Tng Ewallet Payment Methods Cards Adaptis Malaysia Tng Check

Summary

The hosted unsloth/*-NVFP4 diffusion repos now hold only .safetensors files (each one was checked bit-identical to its old .pt through Studio's loader before the .pt files were removed). NVFP4 is still off by default (UNSLOTH_NVFP4_DIFFUSION), so default users see no change. With the flag on, Studio now asks for and loads only the .safetensors and never resolves, prefers or unpickles an NVFP4 pickle.

Changes

studio/backend/core/inference/diffusion_prequant.py

• resolve_prequant_source: for nvfp4 the chain has no .pt / .pth name (no derived <Model>-NVFP4.pt, no legacy transformer_nvfp4.pt, no declared pickle). derived_prequant_filenames derives only <Model>-NVFP4.safetensors, and it counts as declared, so planning does not need the file cached first. UNSLOTH_PREQUANT_SAFETENSORS_TWINS=0 no longer brings back the pickle chain for nvfp4.
• _resolve_checkpoint_path: drops pickle names for nvfp4 even if the source still lists them, so readable or names cannot fall back to a pickle. If this install cannot rebuild NVFP4 safetensors (torchao < 0.17), the load is refused with a clear message.
• restricted_prequant_load_supported("nvfp4", <pickle>) is always False. The scheme-only question for nvfp4 now means "can this install read nvfp4 safetensors".
• NVFP4 classes are removed from _PREQUANT_SAFE_GLOBALS and _SCHEME_REQUIRED_GLOBALS. Since torchao adds them to the global safe list itself, the loader also refuses explicitly: _load_prequant_checkpoint rejects a non-safetensors path when scheme="nvfp4" is passed, and rejects any pickle whose metadata records nvfp4.
• A local nvfp4 .pt override is not used as a source, and the loader refuses it if it gets one anyway. local_prequant_scheme never reports nvfp4 for a pickle. _artifacts_declare_nvfp4 only reads safetensors headers, and declares_nvfp4_checkpoint decides a single .pt file by its name without unpickling it.
• prefer_cached_pickle_twins cannot pick a cached nvfp4 .pt: the pickle is unreadable for nvfp4 and is no longer in the chain.

Tables and tools

• video_families.py: the nvfp4 rows name the hosted files (Wan2.2-TI2V-5B-NVFP4.safetensors, Wan2.2-T2V-A14B-NVFP4.safetensors, Wan2.2-T2V-A14B-transformer_2-NVFP4.safetensors, HunyuanVideo-1.5-Diffusers-480p_t2v-NVFP4.safetensors / 720p). Image families derive <Model>-NVFP4.safetensors, which matches the Hub for all four image repos.
• nvfp4_gate_record.json: each record's filename is now the .safetensors name and checkpoint_sha256 is that file's LFS sha256 from the Hub listing.
• scripts/build_prequant_checkpoint.py refuses a non-safetensors --out (and pickle upload names) for --scheme nvfp4. scripts/convert_prequant_to_safetensors.py still reads a local pickle directly, so it remains the way to turn an old local NVFP4 .pt into a .safetensors.
• Fixed the stale comment in diffusion_families.py and updated the tests that hard-coded NVFP4 .pt names.

INT8, FP8 and MXFP8 are unchanged: same chains, .pt twins and fallbacks, cached-pickle preference and allowlist.

Tests

New studio/backend/tests/test_prequant_nvfp4_safetensors_only.py (38 tests) covers:

• every hosted NVFP4 family chain (image and video, both experts, twins switch on and off)
• a family table that still declares an nvfp4 .pt
• cached-pickle preference
• loader refusal by scheme and by metadata
• local override refusal through the full loader
• detection that never unpickles
• gate records
• int8 / fp8 controls

Plus nvfp4 cases added to test_prequant_prefer_cached_pickle.py and test_video_prequant_safetensors_twins.py, and refusal tests in test_build_prequant_checkpoint.py.

• Base (7ec4b20bc2) with the new test file: 30 failed, 8 passed. The 8 that pass are the int8 / fp8 controls. Two of the 30 call the new scheme= keyword, which base does not have.
• Head: 38 passed.
• 51 related test files (prequant, nvfp4, families, comfy quant, video backend, companion assets): base 3163 passed / 1 failed, head 3179 passed / 1 failed. Both arms fail the same test, test_diffusion_more_families.py::test_create_causal_mask_patch_is_self_disabling_and_idempotent, which fails on main too.

All of this ran on CPU. No GPU render was done here; the .safetensors files had already been verified bit-identical to the old .pt files through Studio's loader.