@huggingface/tasks 0.21.54 → 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.
@@ -22,7 +22,7 @@ export const adapters = (model: ModelData): string[] => [
22
22
  `from adapters import AutoAdapterModel
23
23
 
24
24
  model = AutoAdapterModel.from_pretrained("${escapeStringForJson(
25
- model.config?.adapter_transformers?.model_name ?? "fill-in-model-name",
25
+ model.config?.adapter_transformers?.model_name ?? "<base-model-id>",
26
26
  )}")
27
27
  model.load_adapter("${model.id}", set_active=True)`,
28
28
  ];
@@ -83,7 +83,13 @@ result, message = detector.detect_watermark(watermarked_audio, sr)`;
83
83
  };
84
84
 
85
85
  function get_base_diffusers_model(model: ModelData): string {
86
- return escapeStringForJson(model.cardData?.base_model?.toString() ?? "fill-in-base-model");
86
+ const baseModel = model.cardData?.base_model;
87
+ // several base models can be listed (e.g. a LoRA trained for a base and its turbo variant). When an original checkpoint
88
+ // and its Diffusers conversion are both listed (["Wan-AI/Wan2.1-I2V-14B-480P", "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"]),
89
+ // only the conversion has a model_index.json that DiffusionPipeline can load; otherwise the first one is used
90
+ const candidates = Array.isArray(baseModel) ? baseModel : baseModel ? [baseModel] : [];
91
+ const chosen = candidates.find((id) => id.toLowerCase().endsWith("-diffusers")) ?? candidates[0];
92
+ return escapeStringForJson(chosen ?? "fill-in-base-model");
87
93
  }
88
94
 
89
95
  function get_prompt_from_diffusers_model(model: ModelData): string | undefined {
@@ -174,7 +180,7 @@ detector = cvg.NeuralCornerDetector(checkpoint)
174
180
  embedder = cvg.NeuralEmbedder(checkpoint)`,
175
181
  ];
176
182
 
177
- export const colipri = (model: ModelData): string[] => {
183
+ export const colipri = (): string[] => {
178
184
  const installSnippet = `pip install colipri`;
179
185
 
180
186
  const exampleSnippet = `from colipri import get_model
@@ -184,7 +190,7 @@ from colipri import ZeroShotImageClassificationPipeline
184
190
 
185
191
  model = get_model().cuda()
186
192
  processor = get_processor()
187
- pipeline = ZeroShotImageClassificationPipeline("${model.id}", processor)
193
+ pipeline = ZeroShotImageClassificationPipeline(model, processor)
188
194
 
189
195
  image = load_sample_ct()
190
196
 
@@ -308,7 +314,7 @@ export const depth_anything_v2 = (model: ModelData): string[] => {
308
314
  } else if (model.id === "depth-anything/Depth-Anything-V2-Large") {
309
315
  encoder = "vitl";
310
316
  features = "256";
311
- out_channels = "[256, 512, 1024, 1024";
317
+ out_channels = "[256, 512, 1024, 1024]";
312
318
  }
313
319
 
314
320
  return [
@@ -318,6 +324,7 @@ export const depth_anything_v2 = (model: ModelData): string[] => {
318
324
  # Load the model and infer depth from an image
319
325
  import cv2
320
326
  import torch
327
+ from huggingface_hub import hf_hub_download
321
328
 
322
329
  from depth_anything_v2.dpt import DepthAnythingV2
323
330
 
@@ -327,7 +334,8 @@ model = DepthAnythingV2(encoder="${encoder}", features=${features}, out_channels
327
334
  # load the weights
328
335
  filepath = hf_hub_download(repo_id="${model.id}", filename="depth_anything_v2_${encoder}.pth", repo_type="model")
329
336
  state_dict = torch.load(filepath, map_location="cpu")
330
- model.load_state_dict(state_dict).eval()
337
+ model.load_state_dict(state_dict)
338
+ model.eval()
331
339
 
332
340
  raw_img = cv2.imread("your/image/path")
333
341
  depth = model.infer_image(raw_img) # HxW raw depth map in numpy
@@ -337,8 +345,8 @@ depth = model.infer_image(raw_img) # HxW raw depth map in numpy
337
345
 
338
346
  export const depth_pro = (model: ModelData): string[] => {
339
347
  const installSnippet = `# Download checkpoint
340
- pip install huggingface-hub
341
- huggingface-cli download --local-dir checkpoints ${model.id}`;
348
+ pip install huggingface_hub
349
+ hf download ${model.id} --local-dir checkpoints`;
342
350
 
343
351
  const inferenceSnippet = `import depth_pro
344
352
 
@@ -412,7 +420,7 @@ export const describe_anything = (model: ModelData): string[] => [
412
420
  from huggingface_hub import snapshot_download
413
421
  from dam import DescribeAnythingModel
414
422
 
415
- snapshot_download(${model.id}, local_dir="checkpoints")
423
+ snapshot_download("${model.id}", local_dir="checkpoints")
416
424
 
417
425
  dam = DescribeAnythingModel(
418
426
  model_path="checkpoints",
@@ -535,7 +543,7 @@ pipe.load_lora_weights("${model.id}")
535
543
  prompt = "${get_prompt_from_diffusers_model(model) ?? diffusersVideoDefaultPrompt}"
536
544
  input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/guitar-man.png")
537
545
 
538
- image = pipe(image=input_image, prompt=prompt).frames[0]
546
+ output = pipe(image=input_image, prompt=prompt).frames[0]
539
547
  export_to_video(output, "output.mp4")`,
540
548
  ];
541
549
 
@@ -578,7 +586,7 @@ from diffusers import AutoPipelineForInpainting
578
586
  from diffusers.utils import load_image
579
587
 
580
588
  # switch to "mps" for apple devices
581
- pipe = AutoPipelineForInpainting.from_pretrained("${model.id}", dtype=torch.float16, variant="fp16", device_map="cuda")
589
+ pipe = AutoPipelineForInpainting.from_pretrained("${model.id}", dtype=torch.float16, device_map="cuda")
582
590
 
583
591
  img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
584
592
  mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
@@ -601,11 +609,14 @@ image = pipe(
601
609
  ];
602
610
 
603
611
  export const diffusers = (model: ModelData): string[] => {
612
+ // the Hub emits pipeline class tags as `diffusers:<ClassName>`; bare names are kept for manually added tags
613
+ const hasPipelineClass = (className: string): boolean =>
614
+ model.tags.includes(`diffusers:${className}`) ||
615
+ model.tags.includes(className) ||
616
+ model.config?.diffusers?._class_name === className;
617
+
604
618
  let codeSnippets: string[];
605
- if (
606
- model.tags.includes("StableDiffusionInpaintPipeline") ||
607
- model.tags.includes("StableDiffusionXLInpaintPipeline")
608
- ) {
619
+ if (hasPipelineClass("StableDiffusionInpaintPipeline") || hasPipelineClass("StableDiffusionXLInpaintPipeline")) {
609
620
  codeSnippets = diffusers_inpainting(model);
610
621
  } else if (model.tags.includes("controlnet")) {
611
622
  codeSnippets = diffusers_controlnet(model);
@@ -621,7 +632,7 @@ export const diffusers = (model: ModelData): string[] => {
621
632
  }
622
633
  } else if (model.tags.includes("textual_inversion")) {
623
634
  codeSnippets = diffusers_textual_inversion(model);
624
- } else if (model.tags.includes("FluxFillPipeline")) {
635
+ } else if (hasPipelineClass("FluxFillPipeline")) {
625
636
  codeSnippets = diffusers_flux_fill(model);
626
637
  } else if (model.pipeline_tag === "image-to-video") {
627
638
  codeSnippets = diffusers_image_to_video(model);
@@ -641,7 +652,7 @@ from diffusionkit.mlx import DiffusionPipeline
641
652
  pipeline = DiffusionPipeline(
642
653
  shift=3.0,
643
654
  use_t5=False,
644
- model_version=${model.id},
655
+ model_version="${model.id}",
645
656
  low_memory_mode=True,
646
657
  a16=True,
647
658
  w16=True,
@@ -652,7 +663,7 @@ from diffusionkit.mlx import FluxPipeline
652
663
 
653
664
  pipeline = FluxPipeline(
654
665
  shift=1.0,
655
- model_version=${model.id},
666
+ model_version="${model.id}",
656
667
  low_memory_mode=True,
657
668
  a16=True,
658
669
  w16=True,
@@ -677,21 +688,22 @@ image, _ = pipeline.generate_image(
677
688
  };
678
689
 
679
690
  export const cartesia_pytorch = (model: ModelData): string[] => [
680
- `# pip install --no-binary :all: cartesia-pytorch
681
- from cartesia_pytorch import ReneLMHeadModel
691
+ `# The cartesia-pytorch package on PyPI ships no code: clone the repo and put its cartesia-pytorch folder on PYTHONPATH
692
+ # (it needs an Ampere or newer CUDA GPU, torch 2.4, mamba-ssm 2.2.2 and flash-attn 2.6.3)
693
+ # git clone https://github.com/cartesia-ai/edge && export PYTHONPATH="$PWD/edge/cartesia-pytorch:$PYTHONPATH"
694
+ from cartesia_pytorch.Rene import ReneLMHeadModel
682
695
  from transformers import AutoTokenizer
683
696
 
684
- model = ReneLMHeadModel.from_pretrained("${model.id}")
697
+ model = ReneLMHeadModel.from_pretrained("${model.id}").half().cuda()
685
698
  tokenizer = AutoTokenizer.from_pretrained("allenai/OLMo-1B-hf")
686
699
 
687
700
  in_message = ["Rene Descartes was"]
688
701
  inputs = tokenizer(in_message, return_tensors="pt")
689
702
 
690
- outputs = model.generate(inputs.input_ids, max_length=50, top_k=100, top_p=0.99)
703
+ outputs = model.generate(inputs.input_ids.cuda(), max_length=50, top_k=100, top_p=0.99)
691
704
  out_message = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
692
705
 
693
- print(out_message)
694
- )`,
706
+ print(out_message)`,
695
707
  ];
696
708
 
697
709
  export const cartesia_mlx = (model: ModelData): string[] => [
@@ -742,7 +754,8 @@ speech, *_ = model("text to generate speech from")`,
742
754
  ];
743
755
 
744
756
  export const espnetASR = (model: ModelData): string[] => [
745
- `from espnet2.bin.asr_inference import Speech2Text
757
+ `import soundfile
758
+ from espnet2.bin.asr_inference import Speech2Text
746
759
 
747
760
  model = Speech2Text.from_pretrained(
748
761
  "${model.id}"
@@ -772,9 +785,14 @@ models, cfg, task = load_model_ensemble_and_task_from_hf_hub(
772
785
  ];
773
786
 
774
787
  export const flair = (model: ModelData): string[] => [
775
- `from flair.models import SequenceTagger
788
+ `from flair.data import Sentence
789
+ from flair.models import SequenceTagger
776
790
 
777
- tagger = SequenceTagger.load("${model.id}")`,
791
+ tagger = SequenceTagger.load("${model.id}")
792
+
793
+ sentence = Sentence("George Washington went to Washington.")
794
+ tagger.predict(sentence)
795
+ print(sentence)`,
778
796
  ];
779
797
 
780
798
  export const flexray = (model: ModelData): string[] => [
@@ -866,13 +884,20 @@ pipeline = GaussianFormerRenderingPipeline.from_pretrained("${model.id}")`,
866
884
  export const gliner = (model: ModelData): string[] => [
867
885
  `from gliner import GLiNER
868
886
 
869
- model = GLiNER.from_pretrained("${model.id}")`,
887
+ model = GLiNER.from_pretrained("${model.id}")
888
+
889
+ text = "Cristiano Ronaldo dos Santos Aveiro was born on 5 February 1985 in Funchal, Madeira, Portugal."
890
+ labels = ["person", "date", "location"]
891
+
892
+ entities = model.predict_entities(text, labels)
893
+ for entity in entities:
894
+ print(entity["text"], "=>", entity["label"])`,
870
895
  ];
871
896
 
872
897
  export const gliner2 = (model: ModelData): string[] => [
873
- `from gliner2 import GLiNER2
898
+ `from gliner2 import AutoExtractor
874
899
 
875
- model = GLiNER2.from_pretrained("${model.id}")
900
+ extractor = AutoExtractor.from_pretrained("${model.id}")
876
901
 
877
902
  # Extract entities
878
903
  text = "Apple CEO Tim Cook announced iPhone 15 in Cupertino yesterday."
@@ -885,7 +910,7 @@ export const indextts = (model: ModelData): string[] => [
885
910
  `# Download model
886
911
  from huggingface_hub import snapshot_download
887
912
 
888
- snapshot_download(${model.id}, local_dir="checkpoints")
913
+ snapshot_download("${model.id}", local_dir="checkpoints")
889
914
 
890
915
  from indextts.infer import IndexTTS
891
916
 
@@ -941,9 +966,11 @@ print(out["answers"]["team"]["choice"])`,
941
966
  };
942
967
 
943
968
  export const keras = (model: ModelData): string[] => [
944
- `# Available backend options are: "jax", "torch", "tensorflow".
969
+ `# !pip install -U keras tensorflow huggingface_hub
970
+ # Keras needs TensorFlow installed to read "hf://" paths, so the tensorflow backend is selected here;
971
+ # "jax" and "torch" also work for computation once TensorFlow is installed.
945
972
  import os
946
- os.environ["KERAS_BACKEND"] = "jax"
973
+ os.environ["KERAS_BACKEND"] = "tensorflow"
947
974
 
948
975
  import keras
949
976
 
@@ -1070,13 +1097,15 @@ export const kernels = (model: ModelData): string[] => [
1070
1097
 
1071
1098
  from kernels import get_kernel
1072
1099
 
1073
- kernel = get_kernel("${model.id}")`,
1100
+ # a version (or an explicit revision) is required; see the "Files and versions" tab for the available ones
1101
+ kernel = get_kernel("${model.id}", version=1)`,
1074
1102
  ];
1075
1103
 
1076
1104
  export const kimi_audio = (model: ModelData): string[] => [
1077
1105
  `# Example usage for KimiAudio
1078
- # pip install git+https://github.com/MoonshotAI/Kimi-Audio.git
1106
+ # pip install git+https://github.com/MoonshotAI/Kimi-Audio.git "transformers<5"
1079
1107
 
1108
+ import soundfile as sf
1080
1109
  from kimia_infer.api.kimia import KimiAudio
1081
1110
 
1082
1111
  model = KimiAudio(model_path="${model.id}", load_detokenizer=True)
@@ -1517,14 +1546,17 @@ pip install -U litert-lm
1517
1546
 
1518
1547
  # 2. Download and run this model locally:
1519
1548
  # See: https://ai.google.dev/edge/litert-lm/cli
1549
+ # A single .litertlm file in the repo is picked automatically; otherwise the CLI asks which one to run
1550
+ # (or pass its name right after the repo id).
1520
1551
  litert-lm run \\
1521
1552
  --from-huggingface-repo=${model.id} \\
1522
1553
  --prompt="Write me a poem"`,
1523
1554
  ];
1524
1555
 
1525
1556
  export const tf_keras = (model: ModelData): string[] => [
1526
- `# Note: 'keras<3.x' or 'tf_keras' must be installed (legacy)
1557
+ `# Note: 'keras<3.x' or 'tf_keras' must be installed (legacy), and from_pretrained_keras was removed in huggingface_hub 1.0.
1527
1558
  # See https://github.com/keras-team/tf-keras for more details.
1559
+ # !pip install "huggingface_hub<1.0" tf_keras
1528
1560
  from huggingface_hub import from_pretrained_keras
1529
1561
 
1530
1562
  model = from_pretrained_keras("${model.id}")
@@ -1578,6 +1610,19 @@ from MeshAnything.models.meshanything import MeshAnything
1578
1610
  model = MeshAnything(args)`,
1579
1611
  ];
1580
1612
 
1613
+ export const microduck = (model: ModelData): string[] => {
1614
+ // Model-card tag convention: tags: [microduck-slot:stand].
1615
+ const slots = ["walk", "stand", "sitstand", "ground_pick", "kick_left", "kick_right", "roulade"];
1616
+ const matchingSlots = slots.filter((slot) => model.tags.includes(`microduck-slot:${slot}`));
1617
+ const slot = matchingSlots.length === 1 ? matchingSlots[0] : "SLOT";
1618
+ const prefix =
1619
+ matchingSlots.length === 1
1620
+ ? ""
1621
+ : `# Replace SLOT with the slot specified in the model card (${slots.join(", ")}).\n`;
1622
+
1623
+ return [`${prefix}sudo robotctl policy load ${slot} ${model.id}`];
1624
+ };
1625
+
1581
1626
  export const multimolecule = (model: ModelData): string[] => {
1582
1627
  const widgetExample = model.widgetData?.[0] as WidgetExampleTextInput | undefined;
1583
1628
  const exampleText = escapeStringForJson(widgetExample?.text ?? "");
@@ -1771,14 +1816,18 @@ model = pe.VisionTransformer.from_config("${model.id}", pretrained=True)`;
1771
1816
  };
1772
1817
  export const phantom_wan = (model: ModelData): string[] => [
1773
1818
  `from huggingface_hub import snapshot_download
1774
- from phantom_wan import WANI2V, configs
1819
+ from phantom_wan import WanI2V, configs
1820
+ from PIL import Image
1775
1821
 
1776
1822
  checkpoint_dir = snapshot_download("${model.id}")
1777
1823
  wan_i2v = WanI2V(
1778
- config=configs.WAN_CONFIGS['i2v-14B'],
1779
- checkpoint_dir=checkpoint_dir,
1780
- )
1781
- video = wan_i2v.generate(text_prompt, image_prompt)`,
1824
+ config=configs.WAN_CONFIGS['i2v-14B'],
1825
+ checkpoint_dir=checkpoint_dir,
1826
+ )
1827
+
1828
+ text_prompt = "A cat playing with a ball of yarn"
1829
+ image_prompt = Image.open("path/to/image.jpg").convert("RGB")
1830
+ video = wan_i2v.generate(text_prompt, image_prompt)`,
1782
1831
  ];
1783
1832
 
1784
1833
  export const pocket_tts = (model: ModelData): string[] => [
@@ -1903,7 +1952,7 @@ export const tensorflowtts = (model: ModelData): string[] => {
1903
1952
  export const timm = (model: ModelData): string[] => [
1904
1953
  `import timm
1905
1954
 
1906
- model = timm.create_model("hf_hub:${model.id}", pretrained=True)`,
1955
+ model = timm.create_model("hf-hub:${model.id}", pretrained=True)`,
1907
1956
  ];
1908
1957
 
1909
1958
  export const saelens = (/* model: ModelData */): string[] => [
@@ -1931,13 +1980,15 @@ seed_story_cfg = OmegaConf.load(seed_story_cfg_path)
1931
1980
  seed_story = hydra.utils.instantiate(seed_story_cfg, llm=llm) `,
1932
1981
  ];
1933
1982
 
1934
- const skopsPickle = (model: ModelData, modelFile: string) => {
1983
+ const SKLEARN_PICKLE_FILE = /\.(joblib|pkl|pickle)$/i;
1984
+
1985
+ // skops.hub_utils was removed in skops 0.12, so files are downloaded with huggingface_hub
1986
+ const sklearnPickle = (model: ModelData, modelFile: string) => {
1935
1987
  return [
1936
- `import joblib
1937
- from skops.hub_utils import download
1938
- download("${model.id}", "path_to_folder")
1988
+ `from huggingface_hub import hf_hub_download
1989
+ import joblib
1939
1990
  model = joblib.load(
1940
- "${escapeStringForJson(modelFile)}"
1991
+ hf_hub_download("${model.id}", "${escapeStringForJson(modelFile)}")
1941
1992
  )
1942
1993
  # only load pickle files from sources you trust
1943
1994
  # read more about it here https://skops.readthedocs.io/en/stable/persistence.html`,
@@ -1946,42 +1997,32 @@ model = joblib.load(
1946
1997
 
1947
1998
  const skopsFormat = (model: ModelData, modelFile: string) => {
1948
1999
  return [
1949
- `from skops.hub_utils import download
2000
+ `from huggingface_hub import hf_hub_download
1950
2001
  from skops.io import load
1951
- download("${model.id}", "path_to_folder")
1952
2002
  # make sure model file is in skops format
1953
2003
  # if model is a pickle file, make sure it's from a source you trust
1954
- model = load("path_to_folder/${escapeStringForJson(modelFile)}")`,
1955
- ];
1956
- };
1957
-
1958
- const skopsJobLib = (model: ModelData) => {
1959
- return [
1960
- `from huggingface_hub import hf_hub_download
1961
- import joblib
1962
- model = joblib.load(
1963
- hf_hub_download("${model.id}", "sklearn_model.joblib")
1964
- )
1965
- # only load pickle files from sources you trust
1966
- # read more about it here https://skops.readthedocs.io/en/stable/persistence.html`,
2004
+ model = load(hf_hub_download("${model.id}", "${escapeStringForJson(modelFile)}"))`,
1967
2005
  ];
1968
2006
  };
1969
2007
 
1970
2008
  export const sklearn = (model: ModelData): string[] => {
2009
+ const declaredFile = model.config?.sklearn?.model?.file;
1971
2010
  if (model.tags.includes("skops")) {
1972
- const skopsmodelFile = model.config?.sklearn?.model?.file;
1973
- const skopssaveFormat = model.config?.sklearn?.model_format;
1974
- if (!skopsmodelFile) {
2011
+ if (!declaredFile) {
1975
2012
  return [`# ⚠️ Model filename not specified in config.json`];
1976
2013
  }
1977
- if (skopssaveFormat === "pickle") {
1978
- return skopsPickle(model, skopsmodelFile);
1979
- } else {
1980
- return skopsFormat(model, skopsmodelFile);
2014
+ // some skops repos declare no model_format for a pickled file (e.g. julien-c/skops-digits), which skops.io cannot read
2015
+ if (model.config?.sklearn?.model_format === "pickle" || SKLEARN_PICKLE_FILE.test(declaredFile)) {
2016
+ return sklearnPickle(model, declaredFile);
1981
2017
  }
1982
- } else {
1983
- return skopsJobLib(model);
2018
+ return skopsFormat(model, declaredFile);
1984
2019
  }
2020
+ // repos pushed without skops often still declare their filename in config.json; it is only trusted when it names a
2021
+ // pickle, since some hand-written configs carry typos (e.g. "sklean_model.jpblib" next to a real sklearn_model.joblib)
2022
+ return sklearnPickle(
2023
+ model,
2024
+ declaredFile && SKLEARN_PICKLE_FILE.test(declaredFile) ? declaredFile : "sklearn_model.joblib",
2025
+ );
1985
2026
  };
1986
2027
 
1987
2028
  export const stable_audio_tools = (model: ModelData): string[] => [
@@ -2032,7 +2073,7 @@ export const sam2 = (model: ModelData): string[] => {
2032
2073
  import torch
2033
2074
  from sam2.sam2_image_predictor import SAM2ImagePredictor
2034
2075
 
2035
- predictor = SAM2ImagePredictor.from_pretrained(${model.id})
2076
+ predictor = SAM2ImagePredictor.from_pretrained("${model.id}")
2036
2077
 
2037
2078
  with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
2038
2079
  predictor.set_image(<your_image>)
@@ -2042,13 +2083,13 @@ with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
2042
2083
  import torch
2043
2084
  from sam2.sam2_video_predictor import SAM2VideoPredictor
2044
2085
 
2045
- predictor = SAM2VideoPredictor.from_pretrained(${model.id})
2086
+ predictor = SAM2VideoPredictor.from_pretrained("${model.id}")
2046
2087
 
2047
2088
  with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
2048
2089
  state = predictor.init_state(<your_video>)
2049
2090
 
2050
2091
  # add new prompts and instantly get the output on the same frame
2051
- frame_idx, object_ids, masks = predictor.add_new_points(state, <your_prompts>):
2092
+ frame_idx, object_ids, masks = predictor.add_new_points(state, <your_prompts>)
2052
2093
 
2053
2094
  # propagate the prompts to get masklets throughout the video
2054
2095
  for frame_idx, object_ids, masks in predictor.propagate_in_video(state):
@@ -2070,11 +2111,18 @@ output = inference(image, mask)`,
2070
2111
  ];
2071
2112
 
2072
2113
  export const sam_3d_body = (model: ModelData): string[] => [
2073
- `from notebook.utils import setup_sam_3d_body
2114
+ `import cv2
2115
+ import numpy as np
2116
+ from notebook.utils import setup_sam_3d_body
2117
+ from tools.vis_utils import visualize_sample_together
2118
+
2119
+ estimator = setup_sam_3d_body("${model.id}")
2120
+
2121
+ img_bgr = cv2.imread("path/to/image.jpg")
2122
+ outputs = estimator.process_one_image(cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB))
2074
2123
 
2075
- estimator = setup_sam_3d_body(${model.id})
2076
- outputs = estimator.process_one_image(image)
2077
- rend_img = visualize_sample_together(image, outputs, estimator.faces)`,
2124
+ rend_img = visualize_sample_together(img_bgr, outputs, estimator.faces)
2125
+ cv2.imwrite("output.jpg", rend_img.astype(np.uint8))`,
2078
2126
  ];
2079
2127
 
2080
2128
  export const sampleFactory = (model: ModelData): string[] => [
@@ -2171,7 +2219,17 @@ docs_emb = model.encode(documents, is_query=False)`,
2171
2219
  ];
2172
2220
  }
2173
2221
 
2174
- if (model.tags.includes("cross-encoder") || model.pipeline_tag == "text-ranking") {
2222
+ // Rerankers saved in the CrossEncoder format are often tagged text-classification rather than text-ranking;
2223
+ // their architecture still identifies them (e.g. BAAI/bge-reranker-v2-m3). SetFit models can share that shape but
2224
+ // are bi-encoders with a separate head.
2225
+ const isCrossEncoder =
2226
+ model.tags.includes("cross-encoder") ||
2227
+ model.pipeline_tag == "text-ranking" ||
2228
+ (model.pipeline_tag === "text-classification" &&
2229
+ !model.tags.includes("setfit") &&
2230
+ (model.config?.architectures ?? []).some((arch) => arch.endsWith("ForSequenceClassification")));
2231
+
2232
+ if (isCrossEncoder) {
2175
2233
  return [
2176
2234
  `from sentence_transformers import CrossEncoder
2177
2235
 
@@ -2190,6 +2248,34 @@ print(scores)`,
2190
2248
  ];
2191
2249
  }
2192
2250
 
2251
+ // sentence-transformers tags SparseEncoder repos (SPLADE & co) with "sparse-encoder" and MultiVectorEncoder repos
2252
+ // (ColBERT & co) with "multi-vector" when pushing them; both encode queries and documents separately
2253
+ const retrievalEncoder = model.tags.includes("sparse-encoder")
2254
+ ? "SparseEncoder"
2255
+ : model.tags.includes("multi-vector")
2256
+ ? "MultiVectorEncoder"
2257
+ : undefined;
2258
+ if (retrievalEncoder) {
2259
+ return [
2260
+ `from sentence_transformers import ${retrievalEncoder}
2261
+
2262
+ model = ${retrievalEncoder}("${model.id}"${remote_code_snippet})
2263
+
2264
+ queries = ["Which planet is known as the Red Planet?"]
2265
+ documents = [
2266
+ "Venus is often called Earth's twin because of its similar size and proximity.",
2267
+ "Mars, known for its reddish appearance, is often referred to as the Red Planet.",
2268
+ "Jupiter, the largest planet in our solar system, has a prominent red spot.",
2269
+ ]
2270
+
2271
+ query_embeddings = model.encode_query(queries)
2272
+ document_embeddings = model.encode_document(documents)
2273
+
2274
+ similarities = model.similarity(query_embeddings, document_embeddings)
2275
+ print(similarities)`,
2276
+ ];
2277
+ }
2278
+
2193
2279
  const exampleSentences = get_widget_examples_from_st_model(model) ?? [
2194
2280
  "The weather is lovely today.",
2195
2281
  "It's so sunny outside!",
@@ -2213,7 +2299,10 @@ print(similarities.shape)
2213
2299
  export const setfit = (model: ModelData): string[] => [
2214
2300
  `from setfit import SetFitModel
2215
2301
 
2216
- model = SetFitModel.from_pretrained("${model.id}")`,
2302
+ model = SetFitModel.from_pretrained("${model.id}")
2303
+
2304
+ preds = model.predict(["i loved the spiderman movie!", "pineapple on pizza is the worst"])
2305
+ print(preds)`,
2217
2306
  ];
2218
2307
 
2219
2308
  export const spacy = (model: ModelData): string[] => [
@@ -2231,7 +2320,10 @@ nlp = ${nameWithoutNamespace(model.id)}.load()`,
2231
2320
  export const span_marker = (model: ModelData): string[] => [
2232
2321
  `from span_marker import SpanMarkerModel
2233
2322
 
2234
- model = SpanMarkerModel.from_pretrained("${model.id}")`,
2323
+ model = SpanMarkerModel.from_pretrained("${model.id}")
2324
+
2325
+ entities = model.predict("Amelia Earhart flew her single engine Lockheed Vega 5B across the Atlantic to Paris.")
2326
+ print(entities)`,
2235
2327
  ];
2236
2328
 
2237
2329
  export const stanza = (model: ModelData): string[] => [
@@ -2252,6 +2344,8 @@ const speechBrainMethod = (speechbrainInterface: string) => {
2252
2344
  return "enhance_file";
2253
2345
  case "SepformerSeparation":
2254
2346
  return "separate_file";
2347
+ case "SpeakerRecognition":
2348
+ return "verify_files";
2255
2349
  default:
2256
2350
  return undefined;
2257
2351
  }
@@ -2268,12 +2362,14 @@ export const speechbrain = (model: ModelData): string[] => {
2268
2362
  return [`# interface in config.json invalid`];
2269
2363
  }
2270
2364
 
2365
+ // speaker verification compares two recordings
2366
+ const methodArgs = speechbrainMethod === "verify_files" ? '"speaker1.wav", "speaker2.wav"' : '"file.wav"';
2271
2367
  return [
2272
2368
  `from speechbrain.pretrained import ${speechbrainInterface}
2273
2369
  model = ${speechbrainInterface}.from_hparams(
2274
2370
  "${model.id}"
2275
2371
  )
2276
- model.${speechbrainMethod}("file.wav")`,
2372
+ model.${speechbrainMethod}(${methodArgs})`,
2277
2373
  ];
2278
2374
  };
2279
2375
 
@@ -2311,6 +2407,7 @@ export const transformers = (model: ModelData): string[] => {
2311
2407
  ? "extractor"
2312
2408
  : "processor";
2313
2409
  autoSnippet.push(
2410
+ "# pip install -U transformers accelerate",
2314
2411
  "# Load model directly",
2315
2412
  `from transformers import ${info.processor}, ${auto_model}`,
2316
2413
  "",
@@ -2344,12 +2441,13 @@ export const transformers = (model: ModelData): string[] => {
2344
2441
  ' return_tensors="pt",',
2345
2442
  ").to(model.device)",
2346
2443
  "",
2347
- "outputs = model.generate(**inputs, max_new_tokens=40)",
2444
+ "outputs = model.generate(**inputs, max_new_tokens=256)",
2348
2445
  `print(${processorVarName}.decode(outputs[0][inputs["input_ids"].shape[-1]:]))`,
2349
2446
  );
2350
2447
  }
2351
2448
  } else {
2352
2449
  autoSnippet.push(
2450
+ "# pip install -U transformers accelerate",
2353
2451
  "# Load model directly",
2354
2452
  `from transformers import ${auto_model}`,
2355
2453
  `model = ${auto_model}.from_pretrained("${model.id}"` + remote_code_snippet + ', device_map="auto")',
@@ -2362,7 +2460,7 @@ export const transformers = (model: ModelData): string[] => {
2362
2460
  pipelineSnippet.push(
2363
2461
  `# Warning: Pipeline type "${model.pipeline_tag}" is no longer supported in transformers v5.`,
2364
2462
  `# You must load the model directly (see below) or downgrade to v4.x with:`,
2365
- `# 'pip install "transformers<5.0.0'`,
2463
+ `# pip install "transformers<5.0.0"`,
2366
2464
  );
2367
2465
  }
2368
2466
 
@@ -2372,7 +2470,7 @@ export const transformers = (model: ModelData): string[] => {
2372
2470
  `pipe = pipeline("${model.pipeline_tag}", model="${model.id}"` + remote_code_snippet + ")",
2373
2471
  );
2374
2472
 
2375
- if (model.tags.includes("conversational")) {
2473
+ if (model.tags.includes("conversational") && hasChatTemplate(model)) {
2376
2474
  if (model.tags.includes("image-text-to-text")) {
2377
2475
  pipelineSnippet.push(
2378
2476
  "messages = [",
@@ -2410,6 +2508,11 @@ export const transformers = (model: ModelData): string[] => {
2410
2508
  return [autoSnippet.join("\n")];
2411
2509
  };
2412
2510
 
2511
+ // Hub tasks that transformers.js exposes under a different pipeline() task name
2512
+ const TRANSFORMERS_JS_TASK_ALIASES: Record<string, string> = {
2513
+ "sentence-similarity": "feature-extraction",
2514
+ };
2515
+
2413
2516
  export const transformersJS = (model: ModelData): string[] => {
2414
2517
  if (!model.pipeline_tag) {
2415
2518
  return [`// ⚠️ Unknown pipeline tag`];
@@ -2417,45 +2520,119 @@ export const transformersJS = (model: ModelData): string[] => {
2417
2520
 
2418
2521
  const libName = "@huggingface/transformers";
2419
2522
 
2523
+ // Rerankers score (query, document) pairs. The text-classification pipeline cannot take pairs and applies a softmax
2524
+ // over their single logit, so every score comes out as 1; the model is called directly instead.
2525
+ if (model.pipeline_tag === "text-ranking") {
2526
+ return [
2527
+ `// npm i ${libName}
2528
+ import { AutoTokenizer, AutoModelForSequenceClassification } from '${libName}';
2529
+
2530
+ const tokenizer = await AutoTokenizer.from_pretrained('${model.id}');
2531
+ const model = await AutoModelForSequenceClassification.from_pretrained('${model.id}');
2532
+
2533
+ const query = 'Which planet is known as the Red Planet?';
2534
+ const documents = [
2535
+ 'Mars, known for its reddish appearance, is often referred to as the Red Planet.',
2536
+ 'Venus is often called the twin of Earth because of its similar size and proximity.',
2537
+ ];
2538
+
2539
+ const inputs = tokenizer(new Array(documents.length).fill(query), { text_pair: documents, padding: true, truncation: true });
2540
+ const { logits } = await model(inputs);
2541
+ console.log(logits.sigmoid().tolist()); // one relevance score per document`,
2542
+ ];
2543
+ }
2544
+
2545
+ const task = TRANSFORMERS_JS_TASK_ALIASES[model.pipeline_tag] ?? model.pipeline_tag;
2546
+
2420
2547
  return [
2421
2548
  `// npm i ${libName}
2422
2549
  import { pipeline } from '${libName}';
2423
2550
 
2424
2551
  // Allocate pipeline
2425
- const pipe = await pipeline('${model.pipeline_tag}', '${model.id}');`,
2552
+ const pipe = await pipeline('${task}', '${model.id}');`,
2426
2553
  ];
2427
2554
  };
2428
2555
 
2429
- const peftTask = (peftTaskType?: string) => {
2430
- switch (peftTaskType) {
2431
- case "CAUSAL_LM":
2432
- return "CausalLM";
2433
- case "SEQ_2_SEQ_LM":
2434
- return "Seq2SeqLM";
2435
- case "TOKEN_CLS":
2436
- return "TokenClassification";
2437
- case "SEQ_CLS":
2438
- return "SequenceClassification";
2439
- default:
2440
- return undefined;
2556
+ const PEFT_TASK_TYPE_TO_AUTO_CLASS: Record<string, string> = {
2557
+ CAUSAL_LM: "AutoModelForCausalLM",
2558
+ SEQ_2_SEQ_LM: "AutoModelForSeq2SeqLM",
2559
+ TOKEN_CLS: "AutoModelForTokenClassification",
2560
+ SEQ_CLS: "AutoModelForSequenceClassification",
2561
+ QUESTION_ANS: "AutoModelForQuestionAnswering",
2562
+ // transformers has no AutoModelForFeatureExtraction; PEFT's own AutoPeftModelForFeatureExtraction resolves to AutoModel
2563
+ FEATURE_EXTRACTION: "AutoModel",
2564
+ };
2565
+
2566
+ // PEFT has no speech-specific task type: Whisper-style adapters are tagged SEQ_2_SEQ_LM, but transformers registers
2567
+ // Whisper and Speech2Text only under AutoModelForSpeechSeq2Seq (SeamlessM4T is under both seq2seq auto classes, SpeechT5
2568
+ // under the speech-to-text and text-to-spectrogram ones).
2569
+ // Other audio seq2seq models (Qwen2-Audio, GLM-ASR) stay under AutoModelForSeq2SeqLM, so the base model name is the
2570
+ // main signal. Matched against Hub repo ids, which are hyphenated (facebook/s2t-small-librispeech-asr,
2571
+ // facebook/seamless-m4t-v2-large), not against transformers model_type identifiers.
2572
+ const PEFT_SPEECH_TO_TEXT_BASE_MODEL = /whisper|\bs2t\b|speech[-_]to[-_]text/i;
2573
+ // SeamlessM4T also has a text-to-text head under AutoModelForSeq2SeqLM, which text translation adapters need; most
2574
+ // SeamlessM4T adapters on the Hub are speech ones, so the speech class stays the default when no text task is declared.
2575
+ const PEFT_SEAMLESS_M4T_BASE_MODEL = /seamless[-_]m4t/i;
2576
+ const PEFT_TEXT_TASKS = ["translation", "text2text-generation", "summarization"];
2577
+ // SpeechT5 checkpoints are overwhelmingly text-to-speech fine-tunes (often without "tts" in their name), which load
2578
+ // with AutoModelForTextToSpectrogram; AutoModelForSpeechSeq2Seq would silently load the speech-to-text head, so only
2579
+ // ASR bases or adapters get it.
2580
+ const PEFT_SPEECHT5_BASE_MODEL = /speecht5/i;
2581
+
2582
+ const declaresTask = (model: ModelData, task: string): boolean =>
2583
+ model.pipeline_tag === task || model.tags.includes(task);
2584
+
2585
+ // the auto class a SEQ_2_SEQ_LM adapter needs when its base is a speech model, undefined for text seq2seq bases
2586
+ const peftSpeechAutoClass = (model: ModelData, baseModel: string): string | undefined => {
2587
+ if (PEFT_SPEECH_TO_TEXT_BASE_MODEL.test(baseModel)) {
2588
+ return "AutoModelForSpeechSeq2Seq";
2589
+ }
2590
+ if (PEFT_SEAMLESS_M4T_BASE_MODEL.test(baseModel)) {
2591
+ return PEFT_TEXT_TASKS.some((task) => declaresTask(model, task)) ? undefined : "AutoModelForSpeechSeq2Seq";
2441
2592
  }
2593
+ if (PEFT_SPEECHT5_BASE_MODEL.test(baseModel)) {
2594
+ const isAsr = /(?:^|[^a-z])asr(?:[^a-z]|$)/i.test(baseModel) || declaresTask(model, "automatic-speech-recognition");
2595
+ return isAsr ? "AutoModelForSpeechSeq2Seq" : "AutoModelForTextToSpectrogram";
2596
+ }
2597
+ return undefined;
2442
2598
  };
2443
2599
 
2444
2600
  export const peft = (model: ModelData): string[] => {
2445
2601
  const { base_model_name_or_path: peftBaseModel, task_type: peftTaskType } = model.config?.peft ?? {};
2446
- const pefttask = peftTask(peftTaskType);
2447
- if (!pefttask) {
2602
+ let autoClass = peftTaskType ? PEFT_TASK_TYPE_TO_AUTO_CLASS[peftTaskType] : undefined;
2603
+ if (!autoClass) {
2448
2604
  return [`Task type is invalid.`];
2449
2605
  }
2450
- if (!peftBaseModel) {
2606
+ // adapter_config.json sometimes omits the base model; the model card's base_model is the next best source
2607
+ const cardBaseModel = Array.isArray(model.cardData?.base_model)
2608
+ ? model.cardData.base_model[0]
2609
+ : model.cardData?.base_model;
2610
+ const baseModel = peftBaseModel || cardBaseModel;
2611
+ if (!baseModel) {
2451
2612
  return [`Base model is not found.`];
2452
2613
  }
2614
+ if (peftTaskType === "SEQ_2_SEQ_LM") {
2615
+ autoClass = peftSpeechAutoClass(model, baseModel) ?? autoClass;
2616
+ }
2617
+
2618
+ if (autoClass === "AutoModelForTextToSpectrogram") {
2619
+ return [
2620
+ `from peft import PeftConfig, PeftModel
2621
+ from transformers import AutoModelForTextToSpectrogram
2622
+
2623
+ base_model = AutoModelForTextToSpectrogram.from_pretrained("${escapeStringForJson(baseModel)}")
2624
+ # PEFT's seq2seq wrapper needs generate(), which SpeechT5's text-to-speech model lacks: load the adapter without its task type
2625
+ config = PeftConfig.from_pretrained("${model.id}")
2626
+ config.task_type = None
2627
+ model = PeftModel.from_pretrained(base_model, "${model.id}", config=config)`,
2628
+ ];
2629
+ }
2453
2630
 
2454
2631
  return [
2455
2632
  `from peft import PeftModel
2456
- from transformers import AutoModelFor${pefttask}
2633
+ from transformers import ${autoClass}
2457
2634
 
2458
- base_model = AutoModelFor${pefttask}.from_pretrained("${escapeStringForJson(peftBaseModel)}")
2635
+ base_model = ${autoClass}.from_pretrained("${escapeStringForJson(baseModel)}")
2459
2636
  model = PeftModel.from_pretrained(base_model, "${model.id}")`,
2460
2637
  ];
2461
2638
  };
@@ -2490,7 +2667,7 @@ transcriptions = asr_model.transcribe(["file.wav"])`,
2490
2667
  };
2491
2668
 
2492
2669
  export const mlAgents = (model: ModelData): string[] => [
2493
- `mlagents-load-from-hf --repo-id="${model.id}" --local-dir="./download: string[]s"`,
2670
+ `mlagents-load-from-hf --repo-id="${model.id}" --local-dir="./downloads"`,
2494
2671
  ];
2495
2672
 
2496
2673
  export const sentis = (/* model: ModelData */): string[] => [
@@ -2565,10 +2742,10 @@ model = Model.from_pretrained("${model.id}")`,
2565
2742
 
2566
2743
  export const lvface = (model: ModelData): string[] => [
2567
2744
  `from huggingface_hub import hf_hub_download
2568
- from inference_onnx import LVFaceONNXInferencer
2745
+ from inference_onnx import LVFaceONNXInferencer
2569
2746
 
2570
2747
  model_path = hf_hub_download("${model.id}", "LVFace-L_Glint360K/LVFace-L_Glint360K.onnx")
2571
- inferencer = LVFaceONNXInferencer(model_path, use_gpu=True, timeout=300)
2748
+ inferencer = LVFaceONNXInferencer(model_path, use_gpu=True)
2572
2749
  img_path = 'path/to/image1.jpg'
2573
2750
  embedding = inferencer.infer_from_image(img_path)`,
2574
2751
  ];
@@ -2607,8 +2784,7 @@ export const vui = (): string[] => [
2607
2784
 
2608
2785
  import torchaudio
2609
2786
 
2610
- from vui.inference import render
2611
- from vui.model import Vui,
2787
+ from vui.legacy import Vui, render
2612
2788
 
2613
2789
  model = Vui.from_pretrained().cuda()
2614
2790
  waveform = render(
@@ -2616,16 +2792,17 @@ waveform = render(
2616
2792
  "Hey, here is some random stuff, usually something quite long as the shorter the text the less likely the model can cope!",
2617
2793
  )
2618
2794
  print(waveform.shape)
2619
- torchaudio.save("out.opus", waveform[0], 22050)
2795
+ torchaudio.save("out.wav", waveform[0], 22050)
2620
2796
  `,
2621
2797
  ];
2622
2798
 
2623
2799
  export const chattts = (): string[] => [
2624
2800
  `import ChatTTS
2801
+ import torch
2625
2802
  import torchaudio
2626
2803
 
2627
2804
  chat = ChatTTS.Chat()
2628
- chat.load_models(compile=False) # Set to True for better performance
2805
+ chat.load(compile=False) # Set to True for better performance
2629
2806
 
2630
2807
  texts = ["PUT YOUR TEXT HERE",]
2631
2808
 
@@ -2635,19 +2812,44 @@ torchaudio.save("output1.wav", torch.from_numpy(wavs[0]), 24000)`,
2635
2812
  ];
2636
2813
 
2637
2814
  export const ultralytics = (model: ModelData): string[] => {
2638
- // ultralytics models must have a version tag (e.g. `yolov8`)
2639
- const versionTag = model.tags.find((tag) => tag.match(/^yolov\d+$/));
2815
+ // the YOLOv10 fork (library "yolov10") ships its own `YOLOv10` class with a `from_pretrained` helper that reads the
2816
+ // model.safetensors it pushes. Mainline ultralytics repos often list "yolov10" among their version tags
2817
+ // (e.g. Ultralytics/YOLOv8) but only ship .pt files, so an ultralytics tag alone keeps them on the mainline loader;
2818
+ // fork pushes still tagged ultralytics are recognized by their safetensors weights.
2819
+ const isYolov10Fork =
2820
+ model.library_name === "yolov10" ||
2821
+ (model.tags.includes("yolov10") && (!model.tags.includes("ultralytics") || model.tags.includes("safetensors")));
2822
+ if (isYolov10Fork) {
2823
+ // from_pretrained reads the model.safetensors that fork pushes ship; older fork repos only ship .pt weights
2824
+ if (model.tags.includes("safetensors")) {
2825
+ return [
2826
+ `from ultralytics import YOLOv10
2827
+
2828
+ model = YOLOv10.from_pretrained("${model.id}")
2829
+ source = 'http://images.cocodataset.org/val2017/000000039769.jpg'
2830
+ model.predict(source=source, save=True)`,
2831
+ ];
2832
+ }
2833
+ return [
2834
+ `from huggingface_hub import hf_hub_download
2835
+ from ultralytics import YOLOv10
2640
2836
 
2641
- const className = versionTag ? `YOLOv${versionTag.slice(4)}` : "YOLOvXX";
2642
- const prefix = versionTag
2643
- ? ""
2644
- : `# Couldn't find a valid YOLO version tag.\n# Replace XX with the correct version.\n`;
2837
+ # pick the weights file from this repo's "Files and versions" tab
2838
+ weights = hf_hub_download("${model.id}", "<weights>.pt")
2839
+ model = YOLOv10(weights)
2840
+ source = 'http://images.cocodataset.org/val2017/000000039769.jpg'
2841
+ model.predict(source=source, save=True)`,
2842
+ ];
2843
+ }
2645
2844
 
2845
+ // mainline ultralytics loads weights from a local file
2646
2846
  return [
2647
- prefix +
2648
- `from ultralytics import ${className}
2847
+ `from huggingface_hub import hf_hub_download
2848
+ from ultralytics import YOLO
2649
2849
 
2650
- model = ${className}.from_pretrained("${model.id}")
2850
+ # pick the weights file from this repo's "Files and versions" tab
2851
+ weights = hf_hub_download("${model.id}", "<weights>.pt")
2852
+ model = YOLO(weights)
2651
2853
  source = 'http://images.cocodataset.org/val2017/000000039769.jpg'
2652
2854
  model.predict(source=source, save=True)`,
2653
2855
  ];
@@ -2726,14 +2928,16 @@ model = SwarmFormerModel.from_pretrained("${model.id}")
2726
2928
 
2727
2929
  export const univa = (model: ModelData): string[] => [
2728
2930
  `# Follow installation instructions at https://github.com/PKU-YuanGroup/UniWorld-V1
2729
-
2931
+ import torch
2932
+ from transformers import AutoProcessor
2730
2933
  from univa.models.qwen2p5vl.modeling_univa_qwen2p5vl import UnivaQwen2p5VLForConditionalGeneration
2731
- model = UnivaQwen2p5VLForConditionalGeneration.from_pretrained(
2732
- "${model.id}",
2733
- torch_dtype=torch.bfloat16,
2734
- attn_implementation="flash_attention_2",
2735
- ).to("cuda")
2736
- processor = AutoProcessor.from_pretrained("${model.id}")
2934
+
2935
+ model = UnivaQwen2p5VLForConditionalGeneration.from_pretrained(
2936
+ "${model.id}",
2937
+ torch_dtype=torch.bfloat16,
2938
+ attn_implementation="flash_attention_2",
2939
+ ).to("cuda")
2940
+ processor = AutoProcessor.from_pretrained("${model.id}")
2737
2941
  `,
2738
2942
  ];
2739
2943
 
@@ -2741,7 +2945,7 @@ const mlx_unknown = (model: ModelData): string[] => [
2741
2945
  `# Download the model from the Hub
2742
2946
  pip install huggingface_hub[hf_xet]
2743
2947
 
2744
- huggingface-cli download --local-dir ${nameWithoutNamespace(model.id)} ${model.id}`,
2948
+ hf download ${model.id} --local-dir ${nameWithoutNamespace(model.id)}`,
2745
2949
  ];
2746
2950
 
2747
2951
  const mlxlm = (model: ModelData): string[] => [
@@ -2802,10 +3006,11 @@ output = generate(model, processor, formatted_prompt, image)
2802
3006
  print(output)`,
2803
3007
  ];
2804
3008
 
3009
+ // create_model() takes a name from the mlx-image registry, not a repo id; the weights live in "mlx-vision/<name>-mlxim"
2805
3010
  export const mlxim = (model: ModelData): string[] => [
2806
3011
  `from mlxim.model import create_model
2807
3012
 
2808
- model = create_model(${model.id})`,
3013
+ model = create_model("${nameWithoutNamespace(model.id).replace(/-mlxim$/, "")}")`,
2809
3014
  ];
2810
3015
 
2811
3016
  export const mlx = (model: ModelData): string[] => {
@@ -2825,7 +3030,10 @@ export const mlx = (model: ModelData): string[] => {
2825
3030
  export const model2vec = (model: ModelData): string[] => [
2826
3031
  `from model2vec import StaticModel
2827
3032
 
2828
- model = StaticModel.from_pretrained("${model.id}")`,
3033
+ model = StaticModel.from_pretrained("${model.id}")
3034
+
3035
+ embeddings = model.encode(["It's dangerous to go alone!", "It's a secret to everybody."])
3036
+ print(embeddings.shape)`,
2829
3037
  ];
2830
3038
 
2831
3039
  export const mobilint = (model: ModelData): string[] => {
@@ -2871,8 +3079,9 @@ export const pruna = (model: ModelData): string[] => {
2871
3079
  snippets = snippets.map(ensurePrunaModelImport);
2872
3080
 
2873
3081
  if (model.tags.includes("pruna_pro-ai")) {
3082
+ // only rewrite the import and the class name, never the repo id (an org name can contain "pruna")
2874
3083
  return snippets.map((snippet) =>
2875
- snippet.replace(/\bpruna\b/g, "pruna_pro").replace(/\bPrunaModel\b/g, "PrunaProModel"),
3084
+ snippet.replace(/^from pruna import /gm, "from pruna_pro import ").replace(/\bPrunaModel\b/g, "PrunaProModel"),
2876
3085
  );
2877
3086
  }
2878
3087
 
@@ -2951,21 +3160,19 @@ export const outetts = (model: ModelData): string[] => {
2951
3160
 
2952
3161
  // v1.0 HF → minimal runnable snippet
2953
3162
  return [
2954
- `
2955
- import outetts
2956
-
2957
- enum = outetts.Models("${model.id}".split("/", 1)[1]) # VERSION_1_0_SIZE_1B
2958
- cfg = outetts.ModelConfig.auto_config(enum, outetts.Backend.HF)
2959
- tts = outetts.Interface(cfg)
2960
-
2961
- speaker = tts.load_default_speaker("EN-FEMALE-1-NEUTRAL")
2962
- tts.generate(
2963
- outetts.GenerationConfig(
2964
- text="Hello there, how are you doing?",
2965
- speaker=speaker,
2966
- )
2967
- ).save("output.wav")
2968
- `,
3163
+ `import outetts
3164
+
3165
+ enum = outetts.Models("${model.id}".split("/", 1)[1]) # VERSION_1_0_SIZE_1B
3166
+ cfg = outetts.ModelConfig.auto_config(enum, outetts.Backend.HF)
3167
+ tts = outetts.Interface(cfg)
3168
+
3169
+ speaker = tts.load_default_speaker("EN-FEMALE-1-NEUTRAL")
3170
+ tts.generate(
3171
+ outetts.GenerationConfig(
3172
+ text="Hello there, how are you doing?",
3173
+ speaker=speaker,
3174
+ )
3175
+ ).save("output.wav")`,
2969
3176
  ];
2970
3177
  };
2971
3178