[fix] Allow ORTQuantizer over models with subfolder ONNX files

[tomaarsen]

Hello!

Pull Request overview

• Allow ORTQuantizer over models with subfolder ONNX files

Details

Currently, if you call ORTQuantizer over a model that was loaded with a subfolder, then it'll break:

from optimum.onnxruntime import ORTModelForFeatureExtraction, ORTQuantizer

model = ORTModelForFeatureExtraction.from_pretrained(
    "sentence-transformers-testing/all-MiniLM-L6-v2",
    subfolder="onnx",
    file_name="model.onnx",
)
quantizer = ORTQuantizer.from_pretrained(model)
print(quantizer)

Traceback (most recent call last):
  File "...\optimum\demo_ort_quantizer.py", line 8, in <module>
    quantizer = ORTQuantizer.from_pretrained(model)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "...\optimum\onnxruntime\quantization.py", line 156, in from_pretrained
    return cls(path)
           ^^^^^^^^^
  File "...\optimum\onnxruntime\quantization.py", line 102, in __init__
    self.config = AutoConfig.from_pretrained(self.onnx_model_path.parent)
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "...\site-packages\transformers\models\auto\configuration_auto.py", line 1049, in from_pretrained
    raise ValueError(
ValueError: Unrecognized model in ...\.cache\huggingface\hub\models--sentence-transformers--all-MiniLM-L6-v2\snapshots\fa97f6e7cb1a59073dff9e6b13e2715cf7475ac9\onnx. Should have a `model_type` key in its config.json, or contain one of the following strings in its name: albert, align, altclip, audio-spectrogram-transformer, autoformer, bark, bart, beit, bert, bert-generation, big_bird, bigbird_pegasus, biogpt, bit, blenderbot, blenderbot-small, blip, blip-2, bloom, bridgetower, bros, camembert, canine, chameleon, chinese_clip, chinese_clip_vision_model, clap, clip, clip_text_model, clip_vision_model, clipseg, clvp, code_llama, codegen, cohere, conditional_detr, convbert, convnext, convnextv2, cpmant, ctrl, cvt, dac, data2vec-audio, data2vec-text, data2vec-vision, dbrx, deberta, deberta-v2, decision_transformer, deformable_detr, deit, depth_anything, deta, detr, dinat, dinov2, distilbert, donut-swin, dpr, dpt, efficientformer, efficientnet, electra, encodec, encoder-decoder, ernie, ernie_m, esm, falcon, falcon_mamba, fastspeech2_conformer, flaubert, flava, fnet, focalnet, fsmt, funnel, fuyu, gemma, gemma2, git, glm, glpn, gpt-sw3, gpt2, gpt_bigcode, gpt_neo, gpt_neox, gpt_neox_japanese, gptj, gptsan-japanese, granite, granitemoe, graphormer, grounding-dino, groupvit, hiera, hubert, ibert, idefics, idefics2, idefics3, imagegpt, informer, instructblip, instructblipvideo, jamba, jetmoe, jukebox, kosmos-2, layoutlm, layoutlmv2, layoutlmv3, led, levit, lilt, llama, llava, llava_next, llava_next_video, llava_onevision, longformer, longt5, luke, lxmert, m2m_100, mamba, mamba2, marian, markuplm, mask2former, maskformer, maskformer-swin, mbart, mctct, mega, megatron-bert, mgp-str, mimi, mistral, mixtral, mllama, mobilebert, mobilenet_v1, mobilenet_v2, mobilevit, mobilevitv2, moshi, mpnet, mpt, mra, mt5, musicgen, musicgen_melody, mvp, nat, nemotron, nezha, nllb-moe, nougat, nystromformer, olmo, olmoe, omdet-turbo, oneformer, open-llama, openai-gpt, opt, owlv2, owlvit, paligemma, patchtsmixer, patchtst, pegasus, pegasus_x, perceiver, persimmon, phi, phi3, phimoe, pix2struct, pixtral, plbart, poolformer, pop2piano, prophetnet, pvt, pvt_v2, qdqbert, qwen2, qwen2_audio, qwen2_audio_encoder, qwen2_moe, qwen2_vl, rag, realm, recurrent_gemma, reformer, regnet, rembert, resnet, retribert, roberta, roberta-prelayernorm, roc_bert, roformer, rt_detr, rt_detr_resnet, rwkv, sam, seamless_m4t, seamless_m4t_v2, segformer, seggpt, sew, sew-d, siglip, siglip_vision_model, speech-encoder-decoder, speech_to_text, speech_to_text_2, speecht5, splinter, squeezebert, stablelm, starcoder2, superpoint, swiftformer, swin, swin2sr, swinv2, switch_transformers, t5, table-transformer, tapas, time_series_transformer, timesformer, timm_backbone, trajectory_transformer, transfo-xl, trocr, tvlt, tvp, udop, umt5, unispeech, unispeech-sat, univnet, upernet, van, video_llava, videomae, vilt, vipllava, vision-encoder-decoder, vision-text-dual-encoder, visual_bert, vit, vit_hybrid, vit_mae, vit_msn, vitdet, vitmatte, vits, vivit, wav2vec2, wav2vec2-bert, wav2vec2-conformer, wavlm, whisper, xclip, xglm, xlm, xlm-prophetnet, xlm-roberta, xlm-roberta-xl, xlnet, xmod, yolos, yoso, zamba, zoedepth, onnx_model

There are 2 issues at play:

1. The config (although known in the ORT Model) is not passed nicely to the ORTQuantizer with from_pretrained
2. If AutoConfig.from_pretrained fails, it often fails with a ValueError rather than an OSError..

An underlying issue is that #2044 added/strengthened the "ONNX in subfolders" support by allowing the config to be in root while the model is in a subfolder - the ORTQuantizer wasn't updated to reflect that the config.json isn't necessarily adjacent to the model.onnx.

FYI, this is breaking https://sbert.net/docs/package_reference/util.html#sentence_transformers.backend.export_dynamic_quantized_onnx_model in some cases.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

@echarlaix as you added/strengthened the "ONNX in subfolders" support & that's how I encountered this.

• Tom Aarsen

[echarlaix]

LGTM, thanks for the fix @tomaarsen !