypasses the gate; the
existing UNSLOTH_DIFFUSION_NO_LATENT_CACHE opt-out is unchanged.
A dual-DiT pipeline (Ideogram 4's unconditional tower) placed its second
denoiser resident under the group tier, which defeats the tier since the
pair rarely fits where one alone did not. Stream transformer_2 and
unconditional_transformer alongside the transformer and keep only the
smaller companions resident.
The perf rewrite dropped the bf16 capability guard, so a pre-Ampere CUDA
device (T4/V100/RTX 20xx) would die deep in model load with an opaque dtype
error instead of a clear message. Restores parity with the SDXL trainer.
- Run the trainer's caption discovery in the start route BEFORE freeing GPU
residents, so a missing or uncaptionable dataset 400s without evicting the
loaded chat/Images model.
- sd.cpp unload now waits out a cancelled one-shot generation on the generate
lock before reporting the device free, matching the diffusers backend.
- Clearing a caption that came from metadata.jsonl writes an empty sidecar
tombstone instead of unlinking (both readers treat an existing sidecar as
authoritative), so the cleared label cannot resurface.
- The ControlNet wrapper pipe is only cached while its load is still current,
closing the unload race the model cache already handled.
The prefetch already scopes the file list (no packaged root singles, no
dtype-variant twins, no ONNX/Flax exports), but from_pretrained was then
called with the hub id, and its own snapshot sweep re-downloaded the
skipped files anyway: 24 GB per FLUX.1 repo and 65 GB on FLUX.2-dev, as
found in the blob cache. Return the snapshot dir from the prefetch (keyed
on the pipeline manifest) and hand it to every pipeline-assembly
from_pretrained site; any prefetch failure keeps the hub id and the old
behavior.
- Free reserved VRAM in the diffusion load worker's failure path: a load-time OOM
never commits _state and the next load's _unload_locked early-returns, so nothing
else reclaimed the half-built pipeline's memory
- Use a monotonic clock for the denoise ETA rate
- Sync _GENERATION_DEFAULTS with the UI table: kontext, flux.2-dev, sdxl-turbo and
SDXL base rows so /v1/images/generations stops falling back to 9 steps / CFG 0
- 400 (not sanitized 500) when /v1/images/generations hits an edit-only model
- Fail fast on pre-Ampere CUDA in the DiT trainer instead of dying in model load
- Run the trainer trust gate in the diffusion training route before freeing GPU
residents so an untrusted base cannot tear down loaded chat/Images models
- Protect native sd.cpp companion VAE/text-encoder repos from cache deletion while
a load is downloading them
- Exempt the task-scoped Images picker from the chat-only GGUF/MLX format gate so
local diffusers pipelines stay selectable on no-GPU hosts
The generic Studio config dict path can deliver these flags as strings, and a
non-empty string like "false" is truthy, so an opt-out silently no-ops (the
latent cache still builds, TF32 stays on). Coerce them the same way
gradient_checkpointing already is.
Measured on the fresh linux x64 prebuilt (z-image Q8_0, sd-cli, 192 CPU
threads, 512x512, 9 steps, steady state): sampling 56.1s vs 51.3s (about
9 percent faster), VAE decode unchanged, peak RSS identical. The sd.cpp
engine only serves the no-GPU tier, so the default profile now matches
max: --diffusion-fa plus --diffusion-conv-direct.
A 28-pair accuracy gate on a B200 (same-seed vs the dense bf16 reference)
found per-row fp8 dynamic quant renders EVERY qwen-image frame black
(mean luma 0.0000, SSIM 0.016), reproduced identically with on-the-fly
quantize_ on the dense transformer, so it is the model's activation range,
not a checkpoint artifact. mxfp8 shows real semantic damage at 1024px
(CLIP delta mean 0.0146, worst cases 0.064/0.102) and nvfp4 measures
LPIPS mean 0.51. int8 dynamic (per-token scales) is excellent on Qwen:
LPIPS mean 0.069, SSIM 0.958.
The per-scheme smoke probe only proves the GEMM kernel runs, so it cannot
catch model-level breakage. Add _FAMILY_SCHEME_DENY consulted by
select_transformer_quant_scheme: auto skips denied schemes (Qwen lands on
int8) and an explicit denied request returns None, the same GGUF-fallback
contract as an unsupported scheme. Family is threaded from the three
diffusion.py call sites; existing behavior is unchanged for every other
family. 4 new tests; 529 diffusion tests green; CI-sim green.
The fp16-on-bf16-family refusal in run_dit_lora_training now fires before the heavy
imports, so a host without diffusers gets the real validation error instead of
ModuleNotFoundError. test_in_progress_returns_409_after_validation_passes pins the
resolved device to cuda because the load route only takes the GPU arbiter for non-CPU
loads, which made the ownership assert host-dependent.
