kiapi 0.7.0__tar.gz → 0.8.0__tar.gz

{kiapi-0.7.0 → kiapi-0.8.0}/.gitignore

@@ -32,3 +32,6 @@ tests/assets/
 
 # Node modules (pnpm-managed dev tooling, e.g. firebase-tools)
 node_modules/
+
+# Web UI build output (`mise run web:build`)
+/src/kiapi/api/ui/static/

{kiapi-0.7.0 → kiapi-0.8.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: kiapi
-Version: 0.7.0
+Version: 0.8.0
 Summary: Unified local inference API server for LLM agents on Apple Silicon (MLX): chat, embedding, image, music, sound-effect, and video — one process, one global memory budget, one single-flight queue, unified jobs and files.
 Project-URL: Homepage, https://github.com/kiarina/kiapi
 Project-URL: Repository, https://github.com/kiarina/kiapi
@@ -27,7 +27,7 @@ Requires-Dist: fastapi>=0.141
 Requires-Dist: httpx
 Requires-Dist: kiarina-utils-app>=2.3.0
 Requires-Dist: kiarina-utils-common>=2.2.0
-Requires-Dist: mflux>=0.18.0
+Requires-Dist: mflux>=0.20.0
 Requires-Dist: mlx-audiocraft>=0.1.0
 Requires-Dist: mlx-embeddings>=0.1.0
 Requires-Dist: mlx-lm>=0.31.3
@@ -53,6 +53,11 @@ Description-Content-Type: text/markdown
 ![Python](https://img.shields.io/badge/Python-3.12-blue.svg)
 [![API Docs](https://img.shields.io/badge/API%20Docs-GitHub%20Pages-green.svg)](https://kiarina.github.io/kiapi/)
 
+<picture>
+  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/kiarina/kiapi/main/docs/images/web-ui/playground-dark.webp">
+  <img alt="The kiapi web UI running Z-Image, with the generated image next to the form" src="https://raw.githubusercontent.com/kiarina/kiapi/main/docs/images/web-ui/playground-light.webp">
+</picture>
+
 ## Summary
 
 kiapi is an API server that uses a Mac Studio M4 Max with 128GB of memory at home to provide the following capabilities.
@@ -72,6 +77,37 @@ See: [API Documents](https://kiarina.github.io/kiapi/)
 > The Mac Studio M4 Max with 128GB of memory is an example.
 > Other Apple Silicon models will also work if they have sufficient unified memory.
 
+## Web UI
+
+Open `http://localhost:8000/` while kiapi runs. The web UI lets you try every
+family without writing a request, and shows what each one can do.
+
+- **Playground** for every family: forms built from the API schema, uploads or
+  stored files as inputs, and results as they finish.
+- **Write with chat** drafts prompts with a local chat model that reads the
+  family's guide.
+- **Ask about a family** from the corner button. The assistant answers from
+  the family's OpenAPI document and can fill in the form for you; you review
+  and submit it.
+- **Models** shows which models are ready and the `kiapi activate` command for
+  anything missing. **Jobs** and **Files** show progress and previews.
+
+<p>
+<picture>
+  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/kiarina/kiapi/main/docs/images/web-ui/overview-dark.webp">
+  <img alt="Overview of server status, jobs, and recent files" src="https://raw.githubusercontent.com/kiarina/kiapi/main/docs/images/web-ui/overview-light.webp" width="49%">
+</picture>
+<picture>
+  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/kiarina/kiapi/main/docs/images/web-ui/models-dark.webp">
+  <img alt="Model setup status with activate commands" src="https://raw.githubusercontent.com/kiarina/kiapi/main/docs/images/web-ui/models-light.webp" width="49%">
+</picture>
+</p>
+
+<picture>
+  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/kiarina/kiapi/main/docs/images/web-ui/assistant-dark.webp">
+  <img alt="The assistant filling in the Z-Image form from a request" src="https://raw.githubusercontent.com/kiarina/kiapi/main/docs/images/web-ui/assistant-light.webp">
+</picture>
+
 ## Resources
 
 To provide its capabilities, kiapi downloads and uses selected resources from the list below.
@@ -83,8 +119,8 @@ Always review the upstream license to confirm the terms and whether commercial u
 | Domain | Family | Resource | Kind | Upstream license | Notes |
 |---|---|---|---|---|---|
 | chat | [chat](src/kiapi/capabilities/chat/README.md) | [mlx-community/Qwen3-Omni-30B-A3B-Instruct-4bit](https://huggingface.co/mlx-community/Qwen3-Omni-30B-A3B-Instruct-4bit) | model weights | Apache-2.0 | MLX-converted Qwen3 Omni model. |
+|  |  | [mlx-community/Qwen3.8-Flash-Next-4bit](https://huggingface.co/mlx-community/Qwen3.8-Flash-Next-4bit) | model weights | Qwen Community License 1.0 | MLX-converted Qwen3.8-Flash-Next. Commercial Model-as-a-Service or AI-work-assistant businesses need a separate license from Qwen. |
 |  |  | [mlx-community/Qwen3.8-27B-4bit](https://huggingface.co/mlx-community/Qwen3.8-27B-4bit) | model weights | Apache-2.0 | MLX-converted Qwen3.8 model. |
-|  |  | [mlx-community/Qwen3.6-27B-4bit](https://huggingface.co/mlx-community/Qwen3.6-27B-4bit) | model weights | Apache-2.0 | MLX-converted Qwen3.6 model. |
 | embedding | [embedding](src/kiapi/capabilities/embedding/README.md) | [mlx-community/Qwen3-Embedding-8B-mxfp8](https://huggingface.co/mlx-community/Qwen3-Embedding-8B-mxfp8) | model weights | Apache-2.0 | Text embedding model. |
 |  |  | [mlx-community/Qwen3-VL-Embedding-2B-mxfp8](https://huggingface.co/mlx-community/Qwen3-VL-Embedding-2B-mxfp8) | model weights | Apache-2.0 | Text + image embedding model. |
 | image | [zimage](src/kiapi/capabilities/zimage/README.md) | [filipstrand/Z-Image-Turbo-mflux-4bit](https://huggingface.co/filipstrand/Z-Image-Turbo-mflux-4bit) | model weights | Tongyi Qianwen License | Quantized MLX-compatible Z-Image Turbo; inherits the original Z-Image Turbo license. |
@@ -94,6 +130,7 @@ Always review the upstream license to confirm the terms and whether commercial u
 |  |  | [black-forest-labs/FLUX.2-klein-base-9B](https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9B) | model weights | FLUX Non-Commercial License | Gated upstream model. Confirm terms before any commercial use. |
 |  | [qwen](src/kiapi/capabilities/qwen/README.md) | [Qwen/Qwen-Image](https://huggingface.co/Qwen/Qwen-Image) | model weights | Apache-2.0 | Text-to-image model. |
 |  |  | [Qwen/Qwen-Image-Edit-2509](https://huggingface.co/Qwen/Qwen-Image-Edit-2509) | model weights | Apache-2.0 | Image editing model. |
+|  |  | [Qwen/Qwen-Image-2.1](https://huggingface.co/Qwen/Qwen-Image-2.1) | model weights | Qwen Research License | Non-commercial (research and evaluation) only; commercial use needs a separate license from Alibaba. Unified generation and editing, RGBA output. Editing is pinned from the [kiarina fork](https://github.com/kiarina/mflux/tree/qwen-image-2.1-edit) of mflux until [#741](https://github.com/mflux-community/mflux/pull/741) is released. |
 |  | [ideogram4](src/kiapi/capabilities/ideogram4/README.md) | [ideogram-ai/ideogram-4-fp8](https://huggingface.co/ideogram-ai/ideogram-4-fp8) | model weights | Ideogram Non-Commercial Model Agreement | Gated upstream model. Confirm hosted-service and commercial-use terms. |
 |  | [ernie](src/kiapi/capabilities/ernie/README.md) | [baidu/ERNIE-Image-Turbo](https://huggingface.co/baidu/ERNIE-Image-Turbo) | model weights | Apache-2.0 | Turbo ERNIE-Image variant. |
 |  |  | [baidu/ERNIE-Image](https://huggingface.co/baidu/ERNIE-Image) | model weights | Apache-2.0 | Base ERNIE-Image variant. |
@@ -165,6 +202,8 @@ Always review the upstream license to confirm the terms and whether commercial u
 |  | openapi | `GET /openapi.json` | Return the common API and each capability documentation URL. |
 |  |  | `GET /v1/{domain}/{family}/openapi.json` | Return detailed input/output specs, usage, tips, and examples for each family. |
 |  | health | `GET /health` | Return server status, warmup status, queue length, and memory usage. |
+|  | setup | `GET /v1/setup` | Return every model with its setup state and the command that activates a missing resource. |
+|  | web UI | `GET /` | Serve the web UI. |
 
 See: [kiapi API Docs](https://kiarina.github.io/kiapi/)
 
@@ -232,6 +271,9 @@ with the theme 'a person walking in the rain'.
 open ~/Downloads/bgm.wav
 ```
 
+**Use from a browser:**
+Open `http://localhost:8000/` while kiapi runs. See [Web UI](#web-ui).
+
 **Run as a background service:**
 ```sh
 # kiapi
@@ -284,6 +326,7 @@ make upgrade  # Upgrade dependencies
 make       # Format, type-check, and regenerate dynamic documentation
 make test  # Unit tests
 make dev   # Start the development server with auto-reload
+make web-dev  # Start the web UI dev server against `make dev` (needs Node and pnpm via mise)
 
 # GPU-backed functional and regression tests
 make verify        # Choose the capability family interactively

{kiapi-0.7.0 → kiapi-0.8.0}/README.md

@@ -4,6 +4,11 @@
 ![Python](https://img.shields.io/badge/Python-3.12-blue.svg)
 [![API Docs](https://img.shields.io/badge/API%20Docs-GitHub%20Pages-green.svg)](https://kiarina.github.io/kiapi/)
 
+<picture>
+  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/kiarina/kiapi/main/docs/images/web-ui/playground-dark.webp">
+  <img alt="The kiapi web UI running Z-Image, with the generated image next to the form" src="https://raw.githubusercontent.com/kiarina/kiapi/main/docs/images/web-ui/playground-light.webp">
+</picture>
+
 ## Summary
 
 kiapi is an API server that uses a Mac Studio M4 Max with 128GB of memory at home to provide the following capabilities.
@@ -23,6 +28,37 @@ See: [API Documents](https://kiarina.github.io/kiapi/)
 > The Mac Studio M4 Max with 128GB of memory is an example.
 > Other Apple Silicon models will also work if they have sufficient unified memory.
 
+## Web UI
+
+Open `http://localhost:8000/` while kiapi runs. The web UI lets you try every
+family without writing a request, and shows what each one can do.
+
+- **Playground** for every family: forms built from the API schema, uploads or
+  stored files as inputs, and results as they finish.
+- **Write with chat** drafts prompts with a local chat model that reads the
+  family's guide.
+- **Ask about a family** from the corner button. The assistant answers from
+  the family's OpenAPI document and can fill in the form for you; you review
+  and submit it.
+- **Models** shows which models are ready and the `kiapi activate` command for
+  anything missing. **Jobs** and **Files** show progress and previews.
+
+<p>
+<picture>
+  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/kiarina/kiapi/main/docs/images/web-ui/overview-dark.webp">
+  <img alt="Overview of server status, jobs, and recent files" src="https://raw.githubusercontent.com/kiarina/kiapi/main/docs/images/web-ui/overview-light.webp" width="49%">
+</picture>
+<picture>
+  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/kiarina/kiapi/main/docs/images/web-ui/models-dark.webp">
+  <img alt="Model setup status with activate commands" src="https://raw.githubusercontent.com/kiarina/kiapi/main/docs/images/web-ui/models-light.webp" width="49%">
+</picture>
+</p>
+
+<picture>
+  <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/kiarina/kiapi/main/docs/images/web-ui/assistant-dark.webp">
+  <img alt="The assistant filling in the Z-Image form from a request" src="https://raw.githubusercontent.com/kiarina/kiapi/main/docs/images/web-ui/assistant-light.webp">
+</picture>
+
 ## Resources
 
 To provide its capabilities, kiapi downloads and uses selected resources from the list below.
@@ -34,8 +70,8 @@ Always review the upstream license to confirm the terms and whether commercial u
 | Domain | Family | Resource | Kind | Upstream license | Notes |
 |---|---|---|---|---|---|
 | chat | [chat](src/kiapi/capabilities/chat/README.md) | [mlx-community/Qwen3-Omni-30B-A3B-Instruct-4bit](https://huggingface.co/mlx-community/Qwen3-Omni-30B-A3B-Instruct-4bit) | model weights | Apache-2.0 | MLX-converted Qwen3 Omni model. |
+|  |  | [mlx-community/Qwen3.8-Flash-Next-4bit](https://huggingface.co/mlx-community/Qwen3.8-Flash-Next-4bit) | model weights | Qwen Community License 1.0 | MLX-converted Qwen3.8-Flash-Next. Commercial Model-as-a-Service or AI-work-assistant businesses need a separate license from Qwen. |
 |  |  | [mlx-community/Qwen3.8-27B-4bit](https://huggingface.co/mlx-community/Qwen3.8-27B-4bit) | model weights | Apache-2.0 | MLX-converted Qwen3.8 model. |
-|  |  | [mlx-community/Qwen3.6-27B-4bit](https://huggingface.co/mlx-community/Qwen3.6-27B-4bit) | model weights | Apache-2.0 | MLX-converted Qwen3.6 model. |
 | embedding | [embedding](src/kiapi/capabilities/embedding/README.md) | [mlx-community/Qwen3-Embedding-8B-mxfp8](https://huggingface.co/mlx-community/Qwen3-Embedding-8B-mxfp8) | model weights | Apache-2.0 | Text embedding model. |
 |  |  | [mlx-community/Qwen3-VL-Embedding-2B-mxfp8](https://huggingface.co/mlx-community/Qwen3-VL-Embedding-2B-mxfp8) | model weights | Apache-2.0 | Text + image embedding model. |
 | image | [zimage](src/kiapi/capabilities/zimage/README.md) | [filipstrand/Z-Image-Turbo-mflux-4bit](https://huggingface.co/filipstrand/Z-Image-Turbo-mflux-4bit) | model weights | Tongyi Qianwen License | Quantized MLX-compatible Z-Image Turbo; inherits the original Z-Image Turbo license. |
@@ -45,6 +81,7 @@ Always review the upstream license to confirm the terms and whether commercial u
 |  |  | [black-forest-labs/FLUX.2-klein-base-9B](https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9B) | model weights | FLUX Non-Commercial License | Gated upstream model. Confirm terms before any commercial use. |
 |  | [qwen](src/kiapi/capabilities/qwen/README.md) | [Qwen/Qwen-Image](https://huggingface.co/Qwen/Qwen-Image) | model weights | Apache-2.0 | Text-to-image model. |
 |  |  | [Qwen/Qwen-Image-Edit-2509](https://huggingface.co/Qwen/Qwen-Image-Edit-2509) | model weights | Apache-2.0 | Image editing model. |
+|  |  | [Qwen/Qwen-Image-2.1](https://huggingface.co/Qwen/Qwen-Image-2.1) | model weights | Qwen Research License | Non-commercial (research and evaluation) only; commercial use needs a separate license from Alibaba. Unified generation and editing, RGBA output. Editing is pinned from the [kiarina fork](https://github.com/kiarina/mflux/tree/qwen-image-2.1-edit) of mflux until [#741](https://github.com/mflux-community/mflux/pull/741) is released. |
 |  | [ideogram4](src/kiapi/capabilities/ideogram4/README.md) | [ideogram-ai/ideogram-4-fp8](https://huggingface.co/ideogram-ai/ideogram-4-fp8) | model weights | Ideogram Non-Commercial Model Agreement | Gated upstream model. Confirm hosted-service and commercial-use terms. |
 |  | [ernie](src/kiapi/capabilities/ernie/README.md) | [baidu/ERNIE-Image-Turbo](https://huggingface.co/baidu/ERNIE-Image-Turbo) | model weights | Apache-2.0 | Turbo ERNIE-Image variant. |
 |  |  | [baidu/ERNIE-Image](https://huggingface.co/baidu/ERNIE-Image) | model weights | Apache-2.0 | Base ERNIE-Image variant. |
@@ -116,6 +153,8 @@ Always review the upstream license to confirm the terms and whether commercial u
 |  | openapi | `GET /openapi.json` | Return the common API and each capability documentation URL. |
 |  |  | `GET /v1/{domain}/{family}/openapi.json` | Return detailed input/output specs, usage, tips, and examples for each family. |
 |  | health | `GET /health` | Return server status, warmup status, queue length, and memory usage. |
+|  | setup | `GET /v1/setup` | Return every model with its setup state and the command that activates a missing resource. |
+|  | web UI | `GET /` | Serve the web UI. |
 
 See: [kiapi API Docs](https://kiarina.github.io/kiapi/)
 
@@ -183,6 +222,9 @@ with the theme 'a person walking in the rain'.
 open ~/Downloads/bgm.wav
 ```
 
+**Use from a browser:**
+Open `http://localhost:8000/` while kiapi runs. See [Web UI](#web-ui).
+
 **Run as a background service:**
 ```sh
 # kiapi
@@ -235,6 +277,7 @@ make upgrade  # Upgrade dependencies
 make       # Format, type-check, and regenerate dynamic documentation
 make test  # Unit tests
 make dev   # Start the development server with auto-reload
+make web-dev  # Start the web UI dev server against `make dev` (needs Node and pnpm via mise)
 
 # GPU-backed functional and regression tests
 make verify        # Choose the capability family interactively

{kiapi-0.7.0 → kiapi-0.8.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "kiapi"
-version = "0.7.0"
+version = "0.8.0"
 description = "Unified local inference API server for LLM agents on Apple Silicon (MLX): chat, embedding, image, music, sound-effect, and video — one process, one global memory budget, one single-flight queue, unified jobs and files."
 readme = "README.md"
 requires-python = ">=3.12,<3.13"
@@ -60,7 +60,7 @@ dependencies = [
     # as a PythonPackageResource to keep PyPI metadata free of direct URLs.
     # --- chat capability (mlx-vlm) ---
     # mlx-vlm supports BOTH model_types we use in one install:
-    #   qwen3_omni_moe (Qwen3-Omni) and qwen3_5 (Qwen3.6 / Qwen3.8-27B).
+    #   qwen3_omni_moe (Qwen3-Omni), qwen3_5 (Qwen3.8-27B) and qwen4_exp (Qwen3.8-Flash-Next).
     # Pinned exactly because the chat capability patches its internals;
     # remaining workarounds tracked in src/kiapi/capabilities/chat/README.md.
     "mlx-vlm==0.7.1",
@@ -69,7 +69,7 @@ dependencies = [
     "torch",
     "torchvision",
     "torchaudio",
-    "mflux>=0.18.0",
+    "mflux>=0.20.0",
     "pydantic-settings-manager>=3.8.0",
 ]
 
@@ -88,10 +88,16 @@ build-backend = "hatchling.build"
 
 [tool.hatch.build.targets.wheel]
 packages = ["src/kiapi"]
-artifacts = ["src/kiapi/capabilities/web/resources/searxng/settings.yml"]
+artifacts = [
+    "src/kiapi/capabilities/web/resources/searxng/settings.yml",
+    # Built by `mise run web:build`; gitignored, so listed explicitly.
+    "src/kiapi/api/ui/static/**",
+]
 
 [tool.hatch.build.targets.sdist]
 only-include = ["src"]
+# `uv build` makes the wheel from the sdist, so the built UI must ride along here too.
+artifacts = ["src/kiapi/api/ui/static/**"]
 
 [tool.hatch.metadata]
 allow-direct-references = true
@@ -100,6 +106,13 @@ allow-direct-references = true
 # The development lockfile resolves on Apple Silicon (kiapi pulls MLX wheels).
 environments = ["sys_platform == 'darwin' and platform_machine == 'arm64'"]
 
+[tool.uv.sources]
+# Opt-in Qwen3.8 image and Omni media-prefix APC, pending upstream releases.
+# Published wheels retain the official 0.7.1 dependency and conservative fallback.
+mlx-vlm = { git = "https://github.com/kiarina/mlx-vlm.git", rev = "6581ba8c2f2dbc8b994cb0a350532a4bec59abc2" }
+# Qwen-Image-2.1 editing and RGBA output (mflux-community/mflux#741), pending upstream release.
+mflux = { git = "https://github.com/kiarina/mflux.git", rev = "144a6becfb587bcaa0139758131faf0a7930ae18" }
+
 [dependency-groups]
 dev = [
     "pytest>=9.0.3",

{kiapi-0.7.0 → kiapi-0.8.0}/src/kiapi/api/app.py

@@ -29,6 +29,10 @@ from .image.seedvr2.router import router as seedvr2_router
 from .image.zimage.router import router as zimage_router
 from .jobs.router import router as jobs_router
 from .models.router import router as models_router
+from .setup.router import router as setup_router
+from .ui.router import ASSETS_PATH as ui_assets_path
+from .ui.router import assets as ui_assets
+from .ui.router import router as ui_router
 from .video.ltx2.router import router as ltx2_router
 from .web.router import router as web_router
 
@@ -38,6 +42,7 @@ COMMON_OPENAPI_PATHS = (
     "/health",
     "/v1/files",
     "/v1/jobs",
+    "/v1/setup",
 )
 CAPABILITY_DESCRIPTION_DOMAIN_ORDER = (
     "chat",
@@ -151,3 +156,6 @@ app.include_router(acestep_router)
 app.include_router(audiogen_router)
 app.include_router(ltx2_router)
 app.include_router(web_router)
+app.include_router(setup_router)
+app.include_router(ui_router)
+app.mount(ui_assets_path, ui_assets, name="ui")

{kiapi-0.7.0 → kiapi-0.8.0}/src/kiapi/api/chat/_views/chat_completion_response.py

@@ -49,10 +49,19 @@ class _Choice(BaseModel):
     )
 
 
+class _PromptTokensDetails(BaseModel):
+    cached_tokens: int = Field(
+        description="Prompt tokens reused from the automatic prefix cache."
+    )
+
+
 class _Usage(BaseModel):
     prompt_tokens: int = Field(description="Tokens in the prompt.")
     completion_tokens: int = Field(description="Tokens generated.")
     total_tokens: int = Field(description="Sum of prompt and completion tokens.")
+    prompt_tokens_details: _PromptTokensDetails = Field(
+        description="Detailed accounting for prompt tokens."
+    )
 
 
 class _Timings(BaseModel):

{kiapi-0.7.0 → kiapi-0.8.0}/src/kiapi/api/chat/router.py

@@ -13,11 +13,12 @@ import json
 import logging
 import re
 from collections.abc import AsyncIterator
+from contextlib import suppress
 from copy import deepcopy
 from typing import Any
 
-from fastapi import APIRouter, Depends, HTTPException
-from fastapi.responses import StreamingResponse
+from fastapi import APIRouter, Depends, HTTPException, Request
+from fastapi.responses import Response, StreamingResponse
 
 from kiapi.api import (
     REQUIRE_AUTH,
@@ -28,6 +29,7 @@ from kiapi.api import (
 from kiapi.api._settings import settings_manager
 from kiapi.capabilities.chat import ChatRequest, handle_chat
 from kiapi.core.app import AppContext
+from kiapi.core.job import Job
 from kiapi.core.memory import MemoryBudgetError
 from kiapi.core.model import UnknownModelError, model_registry
 from kiapi.core.setup import SetupRequiredError
@@ -71,9 +73,10 @@ _DATA_URL_BASE64 = re.compile(
 )
 async def chat_completions(
     req: ChatRequest,
+    request: Request,
     ctx: AppContext = Depends(get_ctx),
     worker: Worker = Depends(get_worker),
-) -> dict | StreamingResponse:
+) -> dict | Response:
     """Generate a chat completion (OpenAI-compatible).
 
     Accepts the OpenAI `chat.completions` request shape with multimodal
@@ -102,8 +105,11 @@ async def chat_completions(
 
         loop = asyncio.get_running_loop()
         queue: asyncio.Queue[dict | None] = asyncio.Queue()
+        last_chunk: dict[str, Any] | None = None
 
         def emit(chunk: dict) -> None:
+            nonlocal last_chunk
+            last_chunk = chunk
             loop.call_soon_threadsafe(queue.put_nowait, chunk)
 
         def finish_stream() -> None:
@@ -111,7 +117,7 @@ async def chat_completions(
 
         def thunk():  # type: ignore
             try:
-                return handle_chat(ctx, req, emit=emit)
+                return handle_chat(ctx, req, emit=emit, job=job)
             finally:
                 finish_stream()
 
@@ -129,14 +135,17 @@ async def chat_completions(
                     yield _sse(item)
 
                 try:
-                    await asyncio.shield(fut)
+                    result = await asyncio.shield(fut)
                 except Exception as exc:
                     yield _sse(_stream_error(exc))
+                else:
+                    if req.stream_options and req.stream_options.include_usage:
+                        yield _sse(_stream_usage_chunk(result, last_chunk))
                 yield _sse("[DONE]")
             except asyncio.CancelledError:
-                # The running worker job is not preemptible; let it finish and
-                # keep its final state in the job store.
-                return
+                job.request_cancel()
+                fut.cancel()
+                raise
 
         return StreamingResponse(
             events(),
@@ -148,10 +157,23 @@ async def chat_completions(
         )
 
     job = ctx.job_store.create(type="chat", params={"model": req.model})
-    fut = await worker.submit(job, lambda: handle_chat(ctx, req))
+    fut = await worker.submit(job, lambda: handle_chat(ctx, req, job=job))
+    disconnect_task = asyncio.create_task(
+        _cancel_on_disconnect(request, job, fut),
+        name=f"chat-disconnect-{job.id}",
+    )
 
     try:
-        return await asyncio.wait_for(fut, timeout=settings.sync_timeout_s)
+        return await asyncio.wait_for(
+            asyncio.shield(fut), timeout=settings.sync_timeout_s
+        )
+    except asyncio.CancelledError:
+        disconnected = job.cancel_requested()
+        job.request_cancel()
+        fut.cancel()
+        if disconnected:
+            return Response(status_code=499)
+        raise
     except TimeoutError:
         raise HTTPException(  # noqa: B904
             status_code=504,
@@ -169,6 +191,10 @@ async def chat_completions(
         if exc.__class__.__name__ in ("MediaError", "CapabilityError"):
             raise HTTPException(status_code=400, detail=str(exc))  # noqa: B904
         raise HTTPException(status_code=500, detail=f"generation failed: {exc}")  # noqa: B904
+    finally:
+        disconnect_task.cancel()
+        with suppress(asyncio.CancelledError):
+            await disconnect_task
 
 
 register_capability_endpoints(router, name="chat", base_path="/v1/chat")
@@ -251,3 +277,26 @@ def _sse(payload) -> str:  # type: ignore
 
 def _stream_error(exc: BaseException) -> dict:
     return {"error": {"message": str(exc), "type": exc.__class__.__name__}}
+
+
+async def _cancel_on_disconnect(
+    request: Request, job: Job, fut: asyncio.Future
+) -> None:
+    while not fut.done():
+        if await request.is_disconnected():
+            job.request_cancel()
+            fut.cancel()
+            return
+        await asyncio.sleep(0.1)
+
+
+def _stream_usage_chunk(result: dict, last_chunk: dict[str, Any] | None) -> dict:
+    last_chunk = last_chunk or {}
+    return {
+        "id": last_chunk.get("id", result.get("id")),
+        "object": "chat.completion.chunk",
+        "created": last_chunk.get("created", result.get("created")),
+        "model": last_chunk.get("model", result.get("model")),
+        "choices": [],
+        "usage": result["usage"],
+    }

{kiapi-0.7.0 → kiapi-0.8.0}/src/kiapi/api/image/qwen/router.py

@@ -47,8 +47,8 @@ async def generate_qwen_endpoint(
     Plain text-to-image by default; supply `init_image` (+ optional
     `image_strength`) to run img2img instead — use `/v1/image/qwen/edit` for
     natural-language single/multi-image editing. The same endpoint serves both
-    `sync` and `async` via `mode`. This endpoint only accepts the `image` model
-    variant.
+    `sync` and `async` via `mode`. This endpoint accepts the `image` (default)
+    and `image-2.1` model variants; `image-2.1` does not take `init_image`.
 
     Sync content negotiation: a single image is produced, so unless the client
     asks for JSON the raw image bytes are returned with `X-Kiapi-File-Id` /
@@ -64,12 +64,13 @@ async def generate_qwen_endpoint(
         spec = model_registry.resolve("qwen", req.model or "image")
     except UnknownModelError as exc:
         raise HTTPException(status_code=400, detail=str(exc))  # noqa: B904
-    if spec.name != "image":
+    if spec.name not in ("image", "image-2.1"):
         raise HTTPException(
-            status_code=400, detail="qwen generate requires model 'image'"
+            status_code=400,
+            detail="qwen generate requires model 'image' or 'image-2.1'",
         )
     try:
-        validate_generate(req)
+        validate_generate(req, variant=spec.name)
     except ValidationError as exc:
         raise HTTPException(status_code=422, detail=str(exc))  # noqa: B904
 
@@ -104,8 +105,8 @@ async def edit_qwen_endpoint(
     Takes a list of `images` (FileRefs, single or multi-image) plus a
     natural-language `prompt`. For plain image-to-image from a single seed image,
     use `init_image` on `/v1/image/qwen/generate` instead. The same endpoint
-    serves both `sync` and `async` via `mode`. This endpoint only accepts the
-    `edit-2509` model variant.
+    serves both `sync` and `async` via `mode`. This endpoint accepts the
+    `edit-2509` (default) and `image-2.1` model variants.
 
     Sync content negotiation, transient-model behavior (`loras` / `quantize`
     override), and the ImageResponse `result` shape match
@@ -115,12 +116,13 @@ async def edit_qwen_endpoint(
         spec = model_registry.resolve("qwen", req.model or "edit-2509")
     except UnknownModelError as exc:
         raise HTTPException(status_code=400, detail=str(exc))  # noqa: B904
-    if spec.name != "edit-2509":
+    if spec.name not in ("edit-2509", "image-2.1"):
         raise HTTPException(
-            status_code=400, detail="qwen edit requires model 'edit-2509'"
+            status_code=400,
+            detail="qwen edit requires model 'edit-2509' or 'image-2.1'",
         )
     try:
-        validate_edit(req)
+        validate_edit(req, variant=spec.name)
     except ValidationError as exc:
         raise HTTPException(status_code=422, detail=str(exc))  # noqa: B904
 

{kiapi-0.7.0 → kiapi-0.8.0}/src/kiapi/api/models/_schemas/openai_model_spec.py

@@ -24,3 +24,11 @@ class OpenAIModelSpec(BaseModel):
         description="Owner marker for OpenAI-compatible model lists.",
         examples=["kiapi"],
     )
+    context_window: int | None = Field(
+        default=None,
+        description=(
+            "Maximum prompt plus output tokens, read from the model's config. "
+            "Null until the model is set up."
+        ),
+        examples=[262144],
+    )

{kiapi-0.7.0 → kiapi-0.8.0}/src/kiapi/api/models/router.py

@@ -2,6 +2,7 @@ import time
 
 from fastapi import APIRouter
 
+from kiapi.capabilities.chat import read_context_window
 from kiapi.core.model import model_registry
 
 from ._schemas.openai_model_spec import OpenAIModelSpec
@@ -23,7 +24,11 @@ async def list_openai_compatible_chat_models() -> ModelListResponse:
     /v1/{domain}/{family}/models.
     """
     data = [
-        OpenAIModelSpec(id=spec.name, created=_CREATED)
+        OpenAIModelSpec(
+            id=spec.name,
+            created=_CREATED,
+            context_window=read_context_window(spec),
+        )
         for spec in model_registry.list_specs("chat")
     ]
     return ModelListResponse(data=data)

kiapi-0.8.0/src/kiapi/api/setup/_helpers/build_setup_response.py

@@ -0,0 +1,65 @@
+from typing import Literal, Protocol
+
+from kiapi.core.model import ModelSpec
+from kiapi.core.setup import SetupResource, SetupStatus
+
+from .._views.setup_model_state import SetupModelState
+from .._views.setup_resource_state import SetupResourceState
+from .._views.setup_response import SetupResponse
+from .._views.setup_summary import SetupSummary
+
+
+class _StatusManager(Protocol):
+    def status(self, resource: SetupResource) -> SetupStatus: ...
+
+
+def build_setup_response(
+    specs: list[ModelSpec], manager: _StatusManager
+) -> SetupResponse:
+    # Models often share a resource (e.g. SeedVR2 3b and 7b), so check each once.
+    states: dict[str, SetupResourceState] = {}
+    models = []
+    for spec in sorted(specs, key=lambda s: (s.domain, s.family)):
+        resources = []
+        for resource in spec.setup_resources:
+            if resource.key not in states:
+                status = manager.status(resource)
+                states[resource.key] = SetupResourceState(
+                    kind=resource.kind,
+                    label=resource.label,
+                    ready=status.ready,
+                    disk_gb=resource.disk_gb,
+                    detail=status.detail,
+                    activate_command=f"kiapi activate --repo {resource.label}",
+                )
+            resources.append(states[resource.key])
+        models.append(
+            SetupModelState(
+                domain=spec.domain,
+                family=spec.family,
+                name=spec.name,
+                default=spec.default,
+                status=_model_status(resources),
+                size_gb=sum(r.disk_gb or 0.0 for r in resources),
+                resources=resources,
+            )
+        )
+
+    unique = list(states.values())
+    summary = SetupSummary(
+        models_total=len(models),
+        models_ready=sum(1 for m in models if m.status != "missing"),
+        resources_total=len(unique),
+        resources_ready=sum(1 for r in unique if r.ready),
+        installed_gb=sum(r.disk_gb or 0.0 for r in unique if r.ready),
+        total_gb=sum(r.disk_gb or 0.0 for r in unique),
+    )
+    return SetupResponse(summary=summary, data=models)
+
+
+def _model_status(
+    resources: list[SetupResourceState],
+) -> Literal["ready", "missing", "none"]:
+    if not resources:
+        return "none"
+    return "ready" if all(r.ready for r in resources) else "missing"

kiapi-0.8.0/src/kiapi/api/setup/_views/setup_model_state.py

@@ -0,0 +1,31 @@
+from typing import Literal
+
+from pydantic import BaseModel, Field
+
+from .setup_resource_state import SetupResourceState
+
+
+class SetupModelState(BaseModel):
+    domain: str = Field(..., description="Modality bucket.", examples=["image"])
+    family: str = Field(..., description="Family identifier.", examples=["zimage"])
+    name: str = Field(..., description="Model variant name.", examples=["base"])
+    default: bool = Field(
+        ...,
+        description="Whether this is the family's default variant.",
+        examples=[False],
+    )
+    status: Literal["ready", "missing", "none"] = Field(
+        ...,
+        description="ready when every resource is present, missing when any is "
+        "absent, none when the model needs no setup.",
+        examples=["missing"],
+    )
+    size_gb: float = Field(
+        ...,
+        description="Approximate total on-disk size of the model's resources in GB.",
+        examples=[19.0],
+    )
+    resources: list[SetupResourceState] = Field(
+        default_factory=list,
+        description="Setup resources the model needs.",
+    )