Merge branch 'main' into webgpu-fixes

Author: xenova

.github/workflows/tests.yml

@@ -13,6 +13,9 @@ on:
       - synchronize
       - ready_for_review
 
+env:
+  TESTING_REMOTELY: true
+
 jobs:
   build:
     if: github.event.pull_request.draft == false

README.md

@@ -401,6 +401,7 @@ You can refine your search by selecting the task you're interested in (e.g., [te
 1. **PyAnnote** released in the repository [pyannote/pyannote-audio](https://github.com/pyannote/pyannote-audio) by Hervé Bredin.
 1. **[Qwen2](https://huggingface.co/docs/transformers/model_doc/qwen2)** (from the Qwen team, Alibaba Group) released with the paper [Qwen Technical Report](https://arxiv.org/abs/2309.16609) by Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou and Tianhang Zhu.
 1. **[Qwen2-VL](https://huggingface.co/docs/transformers/model_doc/qwen2_vl)** (from the Qwen team, Alibaba Group) released with the paper [Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond](https://arxiv.org/abs/2308.12966) by Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, Jingren Zhou.
+1. **[Qwen3](https://huggingface.co/docs/transformers/en/model_doc/qwen3)** (from the Qwen team, Alibaba Group) released with the blog post [Qwen3: Think Deeper, Act Faster](https://qwenlm.github.io/blog/qwen3/) by the Qwen team.
 1. **[ResNet](https://huggingface.co/docs/transformers/model_doc/resnet)** (from Microsoft Research) released with the paper [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) by Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun.
 1. **[RF-DETR](https://huggingface.co/docs/transformers/model_doc/rf_detr)** (from Roboflow) released with the blog post [RF-DETR: A SOTA Real-Time Object Detection Model](https://blog.roboflow.com/rf-detr/) by Peter Robicheaux, James Gallagher, Joseph Nelson, Isaac Robinson.
 1. **[RoBERTa](https://huggingface.co/docs/transformers/model_doc/roberta)** (from Facebook), released together with the paper [RoBERTa: A Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.

docs/snippets/6_supported-models.snippet

@@ -115,6 +115,7 @@
 1. **PyAnnote** released in the repository [pyannote/pyannote-audio](https://github.com/pyannote/pyannote-audio) by Hervé Bredin.
 1. **[Qwen2](https://huggingface.co/docs/transformers/model_doc/qwen2)** (from the Qwen team, Alibaba Group) released with the paper [Qwen Technical Report](https://arxiv.org/abs/2309.16609) by Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou and Tianhang Zhu.
 1. **[Qwen2-VL](https://huggingface.co/docs/transformers/model_doc/qwen2_vl)** (from the Qwen team, Alibaba Group) released with the paper [Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond](https://arxiv.org/abs/2308.12966) by Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, Jingren Zhou.
+1. **[Qwen3](https://huggingface.co/docs/transformers/en/model_doc/qwen3)** (from the Qwen team, Alibaba Group) released with the blog post [Qwen3: Think Deeper, Act Faster](https://qwenlm.github.io/blog/qwen3/) by the Qwen team.
 1. **[ResNet](https://huggingface.co/docs/transformers/model_doc/resnet)** (from Microsoft Research) released with the paper [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) by Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun.
 1. **[RF-DETR](https://huggingface.co/docs/transformers/model_doc/rf_detr)** (from Roboflow) released with the blog post [RF-DETR: A SOTA Real-Time Object Detection Model](https://blog.roboflow.com/rf-detr/) by Peter Robicheaux, James Gallagher, Joseph Nelson, Isaac Robinson.
 1. **[RoBERTa](https://huggingface.co/docs/transformers/model_doc/roberta)** (from Facebook), released together with the paper [RoBERTa: A Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.

package-lock.json

@@ -55,7 +55,7 @@
   },
   "homepage": "https://github.com/huggingface/transformers.js#readme",
   "dependencies": {
-    "@huggingface/jinja": "^0.3.4",
+    "@huggingface/jinja": "^0.4.0",
     "onnxruntime-node": "1.21.0",
     "onnxruntime-web": "1.22.0-dev.20250409-89f8206ba4",
     "sharp": "^0.34.1"

package.json

@@ -619,6 +619,10 @@ export class ImageProcessor extends Callable {
         this.pad_size = config.pad_size;
         // @ts-expect-error TS2339
         this.do_pad = config.do_pad;
+        // @ts-expect-error TS2339
+        this.min_pixels = config.min_pixels;
+        // @ts-expect-error TS2339
+        this.max_pixels = config.max_pixels;
 
         if (this.do_pad && !this.pad_size && this.size && this.size.width !== undefined && this.size.height !== undefined) {
             // Should pad, but no pad size specified
@@ -892,12 +896,11 @@ export class ImageProcessor extends Callable {
 
         } else if (this.size_divisibility !== undefined) {
             return enforce_size_divisibility([srcWidth, srcHeight], this.size_divisibility);
-        } else if (size.min_pixels !== undefined && size.max_pixels !== undefined) {
+        } else if (this.min_pixels !== undefined && this.max_pixels !== undefined) {
             // Custom resize logic for Qwen2-VL models
-            const { min_pixels, max_pixels } = size;
             // @ts-expect-error TS2339
             const factor = this.config.patch_size * this.config.merge_size;
-            return smart_resize(srcHeight, srcWidth, factor, min_pixels, max_pixels);
+            return smart_resize(srcHeight, srcWidth, factor, this.min_pixels, this.max_pixels);
         } else {
             throw new Error(`Could not resize image due to unsupported \`this.size\` option in config: ${JSON.stringify(size)}`);
         }

src/base/image_processors_utils.js

@@ -125,6 +125,7 @@ function getNormalizedConfig(config) {
             mapping['hidden_size'] = 'hidden_size';
             mapping['num_attention_heads'] = 'num_attention_heads';
             break;
+        case 'qwen3':
         case 'gemma':
         case 'gemma2':
         case 'gemma3_text':

src/configs.js

@@ -4588,6 +4588,22 @@ export class Qwen2Model extends Qwen2PreTrainedModel { }
 export class Qwen2ForCausalLM extends Qwen2PreTrainedModel { }
 //////////////////////////////////////////////////
 
+
+//////////////////////////////////////////////////
+// Qwen3 models
+
+/**
+ * The bare Qwen3 Model outputting raw hidden-states without any specific head on top.
+ */
+export class Qwen3PreTrainedModel extends PreTrainedModel { }
+/**
+ * The bare Qwen3 Model outputting raw hidden-states without any specific head on top.
+ */
+export class Qwen3Model extends Qwen3PreTrainedModel { }
+
+export class Qwen3ForCausalLM extends Qwen3PreTrainedModel { }
+//////////////////////////////////////////////////
+
 export class Qwen2VLPreTrainedModel extends PreTrainedModel {
     forward_params = [
         // Text inputs
@@ -7637,6 +7653,7 @@ const MODEL_MAPPING_NAMES_DECODER_ONLY = new Map([
     ['glm', ['GlmModel', GlmModel]],
     ['openelm', ['OpenELMModel', OpenELMModel]],
     ['qwen2', ['Qwen2Model', Qwen2Model]],
+    ['qwen3', ['Qwen3Model', Qwen3Model]],
     ['phi', ['PhiModel', PhiModel]],
     ['phi3', ['Phi3Model', Phi3Model]],
     ['mpt', ['MptModel', MptModel]],
@@ -7737,6 +7754,7 @@ const MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = new Map([
     ['glm', ['GlmForCausalLM', GlmForCausalLM]],
     ['openelm', ['OpenELMForCausalLM', OpenELMForCausalLM]],
     ['qwen2', ['Qwen2ForCausalLM', Qwen2ForCausalLM]],
+    ['qwen3', ['Qwen3ForCausalLM', Qwen3ForCausalLM]],
     ['phi', ['PhiForCausalLM', PhiForCausalLM]],
     ['phi3', ['Phi3ForCausalLM', Phi3ForCausalLM]],
     ['mpt', ['MptForCausalLM', MptForCausalLM]],