Add multimodality! (#1802)

## Motivation

Images!

TODO (in a future PR): Add audio and video support.

## Test Plan

### Manual Testing
<img width="2652" height="1900" alt="image"
src="https://github.com/user-attachments/assets/7d3a7137-542f-4f94-9193-2c73b7c4a5ec"
/>

<img width="2770" height="1956" alt="image"
src="https://github.com/user-attachments/assets/e3c3a096-8029-4409-97a6-aca31a9a3f24"
/>
<img width="2738" height="1768" alt="image"
src="https://github.com/user-attachments/assets/d70ea37f-cd1d-4a4c-ad08-3beb9fafa380"
/>

(And batching also works)

---------

Co-authored-by: David Hind <davehind@yahoo.co.uk>

.mlx_typings/mlx_vlm/prompt_utils.pyi

@@ -0,0 +1,12 @@
+from typing import Any
+
+def get_message_json(
+    model_name: str,
+    prompt: str,
+    role: str = "user",
+    skip_image_token: bool = False,
+    skip_audio_token: bool = False,
+    num_images: int = 0,
+    num_audios: int = 0,
+    **kwargs: Any,
+) -> dict[str, Any]: ...

.mlx_typings/mlx_vlm/utils.pyi

@@ -0,0 +1,15 @@
+from pathlib import Path
+from typing import Any
+
+class ImageProcessor:
+    def preprocess(
+        self, images: list[dict[str, Any]], **kwargs: Any
+    ) -> dict[str, Any]: ...
+    def __call__(self, **kwargs: Any) -> dict[str, Any]: ...
+
+def load_image_processor(
+    model_path: str | Path, **kwargs: Any
+) -> ImageProcessor | None: ...
+def load_processor(
+    model_path: str | Path, add_detokenizer: bool = ..., **kwargs: Any
+) -> ImageProcessor: ...

.mlx_typings/safetensors/__init__.pyi

@@ -0,0 +1,8 @@
+from typing import Any, Self
+
+class safe_open:
+    def __init__(self, filename: str, framework: str = "pt") -> None: ...
+    def __enter__(self) -> Self: ...
+    def __exit__(self, *args: Any) -> None: ...
+    def keys(self) -> list[str]: ...
+    def get_tensor(self, name: str) -> Any: ...

dashboard/src/lib/stores/app.svelte.ts

@@ -2344,7 +2344,49 @@ class AppStore {
       const apiMessages = [
         systemPrompt,
         ...targetConversation.messages.slice(0, -1).map((m) => {
-          // Build content including any text file attachments
+          // Check if this message has image attachments
+          const imageAttachments = m.attachments?.filter(
+            (a) => a.type === "image" && a.preview,
+          );
+
+          if (imageAttachments && imageAttachments.length > 0) {
+            // Build multimodal content array (OpenAI vision format)
+            const contentParts: Array<
+              | { type: "text"; text: string }
+              | { type: "image_url"; image_url: { url: string } }
+            > = [];
+
+            // Add image parts first
+            for (const img of imageAttachments) {
+              if (img.preview) {
+                contentParts.push({
+                  type: "image_url",
+                  image_url: { url: img.preview },
+                });
+              }
+            }
+
+            // Build text content including any text file attachments
+            let textContent = m.content;
+            if (m.attachments) {
+              for (const attachment of m.attachments) {
+                if (attachment.type === "text" && attachment.content) {
+                  textContent += `\n\n[File: ${attachment.name}]\n\`\`\`\n${attachment.content}\n\`\`\``;
+                }
+              }
+            }
+
+            if (textContent) {
+              contentParts.push({ type: "text", text: textContent });
+            }
+
+            return {
+              role: m.role,
+              content: contentParts,
+            };
+          }
+
+          // Text-only message (original path)
           let msgContent = m.content;
 
           // Add text attachments as context

flake.lock

@@ -164,11 +164,11 @@
         ]
       },
       "locked": {
-        "lastModified": 1763662255,
-        "narHash": "sha256-4bocaOyLa3AfiS8KrWjZQYu+IAta05u3gYZzZ6zXbT0=",
+        "lastModified": 1773870109,
+        "narHash": "sha256-ZoTdqZP03DcdoyxvpFHCAek4bkPUTUPUF3oCCgc3dP4=",
         "owner": "pyproject-nix",
         "repo": "build-system-pkgs",
-        "rev": "042904167604c681a090c07eb6967b4dd4dae88c",
+        "rev": "b6e74f433b02fa4b8a7965ee24680f4867e2926f",
         "type": "github"
       },
       "original": {
@@ -184,11 +184,11 @@
         ]
       },
       "locked": {
-        "lastModified": 1764134915,
-        "narHash": "sha256-xaKvtPx6YAnA3HQVp5LwyYG1MaN4LLehpQI8xEdBvBY=",
+        "lastModified": 1774498001,
+        "narHash": "sha256-wTfdyzzrmpuqt4TQQNqilF91v0m5Mh1stNy9h7a/WK4=",
         "owner": "pyproject-nix",
         "repo": "pyproject.nix",
-        "rev": "2c8df1383b32e5443c921f61224b198a2282a657",
+        "rev": "794afa6eb588b498344f2eaa36ab1ceb7e6b0b09",
         "type": "github"
       },
       "original": {
@@ -280,11 +280,11 @@
         ]
       },
       "locked": {
-        "lastModified": 1767701098,
-        "narHash": "sha256-CJhKZnWb3gumR9oTRjFvCg/6lYTGbZRU7xtvcyWIRwU=",
+        "lastModified": 1774490495,
+        "narHash": "sha256-a9WmQWj8fF7BctZGCoyzpUjP6GJw8H+lxl+zxpGnETk=",
         "owner": "pyproject-nix",
         "repo": "uv2nix",
-        "rev": "9d357f0d2ce6f5f35ec7959d7e704452352eb4da",
+        "rev": "18ae62fc5e389e3069854a7c66455c22e31708fc",
         "type": "github"
       },
       "original": {

pyproject.toml

@@ -12,7 +12,7 @@ dependencies = [
   "fastapi>=0.116.1",
   "filelock>=3.18.0",
   "rustworkx>=0.17.1",
-  "huggingface-hub>=0.33.4",
+  "huggingface-hub>=1.8.0",
   "psutil>=7.0.0",
   "loguru>=0.7.3",
   "exo_pyo3_bindings",                         # rust bindings
@@ -29,6 +29,8 @@ dependencies = [
   "python-multipart>=0.0.21",
   "msgspec>=0.19.0",
   "zstandard>=0.23.0",
+  "mlx-vlm>=0.3.11",
+  "transformers>=5.0.0,<5.4.0",
 ]
 
 [project.scripts]
@@ -114,6 +116,7 @@ root = "src"
 required-version = ">=0.8.6"
 prerelease = "allow"
 environments = ["sys_platform == 'darwin'", "sys_platform == 'linux'"]
+extra-build-dependencies = { "miniaudio" = ["setuptools", "cffi", "pycparser"] }
 
 ###
 # ruff configuration

resources/inference_model_cards/mlx-community--Kimi-K2.5.toml

@@ -7,7 +7,13 @@ tasks = ["TextGeneration"]
 family = "kimi"
 quantization = ""
 base_model = "Kimi K2.5"
-capabilities = ["text", "thinking", "thinking_toggle"]
+capabilities = ["text", "thinking", "thinking_toggle", "vision"]
 
 [storage_size]
 in_bytes = 662498705408
+
+[vision]
+image_token_id = 163605
+model_type = "kimi_vl"
+weights_repo = "davehind/Kimi-K2.5-vision"
+processor_repo = "moonshotai/Kimi-K2.5"

resources/inference_model_cards/mlx-community--Qwen3-VL-4B-Instruct-4bit.toml

@@ -0,0 +1,12 @@
+model_id = "mlx-community/Qwen3-VL-4B-Instruct-4bit"
+n_layers = 36
+hidden_size = 2560
+supports_tensor = true
+tasks = ["TextGeneration"]
+family = "qwen"
+quantization = "4bit"
+base_model = "Qwen3-VL 4B"
+capabilities = ["text", "thinking", "vision"]
+
+[storage_size]
+in_bytes = 3340000000

src/exo/api/adapters/chat_completions.py

@@ -1,5 +1,7 @@
 """OpenAI Chat Completions API adapter for converting requests/responses."""
 
+import base64
+import re
 import time
 from collections.abc import AsyncGenerator
 from typing import Any
@@ -7,6 +9,7 @@ from typing import Any
 from exo.api.types import (
     ChatCompletionChoice,
     ChatCompletionMessage,
+    ChatCompletionMessageImageUrl,
     ChatCompletionMessageText,
     ChatCompletionRequest,
     ChatCompletionResponse,
@@ -19,6 +22,7 @@ from exo.api.types import (
     ToolCall,
     Usage,
 )
+from exo.download.download_utils import create_http_session
 from exo.shared.types.chunks import (
     ErrorChunk,
     PrefillProgressChunk,
@@ -33,25 +37,65 @@ from exo.shared.types.text_generation import (
 )
 
 
-def chat_request_to_text_generation(
+def extract_base64_from_data_url(data_url: str) -> str:
+    match = re.match(r"data:[^;]+;base64,(.+)", data_url)
+    if match:
+        return match.group(1)
+    return data_url
+
+
+async def fetch_image_url(url: str) -> str:
+    headers = {"User-Agent": "exo/1.0"}
+    async with (
+        create_http_session(timeout_profile="short") as session,
+        session.get(url, headers=headers) as resp,
+    ):
+        resp.raise_for_status()
+        data = await resp.read()
+        return base64.b64encode(data).decode("ascii")
+
+
+async def chat_request_to_text_generation(
     request: ChatCompletionRequest,
 ) -> TextGenerationTaskParams:
     instructions: str | None = None
     input_messages: list[InputMessage] = []
     chat_template_messages: list[dict[str, Any]] = []
+    images: list[str] = []
 
     for msg in request.messages:
         # Normalize content to string
         content: str
+        has_images = False
         if msg.content is None:
             content = ""
         elif isinstance(msg.content, str):
             content = msg.content
         elif isinstance(msg.content, ChatCompletionMessageText):
             content = msg.content.text
+        elif isinstance(msg.content, ChatCompletionMessageImageUrl):
+            url = msg.content.image_url.get("url", "")
+            if url:
+                if url.startswith(("http://", "https://")):
+                    images.append(await fetch_image_url(url))
                 else:
-            # List of ChatCompletionMessageText
-            content = "\n".join(item.text for item in msg.content)
+                    images.append(extract_base64_from_data_url(url))
+                has_images = True
+            content = ""
+        else:
+            text_parts: list[str] = []
+            for part in msg.content:
+                if isinstance(part, ChatCompletionMessageText):
+                    text_parts.append(part.text)
+                else:
+                    url = part.image_url.get("url", "")
+                    if url:
+                        if url.startswith(("http://", "https://")):
+                            images.append(await fetch_image_url(url))
+                        else:
+                            images.append(extract_base64_from_data_url(url))
+                        has_images = True
+            content = "\n".join(text_parts)
 
         # Extract system message as instructions
         if msg.role == "system":
@@ -75,7 +119,20 @@ def chat_request_to_text_generation(
 
             # Build full message dict for chat template (preserves tool_calls etc.)
             # Normalize content for model_dump
+            if has_images:
+                multimodal_content: list[dict[str, Any]] = []
+                assert isinstance(msg.content, list)
+                for part in msg.content:
+                    if isinstance(part, ChatCompletionMessageText):
+                        multimodal_content.append({"type": "text", "text": part.text})
+                    else:
+                        multimodal_content.append({"type": "image"})
+                chat_template_messages.append(
+                    {"role": msg.role, "content": multimodal_content}
+                )
+                continue
             msg_copy = msg.model_copy(update={"content": content})
+
             dumped: dict[str, Any] = msg_copy.model_dump(exclude_none=True)
             chat_template_messages.append(dumped)
 
@@ -107,6 +164,7 @@ def chat_request_to_text_generation(
         min_p=request.min_p,
         repetition_penalty=request.repetition_penalty,
         repetition_context_size=request.repetition_context_size,
+        images=images,
     )
 
 

src/exo/api/adapters/claude.py

@@ -11,6 +11,7 @@ from exo.api.types.claude_api import (
     ClaudeContentBlockDeltaEvent,
     ClaudeContentBlockStartEvent,
     ClaudeContentBlockStopEvent,
+    ClaudeImageBlock,
     ClaudeInputJsonDelta,
     ClaudeMessageDelta,
     ClaudeMessageDeltaEvent,
@@ -61,7 +62,9 @@ def _extract_tool_result_text(block: ClaudeToolResultBlock) -> str:
         return ""
     if isinstance(block.content, str):
         return block.content
-    return "".join(sub_block.text for sub_block in block.content)
+    return "".join(
+        sub.text for sub in block.content if isinstance(sub, ClaudeTextBlock)
+    )
 
 
 # Matches "x-anthropic-billing-header: ...;" (with optional trailing newline)
@@ -86,6 +89,7 @@ def claude_request_to_text_generation(
     # Handle system message
     instructions: str | None = None
     chat_template_messages: list[dict[str, Any]] = []
+    images: list[str] = []
 
     if request.system:
         if isinstance(request.system, str):
@@ -109,10 +113,18 @@ def claude_request_to_text_generation(
         thinking_parts: list[str] = []
         tool_calls: list[dict[str, Any]] = []
         tool_results: list[ClaudeToolResultBlock] = []
+        has_images = False
 
         for block in msg.content:
             if isinstance(block, ClaudeTextBlock):
                 text_parts.append(block.text)
+            elif isinstance(block, ClaudeImageBlock):
+                if block.source.type == "base64" and block.source.data:
+                    images.append(block.source.data)
+                    has_images = True
+                elif block.source.type == "url" and block.source.url:
+                    images.append(block.source.url)
+                    has_images = True
             elif isinstance(block, ClaudeThinkingBlock):
                 thinking_parts.append(block.thinking)
             elif isinstance(block, ClaudeToolUseBlock):
@@ -126,8 +138,17 @@ def claude_request_to_text_generation(
                         },
                     }
                 )
-            elif isinstance(block, ClaudeToolResultBlock):
+            else:
                 tool_results.append(block)
+                if isinstance(block.content, list):
+                    for sub in block.content:
+                        if isinstance(sub, ClaudeImageBlock):
+                            if sub.source.type == "base64" and sub.source.data:
+                                images.append(sub.source.data)
+                                has_images = True
+                            elif sub.source.type == "url" and sub.source.url:
+                                images.append(sub.source.url)
+                                has_images = True
 
         content = "".join(text_parts)
         reasoning_content = "".join(thinking_parts) if thinking_parts else None
@@ -155,6 +176,17 @@ def claude_request_to_text_generation(
                         "content": _extract_tool_result_text(tr),
                     }
                 )
+        elif has_images:
+            multimodal_content: list[dict[str, Any]] = []
+            for block in msg.content:
+                if isinstance(block, ClaudeTextBlock):
+                    multimodal_content.append({"type": "text", "text": block.text})
+                elif isinstance(block, ClaudeImageBlock):
+                    multimodal_content.append({"type": "image"})
+            chat_msg = {"role": msg.role, "content": multimodal_content}
+            if reasoning_content:
+                chat_msg["reasoning_content"] = reasoning_content
+            chat_template_messages.append(chat_msg)
         else:
             chat_msg = {"role": msg.role, "content": content}
             if reasoning_content:
@@ -197,6 +229,7 @@ def claude_request_to_text_generation(
         chat_template_messages=chat_template_messages
         if chat_template_messages
         else None,
+        images=images,
     )
 
 

src/exo/api/adapters/ollama.py

@@ -82,10 +82,15 @@ def ollama_request_to_text_generation(
     instructions: str | None = None
     input_messages: list[InputMessage] = []
     chat_template_messages: list[dict[str, Any]] = []
+    images: list[str] = []
     tool_message_index = 0
 
     for msg in request.messages:
         content = msg.content or ""
+        has_images = False
+        if msg.images:
+            images.extend(msg.images)
+            has_images = True
 
         if msg.role == "system":
             if instructions is None:
@@ -100,6 +105,16 @@ def ollama_request_to_text_generation(
         ):
             input_messages.append(InputMessage(role=msg.role, content=content))
 
+        if has_images:
+            multimodal: list[dict[str, Any]] = [
+                {"type": "image"} for _ in (msg.images or [])
+            ]
+            if content:
+                multimodal.append({"type": "text", "text": content})
+            chat_template_messages.append({"role": msg.role, "content": multimodal})
+            if msg.role in ("user", "assistant"):
+                input_messages.append(InputMessage(role=msg.role, content=content))
+            continue
         dumped: dict[str, Any] = {"role": msg.role, "content": content}
         if msg.thinking is not None:
             dumped["thinking"] = msg.thinking
@@ -152,6 +167,7 @@ def ollama_request_to_text_generation(
         chat_template_messages=chat_template_messages
         if chat_template_messages
         else None,
+        images=images,
     )
 
 
@@ -311,8 +327,16 @@ def ollama_generate_request_to_text_generation(
 ) -> TextGenerationTaskParams:
     """Convert Ollama generate request to exo's internal text generation format."""
     chat_template_messages: list[dict[str, Any]] = []
+    images: list[str] = []
     if request.system:
         chat_template_messages.append({"role": "system", "content": request.system})
+    if request.images:
+        images.extend(request.images)
+        multimodal: list[dict[str, Any]] = [{"type": "image"} for _ in request.images]
+        if request.prompt:
+            multimodal.append({"type": "text", "text": request.prompt})
+        chat_template_messages.append({"role": "user", "content": multimodal})
+    else:
         chat_template_messages.append({"role": "user", "content": request.prompt})
 
     options = request.options
@@ -331,6 +355,7 @@ def ollama_generate_request_to_text_generation(
         chat_template_messages=chat_template_messages
         if chat_template_messages
         else None,
+        images=images,
     )
 
 

src/exo/api/adapters/responses.py

@@ -4,6 +4,10 @@ from collections.abc import AsyncGenerator
 from itertools import count
 from typing import Any
 
+from exo.api.adapters.chat_completions import (
+    extract_base64_from_data_url,
+    fetch_image_url,
+)
 from exo.api.types import Usage
 from exo.api.types.openai_responses import (
     FunctionCallInputItem,
@@ -16,6 +20,7 @@ from exo.api.types.openai_responses import (
     ResponseFunctionCallArgumentsDoneEvent,
     ResponseFunctionCallItem,
     ResponseInProgressEvent,
+    ResponseInputImagePart,
     ResponseInputMessage,
     ResponseItem,
     ResponseMessageItem,
@@ -58,19 +63,23 @@ def _extract_content(content: str | list[ResponseContentPart]) -> str:
     """Extract plain text from a content field that may be a string or list of parts."""
     if isinstance(content, str):
         return content
-    return "".join(part.text for part in content)
+    return "".join(
+        part.text for part in content if not isinstance(part, ResponseInputImagePart)
+    )
 
 
-def responses_request_to_text_generation(
+async def responses_request_to_text_generation(
     request: ResponsesRequest,
 ) -> TextGenerationTaskParams:
     input_value: list[InputMessage]
     built_chat_template: list[dict[str, Any]] | None = None
+    images: list[str] = []
     if isinstance(request.input, str):
         input_value = [InputMessage(role="user", content=request.input)]
     else:
         input_messages: list[InputMessage] = []
         chat_template_messages: list[dict[str, Any]] = []
+        has_images = False
 
         if request.instructions is not None:
             chat_template_messages.append(
@@ -80,12 +89,33 @@ def responses_request_to_text_generation(
         for item in request.input:
             if isinstance(item, ResponseInputMessage):
                 content = _extract_content(item.content)
+                if isinstance(item.content, list):
+                    for part in item.content:
+                        if isinstance(part, ResponseInputImagePart) and part.image_url:
+                            url = part.image_url
+                            if url.startswith(("http://", "https://")):
+                                images.append(await fetch_image_url(url))
+                            else:
+                                images.append(extract_base64_from_data_url(url))
+                            has_images = True
                 if item.role in ("user", "assistant", "developer"):
                     input_messages.append(InputMessage(role=item.role, content=content))
                 if item.role == "system":
                     chat_template_messages.append(
                         {"role": "system", "content": content}
                     )
+                elif has_images:
+                    multimodal: list[dict[str, Any]] = []
+                    if isinstance(item.content, list):
+                        for part in item.content:
+                            if isinstance(part, ResponseInputImagePart):
+                                multimodal.append({"type": "image"})
+                            elif hasattr(part, "text"):
+                                multimodal.append({"type": "text", "text": part.text})
+                    chat_template_messages.append(
+                        {"role": item.role, "content": multimodal}
+                    )
+                    has_images = False
                 else:
                     chat_template_messages.append(
                         {"role": item.role, "content": content}
@@ -165,6 +195,7 @@ def responses_request_to_text_generation(
         chat_template_messages=built_chat_template or request.chat_template_messages,
         reasoning_effort=resolved_effort,
         enable_thinking=resolved_thinking,
+        images=images,
     )
 
 

src/exo/api/main.py

@@ -1,5 +1,6 @@
 import base64
 import contextlib
+import hashlib
 import json
 import random
 import time
@@ -24,6 +25,7 @@ from loguru import logger
 from exo.api.adapters.chat_completions import (
     chat_request_to_text_generation,
     collect_chat_response,
+    fetch_image_url,
     generate_chat_stream,
 )
 from exo.api.adapters.claude import (
@@ -170,6 +172,7 @@ from exo.shared.types.events import (
 )
 from exo.shared.types.memory import Memory
 from exo.shared.types.state import State
+from exo.shared.types.text_generation import TextGenerationTaskParams
 from exo.shared.types.worker.downloads import DownloadCompleted
 from exo.shared.types.worker.instances import Instance, InstanceId, InstanceMeta
 from exo.shared.types.worker.shards import Sharding
@@ -720,18 +723,79 @@ class API:
             "TODO: we should send a notification to the user to download the model"
         )
 
+    _sent_image_hashes: set[str] = set()
+
+    async def _send_text_generation_with_images(
+        self, task_params: TextGenerationTaskParams
+    ) -> TextGeneration:
+        images = task_params.images
+        if not images:
+            command = TextGeneration(task_params=task_params)
+            await self._send(command)
+            return command
+
+        hashes = [hashlib.sha256(img.encode("ascii")).hexdigest() for img in images]
+
+        cached_hashes: dict[int, str] = {}
+        new_images: list[tuple[int, str]] = []
+        for idx, (img, h) in enumerate(zip(images, hashes, strict=True)):
+            if h in self._sent_image_hashes:
+                cached_hashes[idx] = h
+            else:
+                self._sent_image_hashes.add(h)
+                new_images.append((idx, img))
+
+        if not new_images:
+            task_params = task_params.model_copy(
+                update={"images": [], "image_hashes": cached_hashes}
+            )
+            command = TextGeneration(task_params=task_params)
+            await self._send(command)
+            return command
+
+        all_chunks: list[tuple[int, str]] = []
+        for img_idx, img_data in new_images:
+            for i in range(0, len(img_data), EXO_MAX_CHUNK_SIZE):
+                all_chunks.append((img_idx, img_data[i : i + EXO_MAX_CHUNK_SIZE]))
+
+        task_params = task_params.model_copy(
+            update={
+                "images": [],
+                "image_hashes": cached_hashes,
+                "total_input_chunks": len(all_chunks),
+                "image_count": len(new_images),
+            }
+        )
+        command = TextGeneration(task_params=task_params)
+
+        for global_idx, (img_idx, chunk_data) in enumerate(all_chunks):
+            await self._send(
+                SendInputChunk(
+                    chunk=InputImageChunk(
+                        model=task_params.model,
+                        command_id=command.command_id,
+                        data=chunk_data,
+                        chunk_index=global_idx,
+                        total_chunks=len(all_chunks),
+                        image_index=img_idx,
+                    )
+                )
+            )
+
+        await self._send(command)
+        return command
+
     async def chat_completions(
         self, payload: ChatCompletionRequest
     ) -> ChatCompletionResponse | StreamingResponse:
         """OpenAI Chat Completions API - adapter."""
-        task_params = chat_request_to_text_generation(payload)
+        task_params = await chat_request_to_text_generation(payload)
         resolved_model = await self._resolve_and_validate_text_model(
             ModelId(task_params.model)
         )
         task_params = task_params.model_copy(update={"model": resolved_model})
 
-        command = TextGeneration(task_params=task_params)
-        await self._send(command)
+        command = await self._send_text_generation_with_images(task_params)
 
         if payload.stream:
             return StreamingResponse(
@@ -758,7 +822,7 @@ class API:
     async def bench_chat_completions(
         self, payload: BenchChatCompletionRequest
     ) -> BenchChatCompletionResponse:
-        task_params = chat_request_to_text_generation(payload)
+        task_params = await chat_request_to_text_generation(payload)
         resolved_model = await self._resolve_and_validate_text_model(
             ModelId(task_params.model)
         )
@@ -766,8 +830,7 @@ class API:
 
         task_params = task_params.model_copy(update={"stream": False, "bench": True})
 
-        command = TextGeneration(task_params=task_params)
-        await self._send(command)
+        command = await self._send_text_generation_with_images(task_params)
 
         return await self._collect_text_generation_with_stats(command.command_id)
 
@@ -1324,13 +1387,20 @@ class API:
     ) -> ClaudeMessagesResponse | StreamingResponse:
         """Claude Messages API - adapter."""
         task_params = claude_request_to_text_generation(payload)
+        if task_params.images:
+            resolved_images: list[str] = []
+            for img in task_params.images:
+                if img.startswith(("http://", "https://")):
+                    resolved_images.append(await fetch_image_url(img))
+                else:
+                    resolved_images.append(img)
+            task_params = task_params.model_copy(update={"images": resolved_images})
         resolved_model = await self._resolve_and_validate_text_model(
             ModelId(task_params.model)
         )
         task_params = task_params.model_copy(update={"model": resolved_model})
 
-        command = TextGeneration(task_params=task_params)
-        await self._send(command)
+        command = await self._send_text_generation_with_images(task_params)
 
         if payload.stream:
             return StreamingResponse(
@@ -1360,12 +1430,11 @@ class API:
         self, payload: ResponsesRequest
     ) -> ResponsesResponse | StreamingResponse:
         """OpenAI Responses API."""
-        task_params = responses_request_to_text_generation(payload)
+        task_params = await responses_request_to_text_generation(payload)
         resolved_model = await self._resolve_and_validate_text_model(task_params.model)
         task_params = task_params.model_copy(update={"model": resolved_model})
 
-        command = TextGeneration(task_params=task_params)
-        await self._send(command)
+        command = await self._send_text_generation_with_images(task_params)
 
         if payload.stream:
             return StreamingResponse(
@@ -1408,8 +1477,7 @@ class API:
         )
         task_params = task_params.model_copy(update={"model": resolved_model})
 
-        command = TextGeneration(task_params=task_params)
-        await self._send(command)
+        command = await self._send_text_generation_with_images(task_params)
 
         if payload.stream:
             return StreamingResponse(
@@ -1445,8 +1513,7 @@ class API:
         )
         task_params = task_params.model_copy(update={"model": resolved_model})
 
-        command = TextGeneration(task_params=task_params)
-        await self._send(command)
+        command = await self._send_text_generation_with_images(task_params)
 
         if payload.stream:
             return StreamingResponse(

src/exo/api/types/__init__.py

@@ -6,7 +6,9 @@ from .api import BenchImageGenerationResponse as BenchImageGenerationResponse
 from .api import BenchImageGenerationTaskParams as BenchImageGenerationTaskParams
 from .api import CancelCommandResponse as CancelCommandResponse
 from .api import ChatCompletionChoice as ChatCompletionChoice
+from .api import ChatCompletionContentPart as ChatCompletionContentPart
 from .api import ChatCompletionMessage as ChatCompletionMessage
+from .api import ChatCompletionMessageImageUrl as ChatCompletionMessageImageUrl
 from .api import ChatCompletionMessageText as ChatCompletionMessageText
 from .api import ChatCompletionRequest as ChatCompletionRequest
 from .api import ChatCompletionResponse as ChatCompletionResponse

src/exo/api/types/api.py

@@ -60,6 +60,14 @@ class ChatCompletionMessageText(BaseModel):
     text: str
 
 
+class ChatCompletionMessageImageUrl(BaseModel):
+    type: Literal["image_url"] = "image_url"
+    image_url: dict[str, str]  # {"url": "data:image/png;base64,..."}
+
+
+ChatCompletionContentPart = ChatCompletionMessageText | ChatCompletionMessageImageUrl
+
+
 class ToolCallItem(BaseModel):
     id: str = Field(default_factory=lambda: str(uuid4()))
     name: str
@@ -76,7 +84,7 @@ class ToolCall(BaseModel):
 class ChatCompletionMessage(BaseModel):
     role: Literal["system", "user", "assistant", "developer", "tool", "function"]
     content: (
-        str | ChatCompletionMessageText | list[ChatCompletionMessageText] | None
+        str | ChatCompletionContentPart | list[ChatCompletionContentPart] | None
     ) = None
     reasoning_content: str | None = None
     name: str | None = None

src/exo/api/types/claude_api.py

@@ -69,7 +69,7 @@ class ClaudeToolResultBlock(BaseModel, frozen=True):
 
     type: Literal["tool_result"] = "tool_result"
     tool_use_id: str
-    content: str | list[ClaudeTextBlock] | None = None
+    content: str | list[ClaudeTextBlock | ClaudeImageBlock] | None = None
     is_error: bool | None = None
     cache_control: dict[str, str] | None = None
 

src/exo/api/types/ollama_api.py

@@ -65,6 +65,7 @@ class OllamaGenerateRequest(BaseModel, frozen=True):
     keep_alive: str | int | None = None
     think: bool | None = None
     raw: bool = False
+    images: list[str] | None = None
 
 
 class OllamaGenerateResponse(BaseModel, frozen=True, strict=True):

src/exo/api/types/openai_responses.py

@@ -27,6 +27,12 @@ class ResponseInputTextPart(BaseModel, frozen=True):
     text: str
 
 
+class ResponseInputImagePart(BaseModel, frozen=True):
+    type: Literal["input_image"] = "input_image"
+    image_url: str | None = None
+    detail: str | None = None
+
+
 class ResponseOutputTextPart(BaseModel, frozen=True):
     """Output text content part (used when replaying assistant messages in input)."""
 
@@ -34,7 +40,9 @@ class ResponseOutputTextPart(BaseModel, frozen=True):
     text: str
 
 
-ResponseContentPart = ResponseInputTextPart | ResponseOutputTextPart
+ResponseContentPart = (
+    ResponseInputTextPart | ResponseInputImagePart | ResponseOutputTextPart
+)
 
 
 # Request input item types

src/exo/download/impl_shard_downloader.py

@@ -6,9 +6,18 @@ from typing import AsyncIterator, Callable
 
 from loguru import logger
 
-from exo.download.download_utils import RepoDownloadProgress, download_shard
+from exo.download.download_utils import (
+    RepoDownloadProgress,
+    download_shard,
+)
 from exo.download.shard_downloader import ShardDownloader
-from exo.shared.models.model_cards import ModelCard, ModelId, get_model_cards
+from exo.shared.models.model_cards import (
+    ModelCard,
+    ModelId,
+    ModelTask,
+    get_model_cards,
+)
+from exo.shared.types.memory import Memory
 from exo.shared.types.worker.shards import (
     PipelineShardMetadata,
     ShardMetadata,
@@ -115,6 +124,38 @@ class ResumableShardDownloader(ShardDownloader):
             allow_patterns=allow_patterns,
             skip_internet=self.offline,
         )
+
+        if (
+            not config_only
+            and not self.offline
+            and shard.model_card.vision
+            and shard.model_card.vision.weights_repo != str(shard.model_card.model_id)
+        ):
+            vision_repo = shard.model_card.vision.weights_repo
+            vision_card = ModelCard(
+                model_id=ModelId(vision_repo),
+                storage_size=Memory.from_bytes(0),
+                n_layers=1,
+                hidden_size=1,
+                supports_tensor=False,
+                tasks=[ModelTask.TextGeneration],
+            )
+            vision_shard = PipelineShardMetadata(
+                model_card=vision_card,
+                device_rank=0,
+                world_size=1,
+                start_layer=0,
+                end_layer=1,
+                n_layers=1,
+            )
+            await download_shard(
+                vision_shard,
+                self.on_progress_wrapper,
+                max_parallel_downloads=self.max_parallel_downloads,
+                allow_patterns=["*.safetensors", "config.json"],
+                skip_internet=self.offline,
+            )
+
         return target_dir
 
     async def get_shard_download_status(

src/exo/shared/models/model_cards.py

@@ -1,3 +1,4 @@
+import json
 from enum import Enum
 from typing import Annotated, Any
 
@@ -13,6 +14,7 @@ from pydantic import (
     Field,
     PositiveInt,
     ValidationError,
+    ValidationInfo,
     field_validator,
     model_validator,
 )
@@ -21,6 +23,7 @@ from tomlkit.exceptions import TOMLKitError
 from exo.shared.constants import (
     EXO_CUSTOM_MODEL_CARDS_DIR,
     EXO_ENABLE_IMAGE_MODELS,
+    EXO_MODELS_DIRS,
     RESOURCES_DIR,
 )
 from exo.shared.types.common import ModelId
@@ -38,6 +41,23 @@ _BUILTIN_CARD_DIRS = [
 _card_cache: dict[ModelId, "ModelCard"] = {}
 
 
+def _detect_vision_from_config(model_id: ModelId) -> "VisionCardConfig | None":
+    normalized = model_id.normalize()
+    for model_dir in [d / normalized for d in EXO_MODELS_DIRS]:
+        config_path = model_dir / "config.json"
+        if not config_path.exists():
+            continue
+        try:
+            with open(config_path) as f:
+                raw = json.load(f)  # type: ignore
+            return ConfigData.model_validate(
+                raw, context={"model_id": str(model_id)}
+            ).vision
+        except Exception:
+            continue
+    return None
+
+
 async def _load_cards_from_dir(directory: Path, *, is_custom: bool) -> None:
     """Load all TOML model cards from a directory into the cache."""
     async for toml_file in directory.rglob("*.toml"):
@@ -45,6 +65,10 @@ async def _load_cards_from_dir(directory: Path, *, is_custom: bool) -> None:
             card = await ModelCard.load_from_path(toml_file)
             if is_custom:
                 card = card.model_copy(update={"is_custom": True})
+            if card.vision is None:
+                vision = _detect_vision_from_config(card.model_id)
+                if vision is not None:
+                    card = card.model_copy(update={"vision": vision})
             if card.model_id not in _card_cache:
                 _card_cache[card.model_id] = card
         except (ValidationError, TOMLKitError):
@@ -89,6 +113,14 @@ class ComponentInfo(CamelCaseModel):
     safetensors_index_filename: str | None = None
 
 
+class VisionCardConfig(CamelCaseModel):
+    image_token_id: int
+    model_type: str
+    weights_repo: str = ""
+    image_token: str | None = None
+    processor_repo: str | None = None
+
+
 class ModelCard(CamelCaseModel):
     model_id: ModelId
     storage_size: Memory
@@ -105,6 +137,17 @@ class ModelCard(CamelCaseModel):
     uses_cfg: bool = False
     trust_remote_code: bool = True
     is_custom: bool = False
+    vision: VisionCardConfig | None = None
+
+    @model_validator(mode="after")
+    def _fill_vision_weights_repo(self) -> "ModelCard":
+        if self.vision is not None and not self.vision.weights_repo:
+            object.__setattr__(
+                self,
+                "vision",
+                self.vision.model_copy(update={"weights_repo": str(self.model_id)}),
+            )
+        return self
 
     @field_validator("tasks", mode="before")
     @classmethod
@@ -162,6 +205,7 @@ class ModelCard(CamelCaseModel):
             tasks=[ModelTask.TextGeneration],
             trust_remote_code=False,
             is_custom=True,
+            vision=config_data.vision,
         )
 
 
@@ -196,6 +240,7 @@ class ConfigData(BaseModel):
             "decoder_layers",
         )
     )
+    vision: VisionCardConfig | None = None
 
     @property
     def supports_tensor(self) -> bool:
@@ -217,11 +262,9 @@ class ConfigData(BaseModel):
 
     @model_validator(mode="before")
     @classmethod
-    def defer_to_text_config(cls, data: dict[str, Any]):
+    def defer_to_text_config(cls, data: dict[str, Any], info: ValidationInfo):
         text_config = data.get("text_config")
-        if text_config is None:
-            return data
-
+        if text_config is not None:
             for field in [
                 "architectures",
                 "hidden_size",
@@ -236,6 +279,20 @@ class ConfigData(BaseModel):
                 if (val := text_config.get(field)) is not None:  # pyright: ignore[reportAny]
                     data[field] = val
 
+        vision_config = data.get("vision_config")
+        image_token_id = data.get("image_token_id")
+        if vision_config is not None and image_token_id is not None:
+            model_type = str(
+                vision_config.get("model_type", data.get("model_type", ""))  # pyright: ignore[reportAny]
+            )
+            assert info.context is not None
+
+            data["vision"] = VisionCardConfig(
+                image_token_id=int(image_token_id),  # pyright: ignore[reportAny]
+                model_type=model_type,
+                weights_repo=info.context["model_id"],  # type: ignore
+            )
+
         return data
 
 
@@ -257,7 +314,9 @@ async def fetch_config_data(model_id: ModelId) -> ConfigData:
         ),
     )
     async with aiofiles.open(config_path, "r") as f:
-        return ConfigData.model_validate_json(await f.read())
+        return ConfigData.model_validate_json(
+            await f.read(), context={"model_id": str(model_id)}
+        )
 
 
 async def fetch_safetensors_size(model_id: ModelId) -> Memory:

src/exo/shared/types/chunks.py

@@ -68,6 +68,7 @@ class InputImageChunk(BaseChunk):
     data: str
     chunk_index: int
     total_chunks: int
+    image_index: int = 0
 
     def __repr_args__(self) -> Generator[tuple[str, Any], None, None]:
         for name, value in super().__repr_args__():  # pyright: ignore[reportAny]

src/exo/shared/types/text_generation.py

@@ -6,7 +6,7 @@ are converted to TextGenerationTaskParams at the API boundary via adapters.
 
 from typing import Any, Literal
 
-from pydantic import BaseModel
+from pydantic import BaseModel, Field
 
 from exo.shared.types.common import ModelId
 
@@ -70,3 +70,7 @@ class TextGenerationTaskParams(BaseModel, frozen=True):
     min_p: float | None = None
     repetition_penalty: float | None = None
     repetition_context_size: int | None = None
+    images: list[str] = Field(default_factory=list)
+    image_hashes: dict[int, str] = Field(default_factory=dict)
+    total_input_chunks: int = 0
+    image_count: int = 0

src/exo/worker/engines/mlx/cache.py

@@ -1,5 +1,6 @@
 import os
 from copy import deepcopy
+from typing import TYPE_CHECKING
 
 import mlx.core as mx
 import psutil
@@ -17,6 +18,9 @@ from exo.shared.types.mlx import KVCacheType, Model
 from exo.worker.engines.mlx.constants import CACHE_GROUP_SIZE, KV_CACHE_BITS
 from exo.worker.runner.bootstrap import logger
 
+if TYPE_CHECKING:
+    from exo.worker.engines.mlx.vision import MediaRegion
+
 
 # Fraction of device memory above which LRU eviction kicks in.
 # Smaller machines need more aggressive eviction.
@@ -80,6 +84,7 @@ class KVPrefixCache:
         self.prompts: list[mx.array] = []  # mx array of tokens (ints)
         self.caches: list[KVCacheType] = []
         self._snapshots: list[list[CacheSnapshot] | None] = []
+        self._media_regions: list[list["MediaRegion"]] = []
         self._last_used: list[int] = []  # monotonic counter of last access per entry
         self._access_counter: int = 0
         self._group = group
@@ -89,6 +94,7 @@ class KVPrefixCache:
         self.prompts.clear()
         self.caches.clear()
         self._snapshots.clear()
+        self._media_regions.clear()
         self._last_used.clear()
 
     def add_kv_cache(
@@ -96,12 +102,14 @@ class KVPrefixCache:
         prompt_tokens: mx.array,
         cache: KVCacheType,
         ssm_snapshots: list[CacheSnapshot] | None = None,
+        media_regions: list["MediaRegion"] | None = None,
     ):
         """Add a new cache entry. Evicts LRU entries if memory is high."""
         self._evict_if_needed()
         self.prompts.append(prompt_tokens)
         self.caches.append(deepcopy(cache))
         self._snapshots.append(ssm_snapshots)
+        self._media_regions.append(media_regions or [])
         self._access_counter += 1
         self._last_used.append(self._access_counter)
         logger.info(f"KV cache added: {len(prompt_tokens)} tokens")
@@ -113,6 +121,7 @@ class KVPrefixCache:
         cache: KVCacheType,
         snapshots: list[CacheSnapshot] | None,
         restore_pos: int,
+        media_regions: list["MediaRegion"] | None = None,
     ):
         """Update an existing cache entry in-place."""
         old_snapshots = self._snapshots[index]
@@ -125,6 +134,7 @@ class KVPrefixCache:
         self.prompts[index] = prompt_tokens
         self.caches[index] = deepcopy(cache)
         self._snapshots[index] = merged or None
+        self._media_regions[index] = media_regions or []
         self._access_counter += 1
         self._last_used[index] = self._access_counter
         logger.info(f"KV cache updated (index {index}): {len(prompt_tokens)} tokens")
@@ -149,6 +159,7 @@ class KVPrefixCache:
         self,
         model: Model,
         prompt_tokens: mx.array,
+        media_regions: list["MediaRegion"] | None = None,
     ) -> tuple[KVCacheType, mx.array, int | None]:
         """Get KV cache for prompt, returning remaining tokens to prefill.
 
@@ -161,8 +172,13 @@ class KVPrefixCache:
         For models with SSM layers (which are ArraysCache in mlx), the cache is trimmed to the
         nearest SSM snapshot position at or before the match point for correctness.
         Same for rotating KV Cache.
+
+        Media region validation: if the token-level prefix match extends into
+        a cached media region whose content_hash differs from the query's, the
+        match is truncated to the start of that region.
         """
         max_length = len(prompt_tokens)
+        query_regions = media_regions or []
 
         best_index: int | None = None
         best_length = 0
@@ -171,6 +187,12 @@ class KVPrefixCache:
         # Find best cache match
         for i, cached_prompt in enumerate(self.prompts):
             length = get_prefix_length(prompt_tokens, cached_prompt)
+            if length > 0:
+                length = self._validate_media_match(
+                    length,
+                    self._media_regions[i],
+                    query_regions,
+                )
             if length >= max_length - 1:
                 best_index, best_length = i, length
                 is_exact = True
@@ -208,6 +230,37 @@ class KVPrefixCache:
 
         return prompt_cache, remaining, best_index
 
+    @staticmethod
+    def _validate_media_match(
+        match_length: int,
+        cached_regions: list["MediaRegion"],
+        query_regions: list["MediaRegion"],
+    ) -> int:
+        if not cached_regions:
+            return match_length
+
+        query_by_start: dict[int, "MediaRegion"] = {
+            r.start_pos: r for r in query_regions
+        }
+
+        for cached_r in cached_regions:
+            if cached_r.start_pos >= match_length:
+                break
+            query_r = query_by_start.get(cached_r.start_pos)
+            if query_r is None:
+                continue
+            if query_r.content_hash != cached_r.content_hash:
+                logger.info(
+                    f"Media region mismatch at pos {cached_r.start_pos}: "
+                    f"cached={cached_r.content_hash[:12]}... "
+                    f"query={query_r.content_hash[:12]}... — "
+                    f"truncating match from {match_length} to {cached_r.start_pos}"
+                )
+                match_length = cached_r.start_pos
+                break
+
+        return match_length
+
     def _evict_if_needed(self):
         """Evict least recently used entries while memory usage is high."""
         if len(self.caches) == 0:
@@ -223,6 +276,7 @@ class KVPrefixCache:
             self.prompts.pop(lru_index)
             self.caches.pop(lru_index)
             self._snapshots.pop(lru_index)
+            self._media_regions.pop(lru_index)
             self._last_used.pop(lru_index)
             logger.info(
                 f"KV cache evicted LRU entry ({evicted_tokens} tokens) due to memory usage"

src/exo/worker/engines/mlx/generator/batch_generate.py

@@ -1,3 +1,4 @@
+import contextlib
 import time
 from dataclasses import dataclass, field
 from typing import Callable, cast
@@ -36,9 +37,16 @@ from exo.worker.engines.mlx.generator.generate import (
     ban_token_ids,
     eos_ids_from_tokenizer,
     extract_top_logprobs,
+    patch_embed_tokens,
     prefill,
 )
 from exo.worker.engines.mlx.utils_mlx import fix_unmatched_think_end_tokens
+from exo.worker.engines.mlx.vision import (
+    MediaRegion,
+    VisionProcessor,
+    VisionResult,
+    prepare_vision,
+)
 from exo.worker.runner.bootstrap import logger
 
 _MIN_PREFIX_HIT_RATIO_TO_UPDATE = 0.5
@@ -70,6 +78,7 @@ class _EngineTask:
     in_thinking: bool = False
     reasoning_tokens: int = 0
     prefill_tps: float = 0.0
+    media_regions: list[MediaRegion] = field(default_factory=list)
 
 
 @dataclass(eq=False)
@@ -78,6 +87,7 @@ class ExoBatchGenerator:
     tokenizer: TokenizerWrapper
     group: mx.distributed.Group | None
     kv_prefix_cache: KVPrefixCache | None
+    vision_processor: VisionProcessor | None = None
 
     _mlx_gen: MlxBatchGenerator = field(init=False)
     _active_tasks: dict[int, _EngineTask] = field(default_factory=dict, init=False)
@@ -111,6 +121,27 @@ class ExoBatchGenerator:
             all_prompt_tokens, self.tokenizer
         )
 
+        vision: VisionResult | None = None
+        media_regions: list[MediaRegion] = []
+
+        if self.vision_processor is not None:
+            try:
+                vision = prepare_vision(
+                    images=task_params.images,
+                    chat_template_messages=task_params.chat_template_messages,
+                    vision_processor=self.vision_processor,
+                    tokenizer=self.tokenizer,
+                    model=self.model,
+                )
+            except Exception:
+                logger.opt(exception=True).warning(
+                    "Vision processing failed, falling back to text-only"
+                )
+
+        if vision is not None:
+            all_prompt_tokens = vision.prompt_tokens
+            media_regions = vision.media_regions
+
         is_bench = task_params.bench
 
         prefix_hit_length = 0
@@ -119,7 +150,7 @@ class ExoBatchGenerator:
 
         if self.kv_prefix_cache is not None and not is_bench:
             cache, remaining_tokens, matched_index = self.kv_prefix_cache.get_kv_cache(
-                self.model, all_prompt_tokens
+                self.model, all_prompt_tokens, media_regions=media_regions
             )
             prefix_hit_length = len(all_prompt_tokens) - len(remaining_tokens)
             if prefix_hit_length > 0:
@@ -145,6 +176,14 @@ class ExoBatchGenerator:
             top_k=task_params.top_k if task_params.top_k is not None else 0,
         )
 
+        vision_ctx = (
+            patch_embed_tokens(
+                self.model, vision.embeddings, prefix_hit_length, len(prompt_tokens) - 1
+            )
+            if vision is not None
+            else contextlib.nullcontext()
+        )
+        with vision_ctx:
             _prefill_tps, _prefill_tokens, cache_snapshots = prefill(
                 self.model,
                 self.tokenizer,
@@ -176,6 +215,7 @@ class ExoBatchGenerator:
                 cache_snapshots,
                 prefix_hit_length,
                 matched_index,
+                media_regions,
             )
 
         last_tokens = prompt_tokens[-2:]
@@ -217,6 +257,7 @@ class ExoBatchGenerator:
             generation_start_time=time.perf_counter(),
             prefill_tps=_prefill_tps,
             generation_time_at_start=self._mlx_gen._stats.generation_time,
+            media_regions=media_regions,
         )
 
         return uid
@@ -383,6 +424,7 @@ class ExoBatchGenerator:
         cache_snapshots: list[CacheSnapshot] | None,
         prefix_hit_length: int,
         matched_index: int | None,
+        media_regions: list[MediaRegion] | None = None,
     ) -> None:
         if self.kv_prefix_cache is None:
             return
@@ -402,10 +444,14 @@ class ExoBatchGenerator:
                     cache,
                     cache_snapshots,
                     restore_pos=prefix_hit_length,
+                    media_regions=media_regions,
                 )
             else:
                 self.kv_prefix_cache.add_kv_cache(
-                    all_prompt_tokens, cache, cache_snapshots
+                    all_prompt_tokens,
+                    cache,
+                    cache_snapshots,
+                    media_regions=media_regions,
                 )
         except Exception:
             logger.warning("Failed to save prefix cache", exc_info=True)

src/exo/worker/engines/mlx/generator/generate.py

@@ -1,3 +1,4 @@
+import contextlib
 import functools
 import math
 import time
@@ -55,6 +56,13 @@ from exo.worker.engines.mlx.utils_mlx import (
     fix_unmatched_think_end_tokens,
     mx_barrier,
 )
+from exo.worker.engines.mlx.vision import (
+    MediaRegion,
+    VisionProcessor,
+    VisionResult,
+    get_inner_model,
+    prepare_vision,
+)
 from exo.worker.runner.bootstrap import logger
 
 generation_stream = mx.new_stream(mx.default_device())
@@ -62,6 +70,38 @@ generation_stream = mx.new_stream(mx.default_device())
 _MIN_PREFIX_HIT_RATIO_TO_UPDATE = 0.5
 
 
+@contextlib.contextmanager
+def patch_embed_tokens(
+    model: Model, embeddings: mx.array, start_offset: int = 0, token_count: int = 0
+) -> Generator[None]:
+    inner = get_inner_model(model)  # type: ignore
+    original_embed = inner.embed_tokens  # type: ignore
+    end_offset = start_offset + token_count
+    offset = [start_offset]
+
+    def _inject(input_ids: mx.array) -> mx.array:
+        start = offset[0]
+        if start >= end_offset:
+            return original_embed(input_ids)  # type: ignore
+        chunk_len = input_ids.shape[-1]
+        end = min(start + chunk_len, end_offset)
+        offset[0] = end
+        if end - start < chunk_len:
+            return original_embed(input_ids)  # type: ignore
+        return embeddings[:, start:end, :]
+
+    for attr in dir(original_embed):  # type: ignore
+        if not attr.startswith("_") and not hasattr(_inject, attr):
+            with contextlib.suppress(AttributeError, TypeError):
+                setattr(_inject, attr, getattr(original_embed, attr))  # type: ignore
+
+    inner.embed_tokens = _inject
+    try:
+        yield
+    finally:
+        inner.embed_tokens = original_embed
+
+
 class PrefillCancelled(BaseException):
     """Raised when prefill is cancelled via the progress callback."""
 
@@ -447,6 +487,7 @@ def mlx_generate(
     on_prefill_progress: Callable[[int, int], None] | None = None,
     distributed_prompt_progress_callback: Callable[[], None] | None = None,
     on_generation_token: Callable[[], None] | None = None,
+    vision_processor: VisionProcessor | None = None,
 ) -> Generator[GenerationResponse]:
     # Ensure that generation stats only contains peak memory for this generation
     mx.reset_peak_memory()
@@ -458,6 +499,24 @@ def mlx_generate(
     all_prompt_tokens = encode_prompt(tokenizer, prompt)
     all_prompt_tokens = fix_unmatched_think_end_tokens(all_prompt_tokens, tokenizer)
 
+    vision: VisionResult | None = None
+    if vision_processor is not None:
+        try:
+            vision = prepare_vision(
+                images=task.images,
+                chat_template_messages=task.chat_template_messages,
+                vision_processor=vision_processor,
+                tokenizer=tokenizer,
+                model=model,
+            )
+        except Exception:
+            logger.opt(exception=True).warning(
+                "Vision processing failed, falling back to text-only"
+            )
+    if vision is not None:
+        all_prompt_tokens = vision.prompt_tokens
+    media_regions: list[MediaRegion] = vision.media_regions if vision else []
+
     # Do not use the prefix cache if we are trying to do benchmarks.
     is_bench = task.bench
     if is_bench:
@@ -471,7 +530,7 @@ def mlx_generate(
         prompt_tokens = all_prompt_tokens
     else:
         caches, prompt_tokens, matched_index = kv_prefix_cache.get_kv_cache(
-            model, all_prompt_tokens
+            model, all_prompt_tokens, media_regions=media_regions
         )
         prefix_hit_length = len(all_prompt_tokens) - len(prompt_tokens)
         if prefix_hit_length > 0:
@@ -505,7 +564,14 @@ def mlx_generate(
     )
     max_stop_len = max((len(s) for s in stop_sequences), default=0)
 
-    # Prefill cache with all tokens except the last one
+    maybe_vision_ctx = (
+        patch_embed_tokens(
+            model, vision.embeddings, prefix_hit_length, len(prompt_tokens) - 1
+        )
+        if vision is not None
+        else contextlib.nullcontext()
+    )
+    with maybe_vision_ctx:
         prefill_tps, prefill_tokens, ssm_snapshots_list = prefill(
             model,
             tokenizer,
@@ -655,10 +721,14 @@ def mlx_generate(
                         caches,
                         cache_snapshots,
                         restore_pos=prefix_hit_length,
+                        media_regions=media_regions,
                     )
                 else:
                     kv_prefix_cache.add_kv_cache(
-                        full_prompt_tokens, caches, cache_snapshots
+                        full_prompt_tokens,
+                        caches,
+                        cache_snapshots,
+                        media_regions=media_regions,
                     )
 
         if on_generation_token is not None:

src/exo/worker/engines/mlx/utils_mlx.py

@@ -5,7 +5,10 @@ import sys
 import tempfile
 import time
 from pathlib import Path
-from typing import Any, cast
+from typing import TYPE_CHECKING, Any, cast
+
+if TYPE_CHECKING:
+    from exo.worker.engines.mlx.vision import VisionProcessor
 
 # Monkey-patch for transformers 5.x compatibility
 # Kimi's tokenization_kimi.py imports bytes_to_unicode from the old location
@@ -168,7 +171,7 @@ def load_mlx_items(
     group: Group | None,
     on_timeout: TimeoutCallback | None,
     on_layer_loaded: LayerLoadedCallback | None,
-) -> tuple[Model, TokenizerWrapper]:
+) -> "tuple[Model, TokenizerWrapper, VisionProcessor | None]":
     if group is None:
         logger.info(f"Single device used for {bound_instance.instance}")
         model_path = build_model_path(bound_instance.bound_shard.model_card.model_id)
@@ -210,7 +213,18 @@ def load_mlx_items(
 
     mx.clear_cache()
 
-    return cast(Model, model), tokenizer
+    vision_config = bound_instance.bound_shard.model_card.vision
+
+    if vision_config is not None:
+        from exo.worker.engines.mlx.vision import VisionProcessor
+
+        vision_processor: VisionProcessor | None = VisionProcessor(
+            vision_config, bound_instance.bound_shard.model_card.model_id
+        )
+    else:
+        vision_processor = None
+
+    return cast(Model, model), tokenizer, vision_processor
 
 
 def shard_and_load(

src/exo/worker/engines/mlx/vision.py

@@ -0,0 +1,589 @@
+import base64
+import contextlib
+import hashlib
+import importlib
+import inspect
+import io
+import json
+import re
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any
+
+if TYPE_CHECKING:
+    from mlx_vlm.utils import ImageProcessor
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+from mlx_lm.tokenizer_utils import TokenizerWrapper
+from mlx_vlm.prompt_utils import get_message_json
+from mlx_vlm.utils import load_image_processor
+from PIL import Image
+from safetensors import safe_open
+from transformers import AutoImageProcessor
+
+from exo.download.download_utils import build_model_path
+from exo.shared.models.model_cards import VisionCardConfig
+from exo.shared.types.common import ModelId
+from exo.shared.types.mlx import Model
+from exo.worker.engines.mlx.cache import encode_prompt
+from exo.worker.engines.mlx.utils_mlx import fix_unmatched_think_end_tokens
+from exo.worker.runner.bootstrap import logger
+
+
+def _filter_config(cls: type, d: dict[str, Any]) -> dict[str, Any]:
+    valid = set(inspect.signature(cls.__init__).parameters.keys()) - {"self"}
+    return {k: v for k, v in d.items() if k in valid}  # type: ignore
+
+
+_video_processor_patched = False
+
+
+def _patch_video_processor() -> None:
+    """Patch so we don't crash horribly when torch vision isn't installed"""
+    # TODO: Update if we add torch vision.
+    global _video_processor_patched
+    if _video_processor_patched:
+        return
+    try:
+        from transformers.processing_utils import MODALITY_TO_AUTOPROCESSOR_MAPPING
+
+        mapping = MODALITY_TO_AUTOPROCESSOR_MAPPING._MAPPING_NAMES  # type: ignore
+        mapping.pop("video_processor", None)
+    except (ImportError, AttributeError):
+        pass
+    _video_processor_patched = True
+
+
+def decode_base64_image(b64_data: str) -> Image.Image:
+    raw = base64.b64decode(b64_data)
+    img = Image.open(io.BytesIO(raw))
+    return img.convert("RGB")
+
+
+def _format_vlm_messages(
+    messages: list[dict[str, Any]],
+    model_type: str,
+) -> list[dict[str, Any]]:
+    formatted: list[dict[str, Any]] = []
+    for msg in messages:
+        role: str = str(msg.get("role", "user"))  # type: ignore
+        content: Any = msg.get("content")
+        if not isinstance(content, list):
+            formatted.append(msg)
+            continue
+        parts: list[dict[str, Any]] = content  # type: ignore
+        text_parts = [str(p["text"]) for p in parts if p.get("type") == "text"]  # type: ignore
+        n_images = sum(1 for p in parts if p.get("type") in ("image", "image_url"))
+        result: dict[str, Any] = get_message_json(
+            model_type, " ".join(text_parts), role, num_images=n_images
+        )
+        formatted.append(result)
+    return formatted
+
+
+def build_vision_prompt(
+    tokenizer: TokenizerWrapper,
+    chat_template_messages: list[dict[str, Any]],
+    n_tokens_per_image: list[int],
+    image_token: str,
+) -> str:
+    logger.info(
+        f"Vision prompt messages: {[{k: (v[:50] if isinstance(v, str) else v) for k, v in m.items()} for m in chat_template_messages]}"  # type: ignore
+    )
+    prompt: str = tokenizer.apply_chat_template(
+        chat_template_messages,
+        tokenize=False,
+        add_generation_prompt=True,
+    )
+
+    image_idx = 0
+    result: list[str] = []
+    i = 0
+    pad_len = len(image_token)
+    while i < len(prompt):
+        if prompt[i : i + pad_len] == image_token:
+            n = (
+                n_tokens_per_image[image_idx]
+                if image_idx < len(n_tokens_per_image)
+                else 1
+            )
+            result.append(image_token * n)
+            image_idx += 1
+            i += pad_len
+        else:
+            result.append(prompt[i])
+            i += 1
+
+    return "".join(result)
+
+
+@dataclass
+class MediaRegion:
+    content_hash: str
+    start_pos: int
+    end_pos: int
+
+
+@dataclass
+class VisionResult:
+    prompt: str
+    prompt_tokens: mx.array
+    embeddings: mx.array
+    media_regions: list[MediaRegion]
+
+
+class VisionEncoder:
+    def __init__(self, config: VisionCardConfig, model_id: ModelId):
+        self._config = config
+        self._main_model_path = build_model_path(model_id)
+        self._model_path = build_model_path(ModelId(config.weights_repo))
+        self._vision_tower: nn.Module | None = None
+        self._projector: nn.Module | None = None
+        self._processor: "ImageProcessor | None" = None
+        self._spatial_merge_size: int = 2
+        self._merge_kernel_size: list[int] | None = None
+        self._needs_nhwc: bool = False
+        self._loaded = False
+
+    def _load_config_json(self) -> dict[str, Any]:
+        for candidate in (self._main_model_path, self._model_path):
+            path = candidate / "config.json"
+            if path.exists():
+                with open(path) as f:
+                    return json.load(f)  # type: ignore
+        return {}
+
+    def _import_mlx_vlm(self, *submodules: str) -> Any:  # type: ignore
+        mt = self._config.model_type
+        results: list[Any] = []
+        for sub in submodules:
+            name = f"mlx_vlm.models.{mt}.{sub}"
+            results.append(importlib.import_module(name))
+        return results[0] if len(results) == 1 else tuple(results)
+
+    def ensure_loaded(self) -> None:
+        if self._loaded:
+            return
+        self._load_weights()
+        self._loaded = True
+
+    def _load_weights(self) -> None:
+        _patch_video_processor()
+        logger.info(f"Loading vision weights from {self._model_path}")
+        config = self._load_config_json()
+        if not config:
+            raise FileNotFoundError(f"config.json not found in {self._model_path}")
+
+        vision_cfg = config.get("vision_config", {})  # type: ignore
+
+        config_mod, vision_mod = self._import_mlx_vlm("config", "vision")  # type: ignore
+        vision_config_cls = config_mod.VisionConfig  # type: ignore
+        vision_model_cls = vision_mod.VisionModel  # type: ignore
+
+        vision_config = vision_config_cls(  # type: ignore
+            **_filter_config(vision_config_cls, vision_cfg)  # type: ignore
+        )
+        self._spatial_merge_size = getattr(vision_config, "spatial_merge_size", 2)  # type: ignore
+        self._vision_tower = vision_model_cls(vision_config)
+        model_mod: Any = None
+        with contextlib.suppress(ImportError):
+            model_mod = self._import_mlx_vlm(self._config.model_type)  # type: ignore
+
+        projector_cls = None
+        if model_mod is not None:
+            for attr_name in dir(model_mod):  # type: ignore
+                obj = getattr(model_mod, attr_name)  # type: ignore
+                if (
+                    isinstance(obj, type)
+                    and issubclass(obj, nn.Module)
+                    and "Projector" in attr_name
+                ):
+                    projector_cls = obj
+                    break
+
+        if projector_cls is not None:
+            text_config = config_mod.TextConfig(  # type: ignore
+                **_filter_config(config_mod.TextConfig, config.get("text_config", {}))  # type: ignore
+            )
+            extra = {
+                k: v
+                for k, v in config.items()  # type: ignore
+                if k not in ("text_config", "vision_config")
+            }
+            extra.setdefault("model_type", self._config.model_type)
+            model_config = config_mod.ModelConfig(  # type: ignore
+                text_config=text_config,
+                vision_config=vision_config,
+                **_filter_config(config_mod.ModelConfig, extra),  # type: ignore
+            )
+            self._projector = projector_cls(model_config)  # type: ignore
+
+        processor_repo = self._config.processor_repo
+        if processor_repo:
+            self._load_weights_from_separate_repo()
+        else:
+            self._load_weights_from_model_repo()
+
+        repo = processor_repo or str(self._model_path)
+        image_proc = load_image_processor(repo)
+        if image_proc is not None:
+            self._processor = image_proc
+        else:
+            self._processor = AutoImageProcessor.from_pretrained(  # type: ignore
+                repo, trust_remote_code=True
+            )
+        if processor_repo:
+            self._merge_kernel_size = vision_cfg.get("merge_kernel_size", [2, 2])  # type: ignore
+            self._needs_nhwc = True
+        logger.info(f"HF image processor loaded from {repo}")
+
+    def _load_weights_from_separate_repo(self) -> None:
+        safetensors_files = list(self._model_path.glob("*.safetensors"))
+        if not safetensors_files:
+            raise FileNotFoundError(f"No safetensors files found in {self._model_path}")
+
+        weights: dict[str, mx.array] = {}
+        for sf_path in safetensors_files:
+            with safe_open(str(sf_path), framework="pt") as f:
+                keys = f.keys()
+                for key in keys:
+                    tensor = f.get_tensor(key)  # type: ignore
+                    np_tensor = tensor.float().numpy()  # type: ignore
+                    weights[key] = mx.array(np_tensor, dtype=mx.bfloat16)  # type: ignore
+
+        vision_weights: dict[str, mx.array] = {}
+        projector_weights: dict[str, mx.array] = {}
+        for key, val in weights.items():
+            if key.startswith("vision_tower."):
+                short_key = key[len("vision_tower.") :]
+                if short_key.startswith("encoder."):
+                    short_key = short_key[len("encoder.") :]
+                m = re.match(r"^(blocks\.\d+)\.(wqkv|wo)\.(weight|bias)$", short_key)
+                if m:
+                    short_key = f"{m.group(1)}.attn.{m.group(2)}.{m.group(3)}"
+                if short_key == "patch_embed.proj.weight" and val.ndim == 4:
+                    val = val.transpose(0, 2, 3, 1)
+                vision_weights[short_key] = val
+            elif key.startswith(("mm_projector.", "multi_modal_projector.")):
+                if key.startswith("multi_modal_projector."):
+                    short_key = key[len("multi_modal_projector.") :]
+                    if short_key.startswith("mm_projector."):
+                        short_key = short_key[len("mm_projector.") :]
+                else:
+                    short_key = key[len("mm_projector.") :]
+                short_key = short_key.replace("proj.0.", "linear_1.").replace(
+                    "proj.2.", "linear_2."
+                )
+                projector_weights[short_key] = val
+
+        assert self._vision_tower is not None
+        self._vision_tower.load_weights(list(vision_weights.items()))
+        mx.eval(self._vision_tower.parameters())
+
+        if self._projector is not None and projector_weights:
+            self._projector.load_weights(list(projector_weights.items()))
+            mx.eval(self._projector.parameters())
+
+        n_vision = sum(v.size for _, v in vision_weights.items())
+        n_proj = sum(v.size for _, v in projector_weights.items())
+        logger.info(
+            f"Vision encoder loaded: {n_vision / 1e6:.1f}M params"
+            + (f", projector: {n_proj / 1e6:.1f}M params" if n_proj else "")
+        )
+
+    def _load_weights_from_model_repo(self) -> None:
+        safetensors_files = sorted(self._model_path.glob("*.safetensors"))
+        if not safetensors_files:
+            raise FileNotFoundError(f"No safetensors files found in {self._model_path}")
+
+        vision_prefixes = ["vision_tower.", "model.visual."]
+        vision_weights: dict[str, mx.array] = {}
+        found_raw_prefix = False
+        for sf_path in safetensors_files:
+            file_weights: dict[str, mx.array] = mx.load(str(sf_path))  # type: ignore
+            for key, val in file_weights.items():
+                for prefix in vision_prefixes:
+                    if key.startswith(prefix):
+                        short_key = key[len(prefix) :]
+                        vision_weights[short_key] = val
+                        if prefix == "model.visual.":
+                            found_raw_prefix = True
+                        break
+
+        if not vision_weights:
+            raise ValueError(
+                f"No vision weights found with prefixes {vision_prefixes} in {self._model_path}. "
+                "Ensure the model repo contains bundled vision weights."
+            )
+
+        assert self._vision_tower is not None
+        if found_raw_prefix and hasattr(self._vision_tower, "sanitize"):
+            vision_weights = self._vision_tower.sanitize(vision_weights)  # type: ignore
+
+        self._vision_tower.load_weights(list(vision_weights.items()))  # type: ignore
+        mx.eval(self._vision_tower.parameters())
+
+        n_vision = sum(v.size for _, v in vision_weights.items())  # type: ignore
+        logger.info(f"Vision encoder loaded: {n_vision / 1e6:.1f}M params")
+
+    def encode_images(self, images: list[str]) -> tuple[mx.array, list[int]]:
+        self.ensure_loaded()
+        assert self._vision_tower is not None
+        assert self._processor is not None
+
+        pil_images = [decode_base64_image(b64) for b64 in images]
+        for idx, img in enumerate(pil_images):
+            logger.info(f"Image {idx}: {img.width}x{img.height} mode={img.mode}")
+
+        if self._config.processor_repo:
+            processed = self._processor.preprocess(
+                [{"type": "image", "image": img} for img in pil_images],
+                return_tensors="np",
+            )
+            pixel_values = mx.array(processed["pixel_values"])  # type: ignore
+            grid_thw = mx.array(processed["grid_thws"])  # type: ignore
+            assert self._merge_kernel_size is not None
+            merge_length = int(np.prod(self._merge_kernel_size))
+            n_tokens_per_image = [
+                int(mx.prod(grid_thw[i]).item()) // merge_length
+                for i in range(grid_thw.shape[0])
+            ]
+        else:
+            processed = self._processor(
+                images=pil_images,
+                return_tensors="np",
+            )
+            pixel_values = mx.array(processed["pixel_values"])  # type: ignore
+            grid_thw = mx.array(processed["image_grid_thw"])  # type: ignore
+            merge_unit = self._spatial_merge_size**2
+            n_tokens_per_image = [
+                int(
+                    grid_thw[i, 0].item()
+                    * grid_thw[i, 1].item()
+                    * grid_thw[i, 2].item()
+                )
+                // merge_unit
+                for i in range(grid_thw.shape[0])
+            ]
+
+        if self._needs_nhwc:
+            grid_hw = grid_thw[:, 1:] if grid_thw.shape[-1] == 3 else grid_thw
+            hidden_states = self._vision_tower(
+                pixel_values.transpose(0, 2, 3, 1),
+                output_hidden_states=True,
+                grid_thw=grid_hw,
+            )
+        else:
+            result = self._vision_tower(pixel_values, grid_thw)
+            hidden_states = result[0] if isinstance(result, tuple) else result
+
+        if self._projector is not None:
+            image_features: mx.array = self._projector(hidden_states)
+        else:
+            image_features = hidden_states
+
+        return image_features, n_tokens_per_image
+
+
+def get_inner_model(model: nn.Module) -> Any:  # type: ignore
+    for candidate in (
+        getattr(model, "model", None),
+        getattr(getattr(model, "language_model", None), "model", None),
+    ):
+        if candidate is not None and hasattr(candidate, "embed_tokens"):  # type: ignore
+            return candidate  # type: ignore
+
+    raise ValueError(
+        f"Could not find inner transformer (embed_tokens) in {type(model).__name__}. "
+        "Add a new pattern to _get_inner_model() for this architecture."
+    )
+
+
+def create_vision_embeddings(
+    model: Model,
+    prompt_tokens: mx.array,
+    image_features: mx.array,
+    image_token_id: int,
+) -> mx.array:
+    inner = get_inner_model(model)  # type: ignore
+    embed_tokens = inner.embed_tokens  # type: ignore
+
+    input_embeddings: mx.array = embed_tokens(prompt_tokens[None])  # type: ignore
+
+    is_image: mx.array = mx.equal(prompt_tokens, image_token_id)
+    n_placeholders = int(mx.sum(is_image).item())
+
+    if n_placeholders > 0:
+        if n_placeholders != image_features.shape[0]:
+            logger.warning(
+                f"Placeholder count ({n_placeholders}) != image features "
+                f"({image_features.shape[0]}). Using min of both."
+            )
+            n = min(n_placeholders, image_features.shape[0])
+            image_features = image_features[:n]
+
+        image_indices = mx.cumsum(is_image.astype(mx.int32)) - 1
+        image_indices = mx.clip(image_indices, 0, image_features.shape[0] - 1)
+
+        gathered = image_features[image_indices].astype(input_embeddings.dtype)
+        result = mx.where(is_image[:, None], gathered, input_embeddings[0])
+        input_embeddings = result[None]
+
+    return input_embeddings
+
+
+def _find_media_regions(
+    prompt_tokens: mx.array,
+    images: list[str],
+    image_token_id: int,
+) -> list[MediaRegion]:
+    tokens_np = np.array(prompt_tokens)
+    is_pad = tokens_np == image_token_id  # type: ignore
+
+    regions: list[MediaRegion] = []
+    in_run = False
+    run_start = 0
+    for pos, pad in enumerate(is_pad):  # type: ignore
+        if pad and not in_run:
+            run_start = pos
+            in_run = True
+        elif not pad and in_run:
+            regions.append(
+                MediaRegion(content_hash="", start_pos=run_start, end_pos=pos)
+            )
+            in_run = False
+    if in_run:
+        regions.append(
+            MediaRegion(content_hash="", start_pos=run_start, end_pos=len(tokens_np))
+        )
+
+    for i, region in enumerate(regions):
+        if i < len(images):
+            img = decode_base64_image(images[i])
+            region.content_hash = hashlib.sha256(img.tobytes()).hexdigest()
+        else:
+            logger.warning(f"Media region {i} has no corresponding image")
+
+    return regions
+
+
+class VisionProcessor:
+    """
+    Pipeline for vision models:
+    1. Encode images into features (or grab from cache)
+    2. Replace image placeholders with the features
+    3. Build vision prompt
+    4. Provide media regions for prefix caching
+    """
+
+    def __init__(self, config: VisionCardConfig, model_id: ModelId):
+        self.vision_config = config
+        self._encoder = VisionEncoder(config, model_id)
+        self._feature_cache: dict[str, tuple[mx.array, list[int]]] = {}
+        self._feature_cache_max = 32
+
+    def load(self) -> None:
+        self._encoder.ensure_loaded()
+
+    def _image_cache_key(self, images: list[str]) -> str:
+        h = hashlib.sha256()
+        for img in images:
+            pil = decode_base64_image(img)
+            h.update(pil.tobytes())
+        return h.hexdigest()
+
+    def process(
+        self,
+        images: list[str],
+        chat_template_messages: list[dict[str, Any]],
+        tokenizer: TokenizerWrapper,
+        model: Model,
+    ) -> VisionResult:
+        logger.info(f"Vision pipeline: {len(images)} image(s)")
+
+        cache_key = self._image_cache_key(images)
+        cached = self._feature_cache.pop(cache_key, None)
+        if cached is not None:
+            self._feature_cache[cache_key] = cached
+            image_features, n_tokens_per_image = cached
+        else:
+            image_features, n_tokens_per_image = self._encoder.encode_images(images)
+            self._feature_cache[cache_key] = (image_features, n_tokens_per_image)
+            while len(self._feature_cache) > self._feature_cache_max:
+                del self._feature_cache[next(iter(self._feature_cache))]
+        logger.info(
+            f"Vision features: {image_features.shape} "
+            f"({image_features.shape[0]} tokens, per-image: {n_tokens_per_image})"
+        )
+
+        image_token = self.vision_config.image_token
+        if image_token is None:
+            image_token = tokenizer.decode([self.vision_config.image_token_id])
+
+        formatted_messages = _format_vlm_messages(
+            chat_template_messages, self.vision_config.model_type
+        )
+
+        prompt = build_vision_prompt(
+            tokenizer,
+            formatted_messages,
+            n_tokens_per_image,
+            image_token,
+        )
+
+        logger.info(
+            f"Expanded prompt has {prompt.count(image_token)} image_token occurrences, total len={len(prompt)}"
+        )
+
+        prompt_tokens: mx.array = encode_prompt(tokenizer, prompt)
+        prompt_tokens = fix_unmatched_think_end_tokens(prompt_tokens, tokenizer)
+        n_image_tokens = int(
+            mx.sum(mx.equal(prompt_tokens, self.vision_config.image_token_id)).item()
+        )
+        logger.info(
+            f"Encoded prompt: {len(prompt_tokens)} tokens, {n_image_tokens} image pad tokens"
+        )
+
+        embeddings = create_vision_embeddings(
+            model,
+            prompt_tokens,
+            image_features,
+            self.vision_config.image_token_id,
+        )
+        mx.eval(embeddings)
+
+        media_regions = _find_media_regions(
+            prompt_tokens,
+            images,
+            self.vision_config.image_token_id,
+        )
+
+        return VisionResult(
+            prompt=prompt,
+            prompt_tokens=prompt_tokens,
+            embeddings=embeddings,
+            media_regions=media_regions,
+        )
+
+
+def prepare_vision(
+    images: list[str] | None,
+    chat_template_messages: list[dict[str, Any]] | None,
+    vision_processor: VisionProcessor,
+    tokenizer: TokenizerWrapper,
+    model: Model,
+) -> VisionResult | None:
+    if not images:
+        return None
+    if chat_template_messages is None:
+        logger.warning(
+            "Vision request missing chat_template_messages — ignoring images"
+        )
+        return None
+
+    return vision_processor.process(
+        images=images,
+        chat_template_messages=chat_template_messages,
+        tokenizer=tokenizer,
+        model=model,
+    )

src/exo/worker/main.py

@@ -1,3 +1,4 @@
+import hashlib
 from collections import defaultdict
 from datetime import datetime, timezone
 
@@ -9,6 +10,7 @@ from exo.api.types import ImageEditsTaskParams
 from exo.download.download_utils import is_read_only_model_dir, resolve_existing_model
 from exo.shared.apply import apply
 from exo.shared.models.model_cards import ModelId, add_to_card_cache, delete_custom_card
+from exo.shared.types.chunks import InputImageChunk
 from exo.shared.types.commands import (
     ForwarderCommand,
     ForwarderDownloadCommand,
@@ -38,6 +40,7 @@ from exo.shared.types.tasks import (
     Shutdown,
     Task,
     TaskStatus,
+    TextGeneration,
 )
 from exo.shared.types.topology import Connection, SocketConnection
 from exo.shared.types.worker.downloads import DownloadCompleted
@@ -76,8 +79,9 @@ class Worker:
         self._system_id = SystemId()
 
         # Buffer for input image chunks (for image editing)
-        self.input_chunk_buffer: dict[CommandId, dict[int, str]] = {}
+        self.input_chunk_buffer: dict[CommandId, dict[int, InputImageChunk]] = {}
         self.input_chunk_counts: dict[CommandId, int] = {}
+        self.image_cache: dict[str, str] = {}
 
         self._download_backoff: KeyedBackoff[ModelId] = KeyedBackoff(base=0.5, cap=10.0)
         self._stopped: anyio.Event = anyio.Event()
@@ -131,7 +135,7 @@ class Worker:
                         self.input_chunk_counts[cmd_id] = event.chunk.total_chunks
 
                     self.input_chunk_buffer[cmd_id][event.chunk.chunk_index] = (
-                        event.chunk.data
+                        event.chunk
                     )
 
                 if isinstance(event, CustomModelCardAdded):
@@ -152,7 +156,6 @@ class Worker:
                 self.state.runners,
                 self.state.tasks,
                 self.input_chunk_buffer,
-                self.input_chunk_counts,
             )
             if task is None:
                 continue
@@ -245,7 +248,7 @@ class Worker:
                     # Assemble image from chunks and inject into task
                     cmd_id = task.command_id
                     chunks = self.input_chunk_buffer.get(cmd_id, {})
-                    assembled = "".join(chunks[i] for i in range(len(chunks)))
+                    assembled = "".join(chunks[i].data for i in range(len(chunks)))
                     logger.info(
                         f"Assembled input image from {len(chunks)} chunks, "
                         f"total size: {len(assembled)} bytes"
@@ -279,6 +282,52 @@ class Worker:
                     if cmd_id in self.input_chunk_counts:
                         del self.input_chunk_counts[cmd_id]
                     await self._start_runner_task(modified_task)
+
+                case TextGeneration() if (
+                    task.task_params.image_hashes
+                    or task.task_params.total_input_chunks > 0
+                ):
+                    cmd_id = task.command_id
+                    by_index: dict[int, str] = {}
+
+                    for idx, h in task.task_params.image_hashes.items():
+                        assert h in self.image_cache
+                        by_index[idx] = self.image_cache[h]
+
+                    if task.task_params.total_input_chunks > 0:
+                        chunk_buffer = self.input_chunk_buffer.get(cmd_id, {})
+                        per_image: defaultdict[int, list[InputImageChunk]] = (
+                            defaultdict(list)
+                        )
+                        for chunk in chunk_buffer.values():
+                            per_image[chunk.image_index].append(chunk)
+                        for img_idx in sorted(per_image):
+                            sorted_chunks = sorted(
+                                per_image[img_idx], key=lambda c: c.chunk_index
+                            )
+                            img = "".join(c.data for c in sorted_chunks)
+                            self.image_cache[
+                                hashlib.sha256(img.encode("ascii")).hexdigest()
+                            ] = img
+                            by_index[img_idx] = img
+                        logger.info(
+                            f"Assembled {len(per_image)} VLM image(s) "
+                            f"from {len(chunk_buffer)} chunks"
+                        )
+
+                    resolved_images = [by_index[i] for i in sorted(by_index)]
+                    modified_task = task.model_copy(
+                        update={
+                            "task_params": task.task_params.model_copy(
+                                update={"images": resolved_images}
+                            )
+                        }
+                    )
+                    if cmd_id in self.input_chunk_buffer:
+                        del self.input_chunk_buffer[cmd_id]
+                    if cmd_id in self.input_chunk_counts:
+                        del self.input_chunk_counts[cmd_id]
+                    await self._start_runner_task(modified_task)
                 case task:
                     await self._start_runner_task(task)
 

src/exo/worker/plan.py

@@ -2,6 +2,7 @@
 
 from collections.abc import Mapping, Sequence
 
+from exo.shared.types.chunks import InputImageChunk
 from exo.shared.types.common import CommandId, NodeId
 from exo.shared.types.tasks import (
     CancelTask,
@@ -49,8 +50,7 @@ def plan(
     instances: Mapping[InstanceId, Instance],
     all_runners: Mapping[RunnerId, RunnerStatus],  # all global
     tasks: Mapping[TaskId, Task],
-    input_chunk_buffer: Mapping[CommandId, dict[int, str]] | None = None,
-    input_chunk_counts: Mapping[CommandId, int] | None = None,
+    input_chunk_buffer: Mapping[CommandId, Mapping[int, InputImageChunk]] | None = None,
 ) -> Task | None:
     # Python short circuiting OR logic should evaluate these sequentially.
     return (
@@ -272,7 +272,7 @@ def _pending_tasks(
     runners: Mapping[RunnerId, RunnerSupervisor],
     tasks: Mapping[TaskId, Task],
     all_runners: Mapping[RunnerId, RunnerStatus],
-    input_chunk_buffer: Mapping[CommandId, dict[int, str]],
+    input_chunk_buffer: Mapping[CommandId, Mapping[int, InputImageChunk]] | None,
 ) -> Task | None:
     for task in tasks.values():
         # for now, just forward chat completions
@@ -282,12 +282,15 @@ def _pending_tasks(
         if task.task_status not in (TaskStatus.Pending, TaskStatus.Running):
             continue
 
-        # For ImageEdits tasks, verify all input chunks have been received
-        if isinstance(task, ImageEdits) and task.task_params.total_input_chunks > 0:
+        # For tasks with images, verify all input chunks have been received
+        expected_image_chunks = 0
+        if isinstance(task, (ImageEdits, TextGeneration)):
+            expected_image_chunks = task.task_params.total_input_chunks
+        if expected_image_chunks > 0:
+            assert input_chunk_buffer is not None
             cmd_id = task.command_id
-            expected = task.task_params.total_input_chunks
             received = len(input_chunk_buffer.get(cmd_id, {}))
-            if received < expected:
+            if received < expected_image_chunks:
                 continue  # Wait for all chunks to arrive
 
         for runner in runners.values():

src/exo/worker/runner/llm_inference/batch_generator.py

@@ -29,6 +29,7 @@ from exo.worker.engines.mlx.utils_mlx import (
     mx_all_gather_tasks,
     mx_any,
 )
+from exo.worker.engines.mlx.vision import VisionProcessor
 from exo.worker.runner.bootstrap import logger
 
 from .model_output_parsers import apply_all_parsers
@@ -120,6 +121,7 @@ class SequentialGenerator(InferenceGenerator):
     device_rank: int
     cancel_receiver: MpReceiver[TaskId]
     event_sender: MpSender[Event]
+    vision_processor: VisionProcessor | None = None
     check_for_cancel_every: int = 50
 
     _cancelled_tasks: set[TaskId] = field(default_factory=set, init=False)
@@ -304,6 +306,7 @@ class SequentialGenerator(InferenceGenerator):
             distributed_prompt_progress_callback=distributed_prompt_progress_callback,
             on_generation_token=on_generation_token,
             group=self.group,
+            vision_processor=self.vision_processor,
         )
 
     def close(self) -> None:
@@ -322,6 +325,7 @@ class BatchGenerator(InferenceGenerator):
     cancel_receiver: MpReceiver[TaskId]
     event_sender: MpSender[Event]
     check_for_cancel_every: int = 50
+    vision_processor: VisionProcessor | None = None
 
     _cancelled_tasks: set[TaskId] = field(default_factory=set, init=False)
     _maybe_queue: list[TextGeneration] = field(default_factory=list, init=False)
@@ -344,6 +348,7 @@ class BatchGenerator(InferenceGenerator):
             tokenizer=self.tokenizer,
             group=self.group,
             kv_prefix_cache=self.kv_prefix_cache,
+            vision_processor=self.vision_processor,
         )
 
     def warmup(self):

src/exo/worker/runner/llm_inference/runner.py

@@ -57,6 +57,7 @@ from exo.worker.engines.mlx.utils_mlx import (
     initialize_mlx,
     load_mlx_items,
 )
+from exo.worker.engines.mlx.vision import VisionProcessor
 from exo.worker.runner.bootstrap import logger
 from exo.worker.runner.llm_inference.batch_generator import (
     BatchGenerator,
@@ -103,7 +104,9 @@ class Runner:
         self.setup_start_time = time.time()
 
         self.generator: Builder | InferenceGenerator = Builder(
-            self.model_id, self.event_sender, self.cancel_receiver
+            self.model_id,
+            self.event_sender,
+            self.cancel_receiver,
         )
 
         self.seen: set[TaskId] = set()
@@ -195,14 +198,16 @@ class Runner:
                 assert (
                     ModelTask.TextGeneration in self.shard_metadata.model_card.tasks
                 ), f"Incorrect model task(s): {self.shard_metadata.model_card.tasks}"
-                self.generator.inference_model, self.generator.tokenizer = (
-                    load_mlx_items(
+                (
+                    self.generator.inference_model,
+                    self.generator.tokenizer,
+                    self.generator.vision_processor,
+                ) = load_mlx_items(
                     self.bound_instance,
                     self.generator.group,
                     on_timeout=on_model_load_timeout,
                     on_layer_loaded=on_layer_loaded,
                 )
-                )
 
                 self.generator = self.generator.build()
 
@@ -381,6 +386,7 @@ class Builder:
     inference_model: Model | None = None
     tokenizer: TokenizerWrapper | None = None
     group: mx.distributed.Group | None = None
+    vision_processor: VisionProcessor | None = None
 
     def build(
         self,
@@ -389,6 +395,8 @@ class Builder:
         assert self.inference_model
         assert self.tokenizer
 
+        vision_processor = self.vision_processor
+
         tool_parser = None
         logger.info(
             f"model has_tool_calling={self.tokenizer.has_tool_calling} using tokens {self.tokenizer.tool_call_start}, {self.tokenizer.tool_call_end}"
@@ -419,6 +427,7 @@ class Builder:
                 device_rank=device_rank,
                 cancel_receiver=self.cancel_receiver,
                 event_sender=self.event_sender,
+                vision_processor=vision_processor,
             )
         logger.info("using BatchGenerator")
         return BatchGenerator(
@@ -431,4 +440,5 @@ class Builder:
             device_rank=device_rank,
             cancel_receiver=self.cancel_receiver,
             event_sender=self.event_sender,
+            vision_processor=vision_processor,
         )

src/exo/worker/tests/unittests/test_mlx/test_tokenizers.py

@@ -137,7 +137,9 @@ async def test_tokenizer_encode_decode(model_card: ModelCard) -> None:
     # Test decoding
     decoded = tokenizer.decode(encoded)
     assert isinstance(decoded, str), f"decode() should return a string for {model_id}"
-    assert test_text in decoded or decoded.strip() == test_text.strip(), (
+    normalized_decoded = decoded.replace(" ", "").lower()
+    normalized_expected = test_text.replace(" ", "").lower()
+    assert normalized_expected in normalized_decoded, (
         f"decode(encode(x)) should preserve text for {model_id}: got {decoded!r}"
     )
 

src/exo/worker/tests/unittests/test_runner/test_event_ordering.py

@@ -115,7 +115,9 @@ def assert_events_equal(test_events: Iterable[Event], true_events: Iterable[Even
 def patch_out_mlx(monkeypatch: pytest.MonkeyPatch):
     # initialize_mlx returns a mock group
     monkeypatch.setattr(mlx_runner, "initialize_mlx", make_nothin(MockGroup()))
-    monkeypatch.setattr(mlx_runner, "load_mlx_items", make_nothin((1, MockTokenizer)))
+    monkeypatch.setattr(
+        mlx_runner, "load_mlx_items", make_nothin((1, MockTokenizer, None))
+    )
     monkeypatch.setattr(mlx_batch_generator, "warmup_inference", make_nothin(1))
     monkeypatch.setattr(mlx_batch_generator, "_check_for_debug_prompts", nothin)
     monkeypatch.setattr(mlx_batch_generator, "mx_any", make_nothin(False))

tests/test_vision_cache.py

@@ -0,0 +1,63 @@
+from exo.worker.engines.mlx.cache import KVPrefixCache
+from exo.worker.engines.mlx.vision import MediaRegion
+
+validate = KVPrefixCache._validate_media_match
+
+
+class TestValidateMediaMatch:
+    def test_text_only_no_truncation(self):
+        assert validate(8000, [], []) == 8000
+
+    def test_text_prefix_before_image(self):
+        cached = [MediaRegion("hashA", 5000, 8600)]
+        assert validate(5000, cached, []) == 5000
+
+    def test_same_image_same_position(self):
+        cached = [MediaRegion("hashA", 5000, 8600)]
+        query = [MediaRegion("hashA", 5000, 8600)]
+        assert validate(9000, cached, query) == 9000
+
+    def test_different_image_truncates(self):
+        cached = [MediaRegion("hashA", 5000, 8600)]
+        query = [MediaRegion("hashB", 5000, 8600)]
+        assert validate(9000, cached, query) == 5000
+
+    def test_match_below_region_start(self):
+        cached = [MediaRegion("hashA", 5000, 8600)]
+        query = [MediaRegion("hashB", 5000, 8600)]
+        assert validate(4000, cached, query) == 4000
+
+    def test_text_followup_no_images_in_query(self):
+        cached = [MediaRegion("hashA", 5000, 8600)]
+        assert validate(9000, cached, []) == 9000
+
+    def test_multiple_images_first_mismatch_truncates(self):
+        cached = [
+            MediaRegion("hashA", 2000, 4000),
+            MediaRegion("hashB", 6000, 8000),
+        ]
+        query = [
+            MediaRegion("hashA", 2000, 4000),
+            MediaRegion("hashC", 6000, 8000),
+        ]
+        assert validate(9000, cached, query) == 6000
+
+    def test_multiple_images_all_match(self):
+        cached = [
+            MediaRegion("hashA", 2000, 4000),
+            MediaRegion("hashB", 6000, 8000),
+        ]
+        query = [
+            MediaRegion("hashA", 2000, 4000),
+            MediaRegion("hashB", 6000, 8000),
+        ]
+        assert validate(9000, cached, query) == 9000
+
+    def test_no_cached_regions(self):
+        query = [MediaRegion("hashA", 100, 200)]
+        assert validate(500, [], query) == 500
+
+    def test_cached_region_beyond_match(self):
+        cached = [MediaRegion("hashA", 10000, 12000)]
+        query = [MediaRegion("hashB", 10000, 12000)]
+        assert validate(5000, cached, query) == 5000

uv.lock

@@ -207,20 +207,32 @@ name = "cffi"
 version = "2.0.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "pycparser", marker = "implementation_name != 'PyPy' and sys_platform == 'linux'" },
+    { name = "pycparser", marker = "(implementation_name != 'PyPy' and sys_platform == 'darwin') or (implementation_name != 'PyPy' and sys_platform == 'linux')" },
 ]
 sdist = { url = "https://files.pythonhosted.org/packages/eb/56/b1ba7935a17738ae8453301356628e8147c79dbb825bcbc73dc7401f9846/cffi-2.0.0.tar.gz", hash = "sha256:44d1b5909021139fe36001ae048dbdde8214afa20200eda0f64c068cac5d5529", size = 523588, upload-time = "2025-09-08T23:24:04.541Z" }
 wheels = [
+    { url = "https://files.pythonhosted.org/packages/4b/8d/a0a47a0c9e413a658623d014e91e74a50cdd2c423f7ccfd44086ef767f90/cffi-2.0.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:00bdf7acc5f795150faa6957054fbbca2439db2f775ce831222b66f192f03beb", size = 185230, upload-time = "2025-09-08T23:23:00.879Z" },
+    { url = "https://files.pythonhosted.org/packages/4a/d2/a6c0296814556c68ee32009d9c2ad4f85f2707cdecfd7727951ec228005d/cffi-2.0.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:45d5e886156860dc35862657e1494b9bae8dfa63bf56796f2fb56e1679fc0bca", size = 181043, upload-time = "2025-09-08T23:23:02.231Z" },
     { url = "https://files.pythonhosted.org/packages/b0/1e/d22cc63332bd59b06481ceaac49d6c507598642e2230f201649058a7e704/cffi-2.0.0-cp313-cp313-manylinux1_i686.manylinux2014_i686.manylinux_2_17_i686.manylinux_2_5_i686.whl", hash = "sha256:07b271772c100085dd28b74fa0cd81c8fb1a3ba18b21e03d7c27f3436a10606b", size = 212446, upload-time = "2025-09-08T23:23:03.472Z" },
     { url = "https://files.pythonhosted.org/packages/a9/f5/a2c23eb03b61a0b8747f211eb716446c826ad66818ddc7810cc2cc19b3f2/cffi-2.0.0-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:d48a880098c96020b02d5a1f7d9251308510ce8858940e6fa99ece33f610838b", size = 220101, upload-time = "2025-09-08T23:23:04.792Z" },
+    { url = "https://files.pythonhosted.org/packages/f2/7f/e6647792fc5850d634695bc0e6ab4111ae88e89981d35ac269956605feba/cffi-2.0.0-cp313-cp313-manylinux2014_ppc64le.manylinux_2_17_ppc64le.whl", hash = "sha256:f93fd8e5c8c0a4aa1f424d6173f14a892044054871c771f8566e4008eaa359d2", size = 207948, upload-time = "2025-09-08T23:23:06.127Z" },
+    { url = "https://files.pythonhosted.org/packages/cb/1e/a5a1bd6f1fb30f22573f76533de12a00bf274abcdc55c8edab639078abb6/cffi-2.0.0-cp313-cp313-manylinux2014_s390x.manylinux_2_17_s390x.whl", hash = "sha256:dd4f05f54a52fb558f1ba9f528228066954fee3ebe629fc1660d874d040ae5a3", size = 206422, upload-time = "2025-09-08T23:23:07.753Z" },
     { url = "https://files.pythonhosted.org/packages/98/df/0a1755e750013a2081e863e7cd37e0cdd02664372c754e5560099eb7aa44/cffi-2.0.0-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:c8d3b5532fc71b7a77c09192b4a5a200ea992702734a2e9279a37f2478236f26", size = 219499, upload-time = "2025-09-08T23:23:09.648Z" },
     { url = "https://files.pythonhosted.org/packages/50/e1/a969e687fcf9ea58e6e2a928ad5e2dd88cc12f6f0ab477e9971f2309b57c/cffi-2.0.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:d9b29c1f0ae438d5ee9acb31cadee00a58c46cc9c0b2f9038c6b0b3470877a8c", size = 222928, upload-time = "2025-09-08T23:23:10.928Z" },
     { url = "https://files.pythonhosted.org/packages/36/54/0362578dd2c9e557a28ac77698ed67323ed5b9775ca9d3fe73fe191bb5d8/cffi-2.0.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:6d50360be4546678fc1b79ffe7a66265e28667840010348dd69a314145807a1b", size = 221302, upload-time = "2025-09-08T23:23:12.42Z" },
+    { url = "https://files.pythonhosted.org/packages/92/c4/3ce07396253a83250ee98564f8d7e9789fab8e58858f35d07a9a2c78de9f/cffi-2.0.0-cp314-cp314-macosx_10_13_x86_64.whl", hash = "sha256:fc33c5141b55ed366cfaad382df24fe7dcbc686de5be719b207bb248e3053dc5", size = 185320, upload-time = "2025-09-08T23:23:18.087Z" },
+    { url = "https://files.pythonhosted.org/packages/59/dd/27e9fa567a23931c838c6b02d0764611c62290062a6d4e8ff7863daf9730/cffi-2.0.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:c654de545946e0db659b3400168c9ad31b5d29593291482c43e3564effbcee13", size = 181487, upload-time = "2025-09-08T23:23:19.622Z" },
     { url = "https://files.pythonhosted.org/packages/d6/43/0e822876f87ea8a4ef95442c3d766a06a51fc5298823f884ef87aaad168c/cffi-2.0.0-cp314-cp314-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:24b6f81f1983e6df8db3adc38562c83f7d4a0c36162885ec7f7b77c7dcbec97b", size = 220049, upload-time = "2025-09-08T23:23:20.853Z" },
+    { url = "https://files.pythonhosted.org/packages/b4/89/76799151d9c2d2d1ead63c2429da9ea9d7aac304603de0c6e8764e6e8e70/cffi-2.0.0-cp314-cp314-manylinux2014_ppc64le.manylinux_2_17_ppc64le.whl", hash = "sha256:12873ca6cb9b0f0d3a0da705d6086fe911591737a59f28b7936bdfed27c0d47c", size = 207793, upload-time = "2025-09-08T23:23:22.08Z" },
+    { url = "https://files.pythonhosted.org/packages/bb/dd/3465b14bb9e24ee24cb88c9e3730f6de63111fffe513492bf8c808a3547e/cffi-2.0.0-cp314-cp314-manylinux2014_s390x.manylinux_2_17_s390x.whl", hash = "sha256:d9b97165e8aed9272a6bb17c01e3cc5871a594a446ebedc996e2397a1c1ea8ef", size = 206300, upload-time = "2025-09-08T23:23:23.314Z" },
     { url = "https://files.pythonhosted.org/packages/47/d9/d83e293854571c877a92da46fdec39158f8d7e68da75bf73581225d28e90/cffi-2.0.0-cp314-cp314-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:afb8db5439b81cf9c9d0c80404b60c3cc9c3add93e114dcae767f1477cb53775", size = 219244, upload-time = "2025-09-08T23:23:24.541Z" },
     { url = "https://files.pythonhosted.org/packages/2b/0f/1f177e3683aead2bb00f7679a16451d302c436b5cbf2505f0ea8146ef59e/cffi-2.0.0-cp314-cp314-musllinux_1_2_aarch64.whl", hash = "sha256:737fe7d37e1a1bffe70bd5754ea763a62a066dc5913ca57e957824b72a85e205", size = 222828, upload-time = "2025-09-08T23:23:26.143Z" },
     { url = "https://files.pythonhosted.org/packages/c6/0f/cafacebd4b040e3119dcb32fed8bdef8dfe94da653155f9d0b9dc660166e/cffi-2.0.0-cp314-cp314-musllinux_1_2_x86_64.whl", hash = "sha256:38100abb9d1b1435bc4cc340bb4489635dc2f0da7456590877030c9b3d40b0c1", size = 220926, upload-time = "2025-09-08T23:23:27.873Z" },
+    { url = "https://files.pythonhosted.org/packages/3e/61/c768e4d548bfa607abcda77423448df8c471f25dbe64fb2ef6d555eae006/cffi-2.0.0-cp314-cp314t-macosx_10_13_x86_64.whl", hash = "sha256:9a67fc9e8eb39039280526379fb3a70023d77caec1852002b4da7e8b270c4dd9", size = 188773, upload-time = "2025-09-08T23:23:29.347Z" },
+    { url = "https://files.pythonhosted.org/packages/2c/ea/5f76bce7cf6fcd0ab1a1058b5af899bfbef198bea4d5686da88471ea0336/cffi-2.0.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:7a66c7204d8869299919db4d5069a82f1561581af12b11b3c9f48c584eb8743d", size = 185013, upload-time = "2025-09-08T23:23:30.63Z" },
     { url = "https://files.pythonhosted.org/packages/be/b4/c56878d0d1755cf9caa54ba71e5d049479c52f9e4afc230f06822162ab2f/cffi-2.0.0-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:7cc09976e8b56f8cebd752f7113ad07752461f48a58cbba644139015ac24954c", size = 221593, upload-time = "2025-09-08T23:23:31.91Z" },
+    { url = "https://files.pythonhosted.org/packages/e0/0d/eb704606dfe8033e7128df5e90fee946bbcb64a04fcdaa97321309004000/cffi-2.0.0-cp314-cp314t-manylinux2014_ppc64le.manylinux_2_17_ppc64le.whl", hash = "sha256:92b68146a71df78564e4ef48af17551a5ddd142e5190cdf2c5624d0c3ff5b2e8", size = 209354, upload-time = "2025-09-08T23:23:33.214Z" },
+    { url = "https://files.pythonhosted.org/packages/d8/19/3c435d727b368ca475fb8742ab97c9cb13a0de600ce86f62eab7fa3eea60/cffi-2.0.0-cp314-cp314t-manylinux2014_s390x.manylinux_2_17_s390x.whl", hash = "sha256:b1e74d11748e7e98e2f426ab176d4ed720a64412b6a15054378afdb71e0f37dc", size = 208480, upload-time = "2025-09-08T23:23:34.495Z" },
     { url = "https://files.pythonhosted.org/packages/d0/44/681604464ed9541673e486521497406fadcc15b5217c3e326b061696899a/cffi-2.0.0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:28a3a209b96630bca57cce802da70c266eb08c6e97e5afd61a75611ee6c64592", size = 221584, upload-time = "2025-09-08T23:23:36.096Z" },
     { url = "https://files.pythonhosted.org/packages/25/8e/342a504ff018a2825d395d44d63a767dd8ebc927ebda557fecdaca3ac33a/cffi-2.0.0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:7553fb2090d71822f02c629afe6042c299edf91ba1bf94951165613553984512", size = 224443, upload-time = "2025-09-08T23:23:37.328Z" },
     { url = "https://files.pythonhosted.org/packages/e1/5e/b666bacbbc60fbf415ba9988324a132c9a7a0448a9a8f125074671c0f2c3/cffi-2.0.0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:6c6c373cfc5c83a975506110d17457138c8c63016b563cc9ed6e056a82f13ce4", size = 223437, upload-time = "2025-09-08T23:23:38.945Z" },
@@ -344,8 +356,10 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/5c/49/498c86566a1d80e978b42f0d702795f69887005548c041636df6ae1ca64c/cryptography-46.0.3-cp311-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:01ca9ff2885f3acc98c29f1860552e37f6d7c7d013d7334ff2a9de43a449315d", size = 4450807, upload-time = "2025-10-15T23:16:56.414Z" },
     { url = "https://files.pythonhosted.org/packages/4b/0a/863a3604112174c8624a2ac3c038662d9e59970c7f926acdcfaed8d61142/cryptography-46.0.3-cp311-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:6eae65d4c3d33da080cff9c4ab1f711b15c1d9760809dad6ea763f3812d254cb", size = 4299615, upload-time = "2025-10-15T23:16:58.442Z" },
     { url = "https://files.pythonhosted.org/packages/64/02/b73a533f6b64a69f3cd3872acb6ebc12aef924d8d103133bb3ea750dc703/cryptography-46.0.3-cp311-abi3-manylinux_2_28_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:e5bf0ed4490068a2e72ac03d786693adeb909981cc596425d09032d372bcc849", size = 4016800, upload-time = "2025-10-15T23:17:00.378Z" },
+    { url = "https://files.pythonhosted.org/packages/25/d5/16e41afbfa450cde85a3b7ec599bebefaef16b5c6ba4ec49a3532336ed72/cryptography-46.0.3-cp311-abi3-manylinux_2_28_ppc64le.whl", hash = "sha256:5ecfccd2329e37e9b7112a888e76d9feca2347f12f37918facbb893d7bb88ee8", size = 4984707, upload-time = "2025-10-15T23:17:01.98Z" },
     { url = "https://files.pythonhosted.org/packages/c9/56/e7e69b427c3878352c2fb9b450bd0e19ed552753491d39d7d0a2f5226d41/cryptography-46.0.3-cp311-abi3-manylinux_2_28_x86_64.whl", hash = "sha256:a2c0cd47381a3229c403062f764160d57d4d175e022c1df84e168c6251a22eec", size = 4482541, upload-time = "2025-10-15T23:17:04.078Z" },
     { url = "https://files.pythonhosted.org/packages/78/f6/50736d40d97e8483172f1bb6e698895b92a223dba513b0ca6f06b2365339/cryptography-46.0.3-cp311-abi3-manylinux_2_34_aarch64.whl", hash = "sha256:549e234ff32571b1f4076ac269fcce7a808d3bf98b76c8dd560e42dbc66d7d91", size = 4299464, upload-time = "2025-10-15T23:17:05.483Z" },
+    { url = "https://files.pythonhosted.org/packages/00/de/d8e26b1a855f19d9994a19c702fa2e93b0456beccbcfe437eda00e0701f2/cryptography-46.0.3-cp311-abi3-manylinux_2_34_ppc64le.whl", hash = "sha256:c0a7bb1a68a5d3471880e264621346c48665b3bf1c3759d682fc0864c540bd9e", size = 4950838, upload-time = "2025-10-15T23:17:07.425Z" },
     { url = "https://files.pythonhosted.org/packages/8f/29/798fc4ec461a1c9e9f735f2fc58741b0daae30688f41b2497dcbc9ed1355/cryptography-46.0.3-cp311-abi3-manylinux_2_34_x86_64.whl", hash = "sha256:10b01676fc208c3e6feeb25a8b83d81767e8059e1fe86e1dc62d10a3018fa926", size = 4481596, upload-time = "2025-10-15T23:17:09.343Z" },
     { url = "https://files.pythonhosted.org/packages/15/8d/03cd48b20a573adfff7652b76271078e3045b9f49387920e7f1f631d125e/cryptography-46.0.3-cp311-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:0abf1ffd6e57c67e92af68330d05760b7b7efb243aab8377e583284dbab72c71", size = 4426782, upload-time = "2025-10-15T23:17:11.22Z" },
     { url = "https://files.pythonhosted.org/packages/fa/b1/ebacbfe53317d55cf33165bda24c86523497a6881f339f9aae5c2e13e57b/cryptography-46.0.3-cp311-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:a04bee9ab6a4da801eb9b51f1b708a1b5b5c9eb48c03f74198464c66f0d344ac", size = 4698381, upload-time = "2025-10-15T23:17:12.829Z" },
@@ -353,8 +367,10 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/c5/fd/bc1daf8230eaa075184cbbf5f8cd00ba9db4fd32d63fb83da4671b72ed8a/cryptography-46.0.3-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:39b6755623145ad5eff1dab323f4eae2a32a77a7abef2c5089a04a3d04366715", size = 4435078, upload-time = "2025-10-15T23:17:23.042Z" },
     { url = "https://files.pythonhosted.org/packages/82/98/d3bd5407ce4c60017f8ff9e63ffee4200ab3e23fe05b765cab805a7db008/cryptography-46.0.3-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:db391fa7c66df6762ee3f00c95a89e6d428f4d60e7abc8328f4fe155b5ac6e54", size = 4293460, upload-time = "2025-10-15T23:17:24.885Z" },
     { url = "https://files.pythonhosted.org/packages/26/e9/e23e7900983c2b8af7a08098db406cf989d7f09caea7897e347598d4cd5b/cryptography-46.0.3-cp314-cp314t-manylinux_2_28_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:78a97cf6a8839a48c49271cdcbd5cf37ca2c1d6b7fdd86cc864f302b5e9bf459", size = 3995237, upload-time = "2025-10-15T23:17:26.449Z" },
+    { url = "https://files.pythonhosted.org/packages/91/15/af68c509d4a138cfe299d0d7ddb14afba15233223ebd933b4bbdbc7155d3/cryptography-46.0.3-cp314-cp314t-manylinux_2_28_ppc64le.whl", hash = "sha256:dfb781ff7eaa91a6f7fd41776ec37c5853c795d3b358d4896fdbb5df168af422", size = 4967344, upload-time = "2025-10-15T23:17:28.06Z" },
     { url = "https://files.pythonhosted.org/packages/ca/e3/8643d077c53868b681af077edf6b3cb58288b5423610f21c62aadcbe99f4/cryptography-46.0.3-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:6f61efb26e76c45c4a227835ddeae96d83624fb0d29eb5df5b96e14ed1a0afb7", size = 4466564, upload-time = "2025-10-15T23:17:29.665Z" },
     { url = "https://files.pythonhosted.org/packages/0e/43/c1e8726fa59c236ff477ff2b5dc071e54b21e5a1e51aa2cee1676f1c986f/cryptography-46.0.3-cp314-cp314t-manylinux_2_34_aarch64.whl", hash = "sha256:23b1a8f26e43f47ceb6d6a43115f33a5a37d57df4ea0ca295b780ae8546e8044", size = 4292415, upload-time = "2025-10-15T23:17:31.686Z" },
+    { url = "https://files.pythonhosted.org/packages/42/f9/2f8fefdb1aee8a8e3256a0568cffc4e6d517b256a2fe97a029b3f1b9fe7e/cryptography-46.0.3-cp314-cp314t-manylinux_2_34_ppc64le.whl", hash = "sha256:b419ae593c86b87014b9be7396b385491ad7f320bde96826d0dd174459e54665", size = 4931457, upload-time = "2025-10-15T23:17:33.478Z" },
     { url = "https://files.pythonhosted.org/packages/79/30/9b54127a9a778ccd6d27c3da7563e9f2d341826075ceab89ae3b41bf5be2/cryptography-46.0.3-cp314-cp314t-manylinux_2_34_x86_64.whl", hash = "sha256:50fc3343ac490c6b08c0cf0d704e881d0d660be923fd3076db3e932007e726e3", size = 4466074, upload-time = "2025-10-15T23:17:35.158Z" },
     { url = "https://files.pythonhosted.org/packages/ac/68/b4f4a10928e26c941b1b6a179143af9f4d27d88fe84a6a3c53592d2e76bf/cryptography-46.0.3-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:22d7e97932f511d6b0b04f2bfd818d73dcd5928db509460aaf48384778eb6d20", size = 4420569, upload-time = "2025-10-15T23:17:37.188Z" },
     { url = "https://files.pythonhosted.org/packages/a3/49/3746dab4c0d1979888f125226357d3262a6dd40e114ac29e3d2abdf1ec55/cryptography-46.0.3-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:d55f3dffadd674514ad19451161118fd010988540cee43d8bc20675e775925de", size = 4681941, upload-time = "2025-10-15T23:17:39.236Z" },
@@ -362,13 +378,82 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/26/42/fa8389d4478368743e24e61eea78846a0006caffaf72ea24a15159215a14/cryptography-46.0.3-cp38-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:15ab9b093e8f09daab0f2159bb7e47532596075139dd74365da52ecc9cb46c5d", size = 4440029, upload-time = "2025-10-15T23:17:49.837Z" },
     { url = "https://files.pythonhosted.org/packages/5f/eb/f483db0ec5ac040824f269e93dd2bd8a21ecd1027e77ad7bdf6914f2fd80/cryptography-46.0.3-cp38-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:46acf53b40ea38f9c6c229599a4a13f0d46a6c3fa9ef19fc1a124d62e338dfa0", size = 4297222, upload-time = "2025-10-15T23:17:51.357Z" },
     { url = "https://files.pythonhosted.org/packages/fd/cf/da9502c4e1912cb1da3807ea3618a6829bee8207456fbbeebc361ec38ba3/cryptography-46.0.3-cp38-abi3-manylinux_2_28_armv7l.manylinux_2_31_armv7l.whl", hash = "sha256:10ca84c4668d066a9878890047f03546f3ae0a6b8b39b697457b7757aaf18dbc", size = 4012280, upload-time = "2025-10-15T23:17:52.964Z" },
+    { url = "https://files.pythonhosted.org/packages/6b/8f/9adb86b93330e0df8b3dcf03eae67c33ba89958fc2e03862ef1ac2b42465/cryptography-46.0.3-cp38-abi3-manylinux_2_28_ppc64le.whl", hash = "sha256:36e627112085bb3b81b19fed209c05ce2a52ee8b15d161b7c643a7d5a88491f3", size = 4978958, upload-time = "2025-10-15T23:17:54.965Z" },
     { url = "https://files.pythonhosted.org/packages/d1/a0/5fa77988289c34bdb9f913f5606ecc9ada1adb5ae870bd0d1054a7021cc4/cryptography-46.0.3-cp38-abi3-manylinux_2_28_x86_64.whl", hash = "sha256:1000713389b75c449a6e979ffc7dcc8ac90b437048766cef052d4d30b8220971", size = 4473714, upload-time = "2025-10-15T23:17:56.754Z" },
     { url = "https://files.pythonhosted.org/packages/14/e5/fc82d72a58d41c393697aa18c9abe5ae1214ff6f2a5c18ac470f92777895/cryptography-46.0.3-cp38-abi3-manylinux_2_34_aarch64.whl", hash = "sha256:b02cf04496f6576afffef5ddd04a0cb7d49cf6be16a9059d793a30b035f6b6ac", size = 4296970, upload-time = "2025-10-15T23:17:58.588Z" },
+    { url = "https://files.pythonhosted.org/packages/78/06/5663ed35438d0b09056973994f1aec467492b33bd31da36e468b01ec1097/cryptography-46.0.3-cp38-abi3-manylinux_2_34_ppc64le.whl", hash = "sha256:71e842ec9bc7abf543b47cf86b9a743baa95f4677d22baa4c7d5c69e49e9bc04", size = 4940236, upload-time = "2025-10-15T23:18:00.897Z" },
     { url = "https://files.pythonhosted.org/packages/fc/59/873633f3f2dcd8a053b8dd1d38f783043b5fce589c0f6988bf55ef57e43e/cryptography-46.0.3-cp38-abi3-manylinux_2_34_x86_64.whl", hash = "sha256:402b58fc32614f00980b66d6e56a5b4118e6cb362ae8f3fda141ba4689bd4506", size = 4472642, upload-time = "2025-10-15T23:18:02.749Z" },
     { url = "https://files.pythonhosted.org/packages/3d/39/8e71f3930e40f6877737d6f69248cf74d4e34b886a3967d32f919cc50d3b/cryptography-46.0.3-cp38-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:ef639cb3372f69ec44915fafcd6698b6cc78fbe0c2ea41be867f6ed612811963", size = 4423126, upload-time = "2025-10-15T23:18:04.85Z" },
     { url = "https://files.pythonhosted.org/packages/cd/c7/f65027c2810e14c3e7268353b1681932b87e5a48e65505d8cc17c99e36ae/cryptography-46.0.3-cp38-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:3b51b8ca4f1c6453d8829e1eb7299499ca7f313900dd4d89a24b8b87c0a780d4", size = 4686573, upload-time = "2025-10-15T23:18:06.908Z" },
 ]
 
+[[package]]
+name = "cuda-bindings"
+version = "13.2.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "cuda-pathfinder", marker = "sys_platform == 'linux'" },
+]
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/df/93/eef988860a3ca985f82c4f3174fc0cdd94e07331ba9a92e8e064c260337f/cuda_bindings-13.2.0-cp313-cp313-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:6629ca2df6f795b784752409bcaedbd22a7a651b74b56a165ebc0c9dcbd504d0", size = 5614610, upload-time = "2026-03-11T00:12:50.337Z" },
+    { url = "https://files.pythonhosted.org/packages/18/23/6db3aba46864aee357ab2415135b3fe3da7e9f1fa0221fa2a86a5968099c/cuda_bindings-13.2.0-cp313-cp313-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:7dca0da053d3b4cc4869eff49c61c03f3c5dbaa0bcd712317a358d5b8f3f385d", size = 6149914, upload-time = "2026-03-11T00:12:52.374Z" },
+    { url = "https://files.pythonhosted.org/packages/c0/87/87a014f045b77c6de5c8527b0757fe644417b184e5367db977236a141602/cuda_bindings-13.2.0-cp314-cp314-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:a6464b30f46692d6c7f65d4a0e0450d81dd29de3afc1bb515653973d01c2cd6e", size = 5685673, upload-time = "2026-03-11T00:12:56.371Z" },
+    { url = "https://files.pythonhosted.org/packages/ee/5e/c0fe77a73aaefd3fff25ffaccaac69c5a63eafdf8b9a4c476626ef0ac703/cuda_bindings-13.2.0-cp314-cp314-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f4af9f3e1be603fa12d5ad6cfca7844c9d230befa9792b5abdf7dd79979c3626", size = 6191386, upload-time = "2026-03-11T00:12:58.965Z" },
+    { url = "https://files.pythonhosted.org/packages/5f/58/ed2c3b39c8dd5f96aa7a4abef0d47a73932c7a988e30f5fa428f00ed0da1/cuda_bindings-13.2.0-cp314-cp314t-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:df850a1ff8ce1b3385257b08e47b70e959932f5f432d0a4e46a355962b4e4771", size = 5507469, upload-time = "2026-03-11T00:13:04.063Z" },
+    { url = "https://files.pythonhosted.org/packages/1f/01/0c941b112ceeb21439b05895eace78ca1aa2eaaf695c8521a068fd9b4c00/cuda_bindings-13.2.0-cp314-cp314t-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e8a16384c6494e5485f39314b0b4afb04bee48d49edb16d5d8593fd35bbd231b", size = 6059693, upload-time = "2026-03-11T00:13:06.003Z" },
+]
+
+[[package]]
+name = "cuda-pathfinder"
+version = "1.5.0"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/93/66/0c02bd330e7d976f83fa68583d6198d76f23581bcbb5c0e98a6148f326e5/cuda_pathfinder-1.5.0-py3-none-any.whl", hash = "sha256:498f90a9e9de36044a7924742aecce11c50c49f735f1bc53e05aa46de9ea4110", size = 49739, upload-time = "2026-03-24T21:14:30.869Z" },
+]
+
+[[package]]
+name = "cuda-toolkit"
+version = "13.0.2"
+source = { registry = "https://pypi.org/simple" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/57/b2/453099f5f3b698d7d0eab38916aac44c7f76229f451709e2eb9db6615dcd/cuda_toolkit-13.0.2-py2.py3-none-any.whl", hash = "sha256:b198824cf2f54003f50d64ada3a0f184b42ca0846c1c94192fa269ecd97a66eb", size = 2364, upload-time = "2025-12-19T23:24:07.328Z" },
+]
+
+[package.optional-dependencies]
+cublas = [
+    { name = "nvidia-cublas", marker = "sys_platform == 'linux'" },
+]
+cudart = [
+    { name = "nvidia-cuda-runtime", marker = "sys_platform == 'linux'" },
+]
+cufft = [
+    { name = "nvidia-cufft", marker = "sys_platform == 'linux'" },
+]
+cufile = [
+    { name = "nvidia-cufile", marker = "sys_platform == 'linux'" },
+]
+cupti = [
+    { name = "nvidia-cuda-cupti", marker = "sys_platform == 'linux'" },
+]
+curand = [
+    { name = "nvidia-curand", marker = "sys_platform == 'linux'" },
+]
+cusolver = [
+    { name = "nvidia-cusolver", marker = "sys_platform == 'linux'" },
+]
+cusparse = [
+    { name = "nvidia-cusparse", marker = "sys_platform == 'linux'" },
+]
+nvjitlink = [
+    { name = "nvidia-nvjitlink", marker = "sys_platform == 'linux'" },
+]
+nvrtc = [
+    { name = "nvidia-cuda-nvrtc", marker = "sys_platform == 'linux'" },
+]
+nvtx = [
+    { name = "nvidia-nvtx", marker = "sys_platform == 'linux'" },
+]
+
 [[package]]
 name = "cycler"
 version = "0.12.1"
@@ -473,8 +558,9 @@ dependencies = [
     { name = "loguru", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "mflux", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "mlx", version = "0.30.6", source = { registry = "https://pypi.org/simple" }, extra = ["cpu"], marker = "sys_platform == 'linux'" },
-    { name = "mlx", version = "0.31.2.dev20260327+e5e64331", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#e5e64331830d9b04ae9082b843073f9c1fa7705e" }, marker = "sys_platform == 'darwin'" },
+    { name = "mlx", version = "0.31.2.dev20260324+e5e64331", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#e5e64331830d9b04ae9082b843073f9c1fa7705e" }, marker = "sys_platform == 'darwin'" },
     { name = "mlx-lm", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "mlx-vlm", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "msgspec", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "openai-harmony", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "psutil", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
@@ -483,6 +569,7 @@ dependencies = [
     { name = "rustworkx", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "tiktoken", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "tomlkit", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "transformers", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "types-aiofiles", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "zstandard", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
 ]
@@ -506,13 +593,14 @@ requires-dist = [
     { name = "fastapi", specifier = ">=0.116.1" },
     { name = "filelock", specifier = ">=3.18.0" },
     { name = "httpx", specifier = ">=0.28.1" },
-    { name = "huggingface-hub", specifier = ">=0.33.4" },
+    { name = "huggingface-hub", specifier = ">=1.8.0" },
     { name = "hypercorn", specifier = ">=0.18.0" },
     { name = "loguru", specifier = ">=0.7.3" },
     { name = "mflux", specifier = "==0.17.2" },
     { name = "mlx", marker = "sys_platform == 'darwin'", git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks" },
     { name = "mlx", extras = ["cpu"], marker = "sys_platform == 'linux'", specifier = "==0.30.6" },
     { name = "mlx-lm", git = "https://github.com/rltakashige/mlx-lm?branch=leo%2Ffix-deepseek-v32-indexer" },
+    { name = "mlx-vlm", specifier = ">=0.3.11" },
     { name = "msgspec", specifier = ">=0.19.0" },
     { name = "openai-harmony", specifier = ">=0.0.8" },
     { name = "psutil", specifier = ">=7.0.0" },
@@ -521,6 +609,7 @@ requires-dist = [
     { name = "rustworkx", specifier = ">=0.17.1" },
     { name = "tiktoken", specifier = ">=0.12.0" },
     { name = "tomlkit", specifier = ">=0.14.0" },
+    { name = "transformers", specifier = ">=5.0.0,<5.4.0" },
     { name = "types-aiofiles", specifier = ">=24.1.0.20250708" },
     { name = "zstandard", specifier = ">=0.23.0" },
 ]
@@ -783,31 +872,28 @@ wheels = [
 
 [[package]]
 name = "hf-xet"
-version = "1.2.1rc0"
+version = "1.4.2"
 source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/9a/48/61907d37a180a1d016cb79396215b1064f075965cf14ac78b4a9682705d7/hf_xet-1.2.1rc0.tar.gz", hash = "sha256:ee6b196855720767283dbbca6d5f3877afdfa6df83e037bbadbed0181ac5972e", size = 518988, upload-time = "2025-11-21T23:26:10.526Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/09/08/23c84a26716382c89151b5b447b4beb19e3345f3a93d3b73009a71a57ad3/hf_xet-1.4.2.tar.gz", hash = "sha256:b7457b6b482d9e0743bd116363239b1fa904a5e65deede350fbc0c4ea67c71ea", size = 672357, upload-time = "2026-03-13T06:58:51.077Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/8c/2b/e9fb76e7dcba1efc0dc881124d0ebbdf0790ad78f90dae9f23a969224c0c/hf_xet-1.2.1rc0-cp313-cp313t-macosx_10_12_x86_64.whl", hash = "sha256:05acfd78c5b515a0c06103c9471208a71ae52c6a72dba73bbcb5b7f79575c530", size = 2973766, upload-time = "2025-11-21T23:25:50.546Z" },
-    { url = "https://files.pythonhosted.org/packages/95/bf/8365816fb0e2dc0db633bed504fdf70b4e4e052aa86caff62e4b0175e7fa/hf_xet-1.2.1rc0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:2e4bbe0e4195c48aebce7c87438df6ba0748001c15cd088d1f41553b9cbf0aa5", size = 2850724, upload-time = "2025-11-21T23:25:48.95Z" },
-    { url = "https://files.pythonhosted.org/packages/4a/52/72ba543089817fdf0e684032c1664fd249602896d52b76f4278b7c830cc8/hf_xet-1.2.1rc0-cp313-cp313t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:66534e7190bafae92c8e3411011220f189fadcc8cba36ebf4bc261e769fb7e49", size = 3342204, upload-time = "2025-11-21T23:25:31.773Z" },
-    { url = "https://files.pythonhosted.org/packages/85/a0/d0f7b4ffb08bdb25db2dbad8e5d97a266a4ada3c7e8dc4429bfe99c86ed2/hf_xet-1.2.1rc0-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:c9d193015364fb9e95d4d295722538b554e9bfaa7b6a167e09e030148c8b15d0", size = 19434060, upload-time = "2025-11-21T23:25:33.89Z" },
-    { url = "https://files.pythonhosted.org/packages/af/b4/c406e62a1895520da504bb9372f7ed26ef65e32e1b39e397d81b7136b5ab/hf_xet-1.2.1rc0-cp313-cp313t-manylinux_2_28_aarch64.whl", hash = "sha256:dda4a029cd30f10ba205d8a74e232070ec75923e4c262a2d7f5d55eb3a3dd4d1", size = 3249296, upload-time = "2025-11-21T23:25:29.504Z" },
-    { url = "https://files.pythonhosted.org/packages/cf/fb/c40487744c12a038e31af75de661938a6e9c2cfb55a544706d9b9d3cc00c/hf_xet-1.2.1rc0-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:fc95e2b7a1a3a613587f407a8292f1240d45febd66a49ee1da0a94414ff3784e", size = 3434401, upload-time = "2025-11-21T23:25:59.747Z" },
-    { url = "https://files.pythonhosted.org/packages/46/37/8b93e82bace53bb650474562487a4fe2aa43e8b8d9ecd01ddffc1b6a63f2/hf_xet-1.2.1rc0-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:4a4e981ef129bdf1af7be559319b017bed0ae997c8bdd696b6c7e50d888e5a51", size = 3520042, upload-time = "2025-11-21T23:26:01.691Z" },
-    { url = "https://files.pythonhosted.org/packages/c7/b7/6ce9f48be8748b2e8599453dec7012d38e4685a5e5587ee3ef4c09fccaf9/hf_xet-1.2.1rc0-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:1d57ee9323fcf87c3fc1840856ad2f767c0f8ee14a55d470ddba3a6fdab40dd2", size = 2973781, upload-time = "2025-11-21T23:25:58.073Z" },
-    { url = "https://files.pythonhosted.org/packages/72/dc/6e1d3b653fdb34ce86f7b94c2388270f8bb5bb18da8590425a30ef0af1be/hf_xet-1.2.1rc0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:6163f7de633ac0f5f88dc24d369b30df4df0f923dc61ebd9c39a9b022497f47f", size = 2850462, upload-time = "2025-11-21T23:25:56.157Z" },
-    { url = "https://files.pythonhosted.org/packages/8c/6b/6e0daf5811badf6c9d60a49cb3f99fe41cc01f147ecae3911d8621fa69c1/hf_xet-1.2.1rc0-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:05b518a2499dafd510e29ff6c14bfb9aae119f66af785fc99eaf9069e0ccda43", size = 3342036, upload-time = "2025-11-21T23:25:44.283Z" },
-    { url = "https://files.pythonhosted.org/packages/b7/21/9dfdf0c66743cbf14f312d196f19367372a89232b2623d733690474008b9/hf_xet-1.2.1rc0-cp314-cp314t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e5ee726b80a1c0b2868bc58302ba1a47d0702f8d67f69aeecb94fe7f30ac1c2b", size = 19431002, upload-time = "2025-11-21T23:25:46.621Z" },
-    { url = "https://files.pythonhosted.org/packages/f4/8c/f798608de78b5aa1cabbf9c1e5e8a0172a93a47267fe1733f7c9780802e2/hf_xet-1.2.1rc0-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:bf8f5439c39a5fa41dec1071f9576ac510180522690771d54c211151e08cdf35", size = 3248725, upload-time = "2025-11-21T23:25:42.387Z" },
-    { url = "https://files.pythonhosted.org/packages/75/75/7035ea757b2ef27c21a7d734da18c1537473f8dcff468872eb9b4281dd33/hf_xet-1.2.1rc0-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:5ca1fae9189095b15c89cd30ce2f6c3a97f2d1cab261e28a73b84690ebc8960a", size = 3433685, upload-time = "2025-11-21T23:26:06.88Z" },
-    { url = "https://files.pythonhosted.org/packages/0e/47/1627f85cb062283edc9f516d61838c88bcdb46828d903b035674b5e0e89c/hf_xet-1.2.1rc0-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:99676d52bbffc7747950d2686bc91f520758f3d83b594988058478be68706862", size = 3519636, upload-time = "2025-11-21T23:26:08.512Z" },
-    { url = "https://files.pythonhosted.org/packages/6e/ce/bfd825a3aa2a22caa78865a6331e3660825b82de24877b08c10d18c45748/hf_xet-1.2.1rc0-cp37-abi3-macosx_10_12_x86_64.whl", hash = "sha256:b6b6455d68f2b4439028c58198e6dc33f3b1b64314ed05b0a5f5f7dace37d711", size = 2977924, upload-time = "2025-11-21T23:25:54.254Z" },
-    { url = "https://files.pythonhosted.org/packages/88/28/d78d7fcf2f3e18177e8dd6bbb4294bb00ef2f6d3addfc2b636a251ec297b/hf_xet-1.2.1rc0-cp37-abi3-macosx_11_0_arm64.whl", hash = "sha256:3d9894128c63478a3f67d7f0288e8f5780c2b3ae7a09f36fc3949be60dcf7ac8", size = 2853755, upload-time = "2025-11-21T23:25:52.222Z" },
-    { url = "https://files.pythonhosted.org/packages/ae/09/637245509430b3dd9d37f676bbe0b993c723e3671ce0b39fdf42c6f05a02/hf_xet-1.2.1rc0-cp37-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:f8b937c5e2a4f43720eca9564b14324ecfa108cc053a1b44890c620f51aac01e", size = 3347297, upload-time = "2025-11-21T23:25:37.9Z" },
-    { url = "https://files.pythonhosted.org/packages/29/b5/bbc98a35ee5229d0cd6c9436ae97f86cf2ab63d6bd463cd5a43282e5c1f8/hf_xet-1.2.1rc0-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:7bd4629e923dd7b12fb9d05312e03ed123db230ae25fd98a3fd5caa739f2357e", size = 19457253, upload-time = "2025-11-21T23:25:40.115Z" },
-    { url = "https://files.pythonhosted.org/packages/0f/c6/ab21fc91f23ca54cdd44e86981d80475d67ee4122128f5ef988a119ebe28/hf_xet-1.2.1rc0-cp37-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:5484ad943ceec043f0c29733cb87e59c86c2c68804c470176f259b1ef339718e", size = 3254771, upload-time = "2025-11-21T23:25:36.213Z" },
-    { url = "https://files.pythonhosted.org/packages/e6/c0/5a2887739722bd5a531769c1e9555e30dd7f470aefaabbe898d939dbba20/hf_xet-1.2.1rc0-cp37-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:2ec943ba2633ed0df48d2c817ce6a13670e96590f9fd4260011c5753afbc5d53", size = 3439600, upload-time = "2025-11-21T23:26:03.318Z" },
-    { url = "https://files.pythonhosted.org/packages/30/c9/c7cd0a64eb2dba1f70fbb78dee33558567404522776328254a7c805ae23e/hf_xet-1.2.1rc0-cp37-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:87e0bdd71172b7cb1621e706bbf70b75f31df5fa7c359ebc0978567b5c21c2cf", size = 3526094, upload-time = "2025-11-21T23:26:05.018Z" },
+    { url = "https://files.pythonhosted.org/packages/18/06/e8cf74c3c48e5485c7acc5a990d0d8516cdfb5fdf80f799174f1287cc1b5/hf_xet-1.4.2-cp313-cp313t-macosx_10_12_x86_64.whl", hash = "sha256:ac8202ae1e664b2c15cdfc7298cbb25e80301ae596d602ef7870099a126fcad4", size = 3796125, upload-time = "2026-03-13T06:58:33.177Z" },
+    { url = "https://files.pythonhosted.org/packages/66/d4/b73ebab01cbf60777323b7de9ef05550790451eb5172a220d6b9845385ec/hf_xet-1.4.2-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:6d2f8ee39fa9fba9af929f8c0d0482f8ee6e209179ad14a909b6ad78ffcb7c81", size = 3555985, upload-time = "2026-03-13T06:58:31.797Z" },
+    { url = "https://files.pythonhosted.org/packages/ff/e7/ded6d1bd041c3f2bca9e913a0091adfe32371988e047dd3a68a2463c15a2/hf_xet-1.4.2-cp313-cp313t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:4642a6cf249c09da8c1f87fe50b24b2a3450b235bf8adb55700b52f0ea6e2eb6", size = 4212085, upload-time = "2026-03-13T06:58:24.323Z" },
+    { url = "https://files.pythonhosted.org/packages/97/c1/a0a44d1f98934f7bdf17f7a915b934f9fca44bb826628c553589900f6df8/hf_xet-1.4.2-cp313-cp313t-manylinux_2_28_aarch64.whl", hash = "sha256:769431385e746c92dc05492dde6f687d304584b89c33d79def8367ace06cb555", size = 3988266, upload-time = "2026-03-13T06:58:22.887Z" },
+    { url = "https://files.pythonhosted.org/packages/7a/82/be713b439060e7d1f1d93543c8053d4ef2fe7e6922c5b31642eaa26f3c4b/hf_xet-1.4.2-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:c9dd1c1bc4cc56168f81939b0e05b4c36dd2d28c13dc1364b17af89aa0082496", size = 4188513, upload-time = "2026-03-13T06:58:40.858Z" },
+    { url = "https://files.pythonhosted.org/packages/21/a6/cbd4188b22abd80ebd0edbb2b3e87f2633e958983519980815fb8314eae5/hf_xet-1.4.2-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:fca58a2ae4e6f6755cc971ac6fcdf777ea9284d7e540e350bb000813b9a3008d", size = 4428287, upload-time = "2026-03-13T06:58:42.601Z" },
+    { url = "https://files.pythonhosted.org/packages/1e/0f/fcd2504015eab26358d8f0f232a1aed6b8d363a011adef83fe130bff88f7/hf_xet-1.4.2-cp314-cp314t-macosx_10_12_x86_64.whl", hash = "sha256:949dcf88b484bb9d9276ca83f6599e4aa03d493c08fc168c124ad10b2e6f75d7", size = 3796493, upload-time = "2026-03-13T06:58:39.267Z" },
+    { url = "https://files.pythonhosted.org/packages/82/56/19c25105ff81731ca6d55a188b5de2aa99d7a2644c7aa9de1810d5d3b726/hf_xet-1.4.2-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:41659966020d59eb9559c57de2cde8128b706a26a64c60f0531fa2318f409418", size = 3555797, upload-time = "2026-03-13T06:58:37.546Z" },
+    { url = "https://files.pythonhosted.org/packages/bf/e3/8933c073186849b5e06762aa89847991d913d10a95d1603eb7f2c3834086/hf_xet-1.4.2-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:5c588e21d80010119458dd5d02a69093f0d115d84e3467efe71ffb2c67c19146", size = 4212127, upload-time = "2026-03-13T06:58:30.539Z" },
+    { url = "https://files.pythonhosted.org/packages/eb/01/f89ebba4e369b4ed699dcb60d3152753870996f41c6d22d3d7cac01310e1/hf_xet-1.4.2-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:a296744d771a8621ad1d50c098d7ab975d599800dae6d48528ba3944e5001ba0", size = 3987788, upload-time = "2026-03-13T06:58:29.139Z" },
+    { url = "https://files.pythonhosted.org/packages/84/4d/8a53e5ffbc2cc33bbf755382ac1552c6d9af13f623ed125fe67cc3e6772f/hf_xet-1.4.2-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:f563f7efe49588b7d0629d18d36f46d1658fe7e08dce3fa3d6526e1c98315e2d", size = 4188315, upload-time = "2026-03-13T06:58:48.017Z" },
+    { url = "https://files.pythonhosted.org/packages/d1/b8/b7a1c1b5592254bd67050632ebbc1b42cc48588bf4757cb03c2ef87e704a/hf_xet-1.4.2-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:5b2e0132c56d7ee1bf55bdb638c4b62e7106f6ac74f0b786fed499d5548c5570", size = 4428306, upload-time = "2026-03-13T06:58:49.502Z" },
+    { url = "https://files.pythonhosted.org/packages/b4/86/b40b83a2ff03ef05c4478d2672b1fc2b9683ff870e2b25f4f3af240f2e7b/hf_xet-1.4.2-cp37-abi3-macosx_10_12_x86_64.whl", hash = "sha256:71f02d6e4cdd07f344f6844845d78518cc7186bd2bc52d37c3b73dc26a3b0bc5", size = 3800339, upload-time = "2026-03-13T06:58:36.245Z" },
+    { url = "https://files.pythonhosted.org/packages/64/2e/af4475c32b4378b0e92a587adb1aa3ec53e3450fd3e5fe0372a874531c00/hf_xet-1.4.2-cp37-abi3-macosx_11_0_arm64.whl", hash = "sha256:e9b38d876e94d4bdcf650778d6ebbaa791dd28de08db9736c43faff06ede1b5a", size = 3559664, upload-time = "2026-03-13T06:58:34.787Z" },
+    { url = "https://files.pythonhosted.org/packages/3c/4c/781267da3188db679e601de18112021a5cb16506fe86b246e22c5401a9c4/hf_xet-1.4.2-cp37-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:77e8c180b7ef12d8a96739a4e1e558847002afe9ea63b6f6358b2271a8bdda1c", size = 4217422, upload-time = "2026-03-13T06:58:27.472Z" },
+    { url = "https://files.pythonhosted.org/packages/68/47/d6cf4a39ecf6c7705f887a46f6ef5c8455b44ad9eb0d391aa7e8a2ff7fea/hf_xet-1.4.2-cp37-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:c3b3c6a882016b94b6c210957502ff7877802d0dbda8ad142c8595db8b944271", size = 3992847, upload-time = "2026-03-13T06:58:25.989Z" },
+    { url = "https://files.pythonhosted.org/packages/2d/ef/e80815061abff54697239803948abc665c6b1d237102c174f4f7a9a5ffc5/hf_xet-1.4.2-cp37-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:9d9a634cc929cfbaf2e1a50c0e532ae8c78fa98618426769480c58501e8c8ac2", size = 4193843, upload-time = "2026-03-13T06:58:44.59Z" },
+    { url = "https://files.pythonhosted.org/packages/54/75/07f6aa680575d9646c4167db6407c41340cbe2357f5654c4e72a1b01ca14/hf_xet-1.4.2-cp37-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:6b0932eb8b10317ea78b7da6bab172b17be03bbcd7809383d8d5abd6a2233e04", size = 4432751, upload-time = "2026-03-13T06:58:46.533Z" },
 ]
 
 [[package]]
@@ -849,7 +935,7 @@ wheels = [
 
 [[package]]
 name = "huggingface-hub"
-version = "1.3.1"
+version = "1.8.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "filelock", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
@@ -858,14 +944,13 @@ dependencies = [
     { name = "httpx", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "packaging", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "pyyaml", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "shellingham", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "tqdm", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "typer-slim", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "typer", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "typing-extensions", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/dd/dd/1cc985c5dda36298b152f75e82a1c81f52243b78fb7e9cad637a29561ad1/huggingface_hub-1.3.1.tar.gz", hash = "sha256:e80e0cfb4a75557c51ab20d575bdea6bb6106c2f97b7c75d8490642f1efb6df5", size = 622356, upload-time = "2026-01-09T14:08:16.888Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/8e/2a/a847fd02261cd051da218baf99f90ee7c7040c109a01833db4f838f25256/huggingface_hub-1.8.0.tar.gz", hash = "sha256:c5627b2fd521e00caf8eff4ac965ba988ea75167fad7ee72e17f9b7183ec63f3", size = 735839, upload-time = "2026-03-25T16:01:28.152Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/90/fb/cb8fe5f71d5622427f20bcab9e06a696a5aaf21bfe7bd0a8a0c63c88abf5/huggingface_hub-1.3.1-py3-none-any.whl", hash = "sha256:efbc7f3153cb84e2bb69b62ed90985e21ecc9343d15647a419fc0ee4b85f0ac3", size = 533351, upload-time = "2026-01-09T14:08:14.519Z" },
+    { url = "https://files.pythonhosted.org/packages/a9/ae/8a3a16ea4d202cb641b51d2681bdd3d482c1c592d7570b3fa264730829ce/huggingface_hub-1.8.0-py3-none-any.whl", hash = "sha256:d3eb5047bd4e33c987429de6020d4810d38a5bef95b3b40df9b17346b7f353f2", size = 625208, upload-time = "2026-03-25T16:01:26.603Z" },
 ]
 
 [[package]]
@@ -1351,7 +1436,7 @@ dependencies = [
     { name = "huggingface-hub", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "matplotlib", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "mlx", version = "0.30.6", source = { registry = "https://pypi.org/simple" }, extra = ["cuda13"], marker = "sys_platform == 'linux'" },
-    { name = "mlx", version = "0.31.2.dev20260327+e5e64331", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#e5e64331830d9b04ae9082b843073f9c1fa7705e" }, marker = "sys_platform == 'darwin'" },
+    { name = "mlx", version = "0.31.2.dev20260324+e5e64331", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#e5e64331830d9b04ae9082b843073f9c1fa7705e" }, marker = "sys_platform == 'darwin'" },
     { name = "numpy", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "opencv-python", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "piexif", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
@@ -1375,6 +1460,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/68/02/f94eca4e77b7d12685060461eb793cbc8c00e96cc7fe0ce376374201aed2/mflux-0.17.2-py3-none-any.whl", hash = "sha256:be1642b04847413c0a8ed1dae82ce1ca023e155b057d82a8301eca9c3fe08339", size = 1037451, upload-time = "2026-03-23T13:08:16.747Z" },
 ]
 
+[[package]]
+name = "miniaudio"
+version = "1.61"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "cffi", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/55/fa/96d4cc7ada283357117f7890418ac065a0a6d81ec59e681cd965a403aba3/miniaudio-1.61.tar.gz", hash = "sha256:e88e97837d031f0fb6982394218b6487de02eaa382ad273b8fca37791a2b4b15", size = 1103527, upload-time = "2024-07-24T18:13:10.037Z" }
+
 [[package]]
 name = "mlx"
 version = "0.30.6"
@@ -1400,7 +1494,7 @@ cuda13 = [
 
 [[package]]
 name = "mlx"
-version = "0.31.2.dev20260327+e5e64331"
+version = "0.31.2.dev20260324+e5e64331"
 source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#e5e64331830d9b04ae9082b843073f9c1fa7705e" }
 resolution-markers = [
     "python_full_version >= '3.14' and sys_platform == 'darwin'",
@@ -1437,7 +1531,7 @@ version = "0.31.2"
 source = { git = "https://github.com/rltakashige/mlx-lm?branch=leo%2Ffix-deepseek-v32-indexer#d388ff77858fec3b5d2e3b1d9502a7e2878b8109" }
 dependencies = [
     { name = "jinja2", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "mlx", version = "0.31.2.dev20260327+e5e64331", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#e5e64331830d9b04ae9082b843073f9c1fa7705e" }, marker = "sys_platform == 'darwin'" },
+    { name = "mlx", version = "0.31.2.dev20260324+e5e64331", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#e5e64331830d9b04ae9082b843073f9c1fa7705e" }, marker = "sys_platform == 'darwin'" },
     { name = "numpy", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "protobuf", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "pyyaml", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
@@ -1445,6 +1539,30 @@ dependencies = [
     { name = "transformers", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
 ]
 
+[[package]]
+name = "mlx-vlm"
+version = "0.4.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "datasets", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "fastapi", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "miniaudio", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "mlx", version = "0.30.6", source = { registry = "https://pypi.org/simple" }, marker = "sys_platform == 'linux'" },
+    { name = "mlx", version = "0.31.2.dev20260324+e5e64331", source = { git = "https://github.com/rltakashige/mlx-jaccl-fix-small-recv.git?branch=address-rdma-gpu-locks#e5e64331830d9b04ae9082b843073f9c1fa7705e" }, marker = "sys_platform == 'darwin'" },
+    { name = "mlx-lm", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "numpy", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "opencv-python", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "pillow", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "requests", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "tqdm", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "transformers", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "uvicorn", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/57/8f/31204f1a8c7404e523a5578949ea668e668e10dd67a0f63336f261014c0f/mlx_vlm-0.4.1.tar.gz", hash = "sha256:4e2d8a232715dbca72d346f43cf54d5738452848855792ffb1b285228ae7c7bd", size = 621840, upload-time = "2026-03-21T14:26:04.586Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/c1/78/856f44c6bdd8791427fa59a093a1d00c91cdbe16238506602fc3968017bb/mlx_vlm-0.4.1-py3-none-any.whl", hash = "sha256:89feca2e8be31609770c0e8a6d88fa21d00ee25bd3d56b4aafce59d35dd63b71", size = 768806, upload-time = "2026-03-21T14:26:03.129Z" },
+]
+
 [[package]]
 name = "more-itertools"
 version = "10.8.0"
@@ -1680,175 +1798,151 @@ wheels = [
 
 [[package]]
 name = "nvidia-cublas"
-version = "13.2.1.1"
+version = "13.1.0.3"
 source = { registry = "https://pypi.org/simple" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/9d/36/0124129e1378e9834e0cbe19781fbe0ffd5f870c2af6f01cdf17a9869c39/nvidia_cublas-13.2.1.1-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:8b4a4cd8b73772fde9ccaa1f3967eb001ae5fde8b1dc37f7442d072b64d6f5da", size = 502470979, upload-time = "2026-01-13T22:39:37.619Z" },
-    { url = "https://files.pythonhosted.org/packages/e2/e7/39e43c0688f9788c88da0b91ea18125448c5f515104aadf65a70243f144f/nvidia_cublas-13.2.1.1-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:8c13c93cf8be4480b4909905c96d2d31575b4af43fcd3af0e84af94762665e4f", size = 401085577, upload-time = "2026-01-13T22:40:18.702Z" },
+    { url = "https://files.pythonhosted.org/packages/e1/a5/fce49e2ae977e0ccc084e5adafceb4f0ac0c8333cb6863501618a7277f67/nvidia_cublas-13.1.0.3-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:c86fc7f7ae36d7528288c5d88098edcb7b02c633d262e7ddbb86b0ad91be5df2", size = 542851226, upload-time = "2025-10-09T08:59:04.818Z" },
+    { url = "https://files.pythonhosted.org/packages/e7/44/423ac00af4dd95a5aeb27207e2c0d9b7118702149bf4704c3ddb55bb7429/nvidia_cublas-13.1.0.3-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:ee8722c1f0145ab246bccb9e452153b5e0515fd094c3678df50b2a0888b8b171", size = 423133236, upload-time = "2025-10-09T08:59:32.536Z" },
 ]
 
 [[package]]
-name = "nvidia-cublas-cu12"
-version = "12.8.4.1"
+name = "nvidia-cuda-cupti"
+version = "13.0.85"
 source = { registry = "https://pypi.org/simple" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/dc/61/e24b560ab2e2eaeb3c839129175fb330dfcfc29e5203196e5541a4c44682/nvidia_cublas_cu12-12.8.4.1-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:8ac4e771d5a348c551b2a426eda6193c19aa630236b418086020df5ba9667142", size = 594346921, upload-time = "2025-03-07T01:44:31.254Z" },
-]
-
-[[package]]
-name = "nvidia-cuda-cupti-cu12"
-version = "12.8.90"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/f8/02/2adcaa145158bf1a8295d83591d22e4103dbfd821bcaf6f3f53151ca4ffa/nvidia_cuda_cupti_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:ea0cb07ebda26bb9b29ba82cda34849e73c166c18162d3913575b0c9db9a6182", size = 10248621, upload-time = "2025-03-07T01:40:21.213Z" },
+    { url = "https://files.pythonhosted.org/packages/2a/2a/80353b103fc20ce05ef51e928daed4b6015db4aaa9162ed0997090fe2250/nvidia_cuda_cupti-13.0.85-py3-none-manylinux_2_25_aarch64.whl", hash = "sha256:796bd679890ee55fb14a94629b698b6db54bcfd833d391d5e94017dd9d7d3151", size = 10310827, upload-time = "2025-09-04T08:26:42.012Z" },
+    { url = "https://files.pythonhosted.org/packages/33/6d/737d164b4837a9bbd202f5ae3078975f0525a55730fe871d8ed4e3b952b0/nvidia_cuda_cupti-13.0.85-py3-none-manylinux_2_25_x86_64.whl", hash = "sha256:4eb01c08e859bf924d222250d2e8f8b8ff6d3db4721288cf35d14252a4d933c8", size = 10715597, upload-time = "2025-09-04T08:26:51.312Z" },
 ]
 
 [[package]]
 name = "nvidia-cuda-nvrtc"
-version = "13.1.115"
+version = "13.0.88"
 source = { registry = "https://pypi.org/simple" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/ff/d8/6fcf0f32d133a7da92efb1e90844d9f7c104627066cc52b13f7f0b128b54/nvidia_cuda_nvrtc-13.1.115-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:d7cf1284ab82f379884decc8813d9d4a729bc96b1ec020a9cf80303f698d73c4", size = 46564545, upload-time = "2026-01-13T22:35:53.834Z" },
-    { url = "https://files.pythonhosted.org/packages/e9/f8/8bab039cbdd87af53f2ca0ca9e93bd676e53393ab4ea43da4735854dc1ce/nvidia_cuda_nvrtc-13.1.115-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:dfbc5e3bb19db41e4a05280b7b0cb9cbb624699f57dab3798455f43345541f99", size = 44308134, upload-time = "2026-01-13T22:35:35.287Z" },
+    { url = "https://files.pythonhosted.org/packages/c3/68/483a78f5e8f31b08fb1bb671559968c0ca3a065ac7acabfc7cee55214fd6/nvidia_cuda_nvrtc-13.0.88-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:ad9b6d2ead2435f11cbb6868809d2adeeee302e9bb94bcf0539c7a40d80e8575", size = 90215200, upload-time = "2025-09-04T08:28:44.204Z" },
+    { url = "https://files.pythonhosted.org/packages/b7/dc/6bb80850e0b7edd6588d560758f17e0550893a1feaf436807d64d2da040f/nvidia_cuda_nvrtc-13.0.88-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:d27f20a0ca67a4bb34268a5e951033496c5b74870b868bacd046b1b8e0c3267b", size = 43015449, upload-time = "2025-09-04T08:28:20.239Z" },
 ]
 
 [[package]]
-name = "nvidia-cuda-nvrtc-cu12"
-version = "12.8.93"
+name = "nvidia-cuda-runtime"
+version = "13.0.96"
 source = { registry = "https://pypi.org/simple" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/05/6b/32f747947df2da6994e999492ab306a903659555dddc0fbdeb9d71f75e52/nvidia_cuda_nvrtc_cu12-12.8.93-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:a7756528852ef889772a84c6cd89d41dfa74667e24cca16bb31f8f061e3e9994", size = 88040029, upload-time = "2025-03-07T01:42:13.562Z" },
-]
-
-[[package]]
-name = "nvidia-cuda-runtime-cu12"
-version = "12.8.90"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/0d/9b/a997b638fcd068ad6e4d53b8551a7d30fe8b404d6f1804abf1df69838932/nvidia_cuda_runtime_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:adade8dcbd0edf427b7204d480d6066d33902cab2a4707dcfc48a2d0fd44ab90", size = 954765, upload-time = "2025-03-07T01:40:01.615Z" },
-]
-
-[[package]]
-name = "nvidia-cudnn-cu12"
-version = "9.10.2.21"
-source = { registry = "https://pypi.org/simple" }
-dependencies = [
-    { name = "nvidia-cublas-cu12", marker = "sys_platform == 'linux'" },
-]
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/ba/51/e123d997aa098c61d029f76663dedbfb9bc8dcf8c60cbd6adbe42f76d049/nvidia_cudnn_cu12-9.10.2.21-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:949452be657fa16687d0930933f032835951ef0892b37d2d53824d1a84dc97a8", size = 706758467, upload-time = "2025-06-06T21:54:08.597Z" },
+    { url = "https://files.pythonhosted.org/packages/87/4f/17d7b9b8e285199c58ce28e31b5c5bbaa4d8271af06a89b6405258245de2/nvidia_cuda_runtime-13.0.96-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:ef9bcbe90493a2b9d810e43d249adb3d02e98dd30200d86607d8d02687c43f55", size = 2261060, upload-time = "2025-10-09T08:55:15.78Z" },
+    { url = "https://files.pythonhosted.org/packages/2e/24/d1558f3b68b1d26e706813b1d10aa1d785e4698c425af8db8edc3dced472/nvidia_cuda_runtime-13.0.96-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:7f82250d7782aa23b6cfe765ecc7db554bd3c2870c43f3d1821f1d18aebf0548", size = 2243632, upload-time = "2025-10-09T08:55:36.117Z" },
 ]
 
 [[package]]
 name = "nvidia-cudnn-cu13"
-version = "9.18.0.77"
+version = "9.19.0.56"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "nvidia-cublas", marker = "sys_platform == 'linux'" },
 ]
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/0b/74/ad8b8edef8b8a54071cb8bd80b63aee7a833b1eabfcdbc0bbec4f0868cc1/nvidia_cudnn_cu13-9.18.0.77-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:a0957ae96e752840033a505bb6505d634e03c4bb4947e3e8fe1fdbe599120ab3", size = 423102639, upload-time = "2026-01-16T20:25:15.997Z" },
-    { url = "https://files.pythonhosted.org/packages/d7/e9/c26231f84cc906dca63f4517e9824c9e8d166838e61fc0c5015cbe11fe59/nvidia_cudnn_cu13-9.18.0.77-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:82b86573bd24bfc1450c2e94866af23f5327982a7a4f14ee2416f8e0dd4631f6", size = 358843113, upload-time = "2026-01-16T20:26:35.975Z" },
+    { url = "https://files.pythonhosted.org/packages/f1/84/26025437c1e6b61a707442184fa0c03d083b661adf3a3eecfd6d21677740/nvidia_cudnn_cu13-9.19.0.56-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:6ed29ffaee1176c612daf442e4dd6cfeb6a0caa43ddcbeb59da94953030b1be4", size = 433781201, upload-time = "2026-02-03T20:40:53.805Z" },
+    { url = "https://files.pythonhosted.org/packages/a3/22/0b4b932655d17a6da1b92fa92ab12844b053bb2ac2475e179ba6f043da1e/nvidia_cudnn_cu13-9.19.0.56-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:d20e1734305e9d68889a96e3f35094d733ff1f83932ebe462753973e53a572bf", size = 366066321, upload-time = "2026-02-03T20:44:52.837Z" },
 ]
 
 [[package]]
-name = "nvidia-cufft-cu12"
-version = "11.3.3.83"
+name = "nvidia-cufft"
+version = "12.0.0.61"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "nvidia-nvjitlink-cu12", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-nvjitlink", marker = "sys_platform == 'linux'" },
 ]
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/1f/13/ee4e00f30e676b66ae65b4f08cb5bcbb8392c03f54f2d5413ea99a5d1c80/nvidia_cufft_cu12-11.3.3.83-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:4d2dd21ec0b88cf61b62e6b43564355e5222e4a3fb394cac0db101f2dd0d4f74", size = 193118695, upload-time = "2025-03-07T01:45:27.821Z" },
+    { url = "https://files.pythonhosted.org/packages/8b/ae/f417a75c0259e85c1d2f83ca4e960289a5f814ed0cea74d18c353d3e989d/nvidia_cufft-12.0.0.61-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:2708c852ef8cd89d1d2068bdbece0aa188813a0c934db3779b9b1faa8442e5f5", size = 214053554, upload-time = "2025-09-04T08:31:38.196Z" },
+    { url = "https://files.pythonhosted.org/packages/a8/2f/7b57e29836ea8714f81e9898409196f47d772d5ddedddf1592eadb8ab743/nvidia_cufft-12.0.0.61-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:6c44f692dce8fd5ffd3e3df134b6cdb9c2f72d99cf40b62c32dde45eea9ddad3", size = 214085489, upload-time = "2025-09-04T08:31:56.044Z" },
 ]
 
 [[package]]
-name = "nvidia-cufile-cu12"
-version = "1.13.1.3"
+name = "nvidia-cufile"
+version = "1.15.1.6"
 source = { registry = "https://pypi.org/simple" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/bb/fe/1bcba1dfbfb8d01be8d93f07bfc502c93fa23afa6fd5ab3fc7c1df71038a/nvidia_cufile_cu12-1.13.1.3-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:1d069003be650e131b21c932ec3d8969c1715379251f8d23a1860554b1cb24fc", size = 1197834, upload-time = "2025-03-07T01:45:50.723Z" },
+    { url = "https://files.pythonhosted.org/packages/3f/70/4f193de89a48b71714e74602ee14d04e4019ad36a5a9f20c425776e72cd6/nvidia_cufile-1.15.1.6-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:08a3ecefae5a01c7f5117351c64f17c7c62efa5fffdbe24fc7d298da19cd0b44", size = 1223672, upload-time = "2025-09-04T08:32:22.779Z" },
+    { url = "https://files.pythonhosted.org/packages/ab/73/cc4a14c9813a8a0d509417cf5f4bdaba76e924d58beb9864f5a7baceefbf/nvidia_cufile-1.15.1.6-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:bdc0deedc61f548bddf7733bdc216456c2fdb101d020e1ab4b88d232d5e2f6d1", size = 1136992, upload-time = "2025-09-04T08:32:14.119Z" },
 ]
 
 [[package]]
-name = "nvidia-curand-cu12"
-version = "10.3.9.90"
+name = "nvidia-curand"
+version = "10.4.0.35"
 source = { registry = "https://pypi.org/simple" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/fb/aa/6584b56dc84ebe9cf93226a5cde4d99080c8e90ab40f0c27bda7a0f29aa1/nvidia_curand_cu12-10.3.9.90-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:b32331d4f4df5d6eefa0554c565b626c7216f87a06a4f56fab27c3b68a830ec9", size = 63619976, upload-time = "2025-03-07T01:46:23.323Z" },
+    { url = "https://files.pythonhosted.org/packages/1e/72/7c2ae24fb6b63a32e6ae5d241cc65263ea18d08802aaae087d9f013335a2/nvidia_curand-10.4.0.35-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:133df5a7509c3e292aaa2b477afd0194f06ce4ea24d714d616ff36439cee349a", size = 61962106, upload-time = "2025-08-04T10:21:41.128Z" },
+    { url = "https://files.pythonhosted.org/packages/a5/9f/be0a41ca4a4917abf5cb9ae0daff1a6060cc5de950aec0396de9f3b52bc5/nvidia_curand-10.4.0.35-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:1aee33a5da6e1db083fe2b90082def8915f30f3248d5896bcec36a579d941bfc", size = 59544258, upload-time = "2025-08-04T10:22:03.992Z" },
 ]
 
 [[package]]
-name = "nvidia-cusolver-cu12"
-version = "11.7.3.90"
+name = "nvidia-cusolver"
+version = "12.0.4.66"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "nvidia-cublas-cu12", marker = "sys_platform == 'linux'" },
-    { name = "nvidia-cusparse-cu12", marker = "sys_platform == 'linux'" },
-    { name = "nvidia-nvjitlink-cu12", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-cublas", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-cusparse", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-nvjitlink", marker = "sys_platform == 'linux'" },
 ]
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/85/48/9a13d2975803e8cf2777d5ed57b87a0b6ca2cc795f9a4f59796a910bfb80/nvidia_cusolver_cu12-11.7.3.90-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:4376c11ad263152bd50ea295c05370360776f8c3427b30991df774f9fb26c450", size = 267506905, upload-time = "2025-03-07T01:47:16.273Z" },
+    { url = "https://files.pythonhosted.org/packages/c8/c3/b30c9e935fc01e3da443ec0116ed1b2a009bb867f5324d3f2d7e533e776b/nvidia_cusolver-12.0.4.66-py3-none-manylinux_2_27_aarch64.whl", hash = "sha256:02c2457eaa9e39de20f880f4bd8820e6a1cfb9f9a34f820eb12a155aa5bc92d2", size = 223467760, upload-time = "2025-09-04T08:33:04.222Z" },
+    { url = "https://files.pythonhosted.org/packages/5f/67/cba3777620cdacb99102da4042883709c41c709f4b6323c10781a9c3aa34/nvidia_cusolver-12.0.4.66-py3-none-manylinux_2_27_x86_64.whl", hash = "sha256:0a759da5dea5c0ea10fd307de75cdeb59e7ea4fcb8add0924859b944babf1112", size = 200941980, upload-time = "2025-09-04T08:33:22.767Z" },
 ]
 
 [[package]]
-name = "nvidia-cusparse-cu12"
-version = "12.5.8.93"
+name = "nvidia-cusparse"
+version = "12.6.3.3"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "nvidia-nvjitlink-cu12", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-nvjitlink", marker = "sys_platform == 'linux'" },
 ]
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/c2/f5/e1854cb2f2bcd4280c44736c93550cc300ff4b8c95ebe370d0aa7d2b473d/nvidia_cusparse_cu12-12.5.8.93-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:1ec05d76bbbd8b61b06a80e1eaf8cf4959c3d4ce8e711b65ebd0443bb0ebb13b", size = 288216466, upload-time = "2025-03-07T01:48:13.779Z" },
+    { url = "https://files.pythonhosted.org/packages/f8/94/5c26f33738ae35276672f12615a64bd008ed5be6d1ebcb23579285d960a9/nvidia_cusparse-12.6.3.3-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:80bcc4662f23f1054ee334a15c72b8940402975e0eab63178fc7e670aa59472c", size = 162155568, upload-time = "2025-09-04T08:33:42.864Z" },
+    { url = "https://files.pythonhosted.org/packages/fa/18/623c77619c31d62efd55302939756966f3ecc8d724a14dab2b75f1508850/nvidia_cusparse-12.6.3.3-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:2b3c89c88d01ee0e477cb7f82ef60a11a4bcd57b6b87c33f789350b59759360b", size = 145942937, upload-time = "2025-09-04T08:33:58.029Z" },
 ]
 
 [[package]]
-name = "nvidia-cusparselt-cu12"
-version = "0.7.1"
+name = "nvidia-cusparselt-cu13"
+version = "0.8.0"
 source = { registry = "https://pypi.org/simple" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/56/79/12978b96bd44274fe38b5dde5cfb660b1d114f70a65ef962bcbbed99b549/nvidia_cusparselt_cu12-0.7.1-py3-none-manylinux2014_x86_64.whl", hash = "sha256:f1bb701d6b930d5a7cea44c19ceb973311500847f81b634d802b7b539dc55623", size = 287193691, upload-time = "2025-02-26T00:15:44.104Z" },
-]
-
-[[package]]
-name = "nvidia-nccl-cu12"
-version = "2.27.5"
-source = { registry = "https://pypi.org/simple" }
-wheels = [
-    { url = "https://files.pythonhosted.org/packages/6e/89/f7a07dc961b60645dbbf42e80f2bc85ade7feb9a491b11a1e973aa00071f/nvidia_nccl_cu12-2.27.5-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:ad730cf15cb5d25fe849c6e6ca9eb5b76db16a80f13f425ac68d8e2e55624457", size = 322348229, upload-time = "2025-06-26T04:11:28.385Z" },
+    { url = "https://files.pythonhosted.org/packages/46/10/8dcd1175260706a2fc92a16a52e306b71d4c1ea0b0cc4a9484183399818a/nvidia_cusparselt_cu13-0.8.0-py3-none-manylinux2014_aarch64.whl", hash = "sha256:400c6ed1cf6780fc6efedd64ec9f1345871767e6a1a0a552a1ea0578117ea77c", size = 220791277, upload-time = "2025-08-13T19:22:40.982Z" },
+    { url = "https://files.pythonhosted.org/packages/fd/53/43b0d71f4e702fa9733f8b4571fdca50a8813f1e450b656c239beff12315/nvidia_cusparselt_cu13-0.8.0-py3-none-manylinux2014_x86_64.whl", hash = "sha256:25e30a8a7323935d4ad0340b95a0b69926eee755767e8e0b1cf8dd85b197d3fd", size = 169884119, upload-time = "2025-08-13T19:23:41.967Z" },
 ]
 
 [[package]]
 name = "nvidia-nccl-cu13"
-version = "2.29.2"
+version = "2.28.9"
 source = { registry = "https://pypi.org/simple" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/cb/e8/b69bfcbf39d71b4166cf1ceb0e58dd73cc4c6ad005164b56e54acb4dbf2f/nvidia_nccl_cu13-2.29.2-py3-none-manylinux_2_18_aarch64.whl", hash = "sha256:9d4f7e24aff66309f0b71bd6a885afa214e1bac3a562c9a77be428f0a4aeb62a", size = 201038683, upload-time = "2026-01-07T00:21:18.002Z" },
-    { url = "https://files.pythonhosted.org/packages/81/15/5e1d022945dd511d453ba5163fedce67d3bd0fb3dcadc021f00c0c8a491b/nvidia_nccl_cu13-2.29.2-py3-none-manylinux_2_18_x86_64.whl", hash = "sha256:86b997b315df0fb2874fd6062f2930d317bfa6434823351f287936d5ed616fc9", size = 201100704, upload-time = "2026-01-07T00:21:31.626Z" },
+    { url = "https://files.pythonhosted.org/packages/39/55/1920646a2e43ffd4fc958536b276197ed740e9e0c54105b4bb3521591fc7/nvidia_nccl_cu13-2.28.9-py3-none-manylinux_2_18_aarch64.whl", hash = "sha256:01c873ba1626b54caa12272ed228dc5b2781545e0ae8ba3f432a8ef1c6d78643", size = 196561677, upload-time = "2025-11-18T05:49:03.45Z" },
+    { url = "https://files.pythonhosted.org/packages/b0/b4/878fefaad5b2bcc6fcf8d474a25e3e3774bc5133e4b58adff4d0bca238bc/nvidia_nccl_cu13-2.28.9-py3-none-manylinux_2_18_x86_64.whl", hash = "sha256:e4553a30f34195f3fa1da02a6da3d6337d28f2003943aa0a3d247bbc25fefc42", size = 196493177, upload-time = "2025-11-18T05:49:17.677Z" },
 ]
 
 [[package]]
-name = "nvidia-nvjitlink-cu12"
-version = "12.8.93"
+name = "nvidia-nvjitlink"
+version = "13.0.88"
 source = { registry = "https://pypi.org/simple" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/f6/74/86a07f1d0f42998ca31312f998bd3b9a7eff7f52378f4f270c8679c77fb9/nvidia_nvjitlink_cu12-12.8.93-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:81ff63371a7ebd6e6451970684f916be2eab07321b73c9d244dc2b4da7f73b88", size = 39254836, upload-time = "2025-03-07T01:49:55.661Z" },
+    { url = "https://files.pythonhosted.org/packages/56/7a/123e033aaff487c77107195fa5a2b8686795ca537935a24efae476c41f05/nvidia_nvjitlink-13.0.88-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl", hash = "sha256:13a74f429e23b921c1109976abefacc69835f2f433ebd323d3946e11d804e47b", size = 40713933, upload-time = "2025-09-04T08:35:43.553Z" },
+    { url = "https://files.pythonhosted.org/packages/ab/2c/93c5250e64df4f894f1cbb397c6fd71f79813f9fd79d7cd61de3f97b3c2d/nvidia_nvjitlink-13.0.88-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:e931536ccc7d467a98ba1d8b89ff7fa7f1fa3b13f2b0069118cd7f47bff07d0c", size = 38768748, upload-time = "2025-09-04T08:35:20.008Z" },
 ]
 
 [[package]]
-name = "nvidia-nvshmem-cu12"
-version = "3.3.20"
+name = "nvidia-nvshmem-cu13"
+version = "3.4.5"
 source = { registry = "https://pypi.org/simple" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/3b/6c/99acb2f9eb85c29fc6f3a7ac4dccfd992e22666dd08a642b303311326a97/nvidia_nvshmem_cu12-3.3.20-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:d00f26d3f9b2e3c3065be895e3059d6479ea5c638a3f38c9fec49b1b9dd7c1e5", size = 124657145, upload-time = "2025-08-04T20:25:19.995Z" },
+    { url = "https://files.pythonhosted.org/packages/dc/0f/05cc9c720236dcd2db9c1ab97fff629e96821be2e63103569da0c9b72f19/nvidia_nvshmem_cu13-3.4.5-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:6dc2a197f38e5d0376ad52cd1a2a3617d3cdc150fd5966f4aee9bcebb1d68fe9", size = 60215947, upload-time = "2025-09-06T00:32:20.022Z" },
+    { url = "https://files.pythonhosted.org/packages/3c/35/a9bf80a609e74e3b000fef598933235c908fcefcef9026042b8e6dfde2a9/nvidia_nvshmem_cu13-3.4.5-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:290f0a2ee94c9f3687a02502f3b9299a9f9fe826e6d0287ee18482e78d495b80", size = 60412546, upload-time = "2025-09-06T00:32:41.564Z" },
 ]
 
 [[package]]
-name = "nvidia-nvtx-cu12"
-version = "12.8.90"
+name = "nvidia-nvtx"
+version = "13.0.85"
 source = { registry = "https://pypi.org/simple" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/a2/eb/86626c1bbc2edb86323022371c39aa48df6fd8b0a1647bc274577f72e90b/nvidia_nvtx_cu12-12.8.90-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:5b17e2001cc0d751a5bc2c6ec6d26ad95913324a4adb86788c944f8ce9ba441f", size = 89954, upload-time = "2025-03-07T01:42:44.131Z" },
+    { url = "https://files.pythonhosted.org/packages/c2/f3/d86c845465a2723ad7e1e5c36dcd75ddb82898b3f53be47ebd429fb2fa5d/nvidia_nvtx-13.0.85-py3-none-manylinux1_x86_64.manylinux_2_5_x86_64.whl", hash = "sha256:4936d1d6780fbe68db454f5e72a42ff64d1fd6397df9f363ae786930fd5c1cd4", size = 148047, upload-time = "2025-09-04T08:29:01.761Z" },
+    { url = "https://files.pythonhosted.org/packages/a8/64/3708a90d1ebe202ffdeb7185f878a3c84d15c2b2c31858da2ce0583e2def/nvidia_nvtx-13.0.85-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:cb7780edb6b14107373c835bf8b72e7a178bac7367e23da7acb108f973f157a6", size = 148878, upload-time = "2025-09-04T08:28:53.627Z" },
 ]
 
 [[package]]
@@ -2918,46 +3012,37 @@ wheels = [
 
 [[package]]
 name = "torch"
-version = "2.9.1"
+version = "2.11.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
+    { name = "cuda-bindings", marker = "sys_platform == 'linux'" },
+    { name = "cuda-toolkit", extra = ["cublas", "cudart", "cufft", "cufile", "cupti", "curand", "cusolver", "cusparse", "nvjitlink", "nvrtc", "nvtx"], marker = "sys_platform == 'linux'" },
     { name = "filelock", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "fsspec", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "jinja2", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "networkx", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "nvidia-cublas-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
-    { name = "nvidia-cuda-cupti-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
-    { name = "nvidia-cuda-nvrtc-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
-    { name = "nvidia-cuda-runtime-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
-    { name = "nvidia-cudnn-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
-    { name = "nvidia-cufft-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
-    { name = "nvidia-cufile-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
-    { name = "nvidia-curand-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
-    { name = "nvidia-cusolver-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
-    { name = "nvidia-cusparse-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
-    { name = "nvidia-cusparselt-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
-    { name = "nvidia-nccl-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
-    { name = "nvidia-nvjitlink-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
-    { name = "nvidia-nvshmem-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
-    { name = "nvidia-nvtx-cu12", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "nvidia-cudnn-cu13", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-cusparselt-cu13", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-nccl-cu13", marker = "sys_platform == 'linux'" },
+    { name = "nvidia-nvshmem-cu13", marker = "sys_platform == 'linux'" },
     { name = "setuptools", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "sympy", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "triton", marker = "platform_machine == 'x86_64' and sys_platform == 'linux'" },
+    { name = "triton", marker = "sys_platform == 'linux'" },
     { name = "typing-extensions", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
 ]
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/20/60/8fc5e828d050bddfab469b3fe78e5ab9a7e53dda9c3bdc6a43d17ce99e63/torch-2.9.1-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:c29455d2b910b98738131990394da3e50eea8291dfeb4b12de71ecf1fdeb21cb", size = 104135743, upload-time = "2025-11-12T15:21:34.936Z" },
-    { url = "https://files.pythonhosted.org/packages/f2/b7/6d3f80e6918213babddb2a37b46dbb14c15b14c5f473e347869a51f40e1f/torch-2.9.1-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:524de44cd13931208ba2c4bde9ec7741fd4ae6bfd06409a604fc32f6520c2bc9", size = 899749493, upload-time = "2025-11-12T15:24:36.356Z" },
-    { url = "https://files.pythonhosted.org/packages/28/0e/2a37247957e72c12151b33a01e4df651d9d155dd74d8cfcbfad15a79b44a/torch-2.9.1-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:5be4bf7496f1e3ffb1dd44b672adb1ac3f081f204c5ca81eba6442f5f634df8e", size = 74830751, upload-time = "2025-11-12T15:21:43.792Z" },
-    { url = "https://files.pythonhosted.org/packages/4b/f7/7a18745edcd7b9ca2381aa03353647bca8aace91683c4975f19ac233809d/torch-2.9.1-cp313-cp313t-manylinux_2_28_aarch64.whl", hash = "sha256:30a3e170a84894f3652434b56d59a64a2c11366b0ed5776fab33c2439396bf9a", size = 104142929, upload-time = "2025-11-12T15:21:48.319Z" },
-    { url = "https://files.pythonhosted.org/packages/f4/dd/f1c0d879f2863ef209e18823a988dc7a1bf40470750e3ebe927efdb9407f/torch-2.9.1-cp313-cp313t-manylinux_2_28_x86_64.whl", hash = "sha256:8301a7b431e51764629208d0edaa4f9e4c33e6df0f2f90b90e261d623df6a4e2", size = 899748978, upload-time = "2025-11-12T15:23:04.568Z" },
-    { url = "https://files.pythonhosted.org/packages/40/60/71c698b466dd01e65d0e9514b5405faae200c52a76901baf6906856f17e4/torch-2.9.1-cp313-none-macosx_11_0_arm64.whl", hash = "sha256:2c14b3da5df416cf9cb5efab83aa3056f5b8cd8620b8fde81b4987ecab730587", size = 74480347, upload-time = "2025-11-12T15:21:57.648Z" },
-    { url = "https://files.pythonhosted.org/packages/48/50/c4b5112546d0d13cc9eaa1c732b823d676a9f49ae8b6f97772f795874a03/torch-2.9.1-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:1edee27a7c9897f4e0b7c14cfc2f3008c571921134522d5b9b5ec4ebbc69041a", size = 74433245, upload-time = "2025-11-12T15:22:39.027Z" },
-    { url = "https://files.pythonhosted.org/packages/81/c9/2628f408f0518b3bae49c95f5af3728b6ab498c8624ab1e03a43dd53d650/torch-2.9.1-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:19d144d6b3e29921f1fc70503e9f2fc572cde6a5115c0c0de2f7ca8b1483e8b6", size = 104134804, upload-time = "2025-11-12T15:22:35.222Z" },
-    { url = "https://files.pythonhosted.org/packages/28/fc/5bc91d6d831ae41bf6e9e6da6468f25330522e92347c9156eb3f1cb95956/torch-2.9.1-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:c432d04376f6d9767a9852ea0def7b47a7bbc8e7af3b16ac9cf9ce02b12851c9", size = 899747132, upload-time = "2025-11-12T15:23:36.068Z" },
-    { url = "https://files.pythonhosted.org/packages/bd/b2/2d15a52516b2ea3f414643b8de68fa4cb220d3877ac8b1028c83dc8ca1c4/torch-2.9.1-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:cb10896a1f7fedaddbccc2017ce6ca9ecaaf990f0973bdfcf405439750118d2c", size = 74823558, upload-time = "2025-11-12T15:22:43.392Z" },
-    { url = "https://files.pythonhosted.org/packages/86/5c/5b2e5d84f5b9850cd1e71af07524d8cbb74cba19379800f1f9f7c997fc70/torch-2.9.1-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:0a2bd769944991c74acf0c4ef23603b9c777fdf7637f115605a4b2d8023110c7", size = 104145788, upload-time = "2025-11-12T15:23:52.109Z" },
-    { url = "https://files.pythonhosted.org/packages/a9/8c/3da60787bcf70add986c4ad485993026ac0ca74f2fc21410bc4eb1bb7695/torch-2.9.1-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:07c8a9660bc9414c39cac530ac83b1fb1b679d7155824144a40a54f4a47bfa73", size = 899735500, upload-time = "2025-11-12T15:24:08.788Z" },
+    { url = "https://files.pythonhosted.org/packages/87/89/5ea6722763acee56b045435fb84258db7375c48165ec8be7880ab2b281c5/torch-2.11.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:1e6debd97ccd3205bbb37eb806a9d8219e1139d15419982c09e23ef7d4369d18", size = 80606801, upload-time = "2026-03-23T18:10:18.649Z" },
+    { url = "https://files.pythonhosted.org/packages/32/d1/8ed2173589cbfe744ed54e5a73efc107c0085ba5777ee93a5f4c1ab90553/torch-2.11.0-cp313-cp313-manylinux_2_28_aarch64.whl", hash = "sha256:63a68fa59de8f87acc7e85a5478bb2dddbb3392b7593ec3e78827c793c4b73fd", size = 419732382, upload-time = "2026-03-23T18:08:30.835Z" },
+    { url = "https://files.pythonhosted.org/packages/3d/e1/b73f7c575a4b8f87a5928f50a1e35416b5e27295d8be9397d5293e7e8d4c/torch-2.11.0-cp313-cp313-manylinux_2_28_x86_64.whl", hash = "sha256:cc89b9b173d9adfab59fd227f0ab5e5516d9a52b658ae41d64e59d2e55a418db", size = 530711509, upload-time = "2026-03-23T18:08:47.213Z" },
+    { url = "https://files.pythonhosted.org/packages/db/38/8ac78069621b8c2b4979c2f96dc8409ef5e9c4189f6aac629189a78677ca/torch-2.11.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:8b394322f49af4362d4f80e424bcaca7efcd049619af03a4cf4501520bdf0fb4", size = 80959574, upload-time = "2026-03-23T18:10:14.214Z" },
+    { url = "https://files.pythonhosted.org/packages/6d/6c/56bfb37073e7136e6dd86bfc6af7339946dd684e0ecf2155ac0eee687ae1/torch-2.11.0-cp313-cp313t-manylinux_2_28_aarch64.whl", hash = "sha256:2658f34ce7e2dabf4ec73b45e2ca68aedad7a5be87ea756ad656eaf32bf1e1ea", size = 419732324, upload-time = "2026-03-23T18:09:36.604Z" },
+    { url = "https://files.pythonhosted.org/packages/07/f4/1b666b6d61d3394cca306ea543ed03a64aad0a201b6cd159f1d41010aeb1/torch-2.11.0-cp313-cp313t-manylinux_2_28_x86_64.whl", hash = "sha256:98bb213c3084cfe176302949bdc360074b18a9da7ab59ef2edc9d9f742504778", size = 530596026, upload-time = "2026-03-23T18:09:20.842Z" },
+    { url = "https://files.pythonhosted.org/packages/26/0d/8603382f61abd0db35841148ddc1ffd607bf3100b11c6e1dab6d2fc44e72/torch-2.11.0-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:01018087326984a33b64e04c8cb5c2795f9120e0d775ada1f6638840227b04d7", size = 80573442, upload-time = "2026-03-23T18:09:10.117Z" },
+    { url = "https://files.pythonhosted.org/packages/c7/86/7cd7c66cb9cec6be330fff36db5bd0eef386d80c031b581ec81be1d4b26c/torch-2.11.0-cp314-cp314-manylinux_2_28_aarch64.whl", hash = "sha256:2bb3cc54bd0dea126b0060bb1ec9de0f9c7f7342d93d436646516b0330cd5be7", size = 419749385, upload-time = "2026-03-23T18:07:33.77Z" },
+    { url = "https://files.pythonhosted.org/packages/47/e8/b98ca2d39b2e0e4730c0ee52537e488e7008025bc77ca89552ff91021f7c/torch-2.11.0-cp314-cp314-manylinux_2_28_x86_64.whl", hash = "sha256:4dc8b3809469b6c30b411bb8c4cad3828efd26236153d9beb6a3ec500f211a60", size = 530716756, upload-time = "2026-03-23T18:07:50.02Z" },
+    { url = "https://files.pythonhosted.org/packages/bf/46/4419098ed6d801750f26567b478fc185c3432e11e2cad712bc6b4c2ab0d0/torch-2.11.0-cp314-cp314t-macosx_11_0_arm64.whl", hash = "sha256:8245477871c3700d4370352ffec94b103cfcb737229445cf9946cddb7b2ca7cd", size = 80959460, upload-time = "2026-03-23T18:09:00.818Z" },
+    { url = "https://files.pythonhosted.org/packages/fd/66/54a56a4a6ceaffb567231994a9745821d3af922a854ed33b0b3a278e0a99/torch-2.11.0-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:ab9a8482f475f9ba20e12db84b0e55e2f58784bdca43a854a6ccd3fd4b9f75e6", size = 419735835, upload-time = "2026-03-23T18:07:18.974Z" },
+    { url = "https://files.pythonhosted.org/packages/b1/e7/0b6665f533aa9e337662dc190425abc0af1fe3234088f4454c52393ded61/torch-2.11.0-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:563ed3d25542d7e7bbc5b235ccfacfeb97fb470c7fee257eae599adb8005c8a2", size = 530613405, upload-time = "2026-03-23T18:08:07.014Z" },
 ]
 
 [[package]]
@@ -2971,10 +3056,9 @@ wheels = [
 
 [[package]]
 name = "transformers"
-version = "5.0.0"
+version = "5.3.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
-    { name = "filelock", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "huggingface-hub", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "numpy", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "packaging", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
@@ -2983,22 +3067,26 @@ dependencies = [
     { name = "safetensors", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "tokenizers", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "tqdm", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "typer-slim", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "typer", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/bc/79/845941711811789c85fb7e2599cea425a14a07eda40f50896b9d3fda7492/transformers-5.0.0.tar.gz", hash = "sha256:5f5634efed6cf76ad068cc5834c7adbc32db78bbd6211fb70df2325a9c37dec8", size = 8424830, upload-time = "2026-01-26T10:46:46.813Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/fc/1a/70e830d53ecc96ce69cfa8de38f163712d2b43ac52fbd743f39f56025c31/transformers-5.3.0.tar.gz", hash = "sha256:009555b364029da9e2946d41f1c5de9f15e6b1df46b189b7293f33a161b9c557", size = 8830831, upload-time = "2026-03-04T17:41:46.119Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/52/f3/ac976fa8e305c9e49772527e09fbdc27cc6831b8a2f6b6063406626be5dd/transformers-5.0.0-py3-none-any.whl", hash = "sha256:587086f249ce64c817213cf36afdb318d087f790723e9b3d4500b97832afd52d", size = 10142091, upload-time = "2026-01-26T10:46:43.88Z" },
+    { url = "https://files.pythonhosted.org/packages/b8/88/ae8320064e32679a5429a2c9ebbc05c2bf32cefb6e076f9b07f6d685a9b4/transformers-5.3.0-py3-none-any.whl", hash = "sha256:50ac8c89c3c7033444fb3f9f53138096b997ebb70d4b5e50a2e810bf12d3d29a", size = 10661827, upload-time = "2026-03-04T17:41:42.722Z" },
 ]
 
 [[package]]
 name = "triton"
-version = "3.5.1"
+version = "3.6.0"
 source = { registry = "https://pypi.org/simple" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/27/46/8c3bbb5b0a19313f50edcaa363b599e5a1a5ac9683ead82b9b80fe497c8d/triton-3.5.1-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f3f4346b6ebbd4fad18773f5ba839114f4826037c9f2f34e0148894cd5dd3dba", size = 170470410, upload-time = "2025-11-11T17:41:06.319Z" },
-    { url = "https://files.pythonhosted.org/packages/37/92/e97fcc6b2c27cdb87ce5ee063d77f8f26f19f06916aa680464c8104ef0f6/triton-3.5.1-cp313-cp313t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:0b4d2c70127fca6a23e247f9348b8adde979d2e7a20391bfbabaac6aebc7e6a8", size = 170579924, upload-time = "2025-11-11T17:41:12.455Z" },
-    { url = "https://files.pythonhosted.org/packages/a4/e6/c595c35e5c50c4bc56a7bac96493dad321e9e29b953b526bbbe20f9911d0/triton-3.5.1-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d0637b1efb1db599a8e9dc960d53ab6e4637db7d4ab6630a0974705d77b14b60", size = 170480488, upload-time = "2025-11-11T17:41:18.222Z" },
-    { url = "https://files.pythonhosted.org/packages/16/b5/b0d3d8b901b6a04ca38df5e24c27e53afb15b93624d7fd7d658c7cd9352a/triton-3.5.1-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:bac7f7d959ad0f48c0e97d6643a1cc0fd5786fe61cb1f83b537c6b2d54776478", size = 170582192, upload-time = "2025-11-11T17:41:23.963Z" },
+    { url = "https://files.pythonhosted.org/packages/3c/12/34d71b350e89a204c2c7777a9bba0dcf2f19a5bfdd70b57c4dbc5ffd7154/triton-3.6.0-cp313-cp313-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:448e02fe6dc898e9e5aa89cf0ee5c371e99df5aa5e8ad976a80b93334f3494fd", size = 176133521, upload-time = "2026-01-20T16:16:13.321Z" },
+    { url = "https://files.pythonhosted.org/packages/f9/0b/37d991d8c130ce81a8728ae3c25b6e60935838e9be1b58791f5997b24a54/triton-3.6.0-cp313-cp313-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:10c7f76c6e72d2ef08df639e3d0d30729112f47a56b0c81672edc05ee5116ac9", size = 188289450, upload-time = "2026-01-20T16:00:49.136Z" },
+    { url = "https://files.pythonhosted.org/packages/ce/4e/41b0c8033b503fd3cfcd12392cdd256945026a91ff02452bef40ec34bee7/triton-3.6.0-cp313-cp313t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:1722e172d34e32abc3eb7711d0025bb69d7959ebea84e3b7f7a341cd7ed694d6", size = 176276087, upload-time = "2026-01-20T16:16:18.989Z" },
+    { url = "https://files.pythonhosted.org/packages/35/f8/9c66bfc55361ec6d0e4040a0337fb5924ceb23de4648b8a81ae9d33b2b38/triton-3.6.0-cp313-cp313t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:d002e07d7180fd65e622134fbd980c9a3d4211fb85224b56a0a0efbd422ab72f", size = 188400296, upload-time = "2026-01-20T16:00:56.042Z" },
+    { url = "https://files.pythonhosted.org/packages/49/55/5ecf0dcaa0f2fbbd4420f7ef227ee3cb172e91e5fede9d0ecaddc43363b4/triton-3.6.0-cp314-cp314-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:ef5523241e7d1abca00f1d240949eebdd7c673b005edbbce0aca95b8191f1d43", size = 176138577, upload-time = "2026-01-20T16:16:25.426Z" },
+    { url = "https://files.pythonhosted.org/packages/df/3d/9e7eee57b37c80cec63322c0231bb6da3cfe535a91d7a4d64896fcb89357/triton-3.6.0-cp314-cp314-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:a17a5d5985f0ac494ed8a8e54568f092f7057ef60e1b0fa09d3fd1512064e803", size = 188273063, upload-time = "2026-01-20T16:01:07.278Z" },
+    { url = "https://files.pythonhosted.org/packages/48/db/56ee649cab5eaff4757541325aca81f52d02d4a7cd3506776cad2451e060/triton-3.6.0-cp314-cp314t-manylinux_2_27_aarch64.manylinux_2_28_aarch64.whl", hash = "sha256:0b3a97e8ed304dfa9bd23bb41ca04cdf6b2e617d5e782a8653d616037a5d537d", size = 176274804, upload-time = "2026-01-20T16:16:31.528Z" },
+    { url = "https://files.pythonhosted.org/packages/f6/56/6113c23ff46c00aae423333eb58b3e60bdfe9179d542781955a5e1514cb3/triton-3.6.0-cp314-cp314t-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:46bd1c1af4b6704e554cad2eeb3b0a6513a980d470ccfa63189737340c7746a7", size = 188397994, upload-time = "2026-01-20T16:01:14.236Z" },
 ]
 
 [[package]]
@@ -3041,16 +3129,18 @@ datetime = [
 ]
 
 [[package]]
-name = "typer-slim"
-version = "0.21.1"
+name = "typer"
+version = "0.24.1"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
+    { name = "annotated-doc", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
     { name = "click", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
-    { name = "typing-extensions", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "rich", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "shellingham", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/17/d4/064570dec6358aa9049d4708e4a10407d74c99258f8b2136bb8702303f1a/typer_slim-0.21.1.tar.gz", hash = "sha256:73495dd08c2d0940d611c5a8c04e91c2a0a98600cbd4ee19192255a233b6dbfd", size = 110478, upload-time = "2026-01-06T11:21:11.176Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/f5/24/cb09efec5cc954f7f9b930bf8279447d24618bb6758d4f6adf2574c41780/typer-0.24.1.tar.gz", hash = "sha256:e39b4732d65fbdcde189ae76cf7cd48aeae72919dea1fdfc16593be016256b45", size = 118613, upload-time = "2026-02-21T16:54:40.609Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/c8/0a/4aca634faf693e33004796b6cee0ae2e1dba375a800c16ab8d3eff4bb800/typer_slim-0.21.1-py3-none-any.whl", hash = "sha256:6e6c31047f171ac93cc5a973c9e617dbc5ab2bddc4d0a3135dc161b4e2020e0d", size = 47444, upload-time = "2026-01-06T11:21:12.441Z" },
+    { url = "https://files.pythonhosted.org/packages/4a/91/48db081e7a63bb37284f9fbcefda7c44c277b18b0e13fbc36ea2335b71e6/typer-0.24.1-py3-none-any.whl", hash = "sha256:112c1f0ce578bfb4cab9ffdabc68f031416ebcc216536611ba21f04e9aa84c9e", size = 56085, upload-time = "2026-02-21T16:54:41.616Z" },
 ]
 
 [[package]]
@@ -3092,6 +3182,19 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/39/08/aaaad47bc4e9dc8c725e68f9d04865dbcb2052843ff09c97b08904852d84/urllib3-2.6.3-py3-none-any.whl", hash = "sha256:bf272323e553dfb2e87d9bfd225ca7b0f467b919d7bbd355436d3fd37cb0acd4", size = 131584, upload-time = "2026-01-07T16:24:42.685Z" },
 ]
 
+[[package]]
+name = "uvicorn"
+version = "0.42.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "click", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+    { name = "h11", marker = "sys_platform == 'darwin' or sys_platform == 'linux'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/e3/ad/4a96c425be6fb67e0621e62d86c402b4a17ab2be7f7c055d9bd2f638b9e2/uvicorn-0.42.0.tar.gz", hash = "sha256:9b1f190ce15a2dd22e7758651d9b6d12df09a13d51ba5bf4fc33c383a48e1775", size = 85393, upload-time = "2026-03-16T06:19:50.077Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/0a/89/f8827ccff89c1586027a105e5630ff6139a64da2515e24dafe860bd9ae4d/uvicorn-0.42.0-py3-none-any.whl", hash = "sha256:96c30f5c7abe6f74ae8900a70e92b85ad6613b745d4879eb9b16ccad15645359", size = 68830, upload-time = "2026-03-16T06:19:48.325Z" },
+]
+
 [[package]]
 name = "word2number"
 version = "1.1"