Add support for Ministral 3

README.md

@@ -376,7 +376,10 @@ You can refine your search by selecting the task you're interested in (e.g., [te
 1. **[MusicGen](https://huggingface.co/docs/transformers/model_doc/musicgen)** (from Meta) released with the paper [Simple and Controllable Music Generation](https://huggingface.co/papers/2306.05284) by Jade Copet, Felix Kreuk, Itai Gat, Tal Remez, David Kant, Gabriel Synnaeve, Yossi Adi and Alexandre Défossez.
 1. **[MGP-STR](https://huggingface.co/docs/transformers/model_doc/mgp-str)** (from Alibaba Research) released with the paper [Multi-Granularity Prediction for Scene Text Recognition](https://huggingface.co/papers/2209.03592) by Peng Wang, Cheng Da, and Cong Yao.
 1. **[Mimi](https://huggingface.co/docs/transformers/model_doc/mimi)** (from Kyutai) released with the paper [Moshi: a speech-text foundation model for real-time dialogue](https://huggingface.co/papers/2410.00037) by Alexandre Défossez, Laurent Mazaré, Manu Orsini, Amélie Royer, Patrick Pérez, Hervé Jégou, Edouard Grave and Neil Zeghidour.
+1. **[Ministral](https://huggingface.co/docs/transformers/model_doc/ministral)** (from Mistral AI) by The [Mistral AI](https://mistral.ai) team.
+1. **[Ministral3](https://huggingface.co/docs/transformers/model_doc/ministral3)** (from Mistral AI) by The [Mistral AI](https://mistral.ai) team.
 1. **[Mistral](https://huggingface.co/docs/transformers/model_doc/mistral)** (from Mistral AI) by The [Mistral AI](https://mistral.ai) team: Albert Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, William El Sayed.
+1. **[Mistral3](https://huggingface.co/docs/transformers/model_doc/mistral3)** (from Mistral AI) by The [Mistral AI](https://mistral.ai) team.
 1. **[MMS](https://huggingface.co/docs/transformers/model_doc/mms)** (from Facebook) released with the paper [Scaling Speech Technology to 1,000+ Languages](https://huggingface.co/papers/2305.13516) by Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, Michael Auli.
 1. **[MobileBERT](https://huggingface.co/docs/transformers/model_doc/mobilebert)** (from CMU/Google Brain) released with the paper [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://huggingface.co/papers/2004.02984) by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou.
 1. **MobileCLIP** (from Apple) released with the paper [MobileCLIP: Fast Image-Text Models through Multi-Modal Reinforced Training](https://huggingface.co/papers/2311.17049) by Pavan Kumar Anasosalu Vasu, Hadi Pouransari, Fartash Faghri, Raviteja Vemulapalli, Oncel Tuzel.

docs/snippets/6_supported-models.snippet

@@ -90,7 +90,10 @@
 1. **[MusicGen](https://huggingface.co/docs/transformers/model_doc/musicgen)** (from Meta) released with the paper [Simple and Controllable Music Generation](https://huggingface.co/papers/2306.05284) by Jade Copet, Felix Kreuk, Itai Gat, Tal Remez, David Kant, Gabriel Synnaeve, Yossi Adi and Alexandre Défossez.
 1. **[MGP-STR](https://huggingface.co/docs/transformers/model_doc/mgp-str)** (from Alibaba Research) released with the paper [Multi-Granularity Prediction for Scene Text Recognition](https://huggingface.co/papers/2209.03592) by Peng Wang, Cheng Da, and Cong Yao.
 1. **[Mimi](https://huggingface.co/docs/transformers/model_doc/mimi)** (from Kyutai) released with the paper [Moshi: a speech-text foundation model for real-time dialogue](https://huggingface.co/papers/2410.00037) by Alexandre Défossez, Laurent Mazaré, Manu Orsini, Amélie Royer, Patrick Pérez, Hervé Jégou, Edouard Grave and Neil Zeghidour.
+1. **[Ministral](https://huggingface.co/docs/transformers/model_doc/ministral)** (from Mistral AI) by The [Mistral AI](https://mistral.ai) team.
+1. **[Ministral3](https://huggingface.co/docs/transformers/model_doc/ministral3)** (from Mistral AI) by The [Mistral AI](https://mistral.ai) team.
 1. **[Mistral](https://huggingface.co/docs/transformers/model_doc/mistral)** (from Mistral AI) by The [Mistral AI](https://mistral.ai) team: Albert Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, William El Sayed.
+1. **[Mistral3](https://huggingface.co/docs/transformers/model_doc/mistral3)** (from Mistral AI) by The [Mistral AI](https://mistral.ai) team.
 1. **[MMS](https://huggingface.co/docs/transformers/model_doc/mms)** (from Facebook) released with the paper [Scaling Speech Technology to 1,000+ Languages](https://huggingface.co/papers/2305.13516) by Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, Michael Auli.
 1. **[MobileBERT](https://huggingface.co/docs/transformers/model_doc/mobilebert)** (from CMU/Google Brain) released with the paper [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://huggingface.co/papers/2004.02984) by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou.
 1. **MobileCLIP** (from Apple) released with the paper [MobileCLIP: Fast Image-Text Models through Multi-Modal Reinforced Training](https://huggingface.co/papers/2311.17049) by Pavan Kumar Anasosalu Vasu, Hadi Pouransari, Fartash Faghri, Raviteja Vemulapalli, Oncel Tuzel.

package.json

@@ -55,7 +55,7 @@
   },
   "homepage": "https://github.com/huggingface/transformers.js#readme",
   "dependencies": {
-    "@huggingface/jinja": "^0.5.1",
+    "@huggingface/jinja": "^0.5.3",
     "onnxruntime-node": "1.21.0",
     "onnxruntime-web": "1.22.0-dev.20250409-89f8206ba4",
     "sharp": "^0.34.1"

src/configs.js

@@ -75,6 +75,7 @@ function getNormalizedConfig(config) {
         case 'voxtral':
         case 'smolvlm':
         case 'gemma3n':
+        case 'mistral3':
             // @ts-expect-error TS2339
             init_normalized_config = getNormalizedConfig(config.text_config);
             break;
@@ -145,6 +146,8 @@ function getNormalizedConfig(config) {
         case 'glm':
         case 'helium':
         case 'ernie4_5':
+        case 'ministral':
+        case 'ministral3':
             mapping['num_heads'] = 'num_key_value_heads';
             mapping['num_layers'] = 'num_hidden_layers';
             mapping['dim_kv'] = 'head_dim';

src/models.js

@@ -3866,6 +3866,8 @@ export class LlavaQwen2ForCausalLM extends LlavaPreTrainedModel {
     }
 }
 
+export class Mistral3ForConditionalGeneration extends LlavaQwen2ForCausalLM { }
+
 export class Gemma3nPreTrainedModel extends PreTrainedModel {
     forward_params = [
         'input_ids',
@@ -6948,6 +6950,20 @@ export class MistralModel extends MistralPreTrainedModel { }
 export class MistralForCausalLM extends MistralPreTrainedModel { }
 //////////////////////////////////////////////////
 
+//////////////////////////////////////////////////
+// Ministral models
+export class MinistralPreTrainedModel extends PreTrainedModel { }
+export class MinistralModel extends MinistralPreTrainedModel { }
+export class MinistralForCausalLM extends MinistralPreTrainedModel { }
+//////////////////////////////////////////////////
+
+//////////////////////////////////////////////////
+// Ministral3 models
+export class Ministral3PreTrainedModel extends PreTrainedModel { }
+export class Ministral3Model extends Ministral3PreTrainedModel { }
+export class Ministral3ForCausalLM extends Ministral3PreTrainedModel { }
+//////////////////////////////////////////////////
+
 //////////////////////////////////////////////////
 // ERNIE-4.5 models
 export class Ernie4_5PreTrainedModel extends PreTrainedModel { }
@@ -8041,6 +8057,8 @@ const MODEL_MAPPING_NAMES_DECODER_ONLY = new Map([
     ['mpt', ['MptModel', MptModel]],
     ['opt', ['OPTModel', OPTModel]],
     ['mistral', ['MistralModel', MistralModel]],
+    ['ministral', ['MinistralModel', MinistralModel]],
+    ['ministral3', ['Ministral3Model', Ministral3Model]],
     ['ernie4_5', ['Ernie4_5Model', Ernie4_5Model]],
     ['starcoder2', ['Starcoder2Model', Starcoder2Model]],
     ['falcon', ['FalconModel', FalconModel]],
@@ -8155,6 +8173,8 @@ const MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = new Map([
     ['opt', ['OPTForCausalLM', OPTForCausalLM]],
     ['mbart', ['MBartForCausalLM', MBartForCausalLM]],
     ['mistral', ['MistralForCausalLM', MistralForCausalLM]],
+    ['ministral', ['MinistralForCausalLM', MinistralForCausalLM]],
+    ['ministral3', ['Ministral3ForCausalLM', Ministral3ForCausalLM]],
     ['ernie4_5', ['Ernie4_5ForCausalLM', Ernie4_5ForCausalLM]],
     ['starcoder2', ['Starcoder2ForCausalLM', Starcoder2ForCausalLM]],
     ['falcon', ['FalconForCausalLM', FalconForCausalLM]],
@@ -8228,6 +8248,7 @@ const MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES = new Map([
     ['paligemma', ['PaliGemmaForConditionalGeneration', PaliGemmaForConditionalGeneration]],
     ['llava_qwen2', ['LlavaQwen2ForCausalLM', LlavaQwen2ForCausalLM]],
     ['gemma3n', ['Gemma3nForConditionalGeneration', Gemma3nForConditionalGeneration]],
+    ['mistral3', ['Mistral3ForConditionalGeneration', Mistral3ForConditionalGeneration]],
 ]);
 
 const MODEL_FOR_AUDIO_TEXT_TO_TEXT_MAPPING_NAMES = new Map([

src/models/image_processors.js

@@ -27,6 +27,7 @@ export * from './nougat/image_processing_nougat.js'
 export * from './owlv2/image_processing_owlv2.js'
 export * from './owlvit/image_processing_owlvit.js'
 export * from './phi3_v/image_processing_phi3_v.js'
+export * from './pixtral/image_processing_pixtral.js'
 export * from './pvt/image_processing_pvt.js'
 export * from './qwen2_vl/image_processing_qwen2_vl.js'
 export * from './rt_detr/image_processing_rt_detr.js'

src/models/pixtral/image_processing_pixtral.js

@@ -0,0 +1,38 @@
+import {
+    ImageProcessor,
+} from "../../base/image_processors_utils.js";
+
+export class PixtralImageProcessor extends ImageProcessor {
+
+    /** @type {ImageProcessor['get_resize_output_image_size']} */
+    get_resize_output_image_size(image, size) {
+        const { longest_edge } = size;
+        if (longest_edge === undefined) {
+            throw new Error("size must contain 'longest_edge'");
+        }
+
+        const [srcWidth, srcHeight] = image.size;
+
+        const ratio = Math.max(srcWidth, srcHeight) / longest_edge;
+
+        let newWidth = srcWidth;
+        let newHeight = srcHeight;
+        if (ratio > 1) {
+            newWidth = Math.floor(srcWidth / ratio);
+            newHeight = Math.floor(srcHeight / ratio);
+        }
+
+        // @ts-expect-error TS2339
+        const { patch_size, spatial_merge_size } = this.config;
+        if (!spatial_merge_size) {
+            throw new Error("config must contain 'spatial_merge_size'");
+        }
+        const real_patch_size = patch_size * spatial_merge_size;
+
+        // Calculate number of tokens
+        const num_width_tokens = Math.floor((newWidth - 1) / real_patch_size) + 1;
+        const num_height_tokens = Math.floor((newHeight - 1) / real_patch_size) + 1;
+
+        return [num_width_tokens * real_patch_size, num_height_tokens * real_patch_size];
+    }
+}

src/models/pixtral/processing_pixtral.js

@@ -0,0 +1,48 @@
+
+import { Processor } from "../../base/processing_utils.js";
+import { AutoImageProcessor } from "../auto/image_processing_auto.js";
+import { AutoTokenizer } from "../../tokenizers.js";
+
+export class PixtralProcessor extends Processor {
+    static tokenizer_class = AutoTokenizer
+    static image_processor_class = AutoImageProcessor
+    static uses_processor_config = true;
+
+    /**
+     * @typedef {import('../../utils/image.js').RawImage} RawImage
+     */
+
+    // `images` is required, `text` is optional
+    async _call(/** @type {RawImage|RawImage[]} */ images, text = null, kwargs = {}) {
+
+        const image_inputs = await this.image_processor(images, kwargs);
+
+        if (text) {
+            const [height, width] = image_inputs.pixel_values.dims.slice(-2);
+
+            const { image_token, image_break_token, image_end_token, patch_size, spatial_merge_size } = this.config;
+            const real_patch_size = patch_size * spatial_merge_size;
+            const num_height_tokens = Math.floor(height / real_patch_size);
+            const num_width_tokens = Math.floor(width / real_patch_size);
+
+            text = structuredClone(text); // Avoid modifying the original text input
+            if (!Array.isArray(text)) {
+                text = [text];
+            }
+            for (let i = 0; i < text.length; ++i) {
+                const width_tokens = image_token.repeat(num_width_tokens);
+                const row = width_tokens + image_break_token;
+                const finalRow = width_tokens + image_end_token;
+                const full = row.repeat(num_height_tokens - 1) + finalRow;
+                text[i] = text[i].replace(image_token, full);
+            }
+        }
+
+        const text_inputs = text ? this.tokenizer(text, kwargs) : {};
+
+        return {
+            ...image_inputs,
+            ...text_inputs,
+        }
+    }
+}

src/models/processors.js

@@ -8,8 +8,9 @@ export * from './llava/processing_llava.js';
 export * from './mgp_str/processing_mgp_str.js';
 export * from './moonshine/processing_moonshine.js';
 export * from './owlvit/processing_owlvit.js';
-export * from './phi3_v/processing_phi3_v.js';
 export * from './paligemma/processing_paligemma.js';
+export * from './phi3_v/processing_phi3_v.js';
+export * from './pixtral/processing_pixtral.js';
 export * from './pyannote/processing_pyannote.js';
 export * from './qwen2_vl/processing_qwen2_vl.js';
 export * from './sam/processing_sam.js';