Commit History
17 days ago
-
Rebuild tiny fixture with diffusers 0.40.0 / transformers 5.10.4
Verified signatureWeights are byte-identical to the previous 0.39.0 build (all six safetensors match by sha256). Changes are metadata only: - _diffusers_version 0.39.0 -> 0.40.0, transformers_version 5.5.4 -> 5.10.4 - transformer/config.json materialises the four keys 0.40 added to LTX2VideoTransformer3DModel: audio_ff_bias=True, ff_bias=True, use_keyframes_abs_pos_embedding=False, use_prompt_adaln_single=True (same values the real diffusers/LTX-2.3-Diffusers export resolves to) - model_index.json gains duration_head/prompt_enhancer as [null, null] - tiny_model_marker.json marker_version 2 -> 3925baf3
-
Make the fixture loadable on the whole supported transformers range
Verified signatureRegenerated with create_tiny_ltx23_model.py (marker_version 2). The previous revision was saved under transformers 5.x and inherited two version-dependent spellings from save_pretrained, both of which broke transformers 4.57.6: * tokenizer_class was recorded as `GemmaTokenizer` because on 5.x `GemmaTokenizerFast` is a deprecated alias of it. On 4.x that name is the sentencepiece class, which needs a tokenizer.model this fixture does not ship, so loading raised `ValueError: Either model_file or model_proto must be specified.` and the test_exporters_cli_tokenizers text-to-video / image-to-video cases failed. * text_encoder/model.safetensors keys carried an extra leading `model.` segment, which matches no rule in 4.x's Gemma3 _checkpoint_conversion_mapping, so the text encoder loaded silently with randomly initialised weights. Both are now normalised after saving. Verified on transformers 4.57.6, 5.5.4 and 5.10.4: the text encoder loads 51/51 params with 0 missing and 0 unexpected keys…ac56122
-
Add tiny random LTX-2.3 fixture (derived from diffusers/LTX-2.3-Diffusers)
Verified signaturea445c41 -
initial commit
Verified signatured4f8857