@huggingface/tasks 0.21.53 → 0.21.55

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.
@@ -1,8 +1,8 @@
1
1
  "use strict";
2
2
  Object.defineProperty(exports, "__esModule", { value: true });
3
- exports.tf_keras = 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.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.lvface = 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 = 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 = void 0;
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.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.microduck = 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 = exports.vfimamba = 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;
@@ -74,7 +74,13 @@ result, message = detector.detect_watermark(watermarked_audio, sr)`;
74
74
  };
75
75
  exports.audioseal = audioseal;
76
76
  function get_base_diffusers_model(model) {
77
- return escapeStringForJson(model.cardData?.base_model?.toString() ?? "fill-in-base-model");
77
+ const baseModel = model.cardData?.base_model;
78
+ // several base models can be listed (e.g. a LoRA trained for a base and its turbo variant). When an original checkpoint
79
+ // and its Diffusers conversion are both listed (["Wan-AI/Wan2.1-I2V-14B-480P", "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"]),
80
+ // only the conversion has a model_index.json that DiffusionPipeline can load; otherwise the first one is used
81
+ const candidates = Array.isArray(baseModel) ? baseModel : baseModel ? [baseModel] : [];
82
+ const chosen = candidates.find((id) => id.toLowerCase().endsWith("-diffusers")) ?? candidates[0];
83
+ return escapeStringForJson(chosen ?? "fill-in-base-model");
78
84
  }
79
85
  function get_prompt_from_diffusers_model(model) {
80
86
  const prompt = model.widgetData?.[0]?.text ?? model.cardData?.instance_prompt;
@@ -161,7 +167,7 @@ detector = cvg.NeuralCornerDetector(checkpoint)
161
167
  embedder = cvg.NeuralEmbedder(checkpoint)`,
162
168
  ];
163
169
  exports.collectorvision = collectorvision;
164
- const colipri = (model) => {
170
+ const colipri = () => {
165
171
  const installSnippet = `pip install colipri`;
166
172
  const exampleSnippet = `from colipri import get_model
167
173
  from colipri import get_processor
@@ -170,7 +176,7 @@ from colipri import ZeroShotImageClassificationPipeline
170
176
 
171
177
  model = get_model().cuda()
172
178
  processor = get_processor()
173
- pipeline = ZeroShotImageClassificationPipeline("${model.id}", processor)
179
+ pipeline = ZeroShotImageClassificationPipeline(model, processor)
174
180
 
175
181
  image = load_sample_ct()
176
182
 
@@ -291,7 +297,7 @@ const depth_anything_v2 = (model) => {
291
297
  else if (model.id === "depth-anything/Depth-Anything-V2-Large") {
292
298
  encoder = "vitl";
293
299
  features = "256";
294
- out_channels = "[256, 512, 1024, 1024";
300
+ out_channels = "[256, 512, 1024, 1024]";
295
301
  }
296
302
  return [
297
303
  `
@@ -300,6 +306,7 @@ const depth_anything_v2 = (model) => {
300
306
  # Load the model and infer depth from an image
301
307
  import cv2
302
308
  import torch
309
+ from huggingface_hub import hf_hub_download
303
310
 
304
311
  from depth_anything_v2.dpt import DepthAnythingV2
305
312
 
@@ -309,7 +316,8 @@ model = DepthAnythingV2(encoder="${encoder}", features=${features}, out_channels
309
316
  # load the weights
310
317
  filepath = hf_hub_download(repo_id="${model.id}", filename="depth_anything_v2_${encoder}.pth", repo_type="model")
311
318
  state_dict = torch.load(filepath, map_location="cpu")
312
- model.load_state_dict(state_dict).eval()
319
+ model.load_state_dict(state_dict)
320
+ model.eval()
313
321
 
314
322
  raw_img = cv2.imread("your/image/path")
315
323
  depth = model.infer_image(raw_img) # HxW raw depth map in numpy
@@ -319,8 +327,8 @@ depth = model.infer_image(raw_img) # HxW raw depth map in numpy
319
327
  exports.depth_anything_v2 = depth_anything_v2;
320
328
  const depth_pro = (model) => {
321
329
  const installSnippet = `# Download checkpoint
322
- pip install huggingface-hub
323
- huggingface-cli download --local-dir checkpoints ${model.id}`;
330
+ pip install huggingface_hub
331
+ hf download ${model.id} --local-dir checkpoints`;
324
332
  const inferenceSnippet = `import depth_pro
325
333
 
326
334
  # Load model and preprocessing transform
@@ -392,7 +400,7 @@ const describe_anything = (model) => [
392
400
  from huggingface_hub import snapshot_download
393
401
  from dam import DescribeAnythingModel
394
402
 
395
- snapshot_download(${model.id}, local_dir="checkpoints")
403
+ snapshot_download("${model.id}", local_dir="checkpoints")
396
404
 
397
405
  dam = DescribeAnythingModel(
398
406
  model_path="checkpoints",
@@ -504,7 +512,7 @@ pipe.load_lora_weights("${model.id}")
504
512
  prompt = "${get_prompt_from_diffusers_model(model) ?? diffusersVideoDefaultPrompt}"
505
513
  input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/guitar-man.png")
506
514
 
507
- image = pipe(image=input_image, prompt=prompt).frames[0]
515
+ output = pipe(image=input_image, prompt=prompt).frames[0]
508
516
  export_to_video(output, "output.mp4")`,
509
517
  ];
510
518
  const diffusers_textual_inversion = (model) => [
@@ -544,7 +552,7 @@ from diffusers import AutoPipelineForInpainting
544
552
  from diffusers.utils import load_image
545
553
 
546
554
  # switch to "mps" for apple devices
547
- pipe = AutoPipelineForInpainting.from_pretrained("${model.id}", dtype=torch.float16, variant="fp16", device_map="cuda")
555
+ pipe = AutoPipelineForInpainting.from_pretrained("${model.id}", dtype=torch.float16, device_map="cuda")
548
556
 
549
557
  img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
550
558
  mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
@@ -566,9 +574,12 @@ image = pipe(
566
574
  ).images[0]`,
567
575
  ];
568
576
  const diffusers = (model) => {
577
+ // the Hub emits pipeline class tags as `diffusers:<ClassName>`; bare names are kept for manually added tags
578
+ const hasPipelineClass = (className) => model.tags.includes(`diffusers:${className}`) ||
579
+ model.tags.includes(className) ||
580
+ model.config?.diffusers?._class_name === className;
569
581
  let codeSnippets;
570
- if (model.tags.includes("StableDiffusionInpaintPipeline") ||
571
- model.tags.includes("StableDiffusionXLInpaintPipeline")) {
582
+ if (hasPipelineClass("StableDiffusionInpaintPipeline") || hasPipelineClass("StableDiffusionXLInpaintPipeline")) {
572
583
  codeSnippets = diffusers_inpainting(model);
573
584
  }
574
585
  else if (model.tags.includes("controlnet")) {
@@ -591,7 +602,7 @@ const diffusers = (model) => {
591
602
  else if (model.tags.includes("textual_inversion")) {
592
603
  codeSnippets = diffusers_textual_inversion(model);
593
604
  }
594
- else if (model.tags.includes("FluxFillPipeline")) {
605
+ else if (hasPipelineClass("FluxFillPipeline")) {
595
606
  codeSnippets = diffusers_flux_fill(model);
596
607
  }
597
608
  else if (model.pipeline_tag === "image-to-video") {
@@ -613,7 +624,7 @@ from diffusionkit.mlx import DiffusionPipeline
613
624
  pipeline = DiffusionPipeline(
614
625
  shift=3.0,
615
626
  use_t5=False,
616
- model_version=${model.id},
627
+ model_version="${model.id}",
617
628
  low_memory_mode=True,
618
629
  a16=True,
619
630
  w16=True,
@@ -623,7 +634,7 @@ from diffusionkit.mlx import FluxPipeline
623
634
 
624
635
  pipeline = FluxPipeline(
625
636
  shift=1.0,
626
- model_version=${model.id},
637
+ model_version="${model.id}",
627
638
  low_memory_mode=True,
628
639
  a16=True,
629
640
  w16=True,
@@ -645,21 +656,22 @@ image, _ = pipeline.generate_image(
645
656
  };
646
657
  exports.diffusionkit = diffusionkit;
647
658
  const cartesia_pytorch = (model) => [
648
- `# pip install --no-binary :all: cartesia-pytorch
649
- from cartesia_pytorch import ReneLMHeadModel
659
+ `# The cartesia-pytorch package on PyPI ships no code: clone the repo and put its cartesia-pytorch folder on PYTHONPATH
660
+ # (it needs an Ampere or newer CUDA GPU, torch 2.4, mamba-ssm 2.2.2 and flash-attn 2.6.3)
661
+ # git clone https://github.com/cartesia-ai/edge && export PYTHONPATH="$PWD/edge/cartesia-pytorch:$PYTHONPATH"
662
+ from cartesia_pytorch.Rene import ReneLMHeadModel
650
663
  from transformers import AutoTokenizer
651
664
 
652
- model = ReneLMHeadModel.from_pretrained("${model.id}")
665
+ model = ReneLMHeadModel.from_pretrained("${model.id}").half().cuda()
653
666
  tokenizer = AutoTokenizer.from_pretrained("allenai/OLMo-1B-hf")
654
667
 
655
668
  in_message = ["Rene Descartes was"]
656
669
  inputs = tokenizer(in_message, return_tensors="pt")
657
670
 
658
- outputs = model.generate(inputs.input_ids, max_length=50, top_k=100, top_p=0.99)
671
+ outputs = model.generate(inputs.input_ids.cuda(), max_length=50, top_k=100, top_p=0.99)
659
672
  out_message = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
660
673
 
661
- print(out_message)
662
- )`,
674
+ print(out_message)`,
663
675
  ];
664
676
  exports.cartesia_pytorch = cartesia_pytorch;
665
677
  const cartesia_mlx = (model) => [
@@ -710,7 +722,8 @@ speech, *_ = model("text to generate speech from")`,
710
722
  ];
711
723
  exports.espnetTTS = espnetTTS;
712
724
  const espnetASR = (model) => [
713
- `from espnet2.bin.asr_inference import Speech2Text
725
+ `import soundfile
726
+ from espnet2.bin.asr_inference import Speech2Text
714
727
 
715
728
  model = Speech2Text.from_pretrained(
716
729
  "${model.id}"
@@ -740,9 +753,14 @@ models, cfg, task = load_model_ensemble_and_task_from_hf_hub(
740
753
  ];
741
754
  exports.fairseq = fairseq;
742
755
  const flair = (model) => [
743
- `from flair.models import SequenceTagger
756
+ `from flair.data import Sentence
757
+ from flair.models import SequenceTagger
744
758
 
745
- tagger = SequenceTagger.load("${model.id}")`,
759
+ tagger = SequenceTagger.load("${model.id}")
760
+
761
+ sentence = Sentence("George Washington went to Washington.")
762
+ tagger.predict(sentence)
763
+ print(sentence)`,
746
764
  ];
747
765
  exports.flair = flair;
748
766
  const flexray = (model) => [
@@ -820,16 +838,31 @@ print(f'Accuracy {accuracy_score(y_test, predictions):.2%}')
820
838
  return [installSnippet, classificationSnippet, regressionsSnippet, matchingSnippet];
821
839
  };
822
840
  exports.flextab = flextab;
841
+ const gaussianformer = (model) => [
842
+ `# Install from https://github.com/SVLwoof/gaussianformer
843
+
844
+ from gaussianformer.pipelines.rendering_pipeline import GaussianFormerRenderingPipeline
845
+
846
+ pipeline = GaussianFormerRenderingPipeline.from_pretrained("${model.id}")`,
847
+ ];
848
+ exports.gaussianformer = gaussianformer;
823
849
  const gliner = (model) => [
824
850
  `from gliner import GLiNER
825
851
 
826
- model = GLiNER.from_pretrained("${model.id}")`,
852
+ model = GLiNER.from_pretrained("${model.id}")
853
+
854
+ text = "Cristiano Ronaldo dos Santos Aveiro was born on 5 February 1985 in Funchal, Madeira, Portugal."
855
+ labels = ["person", "date", "location"]
856
+
857
+ entities = model.predict_entities(text, labels)
858
+ for entity in entities:
859
+ print(entity["text"], "=>", entity["label"])`,
827
860
  ];
828
861
  exports.gliner = gliner;
829
862
  const gliner2 = (model) => [
830
- `from gliner2 import GLiNER2
863
+ `from gliner2 import AutoExtractor
831
864
 
832
- model = GLiNER2.from_pretrained("${model.id}")
865
+ extractor = AutoExtractor.from_pretrained("${model.id}")
833
866
 
834
867
  # Extract entities
835
868
  text = "Apple CEO Tim Cook announced iPhone 15 in Cupertino yesterday."
@@ -842,7 +875,7 @@ const indextts = (model) => [
842
875
  `# Download model
843
876
  from huggingface_hub import snapshot_download
844
877
 
845
- snapshot_download(${model.id}, local_dir="checkpoints")
878
+ snapshot_download("${model.id}", local_dir="checkpoints")
846
879
 
847
880
  from indextts.infer import IndexTTS
848
881
 
@@ -899,9 +932,11 @@ print(out["answers"]["team"]["choice"])`,
899
932
  };
900
933
  exports.jev_style = jev_style;
901
934
  const keras = (model) => [
902
- `# Available backend options are: "jax", "torch", "tensorflow".
935
+ `# !pip install -U keras tensorflow huggingface_hub
936
+ # Keras needs TensorFlow installed to read "hf://" paths, so the tensorflow backend is selected here;
937
+ # "jax" and "torch" also work for computation once TensorFlow is installed.
903
938
  import os
904
- os.environ["KERAS_BACKEND"] = "jax"
939
+ os.environ["KERAS_BACKEND"] = "tensorflow"
905
940
 
906
941
  import keras
907
942
 
@@ -1018,13 +1053,15 @@ const kernels = (model) => [
1018
1053
 
1019
1054
  from kernels import get_kernel
1020
1055
 
1021
- kernel = get_kernel("${model.id}")`,
1056
+ # a version (or an explicit revision) is required; see the "Files and versions" tab for the available ones
1057
+ kernel = get_kernel("${model.id}", version=1)`,
1022
1058
  ];
1023
1059
  exports.kernels = kernels;
1024
1060
  const kimi_audio = (model) => [
1025
1061
  `# Example usage for KimiAudio
1026
- # pip install git+https://github.com/MoonshotAI/Kimi-Audio.git
1062
+ # pip install git+https://github.com/MoonshotAI/Kimi-Audio.git "transformers<5"
1027
1063
 
1064
+ import soundfile as sf
1028
1065
  from kimia_infer.api.kimia import KimiAudio
1029
1066
 
1030
1067
  model = KimiAudio(model_path="${model.id}", load_detokenizer=True)
@@ -1394,14 +1431,17 @@ pip install -U litert-lm
1394
1431
 
1395
1432
  # 2. Download and run this model locally:
1396
1433
  # See: https://ai.google.dev/edge/litert-lm/cli
1434
+ # A single .litertlm file in the repo is picked automatically; otherwise the CLI asks which one to run
1435
+ # (or pass its name right after the repo id).
1397
1436
  litert-lm run \\
1398
1437
  --from-huggingface-repo=${model.id} \\
1399
1438
  --prompt="Write me a poem"`,
1400
1439
  ];
1401
1440
  exports.litert_lm = litert_lm;
1402
1441
  const tf_keras = (model) => [
1403
- `# Note: 'keras<3.x' or 'tf_keras' must be installed (legacy)
1442
+ `# Note: 'keras<3.x' or 'tf_keras' must be installed (legacy), and from_pretrained_keras was removed in huggingface_hub 1.0.
1404
1443
  # See https://github.com/keras-team/tf-keras for more details.
1444
+ # !pip install "huggingface_hub<1.0" tf_keras
1405
1445
  from huggingface_hub import from_pretrained_keras
1406
1446
 
1407
1447
  model = from_pretrained_keras("${model.id}")
@@ -1453,6 +1493,17 @@ from MeshAnything.models.meshanything import MeshAnything
1453
1493
  model = MeshAnything(args)`,
1454
1494
  ];
1455
1495
  exports.mesh_anything = mesh_anything;
1496
+ const microduck = (model) => {
1497
+ // Model-card tag convention: tags: [microduck-slot:stand].
1498
+ const slots = ["walk", "stand", "sitstand", "ground_pick", "kick_left", "kick_right", "roulade"];
1499
+ const matchingSlots = slots.filter((slot) => model.tags.includes(`microduck-slot:${slot}`));
1500
+ const slot = matchingSlots.length === 1 ? matchingSlots[0] : "SLOT";
1501
+ const prefix = matchingSlots.length === 1
1502
+ ? ""
1503
+ : `# Replace SLOT with the slot specified in the model card (${slots.join(", ")}).\n`;
1504
+ return [`${prefix}sudo robotctl policy load ${slot} ${model.id}`];
1505
+ };
1506
+ exports.microduck = microduck;
1456
1507
  const multimolecule = (model) => {
1457
1508
  const widgetExample = model.widgetData?.[0];
1458
1509
  const exampleText = escapeStringForJson(widgetExample?.text ?? "");
@@ -1633,14 +1684,18 @@ model = pe.VisionTransformer.from_config("${model.id}", pretrained=True)`;
1633
1684
  exports.perception_encoder = perception_encoder;
1634
1685
  const phantom_wan = (model) => [
1635
1686
  `from huggingface_hub import snapshot_download
1636
- from phantom_wan import WANI2V, configs
1687
+ from phantom_wan import WanI2V, configs
1688
+ from PIL import Image
1637
1689
 
1638
1690
  checkpoint_dir = snapshot_download("${model.id}")
1639
1691
  wan_i2v = WanI2V(
1640
- config=configs.WAN_CONFIGS['i2v-14B'],
1641
- checkpoint_dir=checkpoint_dir,
1642
- )
1643
- video = wan_i2v.generate(text_prompt, image_prompt)`,
1692
+ config=configs.WAN_CONFIGS['i2v-14B'],
1693
+ checkpoint_dir=checkpoint_dir,
1694
+ )
1695
+
1696
+ text_prompt = "A cat playing with a ball of yarn"
1697
+ image_prompt = Image.open("path/to/image.jpg").convert("RGB")
1698
+ video = wan_i2v.generate(text_prompt, image_prompt)`,
1644
1699
  ];
1645
1700
  exports.phantom_wan = phantom_wan;
1646
1701
  const pocket_tts = (model) => [
@@ -1762,7 +1817,7 @@ exports.tensorflowtts = tensorflowtts;
1762
1817
  const timm = (model) => [
1763
1818
  `import timm
1764
1819
 
1765
- model = timm.create_model("hf_hub:${model.id}", pretrained=True)`,
1820
+ model = timm.create_model("hf-hub:${model.id}", pretrained=True)`,
1766
1821
  ];
1767
1822
  exports.timm = timm;
1768
1823
  const saelens = ( /* model: ModelData */) => [
@@ -1790,13 +1845,14 @@ seed_story_cfg = OmegaConf.load(seed_story_cfg_path)
1790
1845
  seed_story = hydra.utils.instantiate(seed_story_cfg, llm=llm) `,
1791
1846
  ];
1792
1847
  exports.seed_story = seed_story;
1793
- const skopsPickle = (model, modelFile) => {
1848
+ const SKLEARN_PICKLE_FILE = /\.(joblib|pkl|pickle)$/i;
1849
+ // skops.hub_utils was removed in skops 0.12, so files are downloaded with huggingface_hub
1850
+ const sklearnPickle = (model, modelFile) => {
1794
1851
  return [
1795
- `import joblib
1796
- from skops.hub_utils import download
1797
- download("${model.id}", "path_to_folder")
1852
+ `from huggingface_hub import hf_hub_download
1853
+ import joblib
1798
1854
  model = joblib.load(
1799
- "${escapeStringForJson(modelFile)}"
1855
+ hf_hub_download("${model.id}", "${escapeStringForJson(modelFile)}")
1800
1856
  )
1801
1857
  # only load pickle files from sources you trust
1802
1858
  # read more about it here https://skops.readthedocs.io/en/stable/persistence.html`,
@@ -1804,42 +1860,28 @@ model = joblib.load(
1804
1860
  };
1805
1861
  const skopsFormat = (model, modelFile) => {
1806
1862
  return [
1807
- `from skops.hub_utils import download
1863
+ `from huggingface_hub import hf_hub_download
1808
1864
  from skops.io import load
1809
- download("${model.id}", "path_to_folder")
1810
1865
  # make sure model file is in skops format
1811
1866
  # if model is a pickle file, make sure it's from a source you trust
1812
- model = load("path_to_folder/${escapeStringForJson(modelFile)}")`,
1813
- ];
1814
- };
1815
- const skopsJobLib = (model) => {
1816
- return [
1817
- `from huggingface_hub import hf_hub_download
1818
- import joblib
1819
- model = joblib.load(
1820
- hf_hub_download("${model.id}", "sklearn_model.joblib")
1821
- )
1822
- # only load pickle files from sources you trust
1823
- # read more about it here https://skops.readthedocs.io/en/stable/persistence.html`,
1867
+ model = load(hf_hub_download("${model.id}", "${escapeStringForJson(modelFile)}"))`,
1824
1868
  ];
1825
1869
  };
1826
1870
  const sklearn = (model) => {
1871
+ const declaredFile = model.config?.sklearn?.model?.file;
1827
1872
  if (model.tags.includes("skops")) {
1828
- const skopsmodelFile = model.config?.sklearn?.model?.file;
1829
- const skopssaveFormat = model.config?.sklearn?.model_format;
1830
- if (!skopsmodelFile) {
1873
+ if (!declaredFile) {
1831
1874
  return [`# ⚠️ Model filename not specified in config.json`];
1832
1875
  }
1833
- if (skopssaveFormat === "pickle") {
1834
- return skopsPickle(model, skopsmodelFile);
1835
- }
1836
- else {
1837
- return skopsFormat(model, skopsmodelFile);
1876
+ // some skops repos declare no model_format for a pickled file (e.g. julien-c/skops-digits), which skops.io cannot read
1877
+ if (model.config?.sklearn?.model_format === "pickle" || SKLEARN_PICKLE_FILE.test(declaredFile)) {
1878
+ return sklearnPickle(model, declaredFile);
1838
1879
  }
1880
+ return skopsFormat(model, declaredFile);
1839
1881
  }
1840
- else {
1841
- return skopsJobLib(model);
1842
- }
1882
+ // repos pushed without skops often still declare their filename in config.json; it is only trusted when it names a
1883
+ // pickle, since some hand-written configs carry typos (e.g. "sklean_model.jpblib" next to a real sklearn_model.joblib)
1884
+ return sklearnPickle(model, declaredFile && SKLEARN_PICKLE_FILE.test(declaredFile) ? declaredFile : "sklearn_model.joblib");
1843
1885
  };
1844
1886
  exports.sklearn = sklearn;
1845
1887
  const stable_audio_tools = (model) => [
@@ -1890,7 +1932,7 @@ const sam2 = (model) => {
1890
1932
  import torch
1891
1933
  from sam2.sam2_image_predictor import SAM2ImagePredictor
1892
1934
 
1893
- predictor = SAM2ImagePredictor.from_pretrained(${model.id})
1935
+ predictor = SAM2ImagePredictor.from_pretrained("${model.id}")
1894
1936
 
1895
1937
  with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
1896
1938
  predictor.set_image(<your_image>)
@@ -1899,13 +1941,13 @@ with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
1899
1941
  import torch
1900
1942
  from sam2.sam2_video_predictor import SAM2VideoPredictor
1901
1943
 
1902
- predictor = SAM2VideoPredictor.from_pretrained(${model.id})
1944
+ predictor = SAM2VideoPredictor.from_pretrained("${model.id}")
1903
1945
 
1904
1946
  with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
1905
1947
  state = predictor.init_state(<your_video>)
1906
1948
 
1907
1949
  # add new prompts and instantly get the output on the same frame
1908
- frame_idx, object_ids, masks = predictor.add_new_points(state, <your_prompts>):
1950
+ frame_idx, object_ids, masks = predictor.add_new_points(state, <your_prompts>)
1909
1951
 
1910
1952
  # propagate the prompts to get masklets throughout the video
1911
1953
  for frame_idx, object_ids, masks in predictor.propagate_in_video(state):
@@ -1927,11 +1969,18 @@ output = inference(image, mask)`,
1927
1969
  ];
1928
1970
  exports.sam_3d_objects = sam_3d_objects;
1929
1971
  const sam_3d_body = (model) => [
1930
- `from notebook.utils import setup_sam_3d_body
1972
+ `import cv2
1973
+ import numpy as np
1974
+ from notebook.utils import setup_sam_3d_body
1975
+ from tools.vis_utils import visualize_sample_together
1931
1976
 
1932
- estimator = setup_sam_3d_body(${model.id})
1933
- outputs = estimator.process_one_image(image)
1934
- rend_img = visualize_sample_together(image, outputs, estimator.faces)`,
1977
+ estimator = setup_sam_3d_body("${model.id}")
1978
+
1979
+ img_bgr = cv2.imread("path/to/image.jpg")
1980
+ outputs = estimator.process_one_image(cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB))
1981
+
1982
+ rend_img = visualize_sample_together(img_bgr, outputs, estimator.faces)
1983
+ cv2.imwrite("output.jpg", rend_img.astype(np.uint8))`,
1935
1984
  ];
1936
1985
  exports.sam_3d_body = sam_3d_body;
1937
1986
  const sampleFactory = (model) => [
@@ -2025,7 +2074,15 @@ queries_emb = model.encode(queries, is_query=True)
2025
2074
  docs_emb = model.encode(documents, is_query=False)`,
2026
2075
  ];
2027
2076
  }
2028
- if (model.tags.includes("cross-encoder") || model.pipeline_tag == "text-ranking") {
2077
+ // Rerankers saved in the CrossEncoder format are often tagged text-classification rather than text-ranking;
2078
+ // their architecture still identifies them (e.g. BAAI/bge-reranker-v2-m3). SetFit models can share that shape but
2079
+ // are bi-encoders with a separate head.
2080
+ const isCrossEncoder = model.tags.includes("cross-encoder") ||
2081
+ model.pipeline_tag == "text-ranking" ||
2082
+ (model.pipeline_tag === "text-classification" &&
2083
+ !model.tags.includes("setfit") &&
2084
+ (model.config?.architectures ?? []).some((arch) => arch.endsWith("ForSequenceClassification")));
2085
+ if (isCrossEncoder) {
2029
2086
  return [
2030
2087
  `from sentence_transformers import CrossEncoder
2031
2088
 
@@ -2043,6 +2100,33 @@ scores = model.predict([(query, passage) for passage in passages])
2043
2100
  print(scores)`,
2044
2101
  ];
2045
2102
  }
2103
+ // sentence-transformers tags SparseEncoder repos (SPLADE & co) with "sparse-encoder" and MultiVectorEncoder repos
2104
+ // (ColBERT & co) with "multi-vector" when pushing them; both encode queries and documents separately
2105
+ const retrievalEncoder = model.tags.includes("sparse-encoder")
2106
+ ? "SparseEncoder"
2107
+ : model.tags.includes("multi-vector")
2108
+ ? "MultiVectorEncoder"
2109
+ : undefined;
2110
+ if (retrievalEncoder) {
2111
+ return [
2112
+ `from sentence_transformers import ${retrievalEncoder}
2113
+
2114
+ model = ${retrievalEncoder}("${model.id}"${remote_code_snippet})
2115
+
2116
+ queries = ["Which planet is known as the Red Planet?"]
2117
+ documents = [
2118
+ "Venus is often called Earth's twin because of its similar size and proximity.",
2119
+ "Mars, known for its reddish appearance, is often referred to as the Red Planet.",
2120
+ "Jupiter, the largest planet in our solar system, has a prominent red spot.",
2121
+ ]
2122
+
2123
+ query_embeddings = model.encode_query(queries)
2124
+ document_embeddings = model.encode_document(documents)
2125
+
2126
+ similarities = model.similarity(query_embeddings, document_embeddings)
2127
+ print(similarities)`,
2128
+ ];
2129
+ }
2046
2130
  const exampleSentences = get_widget_examples_from_st_model(model) ?? [
2047
2131
  "The weather is lovely today.",
2048
2132
  "It's so sunny outside!",
@@ -2065,7 +2149,10 @@ exports.sentenceTransformers = sentenceTransformers;
2065
2149
  const setfit = (model) => [
2066
2150
  `from setfit import SetFitModel
2067
2151
 
2068
- model = SetFitModel.from_pretrained("${model.id}")`,
2152
+ model = SetFitModel.from_pretrained("${model.id}")
2153
+
2154
+ preds = model.predict(["i loved the spiderman movie!", "pineapple on pizza is the worst"])
2155
+ print(preds)`,
2069
2156
  ];
2070
2157
  exports.setfit = setfit;
2071
2158
  const spacy = (model) => [
@@ -2083,7 +2170,10 @@ exports.spacy = spacy;
2083
2170
  const span_marker = (model) => [
2084
2171
  `from span_marker import SpanMarkerModel
2085
2172
 
2086
- model = SpanMarkerModel.from_pretrained("${model.id}")`,
2173
+ model = SpanMarkerModel.from_pretrained("${model.id}")
2174
+
2175
+ entities = model.predict("Amelia Earhart flew her single engine Lockheed Vega 5B across the Atlantic to Paris.")
2176
+ print(entities)`,
2087
2177
  ];
2088
2178
  exports.span_marker = span_marker;
2089
2179
  const stanza = (model) => [
@@ -2104,6 +2194,8 @@ const speechBrainMethod = (speechbrainInterface) => {
2104
2194
  return "enhance_file";
2105
2195
  case "SepformerSeparation":
2106
2196
  return "separate_file";
2197
+ case "SpeakerRecognition":
2198
+ return "verify_files";
2107
2199
  default:
2108
2200
  return undefined;
2109
2201
  }
@@ -2117,12 +2209,14 @@ const speechbrain = (model) => {
2117
2209
  if (speechbrainMethod === undefined) {
2118
2210
  return [`# interface in config.json invalid`];
2119
2211
  }
2212
+ // speaker verification compares two recordings
2213
+ const methodArgs = speechbrainMethod === "verify_files" ? '"speaker1.wav", "speaker2.wav"' : '"file.wav"';
2120
2214
  return [
2121
2215
  `from speechbrain.pretrained import ${speechbrainInterface}
2122
2216
  model = ${speechbrainInterface}.from_hparams(
2123
2217
  "${model.id}"
2124
2218
  )
2125
- model.${speechbrainMethod}("file.wav")`,
2219
+ model.${speechbrainMethod}(${methodArgs})`,
2126
2220
  ];
2127
2221
  };
2128
2222
  exports.speechbrain = speechbrain;
@@ -2154,7 +2248,7 @@ const transformers = (model) => {
2154
2248
  : info.processor === "AutoFeatureExtractor"
2155
2249
  ? "extractor"
2156
2250
  : "processor";
2157
- 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")');
2251
+ 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")');
2158
2252
  if (model.tags.includes("conversational") && hasChatTemplate(model)) {
2159
2253
  if (model.tags.includes("image-text-to-text")) {
2160
2254
  autoSnippet.push("messages = [", [
@@ -2170,19 +2264,19 @@ const transformers = (model) => {
2170
2264
  else {
2171
2265
  autoSnippet.push("messages = [", ' {"role": "user", "content": "Who are you?"},', "]");
2172
2266
  }
2173
- 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]:]))`);
2267
+ 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]:]))`);
2174
2268
  }
2175
2269
  }
2176
2270
  else {
2177
- autoSnippet.push("# Load model directly", `from transformers import ${auto_model}`, `model = ${auto_model}.from_pretrained("${model.id}"` + remote_code_snippet + ', device_map="auto")');
2271
+ 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")');
2178
2272
  }
2179
2273
  if (model.pipeline_tag && library_to_tasks_js_1.LIBRARY_TASK_MAPPING.transformers?.includes(model.pipeline_tag)) {
2180
2274
  const pipelineSnippet = ["# Use a pipeline as a high-level helper"];
2181
2275
  if (library_to_tasks_js_1.REMOVED_IN_V5_TRANSFORMERS_PIPELINES.includes(model.pipeline_tag)) {
2182
- pipelineSnippet.push(`# Warning: Pipeline type "${model.pipeline_tag}" is no longer supported in transformers v5.`, `# You must load the model directly (see below) or downgrade to v4.x with:`, `# 'pip install "transformers<5.0.0'`);
2276
+ pipelineSnippet.push(`# Warning: Pipeline type "${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"`);
2183
2277
  }
2184
2278
  pipelineSnippet.push("from transformers import pipeline", "", `pipe = pipeline("${model.pipeline_tag}", model="${model.id}"` + remote_code_snippet + ")");
2185
- if (model.tags.includes("conversational")) {
2279
+ if (model.tags.includes("conversational") && hasChatTemplate(model)) {
2186
2280
  if (model.tags.includes("image-text-to-text")) {
2187
2281
  pipelineSnippet.push("messages = [", [
2188
2282
  " {",
@@ -2211,48 +2305,119 @@ const transformers = (model) => {
2211
2305
  return [autoSnippet.join("\n")];
2212
2306
  };
2213
2307
  exports.transformers = transformers;
2308
+ // Hub tasks that transformers.js exposes under a different pipeline() task name
2309
+ const TRANSFORMERS_JS_TASK_ALIASES = {
2310
+ "sentence-similarity": "feature-extraction",
2311
+ };
2214
2312
  const transformersJS = (model) => {
2215
2313
  if (!model.pipeline_tag) {
2216
2314
  return [`// ⚠️ Unknown pipeline tag`];
2217
2315
  }
2218
2316
  const libName = "@huggingface/transformers";
2317
+ // Rerankers score (query, document) pairs. The text-classification pipeline cannot take pairs and applies a softmax
2318
+ // over their single logit, so every score comes out as 1; the model is called directly instead.
2319
+ if (model.pipeline_tag === "text-ranking") {
2320
+ return [
2321
+ `// npm i ${libName}
2322
+ import { AutoTokenizer, AutoModelForSequenceClassification } from '${libName}';
2323
+
2324
+ const tokenizer = await AutoTokenizer.from_pretrained('${model.id}');
2325
+ const model = await AutoModelForSequenceClassification.from_pretrained('${model.id}');
2326
+
2327
+ const query = 'Which planet is known as the Red Planet?';
2328
+ const documents = [
2329
+ 'Mars, known for its reddish appearance, is often referred to as the Red Planet.',
2330
+ 'Venus is often called the twin of Earth because of its similar size and proximity.',
2331
+ ];
2332
+
2333
+ const inputs = tokenizer(new Array(documents.length).fill(query), { text_pair: documents, padding: true, truncation: true });
2334
+ const { logits } = await model(inputs);
2335
+ console.log(logits.sigmoid().tolist()); // one relevance score per document`,
2336
+ ];
2337
+ }
2338
+ const task = TRANSFORMERS_JS_TASK_ALIASES[model.pipeline_tag] ?? model.pipeline_tag;
2219
2339
  return [
2220
2340
  `// npm i ${libName}
2221
2341
  import { pipeline } from '${libName}';
2222
2342
 
2223
2343
  // Allocate pipeline
2224
- const pipe = await pipeline('${model.pipeline_tag}', '${model.id}');`,
2344
+ const pipe = await pipeline('${task}', '${model.id}');`,
2225
2345
  ];
2226
2346
  };
2227
2347
  exports.transformersJS = transformersJS;
2228
- const peftTask = (peftTaskType) => {
2229
- switch (peftTaskType) {
2230
- case "CAUSAL_LM":
2231
- return "CausalLM";
2232
- case "SEQ_2_SEQ_LM":
2233
- return "Seq2SeqLM";
2234
- case "TOKEN_CLS":
2235
- return "TokenClassification";
2236
- case "SEQ_CLS":
2237
- return "SequenceClassification";
2238
- default:
2239
- return undefined;
2240
- }
2348
+ const PEFT_TASK_TYPE_TO_AUTO_CLASS = {
2349
+ CAUSAL_LM: "AutoModelForCausalLM",
2350
+ SEQ_2_SEQ_LM: "AutoModelForSeq2SeqLM",
2351
+ TOKEN_CLS: "AutoModelForTokenClassification",
2352
+ SEQ_CLS: "AutoModelForSequenceClassification",
2353
+ QUESTION_ANS: "AutoModelForQuestionAnswering",
2354
+ // transformers has no AutoModelForFeatureExtraction; PEFT's own AutoPeftModelForFeatureExtraction resolves to AutoModel
2355
+ FEATURE_EXTRACTION: "AutoModel",
2356
+ };
2357
+ // PEFT has no speech-specific task type: Whisper-style adapters are tagged SEQ_2_SEQ_LM, but transformers registers
2358
+ // Whisper and Speech2Text only under AutoModelForSpeechSeq2Seq (SeamlessM4T is under both seq2seq auto classes, SpeechT5
2359
+ // under the speech-to-text and text-to-spectrogram ones).
2360
+ // Other audio seq2seq models (Qwen2-Audio, GLM-ASR) stay under AutoModelForSeq2SeqLM, so the base model name is the
2361
+ // main signal. Matched against Hub repo ids, which are hyphenated (facebook/s2t-small-librispeech-asr,
2362
+ // facebook/seamless-m4t-v2-large), not against transformers model_type identifiers.
2363
+ const PEFT_SPEECH_TO_TEXT_BASE_MODEL = /whisper|\bs2t\b|speech[-_]to[-_]text/i;
2364
+ // SeamlessM4T also has a text-to-text head under AutoModelForSeq2SeqLM, which text translation adapters need; most
2365
+ // SeamlessM4T adapters on the Hub are speech ones, so the speech class stays the default when no text task is declared.
2366
+ const PEFT_SEAMLESS_M4T_BASE_MODEL = /seamless[-_]m4t/i;
2367
+ const PEFT_TEXT_TASKS = ["translation", "text2text-generation", "summarization"];
2368
+ // SpeechT5 checkpoints are overwhelmingly text-to-speech fine-tunes (often without "tts" in their name), which load
2369
+ // with AutoModelForTextToSpectrogram; AutoModelForSpeechSeq2Seq would silently load the speech-to-text head, so only
2370
+ // ASR bases or adapters get it.
2371
+ const PEFT_SPEECHT5_BASE_MODEL = /speecht5/i;
2372
+ const declaresTask = (model, task) => model.pipeline_tag === task || model.tags.includes(task);
2373
+ // the auto class a SEQ_2_SEQ_LM adapter needs when its base is a speech model, undefined for text seq2seq bases
2374
+ const peftSpeechAutoClass = (model, baseModel) => {
2375
+ if (PEFT_SPEECH_TO_TEXT_BASE_MODEL.test(baseModel)) {
2376
+ return "AutoModelForSpeechSeq2Seq";
2377
+ }
2378
+ if (PEFT_SEAMLESS_M4T_BASE_MODEL.test(baseModel)) {
2379
+ return PEFT_TEXT_TASKS.some((task) => declaresTask(model, task)) ? undefined : "AutoModelForSpeechSeq2Seq";
2380
+ }
2381
+ if (PEFT_SPEECHT5_BASE_MODEL.test(baseModel)) {
2382
+ const isAsr = /(?:^|[^a-z])asr(?:[^a-z]|$)/i.test(baseModel) || declaresTask(model, "automatic-speech-recognition");
2383
+ return isAsr ? "AutoModelForSpeechSeq2Seq" : "AutoModelForTextToSpectrogram";
2384
+ }
2385
+ return undefined;
2241
2386
  };
2242
2387
  const peft = (model) => {
2243
2388
  const { base_model_name_or_path: peftBaseModel, task_type: peftTaskType } = model.config?.peft ?? {};
2244
- const pefttask = peftTask(peftTaskType);
2245
- if (!pefttask) {
2389
+ let autoClass = peftTaskType ? PEFT_TASK_TYPE_TO_AUTO_CLASS[peftTaskType] : undefined;
2390
+ if (!autoClass) {
2246
2391
  return [`Task type is invalid.`];
2247
2392
  }
2248
- if (!peftBaseModel) {
2393
+ // adapter_config.json sometimes omits the base model; the model card's base_model is the next best source
2394
+ const cardBaseModel = Array.isArray(model.cardData?.base_model)
2395
+ ? model.cardData.base_model[0]
2396
+ : model.cardData?.base_model;
2397
+ const baseModel = peftBaseModel || cardBaseModel;
2398
+ if (!baseModel) {
2249
2399
  return [`Base model is not found.`];
2250
2400
  }
2401
+ if (peftTaskType === "SEQ_2_SEQ_LM") {
2402
+ autoClass = peftSpeechAutoClass(model, baseModel) ?? autoClass;
2403
+ }
2404
+ if (autoClass === "AutoModelForTextToSpectrogram") {
2405
+ return [
2406
+ `from peft import PeftConfig, PeftModel
2407
+ from transformers import AutoModelForTextToSpectrogram
2408
+
2409
+ base_model = AutoModelForTextToSpectrogram.from_pretrained("${escapeStringForJson(baseModel)}")
2410
+ # PEFT's seq2seq wrapper needs generate(), which SpeechT5's text-to-speech model lacks: load the adapter without its task type
2411
+ config = PeftConfig.from_pretrained("${model.id}")
2412
+ config.task_type = None
2413
+ model = PeftModel.from_pretrained(base_model, "${model.id}", config=config)`,
2414
+ ];
2415
+ }
2251
2416
  return [
2252
2417
  `from peft import PeftModel
2253
- from transformers import AutoModelFor${pefttask}
2418
+ from transformers import ${autoClass}
2254
2419
 
2255
- base_model = AutoModelFor${pefttask}.from_pretrained("${escapeStringForJson(peftBaseModel)}")
2420
+ base_model = ${autoClass}.from_pretrained("${escapeStringForJson(baseModel)}")
2256
2421
  model = PeftModel.from_pretrained(base_model, "${model.id}")`,
2257
2422
  ];
2258
2423
  };
@@ -2286,7 +2451,7 @@ transcriptions = asr_model.transcribe(["file.wav"])`,
2286
2451
  }
2287
2452
  };
2288
2453
  const mlAgents = (model) => [
2289
- `mlagents-load-from-hf --repo-id="${model.id}" --local-dir="./download: string[]s"`,
2454
+ `mlagents-load-from-hf --repo-id="${model.id}" --local-dir="./downloads"`,
2290
2455
  ];
2291
2456
  exports.mlAgents = mlAgents;
2292
2457
  const sentis = ( /* model: ModelData */) => [
@@ -2361,10 +2526,10 @@ model = Model.from_pretrained("${model.id}")`,
2361
2526
  exports.vfimamba = vfimamba;
2362
2527
  const lvface = (model) => [
2363
2528
  `from huggingface_hub import hf_hub_download
2364
- from inference_onnx import LVFaceONNXInferencer
2529
+ from inference_onnx import LVFaceONNXInferencer
2365
2530
 
2366
2531
  model_path = hf_hub_download("${model.id}", "LVFace-L_Glint360K/LVFace-L_Glint360K.onnx")
2367
- inferencer = LVFaceONNXInferencer(model_path, use_gpu=True, timeout=300)
2532
+ inferencer = LVFaceONNXInferencer(model_path, use_gpu=True)
2368
2533
  img_path = 'path/to/image1.jpg'
2369
2534
  embedding = inferencer.infer_from_image(img_path)`,
2370
2535
  ];
@@ -2403,8 +2568,7 @@ const vui = () => [
2403
2568
 
2404
2569
  import torchaudio
2405
2570
 
2406
- from vui.inference import render
2407
- from vui.model import Vui,
2571
+ from vui.legacy import Vui, render
2408
2572
 
2409
2573
  model = Vui.from_pretrained().cuda()
2410
2574
  waveform = render(
@@ -2412,16 +2576,17 @@ waveform = render(
2412
2576
  "Hey, here is some random stuff, usually something quite long as the shorter the text the less likely the model can cope!",
2413
2577
  )
2414
2578
  print(waveform.shape)
2415
- torchaudio.save("out.opus", waveform[0], 22050)
2579
+ torchaudio.save("out.wav", waveform[0], 22050)
2416
2580
  `,
2417
2581
  ];
2418
2582
  exports.vui = vui;
2419
2583
  const chattts = () => [
2420
2584
  `import ChatTTS
2585
+ import torch
2421
2586
  import torchaudio
2422
2587
 
2423
2588
  chat = ChatTTS.Chat()
2424
- chat.load_models(compile=False) # Set to True for better performance
2589
+ chat.load(compile=False) # Set to True for better performance
2425
2590
 
2426
2591
  texts = ["PUT YOUR TEXT HERE",]
2427
2592
 
@@ -2431,17 +2596,42 @@ torchaudio.save("output1.wav", torch.from_numpy(wavs[0]), 24000)`,
2431
2596
  ];
2432
2597
  exports.chattts = chattts;
2433
2598
  const ultralytics = (model) => {
2434
- // ultralytics models must have a version tag (e.g. `yolov8`)
2435
- const versionTag = model.tags.find((tag) => tag.match(/^yolov\d+$/));
2436
- const className = versionTag ? `YOLOv${versionTag.slice(4)}` : "YOLOvXX";
2437
- const prefix = versionTag
2438
- ? ""
2439
- : `# Couldn't find a valid YOLO version tag.\n# Replace XX with the correct version.\n`;
2599
+ // the YOLOv10 fork (library "yolov10") ships its own `YOLOv10` class with a `from_pretrained` helper that reads the
2600
+ // model.safetensors it pushes. Mainline ultralytics repos often list "yolov10" among their version tags
2601
+ // (e.g. Ultralytics/YOLOv8) but only ship .pt files, so an ultralytics tag alone keeps them on the mainline loader;
2602
+ // fork pushes still tagged ultralytics are recognized by their safetensors weights.
2603
+ const isYolov10Fork = model.library_name === "yolov10" ||
2604
+ (model.tags.includes("yolov10") && (!model.tags.includes("ultralytics") || model.tags.includes("safetensors")));
2605
+ if (isYolov10Fork) {
2606
+ // from_pretrained reads the model.safetensors that fork pushes ship; older fork repos only ship .pt weights
2607
+ if (model.tags.includes("safetensors")) {
2608
+ return [
2609
+ `from ultralytics import YOLOv10
2610
+
2611
+ model = YOLOv10.from_pretrained("${model.id}")
2612
+ source = 'http://images.cocodataset.org/val2017/000000039769.jpg'
2613
+ model.predict(source=source, save=True)`,
2614
+ ];
2615
+ }
2616
+ return [
2617
+ `from huggingface_hub import hf_hub_download
2618
+ from ultralytics import YOLOv10
2619
+
2620
+ # pick the weights file from this repo's "Files and versions" tab
2621
+ weights = hf_hub_download("${model.id}", "<weights>.pt")
2622
+ model = YOLOv10(weights)
2623
+ source = 'http://images.cocodataset.org/val2017/000000039769.jpg'
2624
+ model.predict(source=source, save=True)`,
2625
+ ];
2626
+ }
2627
+ // mainline ultralytics loads weights from a local file
2440
2628
  return [
2441
- prefix +
2442
- `from ultralytics import ${className}
2629
+ `from huggingface_hub import hf_hub_download
2630
+ from ultralytics import YOLO
2443
2631
 
2444
- model = ${className}.from_pretrained("${model.id}")
2632
+ # pick the weights file from this repo's "Files and versions" tab
2633
+ weights = hf_hub_download("${model.id}", "<weights>.pt")
2634
+ model = YOLO(weights)
2445
2635
  source = 'http://images.cocodataset.org/val2017/000000039769.jpg'
2446
2636
  model.predict(source=source, save=True)`,
2447
2637
  ];
@@ -2519,14 +2709,16 @@ model = SwarmFormerModel.from_pretrained("${model.id}")
2519
2709
  exports.swarmformer = swarmformer;
2520
2710
  const univa = (model) => [
2521
2711
  `# Follow installation instructions at https://github.com/PKU-YuanGroup/UniWorld-V1
2522
-
2712
+ import torch
2713
+ from transformers import AutoProcessor
2523
2714
  from univa.models.qwen2p5vl.modeling_univa_qwen2p5vl import UnivaQwen2p5VLForConditionalGeneration
2524
- model = UnivaQwen2p5VLForConditionalGeneration.from_pretrained(
2525
- "${model.id}",
2526
- torch_dtype=torch.bfloat16,
2527
- attn_implementation="flash_attention_2",
2528
- ).to("cuda")
2529
- processor = AutoProcessor.from_pretrained("${model.id}")
2715
+
2716
+ model = UnivaQwen2p5VLForConditionalGeneration.from_pretrained(
2717
+ "${model.id}",
2718
+ torch_dtype=torch.bfloat16,
2719
+ attn_implementation="flash_attention_2",
2720
+ ).to("cuda")
2721
+ processor = AutoProcessor.from_pretrained("${model.id}")
2530
2722
  `,
2531
2723
  ];
2532
2724
  exports.univa = univa;
@@ -2534,7 +2726,7 @@ const mlx_unknown = (model) => [
2534
2726
  `# Download the model from the Hub
2535
2727
  pip install huggingface_hub[hf_xet]
2536
2728
 
2537
- huggingface-cli download --local-dir ${nameWithoutNamespace(model.id)} ${model.id}`,
2729
+ hf download ${model.id} --local-dir ${nameWithoutNamespace(model.id)}`,
2538
2730
  ];
2539
2731
  const mlxlm = (model) => [
2540
2732
  `# Make sure mlx-lm is installed
@@ -2591,10 +2783,11 @@ formatted_prompt = apply_chat_template(
2591
2783
  output = generate(model, processor, formatted_prompt, image)
2592
2784
  print(output)`,
2593
2785
  ];
2786
+ // create_model() takes a name from the mlx-image registry, not a repo id; the weights live in "mlx-vision/<name>-mlxim"
2594
2787
  const mlxim = (model) => [
2595
2788
  `from mlxim.model import create_model
2596
2789
 
2597
- model = create_model(${model.id})`,
2790
+ model = create_model("${nameWithoutNamespace(model.id).replace(/-mlxim$/, "")}")`,
2598
2791
  ];
2599
2792
  exports.mlxim = mlxim;
2600
2793
  const mlx = (model) => {
@@ -2615,7 +2808,10 @@ exports.mlx = mlx;
2615
2808
  const model2vec = (model) => [
2616
2809
  `from model2vec import StaticModel
2617
2810
 
2618
- model = StaticModel.from_pretrained("${model.id}")`,
2811
+ model = StaticModel.from_pretrained("${model.id}")
2812
+
2813
+ embeddings = model.encode(["It's dangerous to go alone!", "It's a secret to everybody."])
2814
+ print(embeddings.shape)`,
2619
2815
  ];
2620
2816
  exports.model2vec = model2vec;
2621
2817
  const mobilint = (model) => {
@@ -2660,7 +2856,8 @@ const pruna = (model) => {
2660
2856
  };
2661
2857
  snippets = snippets.map(ensurePrunaModelImport);
2662
2858
  if (model.tags.includes("pruna_pro-ai")) {
2663
- return snippets.map((snippet) => snippet.replace(/\bpruna\b/g, "pruna_pro").replace(/\bPrunaModel\b/g, "PrunaProModel"));
2859
+ // only rewrite the import and the class name, never the repo id (an org name can contain "pruna")
2860
+ return snippets.map((snippet) => snippet.replace(/^from pruna import /gm, "from pruna_pro import ").replace(/\bPrunaModel\b/g, "PrunaProModel"));
2664
2861
  }
2665
2862
  return snippets;
2666
2863
  };
@@ -2720,21 +2917,19 @@ const outetts = (model) => {
2720
2917
  }
2721
2918
  // v1.0 HF → minimal runnable snippet
2722
2919
  return [
2723
- `
2724
- import outetts
2725
-
2726
- enum = outetts.Models("${model.id}".split("/", 1)[1]) # VERSION_1_0_SIZE_1B
2727
- cfg = outetts.ModelConfig.auto_config(enum, outetts.Backend.HF)
2728
- tts = outetts.Interface(cfg)
2729
-
2730
- speaker = tts.load_default_speaker("EN-FEMALE-1-NEUTRAL")
2731
- tts.generate(
2732
- outetts.GenerationConfig(
2733
- text="Hello there, how are you doing?",
2734
- speaker=speaker,
2735
- )
2736
- ).save("output.wav")
2737
- `,
2920
+ `import outetts
2921
+
2922
+ enum = outetts.Models("${model.id}".split("/", 1)[1]) # VERSION_1_0_SIZE_1B
2923
+ cfg = outetts.ModelConfig.auto_config(enum, outetts.Backend.HF)
2924
+ tts = outetts.Interface(cfg)
2925
+
2926
+ speaker = tts.load_default_speaker("EN-FEMALE-1-NEUTRAL")
2927
+ tts.generate(
2928
+ outetts.GenerationConfig(
2929
+ text="Hello there, how are you doing?",
2930
+ speaker=speaker,
2931
+ )
2932
+ ).save("output.wav")`,
2738
2933
  ];
2739
2934
  };
2740
2935
  exports.outetts = outetts;