@huggingface/tasks 0.21.54 → 0.21.56
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/commonjs/hardware-amd.d.ts.map +1 -1
- package/dist/commonjs/hardware-amd.js +48 -0
- package/dist/commonjs/hardware-nvidia.d.ts +1 -0
- package/dist/commonjs/hardware-nvidia.d.ts.map +1 -1
- package/dist/commonjs/hardware-nvidia.js +122 -0
- package/dist/commonjs/library-to-tasks.d.ts.map +1 -1
- package/dist/commonjs/library-to-tasks.js +8 -1
- package/dist/commonjs/model-libraries-snippets.d.ts +3 -1
- package/dist/commonjs/model-libraries-snippets.d.ts.map +1 -1
- package/dist/commonjs/model-libraries-snippets.js +349 -154
- package/dist/commonjs/model-libraries-snippets.spec.js +300 -0
- package/dist/commonjs/model-libraries.d.ts +11 -2
- package/dist/commonjs/model-libraries.d.ts.map +1 -1
- package/dist/commonjs/model-libraries.js +10 -1
- package/dist/esm/hardware-amd.d.ts.map +1 -1
- package/dist/esm/hardware-amd.js +48 -0
- package/dist/esm/hardware-nvidia.d.ts +1 -0
- package/dist/esm/hardware-nvidia.d.ts.map +1 -1
- package/dist/esm/hardware-nvidia.js +122 -0
- package/dist/esm/library-to-tasks.d.ts.map +1 -1
- package/dist/esm/library-to-tasks.js +8 -1
- package/dist/esm/model-libraries-snippets.d.ts +3 -1
- package/dist/esm/model-libraries-snippets.d.ts.map +1 -1
- package/dist/esm/model-libraries-snippets.js +344 -151
- package/dist/esm/model-libraries-snippets.spec.js +301 -1
- package/dist/esm/model-libraries.d.ts +11 -2
- package/dist/esm/model-libraries.d.ts.map +1 -1
- package/dist/esm/model-libraries.js +10 -1
- package/package.json +1 -1
- package/src/hardware-amd.ts +48 -0
- package/src/hardware-nvidia.ts +122 -0
- package/src/library-to-tasks.ts +8 -1
- package/src/model-libraries-snippets.spec.ts +389 -0
- package/src/model-libraries-snippets.ts +364 -149
- package/src/model-libraries.ts +10 -1
|
@@ -12,7 +12,7 @@ const isValidIdentifier = (str) => /^[A-Za-z_]\w*$/.test(str);
|
|
|
12
12
|
export const adapters = (model) => [
|
|
13
13
|
`from adapters import AutoAdapterModel
|
|
14
14
|
|
|
15
|
-
model = AutoAdapterModel.from_pretrained("${escapeStringForJson(model.config?.adapter_transformers?.model_name ?? "
|
|
15
|
+
model = AutoAdapterModel.from_pretrained("${escapeStringForJson(model.config?.adapter_transformers?.model_name ?? "<base-model-id>")}")
|
|
16
16
|
model.load_adapter("${model.id}", set_active=True)`,
|
|
17
17
|
];
|
|
18
18
|
const allennlpUnknown = (model) => [
|
|
@@ -40,6 +40,13 @@ export const araclip = (model) => [
|
|
|
40
40
|
|
|
41
41
|
model = AraClip.from_pretrained("${model.id}")`,
|
|
42
42
|
];
|
|
43
|
+
export const asimov = (model) => [
|
|
44
|
+
`# Run the policy in a browser simulation of the Asimov 1 humanoid (no GPU needed)
|
|
45
|
+
git clone https://github.com/menloresearch/humanoid-policy-viewer
|
|
46
|
+
cd humanoid-policy-viewer
|
|
47
|
+
npm install
|
|
48
|
+
npm run hf ${model.id}`,
|
|
49
|
+
];
|
|
43
50
|
export const asteroid = (model) => [
|
|
44
51
|
`from asteroid.models import BaseModel
|
|
45
52
|
|
|
@@ -64,7 +71,13 @@ result, message = detector.detect_watermark(watermarked_audio, sr)`;
|
|
|
64
71
|
return [watermarkSnippet, detectorSnippet];
|
|
65
72
|
};
|
|
66
73
|
function get_base_diffusers_model(model) {
|
|
67
|
-
|
|
74
|
+
const baseModel = model.cardData?.base_model;
|
|
75
|
+
// several base models can be listed (e.g. a LoRA trained for a base and its turbo variant). When an original checkpoint
|
|
76
|
+
// and its Diffusers conversion are both listed (["Wan-AI/Wan2.1-I2V-14B-480P", "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers"]),
|
|
77
|
+
// only the conversion has a model_index.json that DiffusionPipeline can load; otherwise the first one is used
|
|
78
|
+
const candidates = Array.isArray(baseModel) ? baseModel : baseModel ? [baseModel] : [];
|
|
79
|
+
const chosen = candidates.find((id) => id.toLowerCase().endsWith("-diffusers")) ?? candidates[0];
|
|
80
|
+
return escapeStringForJson(chosen ?? "fill-in-base-model");
|
|
68
81
|
}
|
|
69
82
|
function get_prompt_from_diffusers_model(model) {
|
|
70
83
|
const prompt = model.widgetData?.[0]?.text ?? model.cardData?.instance_prompt;
|
|
@@ -145,7 +158,7 @@ detector = cvg.NeuralCornerDetector(checkpoint)
|
|
|
145
158
|
# Embedder models, such as Milo:
|
|
146
159
|
embedder = cvg.NeuralEmbedder(checkpoint)`,
|
|
147
160
|
];
|
|
148
|
-
export const colipri = (
|
|
161
|
+
export const colipri = () => {
|
|
149
162
|
const installSnippet = `pip install colipri`;
|
|
150
163
|
const exampleSnippet = `from colipri import get_model
|
|
151
164
|
from colipri import get_processor
|
|
@@ -154,7 +167,7 @@ from colipri import ZeroShotImageClassificationPipeline
|
|
|
154
167
|
|
|
155
168
|
model = get_model().cuda()
|
|
156
169
|
processor = get_processor()
|
|
157
|
-
pipeline = ZeroShotImageClassificationPipeline(
|
|
170
|
+
pipeline = ZeroShotImageClassificationPipeline(model, processor)
|
|
158
171
|
|
|
159
172
|
image = load_sample_ct()
|
|
160
173
|
|
|
@@ -271,7 +284,7 @@ export const depth_anything_v2 = (model) => {
|
|
|
271
284
|
else if (model.id === "depth-anything/Depth-Anything-V2-Large") {
|
|
272
285
|
encoder = "vitl";
|
|
273
286
|
features = "256";
|
|
274
|
-
out_channels = "[256, 512, 1024, 1024";
|
|
287
|
+
out_channels = "[256, 512, 1024, 1024]";
|
|
275
288
|
}
|
|
276
289
|
return [
|
|
277
290
|
`
|
|
@@ -280,6 +293,7 @@ export const depth_anything_v2 = (model) => {
|
|
|
280
293
|
# Load the model and infer depth from an image
|
|
281
294
|
import cv2
|
|
282
295
|
import torch
|
|
296
|
+
from huggingface_hub import hf_hub_download
|
|
283
297
|
|
|
284
298
|
from depth_anything_v2.dpt import DepthAnythingV2
|
|
285
299
|
|
|
@@ -289,7 +303,8 @@ model = DepthAnythingV2(encoder="${encoder}", features=${features}, out_channels
|
|
|
289
303
|
# load the weights
|
|
290
304
|
filepath = hf_hub_download(repo_id="${model.id}", filename="depth_anything_v2_${encoder}.pth", repo_type="model")
|
|
291
305
|
state_dict = torch.load(filepath, map_location="cpu")
|
|
292
|
-
model.load_state_dict(state_dict)
|
|
306
|
+
model.load_state_dict(state_dict)
|
|
307
|
+
model.eval()
|
|
293
308
|
|
|
294
309
|
raw_img = cv2.imread("your/image/path")
|
|
295
310
|
depth = model.infer_image(raw_img) # HxW raw depth map in numpy
|
|
@@ -298,8 +313,8 @@ depth = model.infer_image(raw_img) # HxW raw depth map in numpy
|
|
|
298
313
|
};
|
|
299
314
|
export const depth_pro = (model) => {
|
|
300
315
|
const installSnippet = `# Download checkpoint
|
|
301
|
-
pip install
|
|
302
|
-
|
|
316
|
+
pip install huggingface_hub
|
|
317
|
+
hf download ${model.id} --local-dir checkpoints`;
|
|
303
318
|
const inferenceSnippet = `import depth_pro
|
|
304
319
|
|
|
305
320
|
# Load model and preprocessing transform
|
|
@@ -367,7 +382,7 @@ export const describe_anything = (model) => [
|
|
|
367
382
|
from huggingface_hub import snapshot_download
|
|
368
383
|
from dam import DescribeAnythingModel
|
|
369
384
|
|
|
370
|
-
snapshot_download(${model.id}, local_dir="checkpoints")
|
|
385
|
+
snapshot_download("${model.id}", local_dir="checkpoints")
|
|
371
386
|
|
|
372
387
|
dam = DescribeAnythingModel(
|
|
373
388
|
model_path="checkpoints",
|
|
@@ -478,7 +493,7 @@ pipe.load_lora_weights("${model.id}")
|
|
|
478
493
|
prompt = "${get_prompt_from_diffusers_model(model) ?? diffusersVideoDefaultPrompt}"
|
|
479
494
|
input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/guitar-man.png")
|
|
480
495
|
|
|
481
|
-
|
|
496
|
+
output = pipe(image=input_image, prompt=prompt).frames[0]
|
|
482
497
|
export_to_video(output, "output.mp4")`,
|
|
483
498
|
];
|
|
484
499
|
const diffusers_textual_inversion = (model) => [
|
|
@@ -518,7 +533,7 @@ from diffusers import AutoPipelineForInpainting
|
|
|
518
533
|
from diffusers.utils import load_image
|
|
519
534
|
|
|
520
535
|
# switch to "mps" for apple devices
|
|
521
|
-
pipe = AutoPipelineForInpainting.from_pretrained("${model.id}", dtype=torch.float16,
|
|
536
|
+
pipe = AutoPipelineForInpainting.from_pretrained("${model.id}", dtype=torch.float16, device_map="cuda")
|
|
522
537
|
|
|
523
538
|
img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
|
|
524
539
|
mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
|
|
@@ -540,9 +555,12 @@ image = pipe(
|
|
|
540
555
|
).images[0]`,
|
|
541
556
|
];
|
|
542
557
|
export const diffusers = (model) => {
|
|
558
|
+
// the Hub emits pipeline class tags as `diffusers:<ClassName>`; bare names are kept for manually added tags
|
|
559
|
+
const hasPipelineClass = (className) => model.tags.includes(`diffusers:${className}`) ||
|
|
560
|
+
model.tags.includes(className) ||
|
|
561
|
+
model.config?.diffusers?._class_name === className;
|
|
543
562
|
let codeSnippets;
|
|
544
|
-
if (
|
|
545
|
-
model.tags.includes("StableDiffusionXLInpaintPipeline")) {
|
|
563
|
+
if (hasPipelineClass("StableDiffusionInpaintPipeline") || hasPipelineClass("StableDiffusionXLInpaintPipeline")) {
|
|
546
564
|
codeSnippets = diffusers_inpainting(model);
|
|
547
565
|
}
|
|
548
566
|
else if (model.tags.includes("controlnet")) {
|
|
@@ -565,7 +583,7 @@ export const diffusers = (model) => {
|
|
|
565
583
|
else if (model.tags.includes("textual_inversion")) {
|
|
566
584
|
codeSnippets = diffusers_textual_inversion(model);
|
|
567
585
|
}
|
|
568
|
-
else if (
|
|
586
|
+
else if (hasPipelineClass("FluxFillPipeline")) {
|
|
569
587
|
codeSnippets = diffusers_flux_fill(model);
|
|
570
588
|
}
|
|
571
589
|
else if (model.pipeline_tag === "image-to-video") {
|
|
@@ -586,7 +604,7 @@ from diffusionkit.mlx import DiffusionPipeline
|
|
|
586
604
|
pipeline = DiffusionPipeline(
|
|
587
605
|
shift=3.0,
|
|
588
606
|
use_t5=False,
|
|
589
|
-
model_version
|
|
607
|
+
model_version="${model.id}",
|
|
590
608
|
low_memory_mode=True,
|
|
591
609
|
a16=True,
|
|
592
610
|
w16=True,
|
|
@@ -596,7 +614,7 @@ from diffusionkit.mlx import FluxPipeline
|
|
|
596
614
|
|
|
597
615
|
pipeline = FluxPipeline(
|
|
598
616
|
shift=1.0,
|
|
599
|
-
model_version
|
|
617
|
+
model_version="${model.id}",
|
|
600
618
|
low_memory_mode=True,
|
|
601
619
|
a16=True,
|
|
602
620
|
w16=True,
|
|
@@ -617,21 +635,22 @@ image, _ = pipeline.generate_image(
|
|
|
617
635
|
return [pipelineSnippet, generateSnippet];
|
|
618
636
|
};
|
|
619
637
|
export const cartesia_pytorch = (model) => [
|
|
620
|
-
`#
|
|
621
|
-
|
|
638
|
+
`# The cartesia-pytorch package on PyPI ships no code: clone the repo and put its cartesia-pytorch folder on PYTHONPATH
|
|
639
|
+
# (it needs an Ampere or newer CUDA GPU, torch 2.4, mamba-ssm 2.2.2 and flash-attn 2.6.3)
|
|
640
|
+
# git clone https://github.com/cartesia-ai/edge && export PYTHONPATH="$PWD/edge/cartesia-pytorch:$PYTHONPATH"
|
|
641
|
+
from cartesia_pytorch.Rene import ReneLMHeadModel
|
|
622
642
|
from transformers import AutoTokenizer
|
|
623
643
|
|
|
624
|
-
model = ReneLMHeadModel.from_pretrained("${model.id}")
|
|
644
|
+
model = ReneLMHeadModel.from_pretrained("${model.id}").half().cuda()
|
|
625
645
|
tokenizer = AutoTokenizer.from_pretrained("allenai/OLMo-1B-hf")
|
|
626
646
|
|
|
627
647
|
in_message = ["Rene Descartes was"]
|
|
628
648
|
inputs = tokenizer(in_message, return_tensors="pt")
|
|
629
649
|
|
|
630
|
-
outputs = model.generate(inputs.input_ids, max_length=50, top_k=100, top_p=0.99)
|
|
650
|
+
outputs = model.generate(inputs.input_ids.cuda(), max_length=50, top_k=100, top_p=0.99)
|
|
631
651
|
out_message = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
|
|
632
652
|
|
|
633
|
-
print(out_message)
|
|
634
|
-
)`,
|
|
653
|
+
print(out_message)`,
|
|
635
654
|
];
|
|
636
655
|
export const cartesia_mlx = (model) => [
|
|
637
656
|
`import mlx.core as mx
|
|
@@ -678,7 +697,8 @@ model = Text2Speech.from_pretrained("${model.id}")
|
|
|
678
697
|
speech, *_ = model("text to generate speech from")`,
|
|
679
698
|
];
|
|
680
699
|
export const espnetASR = (model) => [
|
|
681
|
-
`
|
|
700
|
+
`import soundfile
|
|
701
|
+
from espnet2.bin.asr_inference import Speech2Text
|
|
682
702
|
|
|
683
703
|
model = Speech2Text.from_pretrained(
|
|
684
704
|
"${model.id}"
|
|
@@ -705,9 +725,14 @@ models, cfg, task = load_model_ensemble_and_task_from_hf_hub(
|
|
|
705
725
|
)`,
|
|
706
726
|
];
|
|
707
727
|
export const flair = (model) => [
|
|
708
|
-
`from flair.
|
|
728
|
+
`from flair.data import Sentence
|
|
729
|
+
from flair.models import SequenceTagger
|
|
730
|
+
|
|
731
|
+
tagger = SequenceTagger.load("${model.id}")
|
|
709
732
|
|
|
710
|
-
|
|
733
|
+
sentence = Sentence("George Washington went to Washington.")
|
|
734
|
+
tagger.predict(sentence)
|
|
735
|
+
print(sentence)`,
|
|
711
736
|
];
|
|
712
737
|
export const flexray = (model) => [
|
|
713
738
|
`pip install flexray`,
|
|
@@ -783,21 +808,28 @@ print(f'Accuracy {accuracy_score(y_test, predictions):.2%}')
|
|
|
783
808
|
return [installSnippet, classificationSnippet, regressionsSnippet, matchingSnippet];
|
|
784
809
|
};
|
|
785
810
|
export const gaussianformer = (model) => [
|
|
786
|
-
`#
|
|
787
|
-
|
|
788
|
-
from gaussianformer.pipelines.rendering_pipeline import GaussianFormerRenderingPipeline
|
|
811
|
+
`# pip install "gaussianformer @ git+https://github.com/SVLwoof/gaussianformer"
|
|
812
|
+
from gaussianformer import GaussianFormerRenderingPipeline, load_ply
|
|
789
813
|
|
|
790
|
-
pipeline = GaussianFormerRenderingPipeline.from_pretrained("${model.id}")
|
|
814
|
+
pipeline = GaussianFormerRenderingPipeline.from_pretrained("${model.id}").to("cuda")
|
|
815
|
+
gaussians = load_ply("object.ply") # 3D Gaussian Splatting PLY -> [N, 14]`,
|
|
791
816
|
];
|
|
792
817
|
export const gliner = (model) => [
|
|
793
818
|
`from gliner import GLiNER
|
|
794
819
|
|
|
795
|
-
model = GLiNER.from_pretrained("${model.id}")
|
|
820
|
+
model = GLiNER.from_pretrained("${model.id}")
|
|
821
|
+
|
|
822
|
+
text = "Cristiano Ronaldo dos Santos Aveiro was born on 5 February 1985 in Funchal, Madeira, Portugal."
|
|
823
|
+
labels = ["person", "date", "location"]
|
|
824
|
+
|
|
825
|
+
entities = model.predict_entities(text, labels)
|
|
826
|
+
for entity in entities:
|
|
827
|
+
print(entity["text"], "=>", entity["label"])`,
|
|
796
828
|
];
|
|
797
829
|
export const gliner2 = (model) => [
|
|
798
|
-
`from gliner2 import
|
|
830
|
+
`from gliner2 import AutoExtractor
|
|
799
831
|
|
|
800
|
-
|
|
832
|
+
extractor = AutoExtractor.from_pretrained("${model.id}")
|
|
801
833
|
|
|
802
834
|
# Extract entities
|
|
803
835
|
text = "Apple CEO Tim Cook announced iPhone 15 in Cupertino yesterday."
|
|
@@ -809,7 +841,7 @@ export const indextts = (model) => [
|
|
|
809
841
|
`# Download model
|
|
810
842
|
from huggingface_hub import snapshot_download
|
|
811
843
|
|
|
812
|
-
snapshot_download(${model.id}, local_dir="checkpoints")
|
|
844
|
+
snapshot_download("${model.id}", local_dir="checkpoints")
|
|
813
845
|
|
|
814
846
|
from indextts.infer import IndexTTS
|
|
815
847
|
|
|
@@ -863,9 +895,11 @@ print(out["answers"]["team"]["choice"])`,
|
|
|
863
895
|
];
|
|
864
896
|
};
|
|
865
897
|
export const keras = (model) => [
|
|
866
|
-
`#
|
|
898
|
+
`# !pip install -U keras tensorflow huggingface_hub
|
|
899
|
+
# Keras needs TensorFlow installed to read "hf://" paths, so the tensorflow backend is selected here;
|
|
900
|
+
# "jax" and "torch" also work for computation once TensorFlow is installed.
|
|
867
901
|
import os
|
|
868
|
-
os.environ["KERAS_BACKEND"] = "
|
|
902
|
+
os.environ["KERAS_BACKEND"] = "tensorflow"
|
|
869
903
|
|
|
870
904
|
import keras
|
|
871
905
|
|
|
@@ -979,12 +1013,14 @@ export const kernels = (model) => [
|
|
|
979
1013
|
|
|
980
1014
|
from kernels import get_kernel
|
|
981
1015
|
|
|
982
|
-
|
|
1016
|
+
# a version (or an explicit revision) is required; see the "Files and versions" tab for the available ones
|
|
1017
|
+
kernel = get_kernel("${model.id}", version=1)`,
|
|
983
1018
|
];
|
|
984
1019
|
export const kimi_audio = (model) => [
|
|
985
1020
|
`# Example usage for KimiAudio
|
|
986
|
-
# pip install git+https://github.com/MoonshotAI/Kimi-Audio.git
|
|
1021
|
+
# pip install git+https://github.com/MoonshotAI/Kimi-Audio.git "transformers<5"
|
|
987
1022
|
|
|
1023
|
+
import soundfile as sf
|
|
988
1024
|
from kimia_infer.api.kimia import KimiAudio
|
|
989
1025
|
|
|
990
1026
|
model = KimiAudio(model_path="${model.id}", load_detokenizer=True)
|
|
@@ -1348,13 +1384,16 @@ pip install -U litert-lm
|
|
|
1348
1384
|
|
|
1349
1385
|
# 2. Download and run this model locally:
|
|
1350
1386
|
# See: https://ai.google.dev/edge/litert-lm/cli
|
|
1387
|
+
# A single .litertlm file in the repo is picked automatically; otherwise the CLI asks which one to run
|
|
1388
|
+
# (or pass its name right after the repo id).
|
|
1351
1389
|
litert-lm run \\
|
|
1352
1390
|
--from-huggingface-repo=${model.id} \\
|
|
1353
1391
|
--prompt="Write me a poem"`,
|
|
1354
1392
|
];
|
|
1355
1393
|
export const tf_keras = (model) => [
|
|
1356
|
-
`# Note: 'keras<3.x' or 'tf_keras' must be installed (legacy)
|
|
1394
|
+
`# Note: 'keras<3.x' or 'tf_keras' must be installed (legacy), and from_pretrained_keras was removed in huggingface_hub 1.0.
|
|
1357
1395
|
# See https://github.com/keras-team/tf-keras for more details.
|
|
1396
|
+
# !pip install "huggingface_hub<1.0" tf_keras
|
|
1358
1397
|
from huggingface_hub import from_pretrained_keras
|
|
1359
1398
|
|
|
1360
1399
|
model = from_pretrained_keras("${model.id}")
|
|
@@ -1400,6 +1439,16 @@ from MeshAnything.models.meshanything import MeshAnything
|
|
|
1400
1439
|
# and https://github.com/buaacyw/MeshAnything/blob/main/app.py regarding usage
|
|
1401
1440
|
model = MeshAnything(args)`,
|
|
1402
1441
|
];
|
|
1442
|
+
export const microduck = (model) => {
|
|
1443
|
+
// Model-card tag convention: tags: [microduck-slot:stand].
|
|
1444
|
+
const slots = ["walk", "stand", "sitstand", "ground_pick", "kick_left", "kick_right", "roulade"];
|
|
1445
|
+
const matchingSlots = slots.filter((slot) => model.tags.includes(`microduck-slot:${slot}`));
|
|
1446
|
+
const slot = matchingSlots.length === 1 ? matchingSlots[0] : "SLOT";
|
|
1447
|
+
const prefix = matchingSlots.length === 1
|
|
1448
|
+
? ""
|
|
1449
|
+
: `# Replace SLOT with the slot specified in the model card (${slots.join(", ")}).\n`;
|
|
1450
|
+
return [`${prefix}sudo robotctl policy load ${slot} ${model.id}`];
|
|
1451
|
+
};
|
|
1403
1452
|
export const multimolecule = (model) => {
|
|
1404
1453
|
const widgetExample = model.widgetData?.[0];
|
|
1405
1454
|
const exampleText = escapeStringForJson(widgetExample?.text ?? "");
|
|
@@ -1573,14 +1622,18 @@ model = pe.VisionTransformer.from_config("${model.id}", pretrained=True)`;
|
|
|
1573
1622
|
};
|
|
1574
1623
|
export const phantom_wan = (model) => [
|
|
1575
1624
|
`from huggingface_hub import snapshot_download
|
|
1576
|
-
from phantom_wan import
|
|
1625
|
+
from phantom_wan import WanI2V, configs
|
|
1626
|
+
from PIL import Image
|
|
1577
1627
|
|
|
1578
1628
|
checkpoint_dir = snapshot_download("${model.id}")
|
|
1579
1629
|
wan_i2v = WanI2V(
|
|
1580
|
-
|
|
1581
|
-
|
|
1582
|
-
|
|
1583
|
-
|
|
1630
|
+
config=configs.WAN_CONFIGS['i2v-14B'],
|
|
1631
|
+
checkpoint_dir=checkpoint_dir,
|
|
1632
|
+
)
|
|
1633
|
+
|
|
1634
|
+
text_prompt = "A cat playing with a ball of yarn"
|
|
1635
|
+
image_prompt = Image.open("path/to/image.jpg").convert("RGB")
|
|
1636
|
+
video = wan_i2v.generate(text_prompt, image_prompt)`,
|
|
1584
1637
|
];
|
|
1585
1638
|
export const pocket_tts = (model) => [
|
|
1586
1639
|
`from pocket_tts import TTSModel
|
|
@@ -1694,7 +1747,7 @@ export const tensorflowtts = (model) => {
|
|
|
1694
1747
|
export const timm = (model) => [
|
|
1695
1748
|
`import timm
|
|
1696
1749
|
|
|
1697
|
-
model = timm.create_model("
|
|
1750
|
+
model = timm.create_model("hf-hub:${model.id}", pretrained=True)`,
|
|
1698
1751
|
];
|
|
1699
1752
|
export const saelens = ( /* model: ModelData */) => [
|
|
1700
1753
|
`# pip install sae-lens
|
|
@@ -1719,13 +1772,14 @@ llm = hydra.utils.instantiate(llm_cfg, torch_dtype="fp16")
|
|
|
1719
1772
|
seed_story_cfg = OmegaConf.load(seed_story_cfg_path)
|
|
1720
1773
|
seed_story = hydra.utils.instantiate(seed_story_cfg, llm=llm) `,
|
|
1721
1774
|
];
|
|
1722
|
-
const
|
|
1775
|
+
const SKLEARN_PICKLE_FILE = /\.(joblib|pkl|pickle)$/i;
|
|
1776
|
+
// skops.hub_utils was removed in skops 0.12, so files are downloaded with huggingface_hub
|
|
1777
|
+
const sklearnPickle = (model, modelFile) => {
|
|
1723
1778
|
return [
|
|
1724
|
-
`import
|
|
1725
|
-
|
|
1726
|
-
download("${model.id}", "path_to_folder")
|
|
1779
|
+
`from huggingface_hub import hf_hub_download
|
|
1780
|
+
import joblib
|
|
1727
1781
|
model = joblib.load(
|
|
1728
|
-
"${escapeStringForJson(modelFile)}"
|
|
1782
|
+
hf_hub_download("${model.id}", "${escapeStringForJson(modelFile)}")
|
|
1729
1783
|
)
|
|
1730
1784
|
# only load pickle files from sources you trust
|
|
1731
1785
|
# read more about it here https://skops.readthedocs.io/en/stable/persistence.html`,
|
|
@@ -1733,42 +1787,28 @@ model = joblib.load(
|
|
|
1733
1787
|
};
|
|
1734
1788
|
const skopsFormat = (model, modelFile) => {
|
|
1735
1789
|
return [
|
|
1736
|
-
`from
|
|
1790
|
+
`from huggingface_hub import hf_hub_download
|
|
1737
1791
|
from skops.io import load
|
|
1738
|
-
download("${model.id}", "path_to_folder")
|
|
1739
1792
|
# make sure model file is in skops format
|
|
1740
1793
|
# if model is a pickle file, make sure it's from a source you trust
|
|
1741
|
-
model = load("
|
|
1742
|
-
];
|
|
1743
|
-
};
|
|
1744
|
-
const skopsJobLib = (model) => {
|
|
1745
|
-
return [
|
|
1746
|
-
`from huggingface_hub import hf_hub_download
|
|
1747
|
-
import joblib
|
|
1748
|
-
model = joblib.load(
|
|
1749
|
-
hf_hub_download("${model.id}", "sklearn_model.joblib")
|
|
1750
|
-
)
|
|
1751
|
-
# only load pickle files from sources you trust
|
|
1752
|
-
# read more about it here https://skops.readthedocs.io/en/stable/persistence.html`,
|
|
1794
|
+
model = load(hf_hub_download("${model.id}", "${escapeStringForJson(modelFile)}"))`,
|
|
1753
1795
|
];
|
|
1754
1796
|
};
|
|
1755
1797
|
export const sklearn = (model) => {
|
|
1798
|
+
const declaredFile = model.config?.sklearn?.model?.file;
|
|
1756
1799
|
if (model.tags.includes("skops")) {
|
|
1757
|
-
|
|
1758
|
-
const skopssaveFormat = model.config?.sklearn?.model_format;
|
|
1759
|
-
if (!skopsmodelFile) {
|
|
1800
|
+
if (!declaredFile) {
|
|
1760
1801
|
return [`# ⚠️ Model filename not specified in config.json`];
|
|
1761
1802
|
}
|
|
1762
|
-
|
|
1763
|
-
|
|
1764
|
-
|
|
1765
|
-
else {
|
|
1766
|
-
return skopsFormat(model, skopsmodelFile);
|
|
1803
|
+
// some skops repos declare no model_format for a pickled file (e.g. julien-c/skops-digits), which skops.io cannot read
|
|
1804
|
+
if (model.config?.sklearn?.model_format === "pickle" || SKLEARN_PICKLE_FILE.test(declaredFile)) {
|
|
1805
|
+
return sklearnPickle(model, declaredFile);
|
|
1767
1806
|
}
|
|
1807
|
+
return skopsFormat(model, declaredFile);
|
|
1768
1808
|
}
|
|
1769
|
-
|
|
1770
|
-
|
|
1771
|
-
|
|
1809
|
+
// repos pushed without skops often still declare their filename in config.json; it is only trusted when it names a
|
|
1810
|
+
// pickle, since some hand-written configs carry typos (e.g. "sklean_model.jpblib" next to a real sklearn_model.joblib)
|
|
1811
|
+
return sklearnPickle(model, declaredFile && SKLEARN_PICKLE_FILE.test(declaredFile) ? declaredFile : "sklearn_model.joblib");
|
|
1772
1812
|
};
|
|
1773
1813
|
export const stable_audio_tools = (model) => [
|
|
1774
1814
|
`import torch
|
|
@@ -1816,7 +1856,7 @@ export const sam2 = (model) => {
|
|
|
1816
1856
|
import torch
|
|
1817
1857
|
from sam2.sam2_image_predictor import SAM2ImagePredictor
|
|
1818
1858
|
|
|
1819
|
-
predictor = SAM2ImagePredictor.from_pretrained(${model.id})
|
|
1859
|
+
predictor = SAM2ImagePredictor.from_pretrained("${model.id}")
|
|
1820
1860
|
|
|
1821
1861
|
with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
|
|
1822
1862
|
predictor.set_image(<your_image>)
|
|
@@ -1825,13 +1865,13 @@ with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
|
|
|
1825
1865
|
import torch
|
|
1826
1866
|
from sam2.sam2_video_predictor import SAM2VideoPredictor
|
|
1827
1867
|
|
|
1828
|
-
predictor = SAM2VideoPredictor.from_pretrained(${model.id})
|
|
1868
|
+
predictor = SAM2VideoPredictor.from_pretrained("${model.id}")
|
|
1829
1869
|
|
|
1830
1870
|
with torch.inference_mode(), torch.autocast("cuda", dtype=torch.bfloat16):
|
|
1831
1871
|
state = predictor.init_state(<your_video>)
|
|
1832
1872
|
|
|
1833
1873
|
# add new prompts and instantly get the output on the same frame
|
|
1834
|
-
frame_idx, object_ids, masks = predictor.add_new_points(state, <your_prompts>)
|
|
1874
|
+
frame_idx, object_ids, masks = predictor.add_new_points(state, <your_prompts>)
|
|
1835
1875
|
|
|
1836
1876
|
# propagate the prompts to get masklets throughout the video
|
|
1837
1877
|
for frame_idx, object_ids, masks in predictor.propagate_in_video(state):
|
|
@@ -1851,11 +1891,18 @@ mask = load_single_mask("path_to_mask.png", index=14)
|
|
|
1851
1891
|
output = inference(image, mask)`,
|
|
1852
1892
|
];
|
|
1853
1893
|
export const sam_3d_body = (model) => [
|
|
1854
|
-
`
|
|
1894
|
+
`import cv2
|
|
1895
|
+
import numpy as np
|
|
1896
|
+
from notebook.utils import setup_sam_3d_body
|
|
1897
|
+
from tools.vis_utils import visualize_sample_together
|
|
1898
|
+
|
|
1899
|
+
estimator = setup_sam_3d_body("${model.id}")
|
|
1855
1900
|
|
|
1856
|
-
|
|
1857
|
-
outputs = estimator.process_one_image(
|
|
1858
|
-
|
|
1901
|
+
img_bgr = cv2.imread("path/to/image.jpg")
|
|
1902
|
+
outputs = estimator.process_one_image(cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB))
|
|
1903
|
+
|
|
1904
|
+
rend_img = visualize_sample_together(img_bgr, outputs, estimator.faces)
|
|
1905
|
+
cv2.imwrite("output.jpg", rend_img.astype(np.uint8))`,
|
|
1859
1906
|
];
|
|
1860
1907
|
export const sampleFactory = (model) => [
|
|
1861
1908
|
`python -m sample_factory.huggingface.load_from_hub -r ${model.id} -d ./train_dir`,
|
|
@@ -1946,7 +1993,15 @@ queries_emb = model.encode(queries, is_query=True)
|
|
|
1946
1993
|
docs_emb = model.encode(documents, is_query=False)`,
|
|
1947
1994
|
];
|
|
1948
1995
|
}
|
|
1949
|
-
|
|
1996
|
+
// Rerankers saved in the CrossEncoder format are often tagged text-classification rather than text-ranking;
|
|
1997
|
+
// their architecture still identifies them (e.g. BAAI/bge-reranker-v2-m3). SetFit models can share that shape but
|
|
1998
|
+
// are bi-encoders with a separate head.
|
|
1999
|
+
const isCrossEncoder = model.tags.includes("cross-encoder") ||
|
|
2000
|
+
model.pipeline_tag == "text-ranking" ||
|
|
2001
|
+
(model.pipeline_tag === "text-classification" &&
|
|
2002
|
+
!model.tags.includes("setfit") &&
|
|
2003
|
+
(model.config?.architectures ?? []).some((arch) => arch.endsWith("ForSequenceClassification")));
|
|
2004
|
+
if (isCrossEncoder) {
|
|
1950
2005
|
return [
|
|
1951
2006
|
`from sentence_transformers import CrossEncoder
|
|
1952
2007
|
|
|
@@ -1964,6 +2019,33 @@ scores = model.predict([(query, passage) for passage in passages])
|
|
|
1964
2019
|
print(scores)`,
|
|
1965
2020
|
];
|
|
1966
2021
|
}
|
|
2022
|
+
// sentence-transformers tags SparseEncoder repos (SPLADE & co) with "sparse-encoder" and MultiVectorEncoder repos
|
|
2023
|
+
// (ColBERT & co) with "multi-vector" when pushing them; both encode queries and documents separately
|
|
2024
|
+
const retrievalEncoder = model.tags.includes("sparse-encoder")
|
|
2025
|
+
? "SparseEncoder"
|
|
2026
|
+
: model.tags.includes("multi-vector")
|
|
2027
|
+
? "MultiVectorEncoder"
|
|
2028
|
+
: undefined;
|
|
2029
|
+
if (retrievalEncoder) {
|
|
2030
|
+
return [
|
|
2031
|
+
`from sentence_transformers import ${retrievalEncoder}
|
|
2032
|
+
|
|
2033
|
+
model = ${retrievalEncoder}("${model.id}"${remote_code_snippet})
|
|
2034
|
+
|
|
2035
|
+
queries = ["Which planet is known as the Red Planet?"]
|
|
2036
|
+
documents = [
|
|
2037
|
+
"Venus is often called Earth's twin because of its similar size and proximity.",
|
|
2038
|
+
"Mars, known for its reddish appearance, is often referred to as the Red Planet.",
|
|
2039
|
+
"Jupiter, the largest planet in our solar system, has a prominent red spot.",
|
|
2040
|
+
]
|
|
2041
|
+
|
|
2042
|
+
query_embeddings = model.encode_query(queries)
|
|
2043
|
+
document_embeddings = model.encode_document(documents)
|
|
2044
|
+
|
|
2045
|
+
similarities = model.similarity(query_embeddings, document_embeddings)
|
|
2046
|
+
print(similarities)`,
|
|
2047
|
+
];
|
|
2048
|
+
}
|
|
1967
2049
|
const exampleSentences = get_widget_examples_from_st_model(model) ?? [
|
|
1968
2050
|
"The weather is lovely today.",
|
|
1969
2051
|
"It's so sunny outside!",
|
|
@@ -1985,7 +2067,10 @@ print(similarities.shape)
|
|
|
1985
2067
|
export const setfit = (model) => [
|
|
1986
2068
|
`from setfit import SetFitModel
|
|
1987
2069
|
|
|
1988
|
-
model = SetFitModel.from_pretrained("${model.id}")
|
|
2070
|
+
model = SetFitModel.from_pretrained("${model.id}")
|
|
2071
|
+
|
|
2072
|
+
preds = model.predict(["i loved the spiderman movie!", "pineapple on pizza is the worst"])
|
|
2073
|
+
print(preds)`,
|
|
1989
2074
|
];
|
|
1990
2075
|
export const spacy = (model) => [
|
|
1991
2076
|
`!pip install https://huggingface.co/${model.id}/resolve/main/${nameWithoutNamespace(model.id)}-any-py3-none-any.whl
|
|
@@ -2001,7 +2086,10 @@ nlp = ${nameWithoutNamespace(model.id)}.load()`,
|
|
|
2001
2086
|
export const span_marker = (model) => [
|
|
2002
2087
|
`from span_marker import SpanMarkerModel
|
|
2003
2088
|
|
|
2004
|
-
model = SpanMarkerModel.from_pretrained("${model.id}")
|
|
2089
|
+
model = SpanMarkerModel.from_pretrained("${model.id}")
|
|
2090
|
+
|
|
2091
|
+
entities = model.predict("Amelia Earhart flew her single engine Lockheed Vega 5B across the Atlantic to Paris.")
|
|
2092
|
+
print(entities)`,
|
|
2005
2093
|
];
|
|
2006
2094
|
export const stanza = (model) => [
|
|
2007
2095
|
`import stanza
|
|
@@ -2020,6 +2108,8 @@ const speechBrainMethod = (speechbrainInterface) => {
|
|
|
2020
2108
|
return "enhance_file";
|
|
2021
2109
|
case "SepformerSeparation":
|
|
2022
2110
|
return "separate_file";
|
|
2111
|
+
case "SpeakerRecognition":
|
|
2112
|
+
return "verify_files";
|
|
2023
2113
|
default:
|
|
2024
2114
|
return undefined;
|
|
2025
2115
|
}
|
|
@@ -2033,12 +2123,14 @@ export const speechbrain = (model) => {
|
|
|
2033
2123
|
if (speechbrainMethod === undefined) {
|
|
2034
2124
|
return [`# interface in config.json invalid`];
|
|
2035
2125
|
}
|
|
2126
|
+
// speaker verification compares two recordings
|
|
2127
|
+
const methodArgs = speechbrainMethod === "verify_files" ? '"speaker1.wav", "speaker2.wav"' : '"file.wav"';
|
|
2036
2128
|
return [
|
|
2037
2129
|
`from speechbrain.pretrained import ${speechbrainInterface}
|
|
2038
2130
|
model = ${speechbrainInterface}.from_hparams(
|
|
2039
2131
|
"${model.id}"
|
|
2040
2132
|
)
|
|
2041
|
-
model.${speechbrainMethod}(
|
|
2133
|
+
model.${speechbrainMethod}(${methodArgs})`,
|
|
2042
2134
|
];
|
|
2043
2135
|
};
|
|
2044
2136
|
export const terratorch = (model) => [
|
|
@@ -2068,7 +2160,7 @@ export const transformers = (model) => {
|
|
|
2068
2160
|
: info.processor === "AutoFeatureExtractor"
|
|
2069
2161
|
? "extractor"
|
|
2070
2162
|
: "processor";
|
|
2071
|
-
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")');
|
|
2163
|
+
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")');
|
|
2072
2164
|
if (model.tags.includes("conversational") && hasChatTemplate(model)) {
|
|
2073
2165
|
if (model.tags.includes("image-text-to-text")) {
|
|
2074
2166
|
autoSnippet.push("messages = [", [
|
|
@@ -2084,19 +2176,19 @@ export const transformers = (model) => {
|
|
|
2084
2176
|
else {
|
|
2085
2177
|
autoSnippet.push("messages = [", ' {"role": "user", "content": "Who are you?"},', "]");
|
|
2086
2178
|
}
|
|
2087
|
-
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=
|
|
2179
|
+
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]:]))`);
|
|
2088
2180
|
}
|
|
2089
2181
|
}
|
|
2090
2182
|
else {
|
|
2091
|
-
autoSnippet.push("# Load model directly", `from transformers import ${auto_model}`, `model = ${auto_model}.from_pretrained("${model.id}"` + remote_code_snippet + ', device_map="auto")');
|
|
2183
|
+
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")');
|
|
2092
2184
|
}
|
|
2093
2185
|
if (model.pipeline_tag && LIBRARY_TASK_MAPPING.transformers?.includes(model.pipeline_tag)) {
|
|
2094
2186
|
const pipelineSnippet = ["# Use a pipeline as a high-level helper"];
|
|
2095
2187
|
if (REMOVED_IN_V5_TRANSFORMERS_PIPELINES.includes(model.pipeline_tag)) {
|
|
2096
|
-
pipelineSnippet.push(`# Warning: Pipeline type "${model.pipeline_tag}" is no longer supported in transformers v5.`, `# You must load the model directly (see below) or downgrade to v4.x with:`, `#
|
|
2188
|
+
pipelineSnippet.push(`# Warning: Pipeline type "${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"`);
|
|
2097
2189
|
}
|
|
2098
2190
|
pipelineSnippet.push("from transformers import pipeline", "", `pipe = pipeline("${model.pipeline_tag}", model="${model.id}"` + remote_code_snippet + ")");
|
|
2099
|
-
if (model.tags.includes("conversational")) {
|
|
2191
|
+
if (model.tags.includes("conversational") && hasChatTemplate(model)) {
|
|
2100
2192
|
if (model.tags.includes("image-text-to-text")) {
|
|
2101
2193
|
pipelineSnippet.push("messages = [", [
|
|
2102
2194
|
" {",
|
|
@@ -2124,47 +2216,118 @@ export const transformers = (model) => {
|
|
|
2124
2216
|
}
|
|
2125
2217
|
return [autoSnippet.join("\n")];
|
|
2126
2218
|
};
|
|
2219
|
+
// Hub tasks that transformers.js exposes under a different pipeline() task name
|
|
2220
|
+
const TRANSFORMERS_JS_TASK_ALIASES = {
|
|
2221
|
+
"sentence-similarity": "feature-extraction",
|
|
2222
|
+
};
|
|
2127
2223
|
export const transformersJS = (model) => {
|
|
2128
2224
|
if (!model.pipeline_tag) {
|
|
2129
2225
|
return [`// ⚠️ Unknown pipeline tag`];
|
|
2130
2226
|
}
|
|
2131
2227
|
const libName = "@huggingface/transformers";
|
|
2228
|
+
// Rerankers score (query, document) pairs. The text-classification pipeline cannot take pairs and applies a softmax
|
|
2229
|
+
// over their single logit, so every score comes out as 1; the model is called directly instead.
|
|
2230
|
+
if (model.pipeline_tag === "text-ranking") {
|
|
2231
|
+
return [
|
|
2232
|
+
`// npm i ${libName}
|
|
2233
|
+
import { AutoTokenizer, AutoModelForSequenceClassification } from '${libName}';
|
|
2234
|
+
|
|
2235
|
+
const tokenizer = await AutoTokenizer.from_pretrained('${model.id}');
|
|
2236
|
+
const model = await AutoModelForSequenceClassification.from_pretrained('${model.id}');
|
|
2237
|
+
|
|
2238
|
+
const query = 'Which planet is known as the Red Planet?';
|
|
2239
|
+
const documents = [
|
|
2240
|
+
'Mars, known for its reddish appearance, is often referred to as the Red Planet.',
|
|
2241
|
+
'Venus is often called the twin of Earth because of its similar size and proximity.',
|
|
2242
|
+
];
|
|
2243
|
+
|
|
2244
|
+
const inputs = tokenizer(new Array(documents.length).fill(query), { text_pair: documents, padding: true, truncation: true });
|
|
2245
|
+
const { logits } = await model(inputs);
|
|
2246
|
+
console.log(logits.sigmoid().tolist()); // one relevance score per document`,
|
|
2247
|
+
];
|
|
2248
|
+
}
|
|
2249
|
+
const task = TRANSFORMERS_JS_TASK_ALIASES[model.pipeline_tag] ?? model.pipeline_tag;
|
|
2132
2250
|
return [
|
|
2133
2251
|
`// npm i ${libName}
|
|
2134
2252
|
import { pipeline } from '${libName}';
|
|
2135
2253
|
|
|
2136
2254
|
// Allocate pipeline
|
|
2137
|
-
const pipe = await pipeline('${
|
|
2255
|
+
const pipe = await pipeline('${task}', '${model.id}');`,
|
|
2138
2256
|
];
|
|
2139
2257
|
};
|
|
2140
|
-
const
|
|
2141
|
-
|
|
2142
|
-
|
|
2143
|
-
|
|
2144
|
-
|
|
2145
|
-
|
|
2146
|
-
|
|
2147
|
-
|
|
2148
|
-
|
|
2149
|
-
|
|
2150
|
-
|
|
2151
|
-
|
|
2152
|
-
|
|
2258
|
+
const PEFT_TASK_TYPE_TO_AUTO_CLASS = {
|
|
2259
|
+
CAUSAL_LM: "AutoModelForCausalLM",
|
|
2260
|
+
SEQ_2_SEQ_LM: "AutoModelForSeq2SeqLM",
|
|
2261
|
+
TOKEN_CLS: "AutoModelForTokenClassification",
|
|
2262
|
+
SEQ_CLS: "AutoModelForSequenceClassification",
|
|
2263
|
+
QUESTION_ANS: "AutoModelForQuestionAnswering",
|
|
2264
|
+
// transformers has no AutoModelForFeatureExtraction; PEFT's own AutoPeftModelForFeatureExtraction resolves to AutoModel
|
|
2265
|
+
FEATURE_EXTRACTION: "AutoModel",
|
|
2266
|
+
};
|
|
2267
|
+
// PEFT has no speech-specific task type: Whisper-style adapters are tagged SEQ_2_SEQ_LM, but transformers registers
|
|
2268
|
+
// Whisper and Speech2Text only under AutoModelForSpeechSeq2Seq (SeamlessM4T is under both seq2seq auto classes, SpeechT5
|
|
2269
|
+
// under the speech-to-text and text-to-spectrogram ones).
|
|
2270
|
+
// Other audio seq2seq models (Qwen2-Audio, GLM-ASR) stay under AutoModelForSeq2SeqLM, so the base model name is the
|
|
2271
|
+
// main signal. Matched against Hub repo ids, which are hyphenated (facebook/s2t-small-librispeech-asr,
|
|
2272
|
+
// facebook/seamless-m4t-v2-large), not against transformers model_type identifiers.
|
|
2273
|
+
const PEFT_SPEECH_TO_TEXT_BASE_MODEL = /whisper|\bs2t\b|speech[-_]to[-_]text/i;
|
|
2274
|
+
// SeamlessM4T also has a text-to-text head under AutoModelForSeq2SeqLM, which text translation adapters need; most
|
|
2275
|
+
// SeamlessM4T adapters on the Hub are speech ones, so the speech class stays the default when no text task is declared.
|
|
2276
|
+
const PEFT_SEAMLESS_M4T_BASE_MODEL = /seamless[-_]m4t/i;
|
|
2277
|
+
const PEFT_TEXT_TASKS = ["translation", "text2text-generation", "summarization"];
|
|
2278
|
+
// SpeechT5 checkpoints are overwhelmingly text-to-speech fine-tunes (often without "tts" in their name), which load
|
|
2279
|
+
// with AutoModelForTextToSpectrogram; AutoModelForSpeechSeq2Seq would silently load the speech-to-text head, so only
|
|
2280
|
+
// ASR bases or adapters get it.
|
|
2281
|
+
const PEFT_SPEECHT5_BASE_MODEL = /speecht5/i;
|
|
2282
|
+
const declaresTask = (model, task) => model.pipeline_tag === task || model.tags.includes(task);
|
|
2283
|
+
// the auto class a SEQ_2_SEQ_LM adapter needs when its base is a speech model, undefined for text seq2seq bases
|
|
2284
|
+
const peftSpeechAutoClass = (model, baseModel) => {
|
|
2285
|
+
if (PEFT_SPEECH_TO_TEXT_BASE_MODEL.test(baseModel)) {
|
|
2286
|
+
return "AutoModelForSpeechSeq2Seq";
|
|
2287
|
+
}
|
|
2288
|
+
if (PEFT_SEAMLESS_M4T_BASE_MODEL.test(baseModel)) {
|
|
2289
|
+
return PEFT_TEXT_TASKS.some((task) => declaresTask(model, task)) ? undefined : "AutoModelForSpeechSeq2Seq";
|
|
2290
|
+
}
|
|
2291
|
+
if (PEFT_SPEECHT5_BASE_MODEL.test(baseModel)) {
|
|
2292
|
+
const isAsr = /(?:^|[^a-z])asr(?:[^a-z]|$)/i.test(baseModel) || declaresTask(model, "automatic-speech-recognition");
|
|
2293
|
+
return isAsr ? "AutoModelForSpeechSeq2Seq" : "AutoModelForTextToSpectrogram";
|
|
2294
|
+
}
|
|
2295
|
+
return undefined;
|
|
2153
2296
|
};
|
|
2154
2297
|
export const peft = (model) => {
|
|
2155
2298
|
const { base_model_name_or_path: peftBaseModel, task_type: peftTaskType } = model.config?.peft ?? {};
|
|
2156
|
-
|
|
2157
|
-
if (!
|
|
2299
|
+
let autoClass = peftTaskType ? PEFT_TASK_TYPE_TO_AUTO_CLASS[peftTaskType] : undefined;
|
|
2300
|
+
if (!autoClass) {
|
|
2158
2301
|
return [`Task type is invalid.`];
|
|
2159
2302
|
}
|
|
2160
|
-
|
|
2303
|
+
// adapter_config.json sometimes omits the base model; the model card's base_model is the next best source
|
|
2304
|
+
const cardBaseModel = Array.isArray(model.cardData?.base_model)
|
|
2305
|
+
? model.cardData.base_model[0]
|
|
2306
|
+
: model.cardData?.base_model;
|
|
2307
|
+
const baseModel = peftBaseModel || cardBaseModel;
|
|
2308
|
+
if (!baseModel) {
|
|
2161
2309
|
return [`Base model is not found.`];
|
|
2162
2310
|
}
|
|
2311
|
+
if (peftTaskType === "SEQ_2_SEQ_LM") {
|
|
2312
|
+
autoClass = peftSpeechAutoClass(model, baseModel) ?? autoClass;
|
|
2313
|
+
}
|
|
2314
|
+
if (autoClass === "AutoModelForTextToSpectrogram") {
|
|
2315
|
+
return [
|
|
2316
|
+
`from peft import PeftConfig, PeftModel
|
|
2317
|
+
from transformers import AutoModelForTextToSpectrogram
|
|
2318
|
+
|
|
2319
|
+
base_model = AutoModelForTextToSpectrogram.from_pretrained("${escapeStringForJson(baseModel)}")
|
|
2320
|
+
# PEFT's seq2seq wrapper needs generate(), which SpeechT5's text-to-speech model lacks: load the adapter without its task type
|
|
2321
|
+
config = PeftConfig.from_pretrained("${model.id}")
|
|
2322
|
+
config.task_type = None
|
|
2323
|
+
model = PeftModel.from_pretrained(base_model, "${model.id}", config=config)`,
|
|
2324
|
+
];
|
|
2325
|
+
}
|
|
2163
2326
|
return [
|
|
2164
2327
|
`from peft import PeftModel
|
|
2165
|
-
from transformers import
|
|
2328
|
+
from transformers import ${autoClass}
|
|
2166
2329
|
|
|
2167
|
-
base_model =
|
|
2330
|
+
base_model = ${autoClass}.from_pretrained("${escapeStringForJson(baseModel)}")
|
|
2168
2331
|
model = PeftModel.from_pretrained(base_model, "${model.id}")`,
|
|
2169
2332
|
];
|
|
2170
2333
|
};
|
|
@@ -2195,7 +2358,7 @@ transcriptions = asr_model.transcribe(["file.wav"])`,
|
|
|
2195
2358
|
}
|
|
2196
2359
|
};
|
|
2197
2360
|
export const mlAgents = (model) => [
|
|
2198
|
-
`mlagents-load-from-hf --repo-id="${model.id}" --local-dir="./
|
|
2361
|
+
`mlagents-load-from-hf --repo-id="${model.id}" --local-dir="./downloads"`,
|
|
2199
2362
|
];
|
|
2200
2363
|
export const sentis = ( /* model: ModelData */) => [
|
|
2201
2364
|
`string modelName = "[Your model name here].sentis";
|
|
@@ -2264,10 +2427,10 @@ model = Model.from_pretrained("${model.id}")`,
|
|
|
2264
2427
|
];
|
|
2265
2428
|
export const lvface = (model) => [
|
|
2266
2429
|
`from huggingface_hub import hf_hub_download
|
|
2267
|
-
|
|
2430
|
+
from inference_onnx import LVFaceONNXInferencer
|
|
2268
2431
|
|
|
2269
2432
|
model_path = hf_hub_download("${model.id}", "LVFace-L_Glint360K/LVFace-L_Glint360K.onnx")
|
|
2270
|
-
inferencer = LVFaceONNXInferencer(model_path, use_gpu=True
|
|
2433
|
+
inferencer = LVFaceONNXInferencer(model_path, use_gpu=True)
|
|
2271
2434
|
img_path = 'path/to/image1.jpg'
|
|
2272
2435
|
embedding = inferencer.infer_from_image(img_path)`,
|
|
2273
2436
|
];
|
|
@@ -2303,8 +2466,7 @@ export const vui = () => [
|
|
|
2303
2466
|
|
|
2304
2467
|
import torchaudio
|
|
2305
2468
|
|
|
2306
|
-
from vui.
|
|
2307
|
-
from vui.model import Vui,
|
|
2469
|
+
from vui.legacy import Vui, render
|
|
2308
2470
|
|
|
2309
2471
|
model = Vui.from_pretrained().cuda()
|
|
2310
2472
|
waveform = render(
|
|
@@ -2312,15 +2474,16 @@ waveform = render(
|
|
|
2312
2474
|
"Hey, here is some random stuff, usually something quite long as the shorter the text the less likely the model can cope!",
|
|
2313
2475
|
)
|
|
2314
2476
|
print(waveform.shape)
|
|
2315
|
-
torchaudio.save("out.
|
|
2477
|
+
torchaudio.save("out.wav", waveform[0], 22050)
|
|
2316
2478
|
`,
|
|
2317
2479
|
];
|
|
2318
2480
|
export const chattts = () => [
|
|
2319
2481
|
`import ChatTTS
|
|
2482
|
+
import torch
|
|
2320
2483
|
import torchaudio
|
|
2321
2484
|
|
|
2322
2485
|
chat = ChatTTS.Chat()
|
|
2323
|
-
chat.
|
|
2486
|
+
chat.load(compile=False) # Set to True for better performance
|
|
2324
2487
|
|
|
2325
2488
|
texts = ["PUT YOUR TEXT HERE",]
|
|
2326
2489
|
|
|
@@ -2329,17 +2492,42 @@ wavs = chat.infer(texts, )
|
|
|
2329
2492
|
torchaudio.save("output1.wav", torch.from_numpy(wavs[0]), 24000)`,
|
|
2330
2493
|
];
|
|
2331
2494
|
export const ultralytics = (model) => {
|
|
2332
|
-
//
|
|
2333
|
-
|
|
2334
|
-
|
|
2335
|
-
|
|
2336
|
-
|
|
2337
|
-
|
|
2495
|
+
// the YOLOv10 fork (library "yolov10") ships its own `YOLOv10` class with a `from_pretrained` helper that reads the
|
|
2496
|
+
// model.safetensors it pushes. Mainline ultralytics repos often list "yolov10" among their version tags
|
|
2497
|
+
// (e.g. Ultralytics/YOLOv8) but only ship .pt files, so an ultralytics tag alone keeps them on the mainline loader;
|
|
2498
|
+
// fork pushes still tagged ultralytics are recognized by their safetensors weights.
|
|
2499
|
+
const isYolov10Fork = model.library_name === "yolov10" ||
|
|
2500
|
+
(model.tags.includes("yolov10") && (!model.tags.includes("ultralytics") || model.tags.includes("safetensors")));
|
|
2501
|
+
if (isYolov10Fork) {
|
|
2502
|
+
// from_pretrained reads the model.safetensors that fork pushes ship; older fork repos only ship .pt weights
|
|
2503
|
+
if (model.tags.includes("safetensors")) {
|
|
2504
|
+
return [
|
|
2505
|
+
`from ultralytics import YOLOv10
|
|
2506
|
+
|
|
2507
|
+
model = YOLOv10.from_pretrained("${model.id}")
|
|
2508
|
+
source = 'http://images.cocodataset.org/val2017/000000039769.jpg'
|
|
2509
|
+
model.predict(source=source, save=True)`,
|
|
2510
|
+
];
|
|
2511
|
+
}
|
|
2512
|
+
return [
|
|
2513
|
+
`from huggingface_hub import hf_hub_download
|
|
2514
|
+
from ultralytics import YOLOv10
|
|
2515
|
+
|
|
2516
|
+
# pick the weights file from this repo's "Files and versions" tab
|
|
2517
|
+
weights = hf_hub_download("${model.id}", "<weights>.pt")
|
|
2518
|
+
model = YOLOv10(weights)
|
|
2519
|
+
source = 'http://images.cocodataset.org/val2017/000000039769.jpg'
|
|
2520
|
+
model.predict(source=source, save=True)`,
|
|
2521
|
+
];
|
|
2522
|
+
}
|
|
2523
|
+
// mainline ultralytics loads weights from a local file
|
|
2338
2524
|
return [
|
|
2339
|
-
|
|
2340
|
-
|
|
2525
|
+
`from huggingface_hub import hf_hub_download
|
|
2526
|
+
from ultralytics import YOLO
|
|
2341
2527
|
|
|
2342
|
-
|
|
2528
|
+
# pick the weights file from this repo's "Files and versions" tab
|
|
2529
|
+
weights = hf_hub_download("${model.id}", "<weights>.pt")
|
|
2530
|
+
model = YOLO(weights)
|
|
2343
2531
|
source = 'http://images.cocodataset.org/val2017/000000039769.jpg'
|
|
2344
2532
|
model.predict(source=source, save=True)`,
|
|
2345
2533
|
];
|
|
@@ -2412,21 +2600,23 @@ model = SwarmFormerModel.from_pretrained("${model.id}")
|
|
|
2412
2600
|
];
|
|
2413
2601
|
export const univa = (model) => [
|
|
2414
2602
|
`# Follow installation instructions at https://github.com/PKU-YuanGroup/UniWorld-V1
|
|
2415
|
-
|
|
2603
|
+
import torch
|
|
2604
|
+
from transformers import AutoProcessor
|
|
2416
2605
|
from univa.models.qwen2p5vl.modeling_univa_qwen2p5vl import UnivaQwen2p5VLForConditionalGeneration
|
|
2417
|
-
|
|
2418
|
-
|
|
2419
|
-
|
|
2420
|
-
|
|
2421
|
-
|
|
2422
|
-
|
|
2606
|
+
|
|
2607
|
+
model = UnivaQwen2p5VLForConditionalGeneration.from_pretrained(
|
|
2608
|
+
"${model.id}",
|
|
2609
|
+
torch_dtype=torch.bfloat16,
|
|
2610
|
+
attn_implementation="flash_attention_2",
|
|
2611
|
+
).to("cuda")
|
|
2612
|
+
processor = AutoProcessor.from_pretrained("${model.id}")
|
|
2423
2613
|
`,
|
|
2424
2614
|
];
|
|
2425
2615
|
const mlx_unknown = (model) => [
|
|
2426
2616
|
`# Download the model from the Hub
|
|
2427
2617
|
pip install huggingface_hub[hf_xet]
|
|
2428
2618
|
|
|
2429
|
-
|
|
2619
|
+
hf download ${model.id} --local-dir ${nameWithoutNamespace(model.id)}`,
|
|
2430
2620
|
];
|
|
2431
2621
|
const mlxlm = (model) => [
|
|
2432
2622
|
`# Make sure mlx-lm is installed
|
|
@@ -2483,10 +2673,11 @@ formatted_prompt = apply_chat_template(
|
|
|
2483
2673
|
output = generate(model, processor, formatted_prompt, image)
|
|
2484
2674
|
print(output)`,
|
|
2485
2675
|
];
|
|
2676
|
+
// create_model() takes a name from the mlx-image registry, not a repo id; the weights live in "mlx-vision/<name>-mlxim"
|
|
2486
2677
|
export const mlxim = (model) => [
|
|
2487
2678
|
`from mlxim.model import create_model
|
|
2488
2679
|
|
|
2489
|
-
model = create_model(${model.id})`,
|
|
2680
|
+
model = create_model("${nameWithoutNamespace(model.id).replace(/-mlxim$/, "")}")`,
|
|
2490
2681
|
];
|
|
2491
2682
|
export const mlx = (model) => {
|
|
2492
2683
|
if (model.pipeline_tag === "image-text-to-text") {
|
|
@@ -2505,7 +2696,10 @@ export const mlx = (model) => {
|
|
|
2505
2696
|
export const model2vec = (model) => [
|
|
2506
2697
|
`from model2vec import StaticModel
|
|
2507
2698
|
|
|
2508
|
-
model = StaticModel.from_pretrained("${model.id}")
|
|
2699
|
+
model = StaticModel.from_pretrained("${model.id}")
|
|
2700
|
+
|
|
2701
|
+
embeddings = model.encode(["It's dangerous to go alone!", "It's a secret to everybody."])
|
|
2702
|
+
print(embeddings.shape)`,
|
|
2509
2703
|
];
|
|
2510
2704
|
export const mobilint = (model) => {
|
|
2511
2705
|
const modelName = nameWithoutNamespace(model.id);
|
|
@@ -2548,7 +2742,8 @@ export const pruna = (model) => {
|
|
|
2548
2742
|
};
|
|
2549
2743
|
snippets = snippets.map(ensurePrunaModelImport);
|
|
2550
2744
|
if (model.tags.includes("pruna_pro-ai")) {
|
|
2551
|
-
|
|
2745
|
+
// only rewrite the import and the class name, never the repo id (an org name can contain "pruna")
|
|
2746
|
+
return snippets.map((snippet) => snippet.replace(/^from pruna import /gm, "from pruna_pro import ").replace(/\bPrunaModel\b/g, "PrunaProModel"));
|
|
2552
2747
|
}
|
|
2553
2748
|
return snippets;
|
|
2554
2749
|
};
|
|
@@ -2606,21 +2801,19 @@ export const outetts = (model) => {
|
|
|
2606
2801
|
}
|
|
2607
2802
|
// v1.0 HF → minimal runnable snippet
|
|
2608
2803
|
return [
|
|
2609
|
-
`
|
|
2610
|
-
|
|
2611
|
-
|
|
2612
|
-
|
|
2613
|
-
|
|
2614
|
-
|
|
2615
|
-
|
|
2616
|
-
|
|
2617
|
-
|
|
2618
|
-
|
|
2619
|
-
|
|
2620
|
-
|
|
2621
|
-
|
|
2622
|
-
).save("output.wav")
|
|
2623
|
-
`,
|
|
2804
|
+
`import outetts
|
|
2805
|
+
|
|
2806
|
+
enum = outetts.Models("${model.id}".split("/", 1)[1]) # VERSION_1_0_SIZE_1B
|
|
2807
|
+
cfg = outetts.ModelConfig.auto_config(enum, outetts.Backend.HF)
|
|
2808
|
+
tts = outetts.Interface(cfg)
|
|
2809
|
+
|
|
2810
|
+
speaker = tts.load_default_speaker("EN-FEMALE-1-NEUTRAL")
|
|
2811
|
+
tts.generate(
|
|
2812
|
+
outetts.GenerationConfig(
|
|
2813
|
+
text="Hello there, how are you doing?",
|
|
2814
|
+
speaker=speaker,
|
|
2815
|
+
)
|
|
2816
|
+
).save("output.wav")`,
|
|
2624
2817
|
];
|
|
2625
2818
|
};
|
|
2626
2819
|
export const pxia = (model) => [
|