inferna 0.5.0

Changes since the last Release

Added

• SDContextParams.auto_fit, .disable_segmented_compute, .disable_prefetch, .linear_scale, .attn_scale and .tokenizer, plus the matching --auto-fit, --disable-segmented-compute, --disable-prefetch and --tokenizer CLI flags. tokenizer is required for PiD and Lens models. Scheduler.LLADA_IMAGE, Prediction.SENSENOVA_U1_FLOW, SDType.Q2_0/F8_E4M3/F8_E5M2 and LogLevel.VERBOSE follow the header.

Changed

• stable-diffusion.cpp updated to master-898-2bb7294 (from master-816-487de75), matching cyllama (cyllama#19). From master-817 sd.cpp called ops that exist only in leejet's ggml fork, so it no longer compiled against llama.cpp's ggml. Upstream's SD_USE_UPSTREAM_GGML (#1999) compiles those calls out. Shared-ggml builds now pass it, with SD_GGML_SOURCE_DIR pointing at llama.cpp's ggml, instead of swapping that tree into SD's checkout. SD_USE_VENDORED_GGML stays the only switch: both options are derived from it and passed in both modes, because they are CMake cache variables. See Stable Diffusion ggml selection. Features lost in shared mode: INT8 ConvRot model files are rejected, FP8 safetensors load as F16, and SageAttention is unavailable. None of these worked at the old pin. An existing SD checkout fails verify_checkout() at the new pin; run make reset.

• Breaking: SD memory defaults and semantics changed upstream. auto_fit now defaults to True and places modules on the GPU, RAM, another GPU or disk by free memory. A non-empty params_backend, which --offload-to-cpu sets, disables it. max_vram is a per-device budget: "0" uses live free VRAM instead of disabling segmentation; set disable_segmented_compute instead. SDContextParams.stream_layers is removed; prefetch is on by default and disable_prefetch turns it off.

• Breaking: LogLevel values shifted. Upstream inserted SD_LOG_VERBOSE after DEBUG, so INFO, WARN and ERROR are now 2, 3 and 4. Without VERBOSE, set_log_callback would have mislabelled levels and dropped every ERROR message on the LogLevel conversion. The CLI and examples compared levels by number and now compare by name.

• make_xcframework.py fails when SD_USE_VENDORED_GGML=1 instead of silently overriding it: the xcframeworks ship one Ggml.framework. It also points SD at llama.cpp's ggml with SD_GGML_SOURCE_DIR instead of copying that tree over SD's vendored fork.

Removed

• stable-diffusion.cpp-graph-cut-budget-clamp.patch and stable-diffusion.cpp-conditioner-compute-failure.patch. Both defects are fixed upstream at the new pin; see scripts/patches/README.md.