@huggingface/tasks 0.21.54 → 0.21.56

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (35) hide show
  1. package/dist/commonjs/hardware-amd.d.ts.map +1 -1
  2. package/dist/commonjs/hardware-amd.js +48 -0
  3. package/dist/commonjs/hardware-nvidia.d.ts +1 -0
  4. package/dist/commonjs/hardware-nvidia.d.ts.map +1 -1
  5. package/dist/commonjs/hardware-nvidia.js +122 -0
  6. package/dist/commonjs/library-to-tasks.d.ts.map +1 -1
  7. package/dist/commonjs/library-to-tasks.js +8 -1
  8. package/dist/commonjs/model-libraries-snippets.d.ts +3 -1
  9. package/dist/commonjs/model-libraries-snippets.d.ts.map +1 -1
  10. package/dist/commonjs/model-libraries-snippets.js +349 -154
  11. package/dist/commonjs/model-libraries-snippets.spec.js +300 -0
  12. package/dist/commonjs/model-libraries.d.ts +11 -2
  13. package/dist/commonjs/model-libraries.d.ts.map +1 -1
  14. package/dist/commonjs/model-libraries.js +10 -1
  15. package/dist/esm/hardware-amd.d.ts.map +1 -1
  16. package/dist/esm/hardware-amd.js +48 -0
  17. package/dist/esm/hardware-nvidia.d.ts +1 -0
  18. package/dist/esm/hardware-nvidia.d.ts.map +1 -1
  19. package/dist/esm/hardware-nvidia.js +122 -0
  20. package/dist/esm/library-to-tasks.d.ts.map +1 -1
  21. package/dist/esm/library-to-tasks.js +8 -1
  22. package/dist/esm/model-libraries-snippets.d.ts +3 -1
  23. package/dist/esm/model-libraries-snippets.d.ts.map +1 -1
  24. package/dist/esm/model-libraries-snippets.js +344 -151
  25. package/dist/esm/model-libraries-snippets.spec.js +301 -1
  26. package/dist/esm/model-libraries.d.ts +11 -2
  27. package/dist/esm/model-libraries.d.ts.map +1 -1
  28. package/dist/esm/model-libraries.js +10 -1
  29. package/package.json +1 -1
  30. package/src/hardware-amd.ts +48 -0
  31. package/src/hardware-nvidia.ts +122 -0
  32. package/src/library-to-tasks.ts +8 -1
  33. package/src/model-libraries-snippets.spec.ts +389 -0
  34. package/src/model-libraries-snippets.ts +364 -149
  35. package/src/model-libraries.ts +10 -1
@@ -1,8 +1,8 @@
1
1
  "use strict";
2
2
  Object.defineProperty(exports, "__esModule", { value: true });
3
- exports.litert_lm = exports.lerobot = exports.llama_cpp_python = exports.lightning_ir = exports.ltx = exports.kittentts = exports.kimi_audio = exports.kernels = exports.keras_hub = exports.zeromodels = exports.keras = exports.jev_style = exports.htrflow = exports.indextts = exports.gliner2 = exports.gliner = exports.gaussianformer = exports.flextab = exports.flexray = exports.flair = exports.fairseq = exports.espnet = exports.espnetASR = exports.espnetTTS = exports.edsnlp = exports.cartesia_mlx = exports.cartesia_pytorch = exports.diffusionkit = exports.diffusers = exports.describe_anything = exports.dia2 = exports.dia = exports.derm_foundation = exports.depth_pro = exports.depth_anything_v2 = exports.cxr_foundation = exports.cortiq = exports.sap_rpt_one_oss = exports.colipri = exports.collectorvision = exports.chronos_forecasting = exports.chatterbox = exports.bm25s = exports.bertopic = exports.ben2 = exports.audioseal = exports.asteroid = exports.araclip = exports.allennlp = exports.adapters = void 0;
4
- exports.vfimamba = exports.videoprism = exports.vibevoice = exports.sana = exports.sentis = exports.mlAgents = exports.stableBaselines3 = exports.fasttext = exports.peft = exports.transformersJS = exports.transformers = exports.terratorch = exports.speechbrain = exports.stanza = exports.span_marker = exports.spacy = exports.setfit = exports.sentenceTransformers = exports.aneforge = exports.sampleFactory = exports.sam_3d_body = exports.sam_3d_objects = exports.sam2 = exports.fastai = exports.stable_audio_tools = exports.sklearn = exports.seed_story = exports.saelens = exports.timm = exports.tensorflowtts = exports.routee_powertrain = exports.renderformer = exports.relik = exports.pyannote_audio = exports.pyannote_audio_pipeline = exports.pocket_tts = exports.phantom_wan = exports.perception_encoder = exports.paddleocr = exports.paddlenlp = exports.opendde = exports.openasr = exports.open_clip = exports.multimolecule = exports.mesh_anything = exports.memra = exports.matanyone = exports.mars5_tts = exports.mamba_ssm = exports.tf_keras = void 0;
5
- exports.moshi = exports.zonos = exports.hezar = exports.threedtopia_xl = exports.whisperkit = exports.audiocraft = exports.anemoi = exports.qwen3_tts = exports.pythae = exports.pxia = exports.outetts = exports.nemo = exports.pruna = exports.mobilint = exports.model2vec = exports.mlx = exports.mlxim = exports.univa = exports.swarmformer = exports.supertonic = exports.nobg = exports.birefnet = exports.ultralytics = exports.chattts = exports.vui = exports.voxcpm = exports.voicecraft = exports.lvface = void 0;
3
+ exports.lerobot = exports.llama_cpp_python = exports.lightning_ir = exports.ltx = exports.kittentts = exports.kimi_audio = exports.kernels = exports.keras_hub = exports.zeromodels = exports.keras = exports.jev_style = exports.htrflow = exports.indextts = exports.gliner2 = exports.gliner = exports.gaussianformer = exports.flextab = exports.flexray = exports.flair = exports.fairseq = exports.espnet = exports.espnetASR = exports.espnetTTS = exports.edsnlp = exports.cartesia_mlx = exports.cartesia_pytorch = exports.diffusionkit = exports.diffusers = exports.describe_anything = exports.dia2 = exports.dia = exports.derm_foundation = exports.depth_pro = exports.depth_anything_v2 = exports.cxr_foundation = exports.cortiq = exports.sap_rpt_one_oss = exports.colipri = exports.collectorvision = exports.chronos_forecasting = exports.chatterbox = exports.bm25s = exports.bertopic = exports.ben2 = exports.audioseal = exports.asteroid = exports.asimov = exports.araclip = exports.allennlp = exports.adapters = void 0;
4
+ exports.vibevoice = exports.sana = exports.sentis = exports.mlAgents = exports.stableBaselines3 = exports.fasttext = exports.peft = exports.transformersJS = exports.transformers = exports.terratorch = exports.speechbrain = exports.stanza = exports.span_marker = exports.spacy = exports.setfit = exports.sentenceTransformers = exports.aneforge = exports.sampleFactory = exports.sam_3d_body = exports.sam_3d_objects = exports.sam2 = exports.fastai = exports.stable_audio_tools = exports.sklearn = exports.seed_story = exports.saelens = exports.timm = exports.tensorflowtts = exports.routee_powertrain = exports.renderformer = exports.relik = exports.pyannote_audio = exports.pyannote_audio_pipeline = exports.pocket_tts = exports.phantom_wan = exports.perception_encoder = exports.paddleocr = exports.paddlenlp = exports.opendde = exports.openasr = exports.open_clip = exports.multimolecule = exports.microduck = exports.mesh_anything = exports.memra = exports.matanyone = exports.mars5_tts = exports.mamba_ssm = exports.tf_keras = exports.litert_lm = void 0;
5
+ exports.moshi = exports.zonos = exports.hezar = exports.threedtopia_xl = exports.whisperkit = exports.audiocraft = exports.anemoi = exports.qwen3_tts = exports.pythae = exports.pxia = exports.outetts = exports.nemo = exports.pruna = exports.mobilint = exports.model2vec = exports.mlx = exports.mlxim = exports.univa = exports.swarmformer = exports.supertonic = exports.nobg = exports.birefnet = exports.ultralytics = exports.chattts = exports.vui = exports.voxcpm = exports.voicecraft = exports.lvface = exports.vfimamba = exports.videoprism = void 0;
6
6
  const library_to_tasks_js_1 = require("./library-to-tasks.js");
7
7
  const inputs_js_1 = require("./snippets/inputs.js");
8
8
  const common_js_1 = require("./snippets/common.js");
@@ -17,7 +17,7 @@ const isValidIdentifier = (str) => /^[A-Za-z_]\w*$/.test(str);
17
17
  const adapters = (model) => [
18
18
  `from adapters import AutoAdapterModel
19
19
 
20
- model = AutoAdapterModel.from_pretrained("${escapeStringForJson(model.config?.adapter_transformers?.model_name ?? "fill-in-model-name")}")
20
+ model = AutoAdapterModel.from_pretrained("${escapeStringForJson(model.config?.adapter_transformers?.model_name ?? "<base-model-id>")}")
21
21
  model.load_adapter("${model.id}", set_active=True)`,
22
22
  ];
23
23
  exports.adapters = adapters;
@@ -48,6 +48,14 @@ const araclip = (model) => [
48
48
  model = AraClip.from_pretrained("${model.id}")`,
49
49
  ];
50
50
  exports.araclip = araclip;
51
+ const asimov = (model) => [
52
+ `# Run the policy in a browser simulation of the Asimov 1 humanoid (no GPU needed)
53
+ git clone https://github.com/menloresearch/humanoid-policy-viewer
54
+ cd humanoid-policy-viewer
55
+ npm install
56
+ npm run hf ${model.id}`,
57
+ ];
58
+ exports.asimov = asimov;
51
59
  const asteroid = (model) => [
52
60
  `from asteroid.models import BaseModel
53
61
 
@@ -74,7 +82,13 @@ result, message = detector.detect_watermark(watermarked_audio, sr)`;
74
82
  };
75
83
  exports.audioseal = audioseal;
76
84
  function get_base_diffusers_model(model) {
77
- return escapeStringForJson(model.cardData?.base_model?.toString() ?? "fill-in-base-model");
85
+ const baseModel = model.cardData?.base_model;
86
+ // several base models can be listed (e.g. a LoRA trained for a base and its turbo variant). When an original checkpoint
87
+ // and its Diffusers conversion are both listed (["Wan-AI/Wan2.1-I2V-14B-480P", "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"]),
88
+ // only the conversion has a model_index.json that DiffusionPipeline can load; otherwise the first one is used
89
+ const candidates = Array.isArray(baseModel) ? baseModel : baseModel ? [baseModel] : [];
90
+ const chosen = candidates.find((id) => id.toLowerCase().endsWith("-diffusers")) ?? candidates[0];
91
+ return escapeStringForJson(chosen ?? "fill-in-base-model");
78
92
  }
79
93
  function get_prompt_from_diffusers_model(model) {
80
94
  const prompt = model.widgetData?.[0]?.text ?? model.cardData?.instance_prompt;
@@ -161,7 +175,7 @@ detector = cvg.NeuralCornerDetector(checkpoint)
161
175
  embedder = cvg.NeuralEmbedder(checkpoint)`,
162
176
  ];
163
177
  exports.collectorvision = collectorvision;
164
- const colipri = (model) => {
178
+ const colipri = () => {
165
179
  const installSnippet = `pip install colipri`;
166
180
  const exampleSnippet = `from colipri import get_model
167
181
  from colipri import get_processor
@@ -170,7 +184,7 @@ from colipri import ZeroShotImageClassificationPipeline
170
184
 
171
185
  model = get_model().cuda()
172
186
  processor = get_processor()
173
- pipeline = ZeroShotImageClassificationPipeline("${model.id}", processor)
187
+ pipeline = ZeroShotImageClassificationPipeline(model, processor)
174
188
 
175
189
  image = load_sample_ct()
176
190
 
@@ -291,7 +305,7 @@ const depth_anything_v2 = (model) => {
291
305
  else if (model.id === "depth-anything/Depth-Anything-V2-Large") {
292
306
  encoder = "vitl";
293
307
  features = "256";
294
- out_channels = "[256, 512, 1024, 1024";
308
+ out_channels = "[256, 512, 1024, 1024]";
295
309
  }
296
310
  return [
297
311
  `
@@ -300,6 +314,7 @@ const depth_anything_v2 = (model) => {
300
314
  # Load the model and infer depth from an image
301
315
  import cv2
302
316
  import torch
317
+ from huggingface_hub import hf_hub_download
303
318
 
304
319
  from depth_anything_v2.dpt import DepthAnythingV2
305
320
 
@@ -309,7 +324,8 @@ model = DepthAnythingV2(encoder="${encoder}", features=${features}, out_channels
309
324
  # load the weights
310
325
  filepath = hf_hub_download(repo_id="${model.id}", filename="depth_anything_v2_${encoder}.pth", repo_type="model")
311
326
  state_dict = torch.load(filepath, map_location="cpu")
312
- model.load_state_dict(state_dict).eval()
327
+ model.load_state_dict(state_dict)
328
+ model.eval()
313
329
 
314
330
  raw_img = cv2.imread("your/image/path")
315
331
  depth = model.infer_image(raw_img) # HxW raw depth map in numpy
@@ -319,8 +335,8 @@ depth = model.infer_image(raw_img) # HxW raw depth map in numpy
319
335
  exports.depth_anything_v2 = depth_anything_v2;
320
336
  const depth_pro = (model) => {
321
337
  const installSnippet = `# Download checkpoint
322
- pip install huggingface-hub
323
- huggingface-cli download --local-dir checkpoints ${model.id}`;
338
+ pip install huggingface_hub
339
+ hf download ${model.id} --local-dir checkpoints`;
324
340
  const inferenceSnippet = `import depth_pro
325
341
 
326
342
  # Load model and preprocessing transform
@@ -392,7 +408,7 @@ const describe_anything = (model) => [
392
408
  from huggingface_hub import snapshot_download
393
409
  from dam import DescribeAnythingModel
394
410
 
395
- snapshot_download(${model.id}, local_dir="checkpoints")
411
+ snapshot_download("${model.id}", local_dir="checkpoints")
396
412
 
397
413
  dam = DescribeAnythingModel(
398
414
  model_path="checkpoints",
@@ -504,7 +520,7 @@ pipe.load_lora_weights("${model.id}")
504
520
  prompt = "${get_prompt_from_diffusers_model(model) ?? diffusersVideoDefaultPrompt}"
505
521
  input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/guitar-man.png")
506
522
 
507
- image = pipe(image=input_image, prompt=prompt).frames[0]
523
+ output = pipe(image=input_image, prompt=prompt).frames[0]
508
524
  export_to_video(output, "output.mp4")`,
509
525
  ];
510
526
  const diffusers_textual_inversion = (model) => [
@@ -544,7 +560,7 @@ from diffusers import AutoPipelineForInpainting
544
560
  from diffusers.utils import load_image
545
561
 
546
562
  # switch to "mps" for apple devices
547
- pipe = AutoPipelineForInpainting.from_pretrained("${model.id}", dtype=torch.float16, variant="fp16", device_map="cuda")
563
+ pipe = AutoPipelineForInpainting.from_pretrained("${model.id}", dtype=torch.float16, device_map="cuda")
548
564
 
549
565
  img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
550
566
  mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
@@ -566,9 +582,12 @@ image = pipe(
566
582
  ).images[0]`,
567
583
  ];
568
584
  const diffusers = (model) => {
585
+ // the Hub emits pipeline class tags as `diffusers:<ClassName>`; bare names are kept for manually added tags
586
+ const hasPipelineClass = (className) => model.tags.includes(`diffusers:${className}`) ||
587
+ model.tags.includes(className) ||
588
+ model.config?.diffusers?._class_name === className;
569
589
  let codeSnippets;
570
- if (model.tags.includes("StableDiffusionInpaintPipeline") ||
571
- model.tags.includes("StableDiffusionXLInpaintPipeline")) {
590
+ if (hasPipelineClass("StableDiffusionInpaintPipeline") || hasPipelineClass("StableDiffusionXLInpaintPipeline")) {
572
591
  codeSnippets = diffusers_inpainting(model);
573
592
  }
574
593
  else if (model.tags.includes("controlnet")) {
@@ -591,7 +610,7 @@ const diffusers = (model) => {
591
610
  else if (model.tags.includes("textual_inversion")) {
592
611
  codeSnippets = diffusers_textual_inversion(model);
593
612
  }
594
- else if (model.tags.includes("FluxFillPipeline")) {
613
+ else if (hasPipelineClass("FluxFillPipeline")) {
595
614
  codeSnippets = diffusers_flux_fill(model);
596
615
  }
597
616
  else if (model.pipeline_tag === "image-to-video") {
@@ -613,7 +632,7 @@ from diffusionkit.mlx import DiffusionPipeline
613
632
  pipeline = DiffusionPipeline(
614
633
  shift=3.0,
615
634
  use_t5=False,
616
- model_version=${model.id},
635
+ model_version="${model.id}",
617
636
  low_memory_mode=True,
618
637
  a16=True,
619
638
  w16=True,
@@ -623,7 +642,7 @@ from diffusionkit.mlx import FluxPipeline
623
642
 
624
643
  pipeline = FluxPipeline(
625
644
  shift=1.0,
626
- model_version=${model.id},
645
+ model_version="${model.id}",
627
646
  low_memory_mode=True,
628
647
  a16=True,
629
648
  w16=True,
@@ -645,21 +664,22 @@ image, _ = pipeline.generate_image(
645
664
  };
646
665
  exports.diffusionkit = diffusionkit;
647
666
  const cartesia_pytorch = (model) => [
648
- `# pip install --no-binary :all: cartesia-pytorch
649
- from cartesia_pytorch import ReneLMHeadModel
667
+ `# The cartesia-pytorch package on PyPI ships no code: clone the repo and put its cartesia-pytorch folder on PYTHONPATH
668
+ # (it needs an Ampere or newer CUDA GPU, torch 2.4, mamba-ssm 2.2.2 and flash-attn 2.6.3)
669
+ # git clone https://github.com/cartesia-ai/edge && export PYTHONPATH="$PWD/edge/cartesia-pytorch:$PYTHONPATH"
670
+ from cartesia_pytorch.Rene import ReneLMHeadModel
650
671
  from transformers import AutoTokenizer
651
672
 
652
- model = ReneLMHeadModel.from_pretrained("${model.id}")
673
+ model = ReneLMHeadModel.from_pretrained("${model.id}").half().cuda()
653
674
  tokenizer = AutoTokenizer.from_pretrained("allenai/OLMo-1B-hf")
654
675
 
655
676
  in_message = ["Rene Descartes was"]
656
677
  inputs = tokenizer(in_message, return_tensors="pt")
657
678
 
658
- outputs = model.generate(inputs.input_ids, max_length=50, top_k=100, top_p=0.99)
679
+ outputs = model.generate(inputs.input_ids.cuda(), max_length=50, top_k=100, top_p=0.99)
659
680
  out_message = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
660
681
 
661
- print(out_message)
662
- )`,
682
+ print(out_message)`,
663
683
  ];
664
684
  exports.cartesia_pytorch = cartesia_pytorch;
665
685
  const cartesia_mlx = (model) => [
@@ -710,7 +730,8 @@ speech, *_ = model("text to generate speech from")`,
710
730
  ];
711
731
  exports.espnetTTS = espnetTTS;
712
732
  const espnetASR = (model) => [
713
- `from espnet2.bin.asr_inference import Speech2Text
733
+ `import soundfile
734
+ from espnet2.bin.asr_inference import Speech2Text
714
735
 
715
736
  model = Speech2Text.from_pretrained(
716
737
  "${model.id}"
@@ -740,9 +761,14 @@ models, cfg, task = load_model_ensemble_and_task_from_hf_hub(
740
761
  ];
741
762
  exports.fairseq = fairseq;
742
763
  const flair = (model) => [
743
- `from flair.models import SequenceTagger
764
+ `from flair.data import Sentence
765
+ from flair.models import SequenceTagger
744
766
 
745
- tagger = SequenceTagger.load("${model.id}")`,
767
+ tagger = SequenceTagger.load("${model.id}")
768
+
769
+ sentence = Sentence("George Washington went to Washington.")
770
+ tagger.predict(sentence)
771
+ print(sentence)`,
746
772
  ];
747
773
  exports.flair = flair;
748
774
  const flexray = (model) => [
@@ -821,23 +847,30 @@ print(f'Accuracy {accuracy_score(y_test, predictions):.2%}')
821
847
  };
822
848
  exports.flextab = flextab;
823
849
  const gaussianformer = (model) => [
824
- `# Install from https://github.com/SVLwoof/gaussianformer
825
-
826
- from gaussianformer.pipelines.rendering_pipeline import GaussianFormerRenderingPipeline
850
+ `# pip install "gaussianformer @ git+https://github.com/SVLwoof/gaussianformer"
851
+ from gaussianformer import GaussianFormerRenderingPipeline, load_ply
827
852
 
828
- pipeline = GaussianFormerRenderingPipeline.from_pretrained("${model.id}")`,
853
+ pipeline = GaussianFormerRenderingPipeline.from_pretrained("${model.id}").to("cuda")
854
+ gaussians = load_ply("object.ply") # 3D Gaussian Splatting PLY -> [N, 14]`,
829
855
  ];
830
856
  exports.gaussianformer = gaussianformer;
831
857
  const gliner = (model) => [
832
858
  `from gliner import GLiNER
833
859
 
834
- model = GLiNER.from_pretrained("${model.id}")`,
860
+ model = GLiNER.from_pretrained("${model.id}")
861
+
862
+ text = "Cristiano Ronaldo dos Santos Aveiro was born on 5 February 1985 in Funchal, Madeira, Portugal."
863
+ labels = ["person", "date", "location"]
864
+
865
+ entities = model.predict_entities(text, labels)
866
+ for entity in entities:
867
+ print(entity["text"], "=>", entity["label"])`,
835
868
  ];
836
869
  exports.gliner = gliner;
837
870
  const gliner2 = (model) => [
838
- `from gliner2 import GLiNER2
871
+ `from gliner2 import AutoExtractor
839
872
 
840
- model = GLiNER2.from_pretrained("${model.id}")
873
+ extractor = AutoExtractor.from_pretrained("${model.id}")
841
874
 
842
875
  # Extract entities
843
876
  text = "Apple CEO Tim Cook announced iPhone 15 in Cupertino yesterday."
@@ -850,7 +883,7 @@ const indextts = (model) => [
850
883
  `# Download model
851
884
  from huggingface_hub import snapshot_download
852
885
 
853
- snapshot_download(${model.id}, local_dir="checkpoints")
886
+ snapshot_download("${model.id}", local_dir="checkpoints")
854
887
 
855
888
  from indextts.infer import IndexTTS
856
889
 
@@ -907,9 +940,11 @@ print(out["answers"]["team"]["choice"])`,
907
940
  };
908
941
  exports.jev_style = jev_style;
909
942
  const keras = (model) => [
910
- `# Available backend options are: "jax", "torch", "tensorflow".
943
+ `# !pip install -U keras tensorflow huggingface_hub
944
+ # Keras needs TensorFlow installed to read "hf://" paths, so the tensorflow backend is selected here;
945
+ # "jax" and "torch" also work for computation once TensorFlow is installed.
911
946
  import os
912
- os.environ["KERAS_BACKEND"] = "jax"
947
+ os.environ["KERAS_BACKEND"] = "tensorflow"
913
948
 
914
949
  import keras
915
950
 
@@ -1026,13 +1061,15 @@ const kernels = (model) => [
1026
1061
 
1027
1062
  from kernels import get_kernel
1028
1063
 
1029
- kernel = get_kernel("${model.id}")`,
1064
+ # a version (or an explicit revision) is required; see the "Files and versions" tab for the available ones
1065
+ kernel = get_kernel("${model.id}", version=1)`,
1030
1066
  ];
1031
1067
  exports.kernels = kernels;
1032
1068
  const kimi_audio = (model) => [
1033
1069
  `# Example usage for KimiAudio
1034
- # pip install git+https://github.com/MoonshotAI/Kimi-Audio.git
1070
+ # pip install git+https://github.com/MoonshotAI/Kimi-Audio.git "transformers<5"
1035
1071
 
1072
+ import soundfile as sf
1036
1073
  from kimia_infer.api.kimia import KimiAudio
1037
1074
 
1038
1075
  model = KimiAudio(model_path="${model.id}", load_detokenizer=True)
@@ -1402,14 +1439,17 @@ pip install -U litert-lm
1402
1439
 
1403
1440
  # 2. Download and run this model locally:
1404
1441
  # See: https://ai.google.dev/edge/litert-lm/cli
1442
+ # A single .litertlm file in the repo is picked automatically; otherwise the CLI asks which one to run
1443
+ # (or pass its name right after the repo id).
1405
1444
  litert-lm run \\
1406
1445
  --from-huggingface-repo=${model.id} \\
1407
1446
  --prompt="Write me a poem"`,
1408
1447
  ];
1409
1448
  exports.litert_lm = litert_lm;
1410
1449
  const tf_keras = (model) => [
1411
- `# Note: 'keras<3.x' or 'tf_keras' must be installed (legacy)
1450
+ `# Note: 'keras<3.x' or 'tf_keras' must be installed (legacy), and from_pretrained_keras was removed in huggingface_hub 1.0.
1412
1451
  # See https://github.com/keras-team/tf-keras for more details.
1452
+ # !pip install "huggingface_hub<1.0" tf_keras
1413
1453
  from huggingface_hub import from_pretrained_keras
1414
1454
 
1415
1455
  model = from_pretrained_keras("${model.id}")
@@ -1461,6 +1501,17 @@ from MeshAnything.models.meshanything import MeshAnything
1461
1501
  model = MeshAnything(args)`,
1462
1502
  ];
1463
1503
  exports.mesh_anything = mesh_anything;
1504
+ const microduck = (model) => {
1505
+ // Model-card tag convention: tags: [microduck-slot:stand].
1506
+ const slots = ["walk", "stand", "sitstand", "ground_pick", "kick_left", "kick_right", "roulade"];
1507
+ const matchingSlots = slots.filter((slot) => model.tags.includes(`microduck-slot:${slot}`));
1508
+ const slot = matchingSlots.length === 1 ? matchingSlots[0] : "SLOT";
1509
+ const prefix = matchingSlots.length === 1
1510
+ ? ""
1511
+ : `# Replace SLOT with the slot specified in the model card (${slots.join(", ")}).\n`;
1512
+ return [`${prefix}sudo robotctl policy load ${slot} ${model.id}`];
1513
+ };
1514
+ exports.microduck = microduck;
1464
1515
  const multimolecule = (model) => {
1465
1516
  const widgetExample = model.widgetData?.[0];
1466
1517
  const exampleText = escapeStringForJson(widgetExample?.text ?? "");
@@ -1641,14 +1692,18 @@ model = pe.VisionTransformer.from_config("${model.id}", pretrained=True)`;
1641
1692
  exports.perception_encoder = perception_encoder;
1642
1693
  const phantom_wan = (model) => [
1643
1694
  `from huggingface_hub import snapshot_download
1644
- from phantom_wan import WANI2V, configs
1695
+ from phantom_wan import WanI2V, configs
1696
+ from PIL import Image
1645
1697
 
1646
1698
  checkpoint_dir = snapshot_download("${model.id}")
1647
1699
  wan_i2v = WanI2V(
1648
- config=configs.WAN_CONFIGS['i2v-14B'],
1649
- checkpoint_dir=checkpoint_dir,
1650
- )
1651
- video = wan_i2v.generate(text_prompt, image_prompt)`,
1700
+ config=configs.WAN_CONFIGS['i2v-14B'],
1701
+ checkpoint_dir=checkpoint_dir,
1702
+ )
1703
+
1704
+ text_prompt = "A cat playing with a ball of yarn"
1705
+ image_prompt = Image.open("path/to/image.jpg").convert("RGB")
1706
+ video = wan_i2v.generate(text_prompt, image_prompt)`,
1652
1707
  ];
1653
1708
  exports.phantom_wan = phantom_wan;
1654
1709
  const pocket_tts = (model) => [
@@ -1770,7 +1825,7 @@ exports.tensorflowtts = tensorflowtts;
1770
1825
  const timm = (model) => [
1771
1826
  `import timm
1772
1827
 
1773
- model = timm.create_model("hf_hub:${model.id}", pretrained=True)`,
1828
+ model = timm.create_model("hf-hub:${model.id}", pretrained=True)`,
1774
1829
  ];
1775
1830
  exports.timm = timm;
1776
1831
  const saelens = ( /* model: ModelData */) => [
@@ -1798,13 +1853,14 @@ seed_story_cfg = OmegaConf.load(seed_story_cfg_path)
1798
1853
  seed_story = hydra.utils.instantiate(seed_story_cfg, llm=llm) `,
1799
1854
  ];
1800
1855
  exports.seed_story = seed_story;
1801
- const skopsPickle = (model, modelFile) => {
1856
+ const SKLEARN_PICKLE_FILE = /\.(joblib|pkl|pickle)$/i;
1857
+ // skops.hub_utils was removed in skops 0.12, so files are downloaded with huggingface_hub
1858
+ const sklearnPickle = (model, modelFile) => {
1802
1859
  return [
1803
- `import joblib
1804
- from skops.hub_utils import download
1805
- download("${model.id}", "path_to_folder")
1860
+ `from huggingface_hub import hf_hub_download
1861
+ import joblib
1806
1862
  model = joblib.load(
1807
- "${escapeStringForJson(modelFile)}"
1863
+ hf_hub_download("${model.id}", "${escapeStringForJson(modelFile)}")
1808
1864
  )
1809
1865
  # only load pickle files from sources you trust
1810
1866
  # read more about it here https://skops.readthedocs.io/en/stable/persistence.html`,
@@ -1812,42 +1868,28 @@ model = joblib.load(
1812
1868
  };
1813
1869
  const skopsFormat = (model, modelFile) => {
1814
1870
  return [
1815
- `from skops.hub_utils import download
1871
+ `from huggingface_hub import hf_hub_download
1816
1872
  from skops.io import load
1817
- download("${model.id}", "path_to_folder")
1818
1873
  # make sure model file is in skops format
1819
1874
  # if model is a pickle file, make sure it's from a source you trust
1820
- model = load("path_to_folder/${escapeStringForJson(modelFile)}")`,
1821
- ];
1822
- };
1823
- const skopsJobLib = (model) => {
1824
- return [
1825
- `from huggingface_hub import hf_hub_download
1826
- import joblib
1827
- model = joblib.load(
1828
- hf_hub_download("${model.id}", "sklearn_model.joblib")
1829
- )
1830
- # only load pickle files from sources you trust
1831
- # read more about it here https://skops.readthedocs.io/en/stable/persistence.html`,
1875
+ model = load(hf_hub_download("${model.id}", "${escapeStringForJson(modelFile)}"))`,
1832
1876
  ];
1833
1877
  };
1834
1878
  const sklearn = (model) => {
1879
+ const declaredFile = model.config?.sklearn?.model?.file;
1835
1880
  if (model.tags.includes("skops")) {
1836
- const skopsmodelFile = model.config?.sklearn?.model?.file;
1837
- const skopssaveFormat = model.config?.sklearn?.model_format;
1838
- if (!skopsmodelFile) {
1881
+ if (!declaredFile) {
1839
1882
  return [`# ⚠️ Model filename not specified in config.json`];
1840
1883
  }
1841
- if (skopssaveFormat === "pickle") {
1842
- return skopsPickle(model, skopsmodelFile);
1843
- }
1844
- else {
1845
- return skopsFormat(model, skopsmodelFile);
1884
+ // some skops repos declare no model_format for a pickled file (e.g. julien-c/skops-digits), which skops.io cannot read
1885
+ if (model.config?.sklearn?.model_format === "pickle" || SKLEARN_PICKLE_FILE.test(declaredFile)) {
1886
+ return sklearnPickle(model, declaredFile);
1846
1887
  }
1888
+ return skopsFormat(model, declaredFile);
1847
1889
  }
1848
- else {
1849
- return skopsJobLib(model);
1850
- }
1890
+ // repos pushed without skops often still declare their filename in config.json; it is only trusted when it names a
1891
+ // pickle, since some hand-written configs carry typos (e.g. "sklean_model.jpblib" next to a real sklearn_model.joblib)
1892
+ return sklearnPickle(model, declaredFile && SKLEARN_PICKLE_FILE.test(declaredFile) ? declaredFile : "sklearn_model.joblib");
1851
1893
  };
1852
1894
  exports.sklearn = sklearn;
1853
1895
  const stable_audio_tools = (model) => [
@@ -1898,7 +1940,7 @@ const sam2 = (model) => {
1898
1940
  import torch
1899
1941
  from sam2.sam2_image_predictor import SAM2ImagePredictor
1900
1942
 
1901
- predictor = SAM2ImagePredictor.from_pretrained(${model.id})
1943
+ predictor = SAM2ImagePredictor.from_pretrained("${model.id}")
1902
1944
 
1903
1945
  with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
1904
1946
  predictor.set_image(<your_image>)
@@ -1907,13 +1949,13 @@ with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
1907
1949
  import torch
1908
1950
  from sam2.sam2_video_predictor import SAM2VideoPredictor
1909
1951
 
1910
- predictor = SAM2VideoPredictor.from_pretrained(${model.id})
1952
+ predictor = SAM2VideoPredictor.from_pretrained("${model.id}")
1911
1953
 
1912
1954
  with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
1913
1955
  state = predictor.init_state(<your_video>)
1914
1956
 
1915
1957
  # add new prompts and instantly get the output on the same frame
1916
- frame_idx, object_ids, masks = predictor.add_new_points(state, <your_prompts>):
1958
+ frame_idx, object_ids, masks = predictor.add_new_points(state, <your_prompts>)
1917
1959
 
1918
1960
  # propagate the prompts to get masklets throughout the video
1919
1961
  for frame_idx, object_ids, masks in predictor.propagate_in_video(state):
@@ -1935,11 +1977,18 @@ output = inference(image, mask)`,
1935
1977
  ];
1936
1978
  exports.sam_3d_objects = sam_3d_objects;
1937
1979
  const sam_3d_body = (model) => [
1938
- `from notebook.utils import setup_sam_3d_body
1980
+ `import cv2
1981
+ import numpy as np
1982
+ from notebook.utils import setup_sam_3d_body
1983
+ from tools.vis_utils import visualize_sample_together
1984
+
1985
+ estimator = setup_sam_3d_body("${model.id}")
1986
+
1987
+ img_bgr = cv2.imread("path/to/image.jpg")
1988
+ outputs = estimator.process_one_image(cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB))
1939
1989
 
1940
- estimator = setup_sam_3d_body(${model.id})
1941
- outputs = estimator.process_one_image(image)
1942
- rend_img = visualize_sample_together(image, outputs, estimator.faces)`,
1990
+ rend_img = visualize_sample_together(img_bgr, outputs, estimator.faces)
1991
+ cv2.imwrite("output.jpg", rend_img.astype(np.uint8))`,
1943
1992
  ];
1944
1993
  exports.sam_3d_body = sam_3d_body;
1945
1994
  const sampleFactory = (model) => [
@@ -2033,7 +2082,15 @@ queries_emb = model.encode(queries, is_query=True)
2033
2082
  docs_emb = model.encode(documents, is_query=False)`,
2034
2083
  ];
2035
2084
  }
2036
- if (model.tags.includes("cross-encoder") || model.pipeline_tag == "text-ranking") {
2085
+ // Rerankers saved in the CrossEncoder format are often tagged text-classification rather than text-ranking;
2086
+ // their architecture still identifies them (e.g. BAAI/bge-reranker-v2-m3). SetFit models can share that shape but
2087
+ // are bi-encoders with a separate head.
2088
+ const isCrossEncoder = model.tags.includes("cross-encoder") ||
2089
+ model.pipeline_tag == "text-ranking" ||
2090
+ (model.pipeline_tag === "text-classification" &&
2091
+ !model.tags.includes("setfit") &&
2092
+ (model.config?.architectures ?? []).some((arch) => arch.endsWith("ForSequenceClassification")));
2093
+ if (isCrossEncoder) {
2037
2094
  return [
2038
2095
  `from sentence_transformers import CrossEncoder
2039
2096
 
@@ -2051,6 +2108,33 @@ scores = model.predict([(query, passage) for passage in passages])
2051
2108
  print(scores)`,
2052
2109
  ];
2053
2110
  }
2111
+ // sentence-transformers tags SparseEncoder repos (SPLADE & co) with "sparse-encoder" and MultiVectorEncoder repos
2112
+ // (ColBERT & co) with "multi-vector" when pushing them; both encode queries and documents separately
2113
+ const retrievalEncoder = model.tags.includes("sparse-encoder")
2114
+ ? "SparseEncoder"
2115
+ : model.tags.includes("multi-vector")
2116
+ ? "MultiVectorEncoder"
2117
+ : undefined;
2118
+ if (retrievalEncoder) {
2119
+ return [
2120
+ `from sentence_transformers import ${retrievalEncoder}
2121
+
2122
+ model = ${retrievalEncoder}("${model.id}"${remote_code_snippet})
2123
+
2124
+ queries = ["Which planet is known as the Red Planet?"]
2125
+ documents = [
2126
+ "Venus is often called Earth's twin because of its similar size and proximity.",
2127
+ "Mars, known for its reddish appearance, is often referred to as the Red Planet.",
2128
+ "Jupiter, the largest planet in our solar system, has a prominent red spot.",
2129
+ ]
2130
+
2131
+ query_embeddings = model.encode_query(queries)
2132
+ document_embeddings = model.encode_document(documents)
2133
+
2134
+ similarities = model.similarity(query_embeddings, document_embeddings)
2135
+ print(similarities)`,
2136
+ ];
2137
+ }
2054
2138
  const exampleSentences = get_widget_examples_from_st_model(model) ?? [
2055
2139
  "The weather is lovely today.",
2056
2140
  "It's so sunny outside!",
@@ -2073,7 +2157,10 @@ exports.sentenceTransformers = sentenceTransformers;
2073
2157
  const setfit = (model) => [
2074
2158
  `from setfit import SetFitModel
2075
2159
 
2076
- model = SetFitModel.from_pretrained("${model.id}")`,
2160
+ model = SetFitModel.from_pretrained("${model.id}")
2161
+
2162
+ preds = model.predict(["i loved the spiderman movie!", "pineapple on pizza is the worst"])
2163
+ print(preds)`,
2077
2164
  ];
2078
2165
  exports.setfit = setfit;
2079
2166
  const spacy = (model) => [
@@ -2091,7 +2178,10 @@ exports.spacy = spacy;
2091
2178
  const span_marker = (model) => [
2092
2179
  `from span_marker import SpanMarkerModel
2093
2180
 
2094
- model = SpanMarkerModel.from_pretrained("${model.id}")`,
2181
+ model = SpanMarkerModel.from_pretrained("${model.id}")
2182
+
2183
+ entities = model.predict("Amelia Earhart flew her single engine Lockheed Vega 5B across the Atlantic to Paris.")
2184
+ print(entities)`,
2095
2185
  ];
2096
2186
  exports.span_marker = span_marker;
2097
2187
  const stanza = (model) => [
@@ -2112,6 +2202,8 @@ const speechBrainMethod = (speechbrainInterface) => {
2112
2202
  return "enhance_file";
2113
2203
  case "SepformerSeparation":
2114
2204
  return "separate_file";
2205
+ case "SpeakerRecognition":
2206
+ return "verify_files";
2115
2207
  default:
2116
2208
  return undefined;
2117
2209
  }
@@ -2125,12 +2217,14 @@ const speechbrain = (model) => {
2125
2217
  if (speechbrainMethod === undefined) {
2126
2218
  return [`# interface in config.json invalid`];
2127
2219
  }
2220
+ // speaker verification compares two recordings
2221
+ const methodArgs = speechbrainMethod === "verify_files" ? '"speaker1.wav", "speaker2.wav"' : '"file.wav"';
2128
2222
  return [
2129
2223
  `from speechbrain.pretrained import ${speechbrainInterface}
2130
2224
  model = ${speechbrainInterface}.from_hparams(
2131
2225
  "${model.id}"
2132
2226
  )
2133
- model.${speechbrainMethod}("file.wav")`,
2227
+ model.${speechbrainMethod}(${methodArgs})`,
2134
2228
  ];
2135
2229
  };
2136
2230
  exports.speechbrain = speechbrain;
@@ -2162,7 +2256,7 @@ const transformers = (model) => {
2162
2256
  : info.processor === "AutoFeatureExtractor"
2163
2257
  ? "extractor"
2164
2258
  : "processor";
2165
- autoSnippet.push("# Load model directly", `from transformers import ${info.processor}, ${auto_model}`, "", `${processorVarName} = ${info.processor}.from_pretrained("${model.id}"` + remote_code_snippet + ")", `model = ${auto_model}.from_pretrained("${model.id}"` + remote_code_snippet + ', device_map="auto")');
2259
+ autoSnippet.push("# pip install -U transformers accelerate", "# Load model directly", `from transformers import ${info.processor}, ${auto_model}`, "", `${processorVarName} = ${info.processor}.from_pretrained("${model.id}"` + remote_code_snippet + ")", `model = ${auto_model}.from_pretrained("${model.id}"` + remote_code_snippet + ', device_map="auto")');
2166
2260
  if (model.tags.includes("conversational") && hasChatTemplate(model)) {
2167
2261
  if (model.tags.includes("image-text-to-text")) {
2168
2262
  autoSnippet.push("messages = [", [
@@ -2178,19 +2272,19 @@ const transformers = (model) => {
2178
2272
  else {
2179
2273
  autoSnippet.push("messages = [", ' {"role": "user", "content": "Who are you?"},', "]");
2180
2274
  }
2181
- autoSnippet.push(`inputs = ${processorVarName}.apply_chat_template(`, " messages,", " add_generation_prompt=True,", " tokenize=True,", " return_dict=True,", ' return_tensors="pt",', ").to(model.device)", "", "outputs = model.generate(**inputs, max_new_tokens=40)", `print(${processorVarName}.decode(outputs[0][inputs["input_ids"].shape[-1]:]))`);
2275
+ autoSnippet.push(`inputs = ${processorVarName}.apply_chat_template(`, " messages,", " add_generation_prompt=True,", " tokenize=True,", " return_dict=True,", ' return_tensors="pt",', ").to(model.device)", "", "outputs = model.generate(**inputs, max_new_tokens=256)", `print(${processorVarName}.decode(outputs[0][inputs["input_ids"].shape[-1]:]))`);
2182
2276
  }
2183
2277
  }
2184
2278
  else {
2185
- autoSnippet.push("# Load model directly", `from transformers import ${auto_model}`, `model = ${auto_model}.from_pretrained("${model.id}"` + remote_code_snippet + ', device_map="auto")');
2279
+ autoSnippet.push("# pip install -U transformers accelerate", "# Load model directly", `from transformers import ${auto_model}`, `model = ${auto_model}.from_pretrained("${model.id}"` + remote_code_snippet + ', device_map="auto")');
2186
2280
  }
2187
2281
  if (model.pipeline_tag && library_to_tasks_js_1.LIBRARY_TASK_MAPPING.transformers?.includes(model.pipeline_tag)) {
2188
2282
  const pipelineSnippet = ["# Use a pipeline as a high-level helper"];
2189
2283
  if (library_to_tasks_js_1.REMOVED_IN_V5_TRANSFORMERS_PIPELINES.includes(model.pipeline_tag)) {
2190
- pipelineSnippet.push(`# Warning: Pipeline type "${model.pipeline_tag}" is no longer supported in transformers v5.`, `# You must load the model directly (see below) or downgrade to v4.x with:`, `# 'pip install "transformers<5.0.0'`);
2284
+ pipelineSnippet.push(`# Warning: Pipeline type "${model.pipeline_tag}" is no longer supported in transformers v5.`, `# You must load the model directly (see below) or downgrade to v4.x with:`, `# pip install "transformers<5.0.0"`);
2191
2285
  }
2192
2286
  pipelineSnippet.push("from transformers import pipeline", "", `pipe = pipeline("${model.pipeline_tag}", model="${model.id}"` + remote_code_snippet + ")");
2193
- if (model.tags.includes("conversational")) {
2287
+ if (model.tags.includes("conversational") && hasChatTemplate(model)) {
2194
2288
  if (model.tags.includes("image-text-to-text")) {
2195
2289
  pipelineSnippet.push("messages = [", [
2196
2290
  " {",
@@ -2219,48 +2313,119 @@ const transformers = (model) => {
2219
2313
  return [autoSnippet.join("\n")];
2220
2314
  };
2221
2315
  exports.transformers = transformers;
2316
+ // Hub tasks that transformers.js exposes under a different pipeline() task name
2317
+ const TRANSFORMERS_JS_TASK_ALIASES = {
2318
+ "sentence-similarity": "feature-extraction",
2319
+ };
2222
2320
  const transformersJS = (model) => {
2223
2321
  if (!model.pipeline_tag) {
2224
2322
  return [`// ⚠️ Unknown pipeline tag`];
2225
2323
  }
2226
2324
  const libName = "@huggingface/transformers";
2325
+ // Rerankers score (query, document) pairs. The text-classification pipeline cannot take pairs and applies a softmax
2326
+ // over their single logit, so every score comes out as 1; the model is called directly instead.
2327
+ if (model.pipeline_tag === "text-ranking") {
2328
+ return [
2329
+ `// npm i ${libName}
2330
+ import { AutoTokenizer, AutoModelForSequenceClassification } from '${libName}';
2331
+
2332
+ const tokenizer = await AutoTokenizer.from_pretrained('${model.id}');
2333
+ const model = await AutoModelForSequenceClassification.from_pretrained('${model.id}');
2334
+
2335
+ const query = 'Which planet is known as the Red Planet?';
2336
+ const documents = [
2337
+ 'Mars, known for its reddish appearance, is often referred to as the Red Planet.',
2338
+ 'Venus is often called the twin of Earth because of its similar size and proximity.',
2339
+ ];
2340
+
2341
+ const inputs = tokenizer(new Array(documents.length).fill(query), { text_pair: documents, padding: true, truncation: true });
2342
+ const { logits } = await model(inputs);
2343
+ console.log(logits.sigmoid().tolist()); // one relevance score per document`,
2344
+ ];
2345
+ }
2346
+ const task = TRANSFORMERS_JS_TASK_ALIASES[model.pipeline_tag] ?? model.pipeline_tag;
2227
2347
  return [
2228
2348
  `// npm i ${libName}
2229
2349
  import { pipeline } from '${libName}';
2230
2350
 
2231
2351
  // Allocate pipeline
2232
- const pipe = await pipeline('${model.pipeline_tag}', '${model.id}');`,
2352
+ const pipe = await pipeline('${task}', '${model.id}');`,
2233
2353
  ];
2234
2354
  };
2235
2355
  exports.transformersJS = transformersJS;
2236
- const peftTask = (peftTaskType) => {
2237
- switch (peftTaskType) {
2238
- case "CAUSAL_LM":
2239
- return "CausalLM";
2240
- case "SEQ_2_SEQ_LM":
2241
- return "Seq2SeqLM";
2242
- case "TOKEN_CLS":
2243
- return "TokenClassification";
2244
- case "SEQ_CLS":
2245
- return "SequenceClassification";
2246
- default:
2247
- return undefined;
2248
- }
2356
+ const PEFT_TASK_TYPE_TO_AUTO_CLASS = {
2357
+ CAUSAL_LM: "AutoModelForCausalLM",
2358
+ SEQ_2_SEQ_LM: "AutoModelForSeq2SeqLM",
2359
+ TOKEN_CLS: "AutoModelForTokenClassification",
2360
+ SEQ_CLS: "AutoModelForSequenceClassification",
2361
+ QUESTION_ANS: "AutoModelForQuestionAnswering",
2362
+ // transformers has no AutoModelForFeatureExtraction; PEFT's own AutoPeftModelForFeatureExtraction resolves to AutoModel
2363
+ FEATURE_EXTRACTION: "AutoModel",
2364
+ };
2365
+ // PEFT has no speech-specific task type: Whisper-style adapters are tagged SEQ_2_SEQ_LM, but transformers registers
2366
+ // Whisper and Speech2Text only under AutoModelForSpeechSeq2Seq (SeamlessM4T is under both seq2seq auto classes, SpeechT5
2367
+ // under the speech-to-text and text-to-spectrogram ones).
2368
+ // Other audio seq2seq models (Qwen2-Audio, GLM-ASR) stay under AutoModelForSeq2SeqLM, so the base model name is the
2369
+ // main signal. Matched against Hub repo ids, which are hyphenated (facebook/s2t-small-librispeech-asr,
2370
+ // facebook/seamless-m4t-v2-large), not against transformers model_type identifiers.
2371
+ const PEFT_SPEECH_TO_TEXT_BASE_MODEL = /whisper|\bs2t\b|speech[-_]to[-_]text/i;
2372
+ // SeamlessM4T also has a text-to-text head under AutoModelForSeq2SeqLM, which text translation adapters need; most
2373
+ // SeamlessM4T adapters on the Hub are speech ones, so the speech class stays the default when no text task is declared.
2374
+ const PEFT_SEAMLESS_M4T_BASE_MODEL = /seamless[-_]m4t/i;
2375
+ const PEFT_TEXT_TASKS = ["translation", "text2text-generation", "summarization"];
2376
+ // SpeechT5 checkpoints are overwhelmingly text-to-speech fine-tunes (often without "tts" in their name), which load
2377
+ // with AutoModelForTextToSpectrogram; AutoModelForSpeechSeq2Seq would silently load the speech-to-text head, so only
2378
+ // ASR bases or adapters get it.
2379
+ const PEFT_SPEECHT5_BASE_MODEL = /speecht5/i;
2380
+ const declaresTask = (model, task) => model.pipeline_tag === task || model.tags.includes(task);
2381
+ // the auto class a SEQ_2_SEQ_LM adapter needs when its base is a speech model, undefined for text seq2seq bases
2382
+ const peftSpeechAutoClass = (model, baseModel) => {
2383
+ if (PEFT_SPEECH_TO_TEXT_BASE_MODEL.test(baseModel)) {
2384
+ return "AutoModelForSpeechSeq2Seq";
2385
+ }
2386
+ if (PEFT_SEAMLESS_M4T_BASE_MODEL.test(baseModel)) {
2387
+ return PEFT_TEXT_TASKS.some((task) => declaresTask(model, task)) ? undefined : "AutoModelForSpeechSeq2Seq";
2388
+ }
2389
+ if (PEFT_SPEECHT5_BASE_MODEL.test(baseModel)) {
2390
+ const isAsr = /(?:^|[^a-z])asr(?:[^a-z]|$)/i.test(baseModel) || declaresTask(model, "automatic-speech-recognition");
2391
+ return isAsr ? "AutoModelForSpeechSeq2Seq" : "AutoModelForTextToSpectrogram";
2392
+ }
2393
+ return undefined;
2249
2394
  };
2250
2395
  const peft = (model) => {
2251
2396
  const { base_model_name_or_path: peftBaseModel, task_type: peftTaskType } = model.config?.peft ?? {};
2252
- const pefttask = peftTask(peftTaskType);
2253
- if (!pefttask) {
2397
+ let autoClass = peftTaskType ? PEFT_TASK_TYPE_TO_AUTO_CLASS[peftTaskType] : undefined;
2398
+ if (!autoClass) {
2254
2399
  return [`Task type is invalid.`];
2255
2400
  }
2256
- if (!peftBaseModel) {
2401
+ // adapter_config.json sometimes omits the base model; the model card's base_model is the next best source
2402
+ const cardBaseModel = Array.isArray(model.cardData?.base_model)
2403
+ ? model.cardData.base_model[0]
2404
+ : model.cardData?.base_model;
2405
+ const baseModel = peftBaseModel || cardBaseModel;
2406
+ if (!baseModel) {
2257
2407
  return [`Base model is not found.`];
2258
2408
  }
2409
+ if (peftTaskType === "SEQ_2_SEQ_LM") {
2410
+ autoClass = peftSpeechAutoClass(model, baseModel) ?? autoClass;
2411
+ }
2412
+ if (autoClass === "AutoModelForTextToSpectrogram") {
2413
+ return [
2414
+ `from peft import PeftConfig, PeftModel
2415
+ from transformers import AutoModelForTextToSpectrogram
2416
+
2417
+ base_model = AutoModelForTextToSpectrogram.from_pretrained("${escapeStringForJson(baseModel)}")
2418
+ # PEFT's seq2seq wrapper needs generate(), which SpeechT5's text-to-speech model lacks: load the adapter without its task type
2419
+ config = PeftConfig.from_pretrained("${model.id}")
2420
+ config.task_type = None
2421
+ model = PeftModel.from_pretrained(base_model, "${model.id}", config=config)`,
2422
+ ];
2423
+ }
2259
2424
  return [
2260
2425
  `from peft import PeftModel
2261
- from transformers import AutoModelFor${pefttask}
2426
+ from transformers import ${autoClass}
2262
2427
 
2263
- base_model = AutoModelFor${pefttask}.from_pretrained("${escapeStringForJson(peftBaseModel)}")
2428
+ base_model = ${autoClass}.from_pretrained("${escapeStringForJson(baseModel)}")
2264
2429
  model = PeftModel.from_pretrained(base_model, "${model.id}")`,
2265
2430
  ];
2266
2431
  };
@@ -2294,7 +2459,7 @@ transcriptions = asr_model.transcribe(["file.wav"])`,
2294
2459
  }
2295
2460
  };
2296
2461
  const mlAgents = (model) => [
2297
- `mlagents-load-from-hf --repo-id="${model.id}" --local-dir="./download: string[]s"`,
2462
+ `mlagents-load-from-hf --repo-id="${model.id}" --local-dir="./downloads"`,
2298
2463
  ];
2299
2464
  exports.mlAgents = mlAgents;
2300
2465
  const sentis = ( /* model: ModelData */) => [
@@ -2369,10 +2534,10 @@ model = Model.from_pretrained("${model.id}")`,
2369
2534
  exports.vfimamba = vfimamba;
2370
2535
  const lvface = (model) => [
2371
2536
  `from huggingface_hub import hf_hub_download
2372
- from inference_onnx import LVFaceONNXInferencer
2537
+ from inference_onnx import LVFaceONNXInferencer
2373
2538
 
2374
2539
  model_path = hf_hub_download("${model.id}", "LVFace-L_Glint360K/LVFace-L_Glint360K.onnx")
2375
- inferencer = LVFaceONNXInferencer(model_path, use_gpu=True, timeout=300)
2540
+ inferencer = LVFaceONNXInferencer(model_path, use_gpu=True)
2376
2541
  img_path = 'path/to/image1.jpg'
2377
2542
  embedding = inferencer.infer_from_image(img_path)`,
2378
2543
  ];
@@ -2411,8 +2576,7 @@ const vui = () => [
2411
2576
 
2412
2577
  import torchaudio
2413
2578
 
2414
- from vui.inference import render
2415
- from vui.model import Vui,
2579
+ from vui.legacy import Vui, render
2416
2580
 
2417
2581
  model = Vui.from_pretrained().cuda()
2418
2582
  waveform = render(
@@ -2420,16 +2584,17 @@ waveform = render(
2420
2584
  "Hey, here is some random stuff, usually something quite long as the shorter the text the less likely the model can cope!",
2421
2585
  )
2422
2586
  print(waveform.shape)
2423
- torchaudio.save("out.opus", waveform[0], 22050)
2587
+ torchaudio.save("out.wav", waveform[0], 22050)
2424
2588
  `,
2425
2589
  ];
2426
2590
  exports.vui = vui;
2427
2591
  const chattts = () => [
2428
2592
  `import ChatTTS
2593
+ import torch
2429
2594
  import torchaudio
2430
2595
 
2431
2596
  chat = ChatTTS.Chat()
2432
- chat.load_models(compile=False) # Set to True for better performance
2597
+ chat.load(compile=False) # Set to True for better performance
2433
2598
 
2434
2599
  texts = ["PUT YOUR TEXT HERE",]
2435
2600
 
@@ -2439,17 +2604,42 @@ torchaudio.save("output1.wav", torch.from_numpy(wavs[0]), 24000)`,
2439
2604
  ];
2440
2605
  exports.chattts = chattts;
2441
2606
  const ultralytics = (model) => {
2442
- // ultralytics models must have a version tag (e.g. `yolov8`)
2443
- const versionTag = model.tags.find((tag) => tag.match(/^yolov\d+$/));
2444
- const className = versionTag ? `YOLOv${versionTag.slice(4)}` : "YOLOvXX";
2445
- const prefix = versionTag
2446
- ? ""
2447
- : `# Couldn't find a valid YOLO version tag.\n# Replace XX with the correct version.\n`;
2607
+ // the YOLOv10 fork (library "yolov10") ships its own `YOLOv10` class with a `from_pretrained` helper that reads the
2608
+ // model.safetensors it pushes. Mainline ultralytics repos often list "yolov10" among their version tags
2609
+ // (e.g. Ultralytics/YOLOv8) but only ship .pt files, so an ultralytics tag alone keeps them on the mainline loader;
2610
+ // fork pushes still tagged ultralytics are recognized by their safetensors weights.
2611
+ const isYolov10Fork = model.library_name === "yolov10" ||
2612
+ (model.tags.includes("yolov10") && (!model.tags.includes("ultralytics") || model.tags.includes("safetensors")));
2613
+ if (isYolov10Fork) {
2614
+ // from_pretrained reads the model.safetensors that fork pushes ship; older fork repos only ship .pt weights
2615
+ if (model.tags.includes("safetensors")) {
2616
+ return [
2617
+ `from ultralytics import YOLOv10
2618
+
2619
+ model = YOLOv10.from_pretrained("${model.id}")
2620
+ source = 'http://images.cocodataset.org/val2017/000000039769.jpg'
2621
+ model.predict(source=source, save=True)`,
2622
+ ];
2623
+ }
2624
+ return [
2625
+ `from huggingface_hub import hf_hub_download
2626
+ from ultralytics import YOLOv10
2627
+
2628
+ # pick the weights file from this repo's "Files and versions" tab
2629
+ weights = hf_hub_download("${model.id}", "<weights>.pt")
2630
+ model = YOLOv10(weights)
2631
+ source = 'http://images.cocodataset.org/val2017/000000039769.jpg'
2632
+ model.predict(source=source, save=True)`,
2633
+ ];
2634
+ }
2635
+ // mainline ultralytics loads weights from a local file
2448
2636
  return [
2449
- prefix +
2450
- `from ultralytics import ${className}
2637
+ `from huggingface_hub import hf_hub_download
2638
+ from ultralytics import YOLO
2451
2639
 
2452
- model = ${className}.from_pretrained("${model.id}")
2640
+ # pick the weights file from this repo's "Files and versions" tab
2641
+ weights = hf_hub_download("${model.id}", "<weights>.pt")
2642
+ model = YOLO(weights)
2453
2643
  source = 'http://images.cocodataset.org/val2017/000000039769.jpg'
2454
2644
  model.predict(source=source, save=True)`,
2455
2645
  ];
@@ -2527,14 +2717,16 @@ model = SwarmFormerModel.from_pretrained("${model.id}")
2527
2717
  exports.swarmformer = swarmformer;
2528
2718
  const univa = (model) => [
2529
2719
  `# Follow installation instructions at https://github.com/PKU-YuanGroup/UniWorld-V1
2530
-
2720
+ import torch
2721
+ from transformers import AutoProcessor
2531
2722
  from univa.models.qwen2p5vl.modeling_univa_qwen2p5vl import UnivaQwen2p5VLForConditionalGeneration
2532
- model = UnivaQwen2p5VLForConditionalGeneration.from_pretrained(
2533
- "${model.id}",
2534
- torch_dtype=torch.bfloat16,
2535
- attn_implementation="flash_attention_2",
2536
- ).to("cuda")
2537
- processor = AutoProcessor.from_pretrained("${model.id}")
2723
+
2724
+ model = UnivaQwen2p5VLForConditionalGeneration.from_pretrained(
2725
+ "${model.id}",
2726
+ torch_dtype=torch.bfloat16,
2727
+ attn_implementation="flash_attention_2",
2728
+ ).to("cuda")
2729
+ processor = AutoProcessor.from_pretrained("${model.id}")
2538
2730
  `,
2539
2731
  ];
2540
2732
  exports.univa = univa;
@@ -2542,7 +2734,7 @@ const mlx_unknown = (model) => [
2542
2734
  `# Download the model from the Hub
2543
2735
  pip install huggingface_hub[hf_xet]
2544
2736
 
2545
- huggingface-cli download --local-dir ${nameWithoutNamespace(model.id)} ${model.id}`,
2737
+ hf download ${model.id} --local-dir ${nameWithoutNamespace(model.id)}`,
2546
2738
  ];
2547
2739
  const mlxlm = (model) => [
2548
2740
  `# Make sure mlx-lm is installed
@@ -2599,10 +2791,11 @@ formatted_prompt = apply_chat_template(
2599
2791
  output = generate(model, processor, formatted_prompt, image)
2600
2792
  print(output)`,
2601
2793
  ];
2794
+ // create_model() takes a name from the mlx-image registry, not a repo id; the weights live in "mlx-vision/<name>-mlxim"
2602
2795
  const mlxim = (model) => [
2603
2796
  `from mlxim.model import create_model
2604
2797
 
2605
- model = create_model(${model.id})`,
2798
+ model = create_model("${nameWithoutNamespace(model.id).replace(/-mlxim$/, "")}")`,
2606
2799
  ];
2607
2800
  exports.mlxim = mlxim;
2608
2801
  const mlx = (model) => {
@@ -2623,7 +2816,10 @@ exports.mlx = mlx;
2623
2816
  const model2vec = (model) => [
2624
2817
  `from model2vec import StaticModel
2625
2818
 
2626
- model = StaticModel.from_pretrained("${model.id}")`,
2819
+ model = StaticModel.from_pretrained("${model.id}")
2820
+
2821
+ embeddings = model.encode(["It's dangerous to go alone!", "It's a secret to everybody."])
2822
+ print(embeddings.shape)`,
2627
2823
  ];
2628
2824
  exports.model2vec = model2vec;
2629
2825
  const mobilint = (model) => {
@@ -2668,7 +2864,8 @@ const pruna = (model) => {
2668
2864
  };
2669
2865
  snippets = snippets.map(ensurePrunaModelImport);
2670
2866
  if (model.tags.includes("pruna_pro-ai")) {
2671
- return snippets.map((snippet) => snippet.replace(/\bpruna\b/g, "pruna_pro").replace(/\bPrunaModel\b/g, "PrunaProModel"));
2867
+ // only rewrite the import and the class name, never the repo id (an org name can contain "pruna")
2868
+ return snippets.map((snippet) => snippet.replace(/^from pruna import /gm, "from pruna_pro import ").replace(/\bPrunaModel\b/g, "PrunaProModel"));
2672
2869
  }
2673
2870
  return snippets;
2674
2871
  };
@@ -2728,21 +2925,19 @@ const outetts = (model) => {
2728
2925
  }
2729
2926
  // v1.0 HF → minimal runnable snippet
2730
2927
  return [
2731
- `
2732
- import outetts
2733
-
2734
- enum = outetts.Models("${model.id}".split("/", 1)[1]) # VERSION_1_0_SIZE_1B
2735
- cfg = outetts.ModelConfig.auto_config(enum, outetts.Backend.HF)
2736
- tts = outetts.Interface(cfg)
2737
-
2738
- speaker = tts.load_default_speaker("EN-FEMALE-1-NEUTRAL")
2739
- tts.generate(
2740
- outetts.GenerationConfig(
2741
- text="Hello there, how are you doing?",
2742
- speaker=speaker,
2743
- )
2744
- ).save("output.wav")
2745
- `,
2928
+ `import outetts
2929
+
2930
+ enum = outetts.Models("${model.id}".split("/", 1)[1]) # VERSION_1_0_SIZE_1B
2931
+ cfg = outetts.ModelConfig.auto_config(enum, outetts.Backend.HF)
2932
+ tts = outetts.Interface(cfg)
2933
+
2934
+ speaker = tts.load_default_speaker("EN-FEMALE-1-NEUTRAL")
2935
+ tts.generate(
2936
+ outetts.GenerationConfig(
2937
+ text="Hello there, how are you doing?",
2938
+ speaker=speaker,
2939
+ )
2940
+ ).save("output.wav")`,
2746
2941
  ];
2747
2942
  };
2748
2943
  exports.outetts = outetts;