Build and run agents you can see, understand and trust.
Python agentchatbotlarge-language-modelsllmllm-agentmcpmulti-agentmulti-modalreact-agentrealtime-agentvoice-agentvoice-assistant
refactor(model): refactor dashscope model to openai compatible (#1617)
qbcec0a6e255912
diff --git a/scripts/model_examples/README.md b/scripts/model_examples/README.mdindex b953326e1..e7b0822fc 100644--- a/scripts/model_examples/README.md+++ b/scripts/model_examples/README.md@@ -212,7 +212,7 @@ Each script typically defines two or more async functions:- `example_simple_call()` – basic text call with streaming- `example_tool_call()` – two-round conversation with tool/function calling-- `example_multimodal_call()` – image + text input (in `_multimodal.py` variants)+- `example_image_url()` – image + text input (in `_multimodal.py` variants)---diff --git a/scripts/model_examples/_utils.py b/scripts/model_examples/_utils.pyindex 7f8400a59..534d0a6d5 100644--- a/scripts/model_examples/_utils.py+++ b/scripts/model_examples/_utils.py@@ -2,7 +2,7 @@"""Shared utility helpers for model-examples scripts."""from typing import AsyncGenerator-from agentscope.message import TextBlock, ThinkingBlock+from agentscope.message import TextBlock, ThinkingBlock, DataBlockfrom agentscope.model import ChatResponse@@ -52,6 +52,17 @@ async def stream_and_collect(print(final_text)if in_thinking:print()+ # Report audio output if present in the final chunk+ if final is not None:+ import base64 as _b64++ for block in final.content:+ if isinstance(block, DataBlock):+ media = block.source.media_type+ byte_size = len(_b64.b64decode(block.source.data))+ print(+ f"[Audio output: {media}, " f"{byte_size} bytes]",+ )print()assert final is not Nonereturn finaldiff --git a/scripts/model_examples/anthropic_multiagent_multimodal.py b/scripts/model_examples/anthropic_multiagent_multimodal.pyindex 25f131b2c..a4d79569d 100644--- a/scripts/model_examples/anthropic_multiagent_multimodal.py+++ b/scripts/model_examples/anthropic_multiagent_multimodal.py@@ -21,7 +21,7 @@ TEST_IMAGE_URL = ()-async def example_multiagent_multimodal() -> None:+async def example_multiagent_image_url() -> None:"""Multi-agent conversation where Alice shares an image for the group."""formatter = AnthropicMultiAgentFormatter()@@ -82,4 +82,4 @@ async def example_multiagent_multimodal() -> None:if __name__ == "__main__":- asyncio.run(example_multiagent_multimodal())+ asyncio.run(example_multiagent_image_url())diff --git a/scripts/model_examples/anthropic_multimodal.py b/scripts/model_examples/anthropic_multimodal.pyindex 9c9920590..9b5bd0db5 100644--- a/scripts/model_examples/anthropic_multimodal.py+++ b/scripts/model_examples/anthropic_multimodal.py@@ -24,7 +24,7 @@ TEST_IMAGE_URL = ()-async def example_multimodal() -> None:+async def example_image_url() -> None:"""Call claude-opus-4-5 with an image URL and ask what is in the image."""model = AnthropicChatModel(credential=AnthropicCredential(@@ -71,7 +71,7 @@ def _build_model() -> AnthropicChatModel:)-async def example_local_path() -> None:+async def example_image_local_path() -> None:"""Call claude-opus-4-5 with a local image using a ``file://`` URL.The formatter automatically reads the file and converts it to base64.@@ -104,7 +104,7 @@ async def example_local_path() -> None:await stream_and_collect(await model(msgs))-async def example_file_url() -> None:+async def example_image_base64() -> None:"""Call claude-opus-4-5 with a local image using explicit base64 encoding.Use ``Base64Source`` when you already have the binary data in memory or@@ -141,6 +141,6 @@ async def example_file_url() -> None:if __name__ == "__main__":- asyncio.run(example_multimodal())- asyncio.run(example_local_path())- asyncio.run(example_file_url())+ asyncio.run(example_image_url())+ asyncio.run(example_image_local_path())+ asyncio.run(example_image_base64())diff --git a/scripts/model_examples/dashscope_call.py b/scripts/model_examples/dashscope_call.pyindex 7e4a7ea47..ac77abbda 100644--- a/scripts/model_examples/dashscope_call.py+++ b/scripts/model_examples/dashscope_call.py@@ -27,7 +27,6 @@ async def example_simple_call() -> None:api_key=os.environ["DASHSCOPE_API_KEY"],),model="qwen3.5-plus",- multimodality=True,stream=True,context_size=1_000_000,parameters=DashScopeChatModel.Parameters(thinking_enable=True),@@ -77,7 +76,6 @@ async def example_tool_call() -> None:),model="qwen3.5-plus",stream=True,- multimodality=True,context_size=1_000_000,parameters=DashScopeChatModel.Parameters(thinking_enable=True),)diff --git a/scripts/model_examples/dashscope_multiagent.py b/scripts/model_examples/dashscope_multiagent.pyindex f03e90436..c2d5a19fa 100644--- a/scripts/model_examples/dashscope_multiagent.py+++ b/scripts/model_examples/dashscope_multiagent.py@@ -28,7 +28,6 @@ async def example_multiagent() -> None:api_key=os.environ["DASHSCOPE_API_KEY"],),model="qwen3.5-plus",- multimodality=True,stream=True,context_size=1_000_000,formatter=formatter,diff --git a/scripts/model_examples/dashscope_multiagent_multimodal.py b/scripts/model_examples/dashscope_multiagent_multimodal.pyindex ceb37e78b..b9189b6a1 100644--- a/scripts/model_examples/dashscope_multiagent_multimodal.py+++ b/scripts/model_examples/dashscope_multiagent_multimodal.py@@ -20,7 +20,7 @@ TEST_IMAGE_URL = ()-async def example_multiagent_multimodal() -> None:+async def example_multiagent_image_url() -> None:"""Multi-agent conversation where Alice shares an image for the group."""formatter = DashScopeMultiAgentFormatter()@@ -29,7 +29,6 @@ async def example_multiagent_multimodal() -> None:api_key=os.environ["DASHSCOPE_API_KEY"],),model="qwen3.5-plus",- multimodality=True,stream=True,context_size=1_000_000,formatter=formatter,@@ -82,4 +81,4 @@ async def example_multiagent_multimodal() -> None:if __name__ == "__main__":- asyncio.run(example_multiagent_multimodal())+ asyncio.run(example_multiagent_image_url())diff --git a/scripts/model_examples/dashscope_multimodal.py b/scripts/model_examples/dashscope_multimodal.pyindex 51f481dc0..973a0de2b 100644--- a/scripts/model_examples/dashscope_multimodal.py+++ b/scripts/model_examples/dashscope_multimodal.py@@ -16,21 +16,32 @@ from agentscope.message import (from agentscope.model import DashScopeChatModelfrom agentscope.credential import DashScopeCredential-# A publicly accessible test image (a simple cat photo)+# A publicly accessible test imageTEST_IMAGE_URL = ("https://help-static-aliyun-doc.aliyuncs.com/file-manage""-files/zh-CN/20241022/emyrja/dog_and_girl.jpeg")+# A publicly accessible test video+TEST_VIDEO_URL = (+ "https://help-static-aliyun-doc.aliyuncs.com/file-manage"+ "-files/zh-CN/20241115/cqqkru/1.mp4"+)-async def example_multimodal() -> None:+# A publicly accessible test audio+TEST_AUDIO_URL = (+ "https://help-static-aliyun-doc.aliyuncs.com/file-manage"+ "-files/zh-CN/20250211/tixcef/cherry.wav"+)+++async def example_image_url() -> None:"""Call qwen3.5-plus with an image URL and ask what is in the image."""model = DashScopeChatModel(credential=DashScopeCredential(api_key=os.environ["DASHSCOPE_API_KEY"],),model="qwen3.5-plus",- multimodality=True,stream=True,context_size=1_000_000,)@@ -66,13 +77,12 @@ def _build_model() -> DashScopeChatModel:api_key=os.environ["DASHSCOPE_API_KEY"],),model="qwen3.5-plus",- multimodality=True,stream=True,context_size=1_000_000,)-async def example_local_path() -> None:+async def example_image_local_path() -> None:"""Call qwen3.5-plus with a local image using a ``file://`` URL.The formatter reads the file from disk and converts it to a base64 data@@ -106,7 +116,7 @@ async def example_local_path() -> None:await stream_and_collect(await model(msgs))-async def example_file_url() -> None:+async def example_image_base64() -> None:"""Call qwen3.5-plus with a local image using explicit base64 encoding.Use ``Base64Source`` when you already have the binary data in memory or@@ -142,7 +152,78 @@ async def example_file_url() -> None:await stream_and_collect(await model(msgs))+async def example_video() -> None:+ """Call qwen3.5-plus with a video URL and ask what is in the video."""+ model = _build_model()++ video_block = DataBlock(+ source=URLSource(+ url=TEST_VIDEO_URL,+ media_type="video/mp4",+ ),+ )++ msgs = [+ Msg(+ name="user",+ content=[+ TextBlock(+ text="What is happening in this video? "+ "Describe it briefly.",+ ),+ video_block,+ ],+ role="user",+ ),+ ]++ print("=== Multimodal Call (Video URL) ===")+ await stream_and_collect(await model(msgs))+++async def example_audio() -> None:+ """Call qwen3.5-omni-plus with an audio URL.++ Audio understanding requires an Omni model (qwen3.5-omni-plus or+ Qwen3-Omni-Flash). Omni models also require stream=True and the+ ``modalities`` parameter.+ """+ model = DashScopeChatModel(+ credential=DashScopeCredential(+ api_key=os.environ["DASHSCOPE_API_KEY"],+ ),+ model="qwen3.5-omni-plus",+ stream=True,+ context_size=1_000_000,+ )++ audio_block = DataBlock(+ source=URLSource(+ url=TEST_AUDIO_URL,+ media_type="audio/wav",+ ),+ )++ msgs = [+ Msg(+ name="user",+ content=[+ TextBlock(text="What is being said in this audio clip?"),+ audio_block,+ ],+ role="user",+ ),+ ]++ print("=== Multimodal Call (Audio URL - Omni) ===")+ await stream_and_collect(+ await model(msgs, modalities=["text", "audio"]),+ )++if __name__ == "__main__":- asyncio.run(example_multimodal())- asyncio.run(example_local_path())- asyncio.run(example_file_url())+ asyncio.run(example_image_url())+ asyncio.run(example_image_local_path())+ asyncio.run(example_image_base64())+ asyncio.run(example_video())+ asyncio.run(example_audio())diff --git a/scripts/model_examples/gemini_multiagent_multimodal.py b/scripts/model_examples/gemini_multiagent_multimodal.pyindex 9e632e510..4ce040a7e 100644--- a/scripts/model_examples/gemini_multiagent_multimodal.py+++ b/scripts/model_examples/gemini_multiagent_multimodal.py@@ -15,7 +15,7 @@ TEST_IMAGE_URL = ()-async def example_multiagent_multimodal() -> None:+async def example_multiagent_image_url() -> None:"""Multi-agent conversation where Alice shares an image for the group."""formatter = GeminiMultiAgentFormatter()@@ -76,4 +76,4 @@ async def example_multiagent_multimodal() -> None:if __name__ == "__main__":- asyncio.run(example_multiagent_multimodal())+ asyncio.run(example_multiagent_image_url())diff --git a/scripts/model_examples/gemini_multimodal.py b/scripts/model_examples/gemini_multimodal.pyindex c2e8814bb..ed78c4fe0 100644--- a/scripts/model_examples/gemini_multimodal.py+++ b/scripts/model_examples/gemini_multimodal.py@@ -23,7 +23,7 @@ TEST_IMAGE_URL = ()-async def example_multimodal() -> None:+async def example_image_url() -> None:"""Call gemini-2.5-flash with an image URL and ask what is in the image."""model = GeminiChatModel(credential=GeminiCredential(@@ -67,7 +67,7 @@ def _build_model() -> GeminiChatModel:)-async def example_local_path() -> None:+async def example_image_local_path() -> None:"""Call gemini-2.5-flash with a local image using a ``file://`` URL.The formatter reads the file from disk and converts it to base64.@@ -98,7 +98,7 @@ async def example_local_path() -> None:await stream_and_collect(await model(msgs))-async def example_file_url() -> None:+async def example_image_base64() -> None:"""Call gemini-2.5-flash with a local image using explicit base64 encoding.Use ``Base64Source`` when you already have the binary data in memory or@@ -130,6 +130,6 @@ async def example_file_url() -> None:if __name__ == "__main__":- asyncio.run(example_multimodal())- asyncio.run(example_local_path())- asyncio.run(example_file_url())+ asyncio.run(example_image_url())+ asyncio.run(example_image_local_path())+ asyncio.run(example_image_base64())diff --git a/scripts/model_examples/kimi_multiagent_multimodal.py b/scripts/model_examples/kimi_multiagent_multimodal.pyindex a4858bd46..489660b38 100644--- a/scripts/model_examples/kimi_multiagent_multimodal.py+++ b/scripts/model_examples/kimi_multiagent_multimodal.py@@ -15,7 +15,7 @@ TEST_IMAGE_URL = ()-async def example_multiagent_multimodal() -> None:+async def example_multiagent_image_url() -> None:"""Multi-agent conversation where Alice shares an image for the group."""formatter = KimiMultiAgentFormatter()@@ -76,4 +76,4 @@ async def example_multiagent_multimodal() -> None:if __name__ == "__main__":- asyncio.run(example_multiagent_multimodal())+ asyncio.run(example_multiagent_image_url())diff --git a/scripts/model_examples/kimi_multimodal.py b/scripts/model_examples/kimi_multimodal.pyindex 8c2b34a63..c0b544fcd 100644--- a/scripts/model_examples/kimi_multimodal.py+++ b/scripts/model_examples/kimi_multimodal.py@@ -23,7 +23,7 @@ TEST_IMAGE_URL = ()-async def example_multimodal() -> None:+async def example_image_url() -> None:"""Call kimi-k2.6 with an image URL and ask what is in the image."""model = KimiChatModel(credential=KimiCredential(@@ -67,7 +67,7 @@ def _build_model() -> KimiChatModel:)-async def example_local_path() -> None:+async def example_image_local_path() -> None:"""Call kimi-k2.6 with a local image using a ``file://`` URL.The formatter reads the file from disk and converts it to a base64 data@@ -99,7 +99,7 @@ async def example_local_path() -> None:await stream_and_collect(await model(msgs))-async def example_file_url() -> None:+async def example_image_base64() -> None:"""Call kimi-k2.6 with a local image using explicit base64 encoding.Use ``Base64Source`` when you already have the binary data in memory or@@ -131,6 +131,6 @@ async def example_file_url() -> None:if __name__ == "__main__":- asyncio.run(example_multimodal())- asyncio.run(example_local_path())- asyncio.run(example_file_url())+ asyncio.run(example_image_url())+ asyncio.run(example_image_local_path())+ asyncio.run(example_image_base64())diff --git a/scripts/model_examples/ollama_multiagent_multimodal.py b/scripts/model_examples/ollama_multiagent_multimodal.pyindex 4e9abd9ed..b55f0aba6 100644--- a/scripts/model_examples/ollama_multiagent_multimodal.py+++ b/scripts/model_examples/ollama_multiagent_multimodal.py@@ -17,7 +17,7 @@ TEST_IMAGE_URL = ()-async def example_multiagent_multimodal() -> None:+async def example_multiagent_image_url() -> None:"""Multi-agent conversation where Alice shares an image for the group.Requires `ollama pull llava` to have been run first.@@ -78,4 +78,4 @@ async def example_multiagent_multimodal() -> None:if __name__ == "__main__":- asyncio.run(example_multiagent_multimodal())+ asyncio.run(example_multiagent_image_url())diff --git a/scripts/model_examples/ollama_multimodal.py b/scripts/model_examples/ollama_multimodal.pyindex 461ae3973..33a333139 100644--- a/scripts/model_examples/ollama_multimodal.py+++ b/scripts/model_examples/ollama_multimodal.py@@ -25,7 +25,7 @@ TEST_IMAGE_URL = ()-async def example_multimodal() -> None:+async def example_image_url() -> None:"""Call llava:7b with an image URL and ask what is in the image.Requires `ollama pull llava` to be run first.@@ -68,7 +68,7 @@ def _build_model() -> OllamaChatModel:)-async def example_local_path() -> None:+async def example_image_local_path() -> None:"""Call llava:7b with a local image using a ``file://`` URL.The formatter reads the file from disk and converts it to base64.@@ -99,7 +99,7 @@ async def example_local_path() -> None:await stream_and_collect(await model(msgs))-async def example_file_url() -> None:+async def example_image_base64() -> None:"""Call llava:7b with a local image using explicit base64 encoding.Use ``Base64Source`` when you already have the binary data in memory or@@ -131,6 +131,6 @@ async def example_file_url() -> None:if __name__ == "__main__":- asyncio.run(example_multimodal())- asyncio.run(example_local_path())- asyncio.run(example_file_url())+ asyncio.run(example_image_url())+ asyncio.run(example_image_local_path())+ asyncio.run(example_image_base64())diff --git a/scripts/model_examples/openai_chat_multiagent_multimodal.py b/scripts/model_examples/openai_chat_multiagent_multimodal.pyindex a74337710..8f58ea2ca 100644--- a/scripts/model_examples/openai_chat_multiagent_multimodal.py+++ b/scripts/model_examples/openai_chat_multiagent_multimodal.py@@ -16,7 +16,7 @@ TEST_IMAGE_URL = ()-async def example_multiagent_multimodal() -> None:+async def example_multiagent_image_url() -> None:"""Multi-agent conversation where Alice shares an image for the group."""formatter = OpenAIMultiAgentFormatter()@@ -77,4 +77,4 @@ async def example_multiagent_multimodal() -> None:if __name__ == "__main__":- asyncio.run(example_multiagent_multimodal())+ asyncio.run(example_multiagent_image_url())diff --git a/scripts/model_examples/openai_chat_multimodal.py b/scripts/model_examples/openai_chat_multimodal.pyindex 5dfc3666f..07ec51dc8 100644--- a/scripts/model_examples/openai_chat_multimodal.py+++ b/scripts/model_examples/openai_chat_multimodal.py@@ -23,7 +23,7 @@ TEST_IMAGE_URL = ()-async def example_multimodal() -> None:+async def example_image_url() -> None:"""Call gpt-4.1 with an image URL and ask what is in the image."""model = OpenAIChatModel(credential=OpenAICredential(@@ -67,7 +67,7 @@ def _build_model() -> OpenAIChatModel:)-async def example_local_path() -> None:+async def example_image_local_path() -> None:"""Call gpt-4.1 with a local image using a ``file://`` URL.The formatter reads the file from disk and converts it to a base64 data@@ -99,7 +99,7 @@ async def example_local_path() -> None:await stream_and_collect(await model(msgs))-async def example_file_url() -> None:+async def example_image_base64() -> None:"""Call gpt-4.1 with a local image using explicit base64 encoding.Use ``Base64Source`` when you already have the binary data in memory or@@ -131,6 +131,6 @@ async def example_file_url() -> None:if __name__ == "__main__":- asyncio.run(example_multimodal())- asyncio.run(example_local_path())- asyncio.run(example_file_url())+ asyncio.run(example_image_url())+ asyncio.run(example_image_local_path())+ asyncio.run(example_image_base64())diff --git a/scripts/model_examples/openai_response_multiagent_multimodal.py b/scripts/model_examples/openai_response_multiagent_multimodal.pyindex 914811806..f6b909393 100644--- a/scripts/model_examples/openai_response_multiagent_multimodal.py+++ b/scripts/model_examples/openai_response_multiagent_multimodal.py@@ -16,7 +16,7 @@ TEST_IMAGE_URL = ()-async def example_multiagent_multimodal() -> None:+async def example_multiagent_image_url() -> None:"""Multi-agent conversation where Alice shares an image for the group."""formatter = OpenAIResponseMultiAgentFormatter()@@ -77,4 +77,4 @@ async def example_multiagent_multimodal() -> None:if __name__ == "__main__":- asyncio.run(example_multiagent_multimodal())+ asyncio.run(example_multiagent_image_url())diff --git a/scripts/model_examples/openai_response_multimodal.py b/scripts/model_examples/openai_response_multimodal.pyindex e9cfe436b..f0a5be805 100644--- a/scripts/model_examples/openai_response_multimodal.py+++ b/scripts/model_examples/openai_response_multimodal.py@@ -24,7 +24,7 @@ TEST_IMAGE_URL = ()-async def example_multimodal() -> None:+async def example_image_url() -> None:"""Call gpt-4.1 (Responses API) with an image URL and ask what is inthe image."""model = OpenAIResponseModel(@@ -69,7 +69,7 @@ def _build_model() -> OpenAIResponseModel:)-async def example_local_path() -> None:+async def example_image_local_path() -> None:"""Call gpt-4.1 (Responses API) with a local image using a ``file://`` URL.The formatter reads the file from disk and converts it to a base64 data@@ -101,7 +101,7 @@ async def example_local_path() -> None:await stream_and_collect(await model(msgs))-async def example_file_url() -> None:+async def example_image_base64() -> None:"""Call gpt-4.1 (Responses API) with a local image using explicit base64.Use ``Base64Source`` when you already have the binary data in memory or@@ -133,6 +133,6 @@ async def example_file_url() -> None:if __name__ == "__main__":- asyncio.run(example_multimodal())- asyncio.run(example_local_path())- asyncio.run(example_file_url())+ asyncio.run(example_image_url())+ asyncio.run(example_image_local_path())+ asyncio.run(example_image_base64())diff --git a/scripts/model_examples/xai_multiagent_multimodal.py b/scripts/model_examples/xai_multiagent_multimodal.pyindex 845d0b96f..ae1678e8d 100644--- a/scripts/model_examples/xai_multiagent_multimodal.py+++ b/scripts/model_examples/xai_multiagent_multimodal.py@@ -19,7 +19,7 @@ TEST_IMAGE_URL = ()-async def example_multiagent_multimodal() -> None:+async def example_multiagent_image_url() -> None:"""Multi-agent conversation where Alice shares an image for the group."""formatter = XAIMultiAgentFormatter()@@ -80,4 +80,4 @@ async def example_multiagent_multimodal() -> None:if __name__ == "__main__":- asyncio.run(example_multiagent_multimodal())+ asyncio.run(example_multiagent_image_url())diff --git a/scripts/model_examples/xai_multimodal.py b/scripts/model_examples/xai_multimodal.pyindex 303d4fa52..bee93d0a7 100644--- a/scripts/model_examples/xai_multimodal.py+++ b/scripts/model_examples/xai_multimodal.py@@ -27,7 +27,7 @@ TEST_IMAGE_URL = ()-async def example_multimodal() -> None:+async def example_image_url() -> None:"""Call grok-4.3 with an image URL and ask what is in the image."""model = XAIChatModel(credential=XAICredential(@@ -71,7 +71,7 @@ def _build_model() -> XAIChatModel:)-async def example_local_path() -> None:+async def example_image_local_path() -> None:"""Call grok-4.3 with a local image using a ``file://`` URL.The XAI formatter automatically reads the file from disk and converts@@ -103,7 +103,7 @@ async def example_local_path() -> None:await stream_and_collect(await model(msgs))-async def example_file_url() -> None:+async def example_image_base64() -> None:"""Call grok-4.3 with a local image using explicit base64 encoding.Use ``Base64Source`` when you already have the binary data in memory or@@ -135,6 +135,6 @@ async def example_file_url() -> None:if __name__ == "__main__":- asyncio.run(example_multimodal())- asyncio.run(example_local_path())- asyncio.run(example_file_url())+ asyncio.run(example_image_url())+ asyncio.run(example_image_local_path())+ asyncio.run(example_image_base64())diff --git a/src/agentscope/credential/_dashscope.py b/src/agentscope/credential/_dashscope.pyindex bdcc62762..dac93481e 100644--- a/src/agentscope/credential/_dashscope.py+++ b/src/agentscope/credential/_dashscope.py@@ -9,6 +9,8 @@ from ._base import CredentialBaseif TYPE_CHECKING:from ..model import ChatModelBase+_DASHSCOPE_BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1"+class DashScopeCredential(CredentialBase):"""The credential for DashScope API."""@@ -21,10 +23,12 @@ class DashScopeCredential(CredentialBase):title="API Key",)- base_http_api_url: str | None = Field(- default=None,+ base_url: str = Field(+ default=_DASHSCOPE_BASE_URL,title="API Base URL",- description="The base URL of the DashScope API.",+ description=(+ "The base URL for the DashScope OpenAI-compatible API endpoint."+ ),)@classmethoddiff --git a/src/agentscope/formatter/_dashscope_formatter.py b/src/agentscope/formatter/_dashscope_formatter.pyindex 56c9fded0..67b5bf5b9 100644--- a/src/agentscope/formatter/_dashscope_formatter.py+++ b/src/agentscope/formatter/_dashscope_formatter.py@@ -1,5 +1,5 @@# -*- coding: utf-8 -*--"""The dashscope formatter module."""+"""The DashScope formatter module (OpenAI-compatible format)."""import base64from typing import Any@@ -25,8 +25,8 @@ from ..message import (class _DashScopeFormatterBase(FormatterBase, ABC):- """Base class for DashScope formatters, providing shared data block- formatting logic."""+ """Base class for DashScope formatters (OpenAI-compatible format),+ providing shared data block formatting logic."""input_types: list[str] = Field(default_factory=lambda: [@@ -64,17 +64,28 @@ class _DashScopeFormatterBase(FormatterBase, ABC):def _format_dashscope_data_block(self,block: DataBlock,+ role: str = "user",) -> dict[str, Any] | None:- """Format a DataBlock into the required format for DashScope API.+ """Format a DataBlock into the OpenAI-compatible format for+ DashScope API.++ Supports:+ - Images: ``{"type": "image_url", "image_url": {"url": ...}}``+ - Videos: ``{"type": "video_url", "video_url": {"url": ...}}``+ - Audio: ``{"type": "input_audio", "input_audio": {...}}``Args:block (`DataBlock`):The DataBlock to format.+ role (`str`, defaults to ``"user"``):+ The role of the message containing this block. Audio blocks+ in assistant messages are skipped to avoid errors in+ subsequent model calls.Returns:`dict[str, Any] | None`:- A dictionary representing the formatted DataBlock for- DashScope API.+ A dictionary representing the formatted DataBlock, or ``None``+ if the media type is unsupported."""if not any(fnmatch(block.source.media_type, pattern)@@ -90,56 +101,142 @@ class _DashScopeFormatterBase(FormatterBase, ABC):main_type = block.source.media_type.split("/")[0]- if isinstance(block.source, URLSource):- url_str = str(block.source.url)+ if main_type == "image":+ return self._format_image_source(block.source)++ if main_type == "video":+ return self._format_video_source(block.source)++ if main_type == "audio":+ if role == "assistant":+ return None+ return self._format_audio_source(block.source)++ logger.warning(+ "Unsupported main media type %s for DashScope API. "+ "This block will be skipped.",+ main_type,+ )+ return None++ @staticmethod+ def _format_image_source(+ source: URLSource | Base64Source,+ ) -> dict[str, Any]:+ """Convert an image source to OpenAI-compatible ``image_url`` format.++ Local ``file://`` URLs are read from disk and converted to base64+ data URIs. Remote URLs are passed through unchanged.+ """+ if isinstance(source, Base64Source):+ url = f"data:{source.media_type};base64,{source.data}"+ elif isinstance(source, URLSource):+ url_str = str(source.url)if url_str.startswith("file://"):- # Local file — read and encode as base64 data URIlocal_path = url_str.removeprefix("file://")with open(local_path, "rb") as f:encoded = base64.b64encode(f.read()).decode("utf-8")- return {- main_type: f"data:{block.source.media_type};"- f"base64,{encoded}",- }- return {main_type: url_str}+ url = f"data:{source.media_type};base64,{encoded}"+ else:+ url = url_str+ else:+ raise ValueError(f"Unsupported image source type: {type(source)}")- if isinstance(block.source, Base64Source):- return {- main_type: f"data:{block.source.media_type};"- f"base64,{block.source.data}",- }+ return {+ "type": "image_url",+ "image_url": {"url": url},+ }- return None+ @staticmethod+ def _format_video_source(+ source: URLSource | Base64Source,+ ) -> dict[str, Any]:+ """Convert a video source to DashScope's ``video_url`` format+ (OpenAI-compatible extension).+ Local ``file://`` URLs are read from disk and converted to base64+ data URIs. Remote URLs are passed through unchanged.+ """+ if isinstance(source, Base64Source):+ url = f"data:{source.media_type};base64,{source.data}"+ elif isinstance(source, URLSource):+ url_str = str(source.url)+ if url_str.startswith("file://"):+ local_path = url_str.removeprefix("file://")+ with open(local_path, "rb") as f:+ encoded = base64.b64encode(f.read()).decode("utf-8")+ url = f"data:{source.media_type};base64,{encoded}"+ else:+ url = url_str+ else:+ raise ValueError(f"Unsupported video source type: {type(source)}")-class DashScopeChatFormatter(_DashScopeFormatterBase):- """The DashScope formatter class for chatbot scenario, where only a user- and an agent are involved. We use the `role` field to identify different- entities in the conversation.+ return {+ "type": "video_url",+ "video_url": {"url": url},+ }- .. warning::- Known Issues with DashScope API:+ @staticmethod+ def _format_audio_source(+ source: URLSource | Base64Source,+ ) -> dict[str, Any]:+ """Convert an audio source to DashScope ``input_audio`` format.- 1. **Missing content field**: When messages lack the 'content' field,- qwen-vl-max models will raise ``KeyError: 'content'``.+ DashScope's compatible API accepts URLs directly in the ``data``+ field (unlike standard OpenAI which requires base64). Local files+ are still read and base64-encoded.+ """+ if isinstance(source, Base64Source):+ fmt = source.media_type.split("/")[-1]+ return {+ "type": "input_audio",+ "input_audio": {+ "data": source.data,+ "format": fmt,+ },+ }- 2. **None content value**: When content is ``None``, qwen-vl-max models- will raise ``TypeError: 'NoneType' object is not iterable``.+ if isinstance(source, URLSource):+ url_str = str(source.url)+ fmt = source.media_type.split("/")[-1]+ if url_str.startswith("file://"):+ local_path = url_str.removeprefix("file://")+ with open(local_path, "rb") as f:+ data = base64.b64encode(f.read()).decode("utf-8")+ return {+ "type": "input_audio",+ "input_audio": {+ "data": data,+ "format": fmt,+ },+ }+ else:+ return {+ "type": "input_audio",+ "input_audio": {+ "data": url_str,+ "format": fmt,+ },+ }- 3. **Empty text in content**: When content contains- ``[{"text": None}]``, qwen3-max may repeatedly invoke tools- multiple times. Note that when qwen3-max initiates tool calls,- the returned message contains ``"content": ""``.+ raise ValueError(f"Unsupported audio source type: {type(source)}")+++class DashScopeChatFormatter(_DashScopeFormatterBase):+ """The DashScope formatter class for chatbot scenario (OpenAI-compatible+ format), where only a user and an agent are involved. We use the ``role``+ field to identify different entities in the conversation.- To avoid these issues, this formatter assigns content as an empty- list ``[]`` for messages without valid content blocks.+ This formatter outputs messages in the OpenAI Chat Completions format,+ with DashScope-specific extensions for video (``video_url``) and+ thinking (``reasoning_content``)."""async def format(self,msgs: list[Msg],) -> list[dict[str, Any]]:- """Format message objects into DashScope API format.+ """Format message objects into DashScope OpenAI-compatible format.Args:msgs (`list[Msg]`):@@ -161,29 +258,25 @@ class DashScopeChatFormatter(_DashScopeFormatterBase):for block in msg.get_content_blocks():if isinstance(block, TextBlock):- content_blocks.append({"text": block.text})+ content_blocks.append({"type": "text", "text": block.text})elif isinstance(block, DataBlock):- formatted_block = self._format_dashscope_data_block(block)+ formatted_block = self._format_dashscope_data_block(+ block,+ role=msg.role,+ )if formatted_block:content_blocks.append(formatted_block)elif isinstance(block, HintBlock):- # Insert a new user message with the hint content right- # after the current message, and go on processing the- # rest of the blocks in the current messageif content_blocks or tool_calls:- formatted_msgs.append(- {- "role": "user",- "content": content_blocks,- "tool_calls": tool_calls- if tool_calls- else None,- },- )- # Refresh content_blocks and tool_calls for the last- # blocks+ msg_openai: dict[str, Any] = {+ "role": msg.role,+ "content": content_blocks or None,+ }+ if tool_calls:+ msg_openai["tool_calls"] = tool_calls+ formatted_msgs.append(msg_openai)content_blocks = []tool_calls = []@@ -209,7 +302,6 @@ class DashScopeChatFormatter(_DashScopeFormatterBase):multimodal_data,) = self.convert_tool_result_to_string(block.output)- # First add the tool result message in DashScope API formatformatted_msgs.append({"role": "tool",@@ -219,8 +311,6 @@ class DashScopeChatFormatter(_DashScopeFormatterBase):},)- # If we have multimodal data that needs to be prompted to- # the out-tool-result messageif multimodal_data:msgs.insert(i + 1,@@ -236,9 +326,9 @@ class DashScopeChatFormatter(_DashScopeFormatterBase):type(block),)- msg_dashscope = {+ msg_dashscope: dict[str, Any] = {"role": msg.role,- "content": content_blocks,+ "content": content_blocks or None,}if tool_calls:@@ -250,50 +340,26 @@ class DashScopeChatFormatter(_DashScopeFormatterBase):if msg_dashscope["content"] or msg_dashscope.get("tool_calls"):formatted_msgs.append(msg_dashscope)- # Move to next messagei += 1- # Merge adjacent text blocks into one to avoid API issues.- # Tool-result messages have content as a plain string (not a list),- # so skip merging for those.- cleaned_msgs: list = []- for formatted_msg in formatted_msgs:- if not isinstance(formatted_msg.get("content"), list):- cleaned_msgs.append(formatted_msg)- continue- new_content = []- for block in formatted_msg["content"]:- if (- block.get("text")- and new_content- and new_content[-1].get("text")- ):- new_content[-1]["text"] += "\n" + block["text"]- else:- new_content.append(block)- formatted_msg["content"] = new_content- cleaned_msgs.append(formatted_msg)-- return cleaned_msgs+ return formatted_msgsclass DashScopeMultiAgentFormatter(_DashScopeFormatterBase):- """DashScope formatter for multi-agent conversations, where more than- a user and an agent are involved.+ """DashScope formatter for multi-agent conversations (OpenAI-compatible+ format), where more than a user and an agent are involved... note:: This formatter will combine previous messages (except toolcalls/results) into a history section in the first system message withthe conversation history prompt... note:: For tool calls/results, they will be presented as separate- messages as required by the DashScope API. Therefore, the tool calls/- results messages are expected to be placed at the end of the input- messages.+ messages as required by the API. Therefore, the tool calls/results+ messages are expected to be placed at the end of the input messages... tip:: Telling the assistant's name in the system prompt is veryimportant in multi-agent conversations. So that LLM can know who itis playing as.-"""conversation_history_prompt: str = Field(@@ -307,7 +373,7 @@ class DashScopeMultiAgentFormatter(_DashScopeFormatterBase):async def format(self, msgs: list[Msg]) -> list[dict]:"""Format input messages into the structure required by the DashScope- API.+ OpenAI-compatible API.To support multi-agent conversations, this formatter processes messagesas follows:@@ -315,8 +381,8 @@ class DashScopeMultiAgentFormatter(_DashScopeFormatterBase):- Prepends an instruction before the first conversation historysection.- Combines conversation turns into a history section, where each entry- is formatted as `{name}: {content}`.- - Wraps the conversation history with `<history>` and `</history>`+ is formatted as ``{name}: {content}``.+ - Wraps the conversation history with ``<history>`` and ``</history>``tags.Returns:@@ -355,16 +421,7 @@ class DashScopeMultiAgentFormatter(_DashScopeFormatterBase):msgs: list[Msg],) -> list[dict[str, Any]]:"""Given a sequence of tool call/result messages, format them into- the required format for the DashScope API.-- Args:- msgs (`list[Msg]`):- The list of messages containing tool calls/results to format.-- Returns:- `list[dict[str, Any]]`:- A list of dictionaries formatted for the DashScope API.- """+ the required format for the DashScope API."""return await DashScopeChatFormatter(input_types=self.input_types,).format(msgs)@@ -375,50 +432,34 @@ class DashScopeMultiAgentFormatter(_DashScopeFormatterBase):is_first: bool = True,) -> list[dict[str, Any]]:"""Given a sequence of messages without tool calls/results, format- them into a user message with conversation history tags. For the- first agent message, it will include the conversation history prompt.-- Args:- msgs (`list[Msg]`):- A list of Msg objects to be formatted.- is_first (`bool`, defaults to `True`):- Whether this is the first agent message in the conversation.- If `True`, the conversation history prompt will be included.-- Returns:- `list[dict[str, Any]]`:- A list of dictionaries formatted for the DashScope API.- """+ them into a user message with conversation history tags."""if is_first:conversation_history_prompt = self.conversation_history_promptelse:conversation_history_prompt = ""- # Format into required DashScope formatformatted_msgs: list[dict] = []-- # Collect the multimodal files- conversation_blocks = []+ conversation_blocks: list = []accumulated_text = []+ media_blocks: list[dict] = []+for msg in msgs:for block in msg.get_content_blocks():if isinstance(block, TextBlock):accumulated_text.append(f"{msg.name}: {block.text}")elif isinstance(block, DataBlock):- # Handle the accumulated text as a single block- if accumulated_text:- conversation_blocks.append(- {"text": "\n".join(accumulated_text)},- )- accumulated_text.clear()-- formatted_block = self._format_dashscope_data_block(block)+ formatted_block = self._format_dashscope_data_block(+ block,+ role=msg.role,+ )if formatted_block is not None:- conversation_blocks.append(formatted_block)+ media_blocks.append(formatted_block)if accumulated_text:- conversation_blocks.append({"text": "\n".join(accumulated_text)})+ conversation_blocks.append(+ {"text": "\n".join(accumulated_text)},+ )if conversation_blocks:if conversation_blocks[0].get("text"):@@ -427,36 +468,30 @@ class DashScopeMultiAgentFormatter(_DashScopeFormatterBase):+ "<history>\n"+ conversation_blocks[0]["text"])-else:conversation_blocks.insert(0,- {- "text": conversation_history_prompt + "<history>\n",- },+ {"text": conversation_history_prompt + "<history>\n"},)- conversation_blocks.append({"text": "</history>"})-- # Merge the adjacent text blocks into one text blocks to avoid API- # issues- new_content = []- for block in conversation_blocks:- if (- block.get("text")- and new_content- and new_content[-1].get("text")- ):- new_content[-1]["text"] += "\n" + block["text"]- else:- new_content.append(block)+ if conversation_blocks[-1].get("text"):+ conversation_blocks[-1]["text"] += "\n</history>"+ else:+ conversation_blocks.append({"text": "</history>"})- formatted_msgs.append(- {- "role": "user",- "content": new_content,- },+ conversation_blocks_text = "\n".join(+ b.get("text", "") for b in conversation_blocks+ )++ content_list: list[dict[str, Any]] = []+ if conversation_blocks_text:+ content_list.append(+ {"type": "text", "text": conversation_blocks_text},)+ content_list.extend(media_blocks)++ if content_list:+ formatted_msgs.append({"role": "user", "content": content_list})return formatted_msgsdiff --git a/src/agentscope/model/_dashscope/_model.py b/src/agentscope/model/_dashscope/_model.pyindex 6705f1f31..5ffdfe520 100644--- a/src/agentscope/model/_dashscope/_model.py+++ b/src/agentscope/model/_dashscope/_model.py@@ -1,36 +1,44 @@# -*- coding: utf-8 -*--"""The DashScope chat model class."""-import uuid+"""The DashScope chat model class (OpenAI-compatible implementation)."""import warnings+from collections import OrderedDictfrom datetime import datetime-from http import HTTPStatus-from typing import Any, AsyncGenerator, Generator, List, Literal, TYPE_CHECKING+from typing import Any, AsyncGenerator, List, Literal, TYPE_CHECKINGfrom pydantic import BaseModel, Field-from aioitertools import iter as giterfrom .._base import ChatModelBase, _TOOL_CHOICE_LITERAL_MODESfrom .._model_response import ChatResponsefrom .._model_usage import ChatUsagefrom ...credential import DashScopeCredentialfrom ...formatter import FormatterBase, DashScopeChatFormatter-from ...message import Msg, TextBlock, ThinkingBlock, ToolCallBlock+from ...message import (+ Msg,+ TextBlock,+ ThinkingBlock,+ ToolCallBlock,+ DataBlock,+ Base64Source,+)from ...tool import ToolChoicefrom ...tracing import trace_llm-if TYPE_CHECKING:- from dashscope.api_entities.dashscope_response import GenerationResponse- from dashscope.api_entities.dashscope_response import (- MultiModalConversationResponse,- )+ from openai.types.chat import ChatCompletion+ from openai import AsyncStreamelse:- GenerationResponse = Any- MultiModalConversationResponse = Any+ ChatCompletion = Any+ AsyncStream = Anyclass DashScopeChatModel(ChatModelBase):- """The DashScope chat model."""+ """The DashScope chat model (OpenAI-compatible implementation).++ This implementation uses the OpenAI Python SDK to call DashScope's+ OpenAI-compatible endpoint (``compatible-mode/v1``), which supports+ both text-only and multimodal (image/video) inputs through the same+ unified API.+ """class Parameters(BaseModel):"""The parameters for DashScope LLM API."""@@ -96,7 +104,6 @@ class DashScopeChatModel(ChatModelBase):stream: bool = True,max_retries: int = 3,context_size: int = 131072,- multimodality: bool | None = None,formatter: FormatterBase | None = None,) -> None:"""Initialize the DashScope chat model.@@ -116,11 +123,6 @@ class DashScopeChatModel(ChatModelBase):The maximum number of retries for the DashScope API.context_size (`int`, defaults to `131072`):The model context size used for context compression.- multimodality (`bool | None`, defaults to `None`):- Whether to call the MultiModalConversation API. ``True``- forces multimodal mode, ``False`` forces text-only mode,- and ``None`` auto-detects from the model name- (e.g. ``qvq`` or ``-vl`` suffix).formatter (`FormatterBase | None`, defaults to `None`):The formatter that converts ``Msg`` objects to the formatrequired by the DashScope API. When ``None``, a@@ -134,7 +136,6 @@ class DashScopeChatModel(ChatModelBase):)self.credential = credentialself.parameters = parameters or self.Parameters()- self.multimodality = multimodalityself.formatter = formatter or DashScopeChatFormatter()@trace_llm@@ -146,367 +147,305 @@ class DashScopeChatModel(ChatModelBase):tool_choice: ToolChoice | None = None,**kwargs: Any,) -> ChatResponse | AsyncGenerator[ChatResponse, None]:- """Get the response from the dashscope- Generation/MultimodalConversation API by the given arguments.-- .. note:: We unify the DashScope generation and multimodal conversation- APIs into one method, since they support similar arguments and share- the same functionality.+ """Call the DashScope chat completions API via OpenAI-compatible+ endpoint.Args:- messages (`list[Msg]`):- The Msg object that will be formatted and sent to the LLM API- as the conversation messages.+ model_name (`str`):+ The model name to use for this call.+ messages (`list`):+ The Msg objects that will be formatted and sent to the API.tools (`list[dict] | None`, default `None`):The tools JSON schemas that the model can use.tool_choice (`ToolChoice | None`, default `None`):Controls which (if any) tool is called by the model.- Note: DashScope API only supports "auto" and "none", so- "required" will be converted to "auto".- For more details, please refer to- https://help.aliyun.com/zh/model-studio/qwen-function-calling**kwargs (`Any`):The keyword arguments for DashScope chat completions API,- e.g. `temperature`, `max_tokens`, `top_p`, etc. Please- refer to `DashScope documentation- <https://help.aliyun.com/zh/dashscope/developer-reference/api-details>`_- for more detailed arguments.+ e.g. ``temperature``, ``max_tokens``, ``top_p``, etc."""- import dashscope+ import openai++ client = openai.AsyncClient(+ api_key=self.credential.api_key.get_secret_value(),+ base_url=self.credential.base_url,+ )- # 1. Format the messages to the format required by DashScope API- formatted_msg = await self.formatter.format(messages)+ formatted_messages = await self.formatter.format(messages)- # 2. Prepare the generation keyword arguments- kwargs = {- "messages": formatted_msg,+ request_kwargs: dict[str, Any] = {"model": model_name,+ "messages": formatted_messages,"stream": self.stream,- "result_format": "message",- # In agentscope, the `incremental_output` must be `True` when- # `self.stream` is True- "incremental_output": self.stream,- **kwargs,}+ if self.parameters.max_tokens is not None:+ request_kwargs["max_tokens"] = self.parameters.max_tokens++ if self.parameters.temperature is not None:+ request_kwargs["temperature"] = self.parameters.temperature++ if self.parameters.top_p is not None:+ request_kwargs["top_p"] = self.parameters.top_p++ request_kwargs.update(kwargs)+fmt_tools, fmt_tool_choice = self._format_tools(tools, tool_choice)if fmt_tools is not None:- kwargs["tools"] = fmt_tools+ request_kwargs["tools"] = fmt_tools+ if not self.parameters.parallel_tool_calls:+ request_kwargs["parallel_tool_calls"] = Falseif fmt_tool_choice is not None:- kwargs["tool_choice"] = fmt_tool_choice+ request_kwargs["tool_choice"] = fmt_tool_choice- # thinking related options — map to DashScope API parameter names- kwargs["enable_thinking"] = self.parameters.thinking_enable+ extra_body: dict[str, Any] = {}+ if self.parameters.thinking_enable is not None:+ extra_body["enable_thinking"] = self.parameters.thinking_enableif self.parameters.thinking_budget is not None:- kwargs["thinking_budget"] = self.parameters.thinking_budget+ extra_body["thinking_budget"] = self.parameters.thinking_budget+ if self.parameters.top_k is not None:+ extra_body["top_k"] = self.parameters.top_k- # 3. Call the API and parse the response- start_datetime = datetime.now()- # Use MultiModalConversation API if multimodality is True, or if the- # model is multimodal based on the model name.- api_key = self.credential.api_key.get_secret_value()- if self.multimodality or (- self.multimodality is None- and (- "qvq" in model_name- or "-vl" in model_name- or "-omni" in model_name- )- ):- response = await dashscope.AioMultiModalConversation.call(- api_key=api_key,- **kwargs,- )+ if extra_body:+ request_kwargs.setdefault("extra_body", {})+ request_kwargs["extra_body"].update(extra_body)- else:- response = await dashscope.aigc.generation.AioGeneration.call(- api_key=api_key,- **kwargs,- )+ if self.stream:+ request_kwargs["stream_options"] = {"include_usage": True}++ start_datetime = datetime.now()+ response = await client.chat.completions.create(**request_kwargs)if self.stream:- return self._parse_dashscope_stream_response(+ audio_cfg = request_kwargs.get("audio")+ audio_fmt = (+ audio_cfg.get("format", "wav")+ if isinstance(audio_cfg, dict)+ else "wav"+ )+ return self._parse_stream_response(start_datetime,response,+ audio_fmt,)- parsed_response = await self._parse_dashscope_generation_response(- start_datetime,- response,- )-- return parsed_response+ return self._parse_completion_response(start_datetime, response)- async def _parse_dashscope_stream_response(+ async def _parse_stream_response(self,start_datetime: datetime,- response: AsyncGenerator[GenerationResponse, None]- | AsyncGenerator[MultiModalConversationResponse, None]- | Generator[MultiModalConversationResponse, None, None],- ) -> AsyncGenerator[ChatResponse, Any]:- """Given a DashScope streaming response generator, extract the content- blocks and usages from it and yield ChatResponse objects.+ response: AsyncStream,+ audio_format: str = "wav",+ ) -> AsyncGenerator[ChatResponse, None]:+ """Parse the DashScope streaming response (OpenAI-compatible format).Args:start_datetime (`datetime`):The start datetime of the response generation.- response (- `AsyncGenerator[GenerationResponse, None] | \- AsyncGenerator[MultiModalConversationResponse, None] | \- Generator[MultiModalConversationResponse, None, None]`- ):- DashScope streaming response (async) generator- (GenerationResponse or MultiModalConversationResponse).+ response (`AsyncStream`):+ The OpenAI-compatible async stream object.+ audio_format (`str`, defaults to ``"wav"``):+ The audio format requested (used to set the media type on+ the output ``DataBlock``).- Returns:- `AsyncGenerator[ChatResponse, Any]`:- An async generator that yields ChatResponse objects containing- the content blocks and usage information for each chunk in the- streaming response.+ Yields:+ `ChatResponse`:+ Incremental ``ChatResponse`` objects with ``is_last=False``+ followed by a final one with ``is_last=True``."""- # All delta should have the same block identifier- acc_text = TextBlock(text="")- acc_thinking = ThinkingBlock(thinking="")- acc_tool_calls = {}usage = Noneresponse_id: str | None = None+ acc_text = TextBlock(text="")+ acc_thinking = ThinkingBlock(thinking="")+ acc_tool_calls: OrderedDict = OrderedDict()+ acc_audio_data: str = ""++ async with response as stream:+ async for chunk in stream:+ if chunk.usage:+ u = chunk.usage+ ptd = getattr(u, "prompt_tokens_details", None)+ if ptd and hasattr(ptd, "cached_tokens"):+ cache_read = ptd.cached_tokens or 0+ else:+ cache_read = 0+ usage = ChatUsage(+ input_tokens=u.prompt_tokens or 0,+ output_tokens=u.completion_tokens or 0,+ time=(datetime.now() - start_datetime).total_seconds(),+ cache_input_tokens=cache_read,+ )- async for chunk in giter(response):- # The yield delta content blocks in the current chunk- delta_content: list = []-- if chunk.status_code != HTTPStatus.OK:- raise RuntimeError(- f"Failed to get response from DashScope API: {chunk}",- )-- # Capture response_id from the first chunk if not already set- response_id = response_id or chunk.request_id+ response_id = response_id or getattr(chunk, "id", None)- # The message field in the chunk- message = chunk.output.choices[0].message+ if not chunk.choices:+ continue- # Update reasoning content- if isinstance(message.get("reasoning_content"), str):- acc_thinking.thinking += message["reasoning_content"]- delta_content.append(- ThinkingBlock(- id=acc_thinking.id,- thinking=message["reasoning_content"],- ),- )+ choice = chunk.choices[0]+ delta = choice.delta- # Update text content- if isinstance(message.content, str):- acc_text.text += message.content- delta_content.append(- TextBlock(- id=acc_text.id,- text=message.content,- ),- )- elif isinstance(message.content, list):- delta_text = "".join(- [- _["text"]- for _ in message.content- if isinstance(_, dict)- and isinstance(_.get("text"), str)- ],+ delta_thinking = (+ getattr(delta, "reasoning_content", None) or "")+ delta_text = getattr(delta, "content", None) or ""++ # Collect audio output from Omni models (delta.audio.data)+ delta_audio = getattr(delta, "audio", None)+ if delta_audio is not None:+ if isinstance(delta_audio, dict):+ audio_chunk = delta_audio.get("data", "")+ else:+ audio_chunk = getattr(delta_audio, "data", "") or ""+ if audio_chunk:+ acc_audio_data += audio_chunk++ acc_thinking.thinking += delta_thinkingacc_text.text += delta_text- delta_content.append(- ThinkingBlock(id=acc_text.id, thinking=delta_text),- )- # Update tool calls- for tool_call in message.get("tool_calls") or []:- # Must be a dictionary- if not isinstance(tool_call, dict):- continue-- # Use the index to identify different tool calls- index = tool_call.get("index", 0)-- # Initialize accumulator for a new tool call index- acc_tool_calls.setdefault(index, {})-- # Avoid appending duplicate id- if "id" in tool_call and tool_call["id"] != acc_tool_calls[- index- ].get("id"):- acc_tool_calls[index]["id"] = tool_call["id"]-- # Arguments- if isinstance(tool_call.get("function"), dict):- func = tool_call["function"]-- # Function name- if "name" in func:- acc_tool_calls[index]["name"] = func["name"]-- # Function input arguments- if isinstance(func.get("arguments"), str):- acc_tool_calls[index]["arguments"] = (- acc_tool_calls[index].get("arguments", "")- + func["arguments"]- )-- # Append the tool call block for this chunk- delta_content.append(- ToolCallBlock(- id=acc_tool_calls[index].get(- "id",- uuid.uuid4().hex,+ for tool_call in getattr(delta, "tool_calls", None) or []:+ idx = tool_call.index+ if idx in acc_tool_calls:+ if tool_call.function and tool_call.function.arguments:+ acc_tool_calls[idx][+ "input"+ ] += tool_call.function.arguments+ else:+ acc_tool_calls[idx] = {+ "id": tool_call.id or "",+ "name": (+ tool_call.function.name+ if tool_call.function+ else ""),- name=func.get("name", ""),- input=func.get("arguments", ""),+ "input": (+ tool_call.function.arguments+ if tool_call.function+ else ""+ )+ or "",+ }++ delta_contents: List[+ TextBlock | ToolCallBlock | ThinkingBlock+ ] = []+ if delta_thinking:+ delta_contents.append(+ ThinkingBlock(+ id=acc_thinking.id,+ thinking=delta_thinking,),)+ if delta_text:+ delta_contents.append(+ TextBlock(id=acc_text.id, text=delta_text),+ )- # The chunk usage- if chunk.usage:- u = chunk.usage- ptd = u.get("prompt_tokens_details", None)- if isinstance(ptd, dict):- cache_creation = ptd.get("cache_creation_input_tokens", 0)- cache_read = ptd.get("cached_tokens", 0)- else:- cache_creation = 0- cache_read = u.get("cached_tokens", 0)- usage = ChatUsage(- input_tokens=u.input_tokens,- output_tokens=u.output_tokens,- time=(datetime.now() - start_datetime).total_seconds(),- cache_creation_input_tokens=cache_creation,- cache_input_tokens=cache_read,- )-- # Yield the chat response- yield ChatResponse(- id=response_id,- content=delta_content,- is_last=False,- usage=usage,- )-- # Yield a final complete response with the accumulated content and- # final usage. Thinking comes before text to preserve the natural- # generation order (reasoning first, then answer).- content = []+ if delta_contents:+ _kwargs: dict[str, Any] = {+ "content": delta_contents,+ "usage": usage,+ "is_last": False,+ }+ if response_id:+ _kwargs["id"] = response_id+ yield ChatResponse(**_kwargs)++ final_contents: List[+ TextBlock | ToolCallBlock | ThinkingBlock | DataBlock+ ] = []if acc_thinking.thinking:- content.append(acc_thinking)-+ final_contents.append(acc_thinking)if acc_text.text:- content.append(acc_text)-- if acc_tool_calls:- content.extend(+ final_contents.append(acc_text)+ for tc in acc_tool_calls.values():+ final_contents.append(ToolCallBlock(- id=tc.get("id", uuid.uuid4().hex),- name=tc.get("name", ""),- input=tc.get("arguments", "{}"),- )- for tc in acc_tool_calls.values()+ id=tc["id"],+ name=tc["name"],+ input=tc["input"],+ ),+ )+ if acc_audio_data:+ final_contents.append(+ DataBlock(+ source=Base64Source(+ data=acc_audio_data,+ media_type=f"audio/{audio_format}",+ ),+ ),)- yield ChatResponse(- id=response_id or uuid.uuid4().hex,- content=content,- is_last=True,- usage=usage,- )+ _final_kwargs: dict[str, Any] = {+ "content": final_contents,+ "usage": usage,+ "is_last": True,+ }+ if response_id:+ _final_kwargs["id"] = response_id+ yield ChatResponse(**_final_kwargs)- async def _parse_dashscope_generation_response(+ def _parse_completion_response(self,start_datetime: datetime,- response: GenerationResponse | MultiModalConversationResponse,+ response: ChatCompletion,) -> ChatResponse:- """Given a DashScope GenerationResponse object, extract the content- blocks and usages from it.+ """Parse the DashScope non-streaming response (OpenAI-compatible+ format).Args:start_datetime (`datetime`):The start datetime of the response generation.- response (- `GenerationResponse | MultiModalConversationResponse`- ):- Dashscope GenerationResponse | MultiModalConversationResponse- object to parse.+ response (`ChatCompletion`):+ The OpenAI-compatible chat completion object.Returns:`ChatResponse`:- A ChatResponse object containing the content blocks and usage.+ A single ``ChatResponse`` with ``is_last=True``."""- # Collect the content blocks from the response.- if response.status_code != HTTPStatus.OK:- raise RuntimeError(- f"Failed to get response from DashScope API: {response}",- )-- content_blocks: List[TextBlock | ToolCallBlock] = []+ content_blocks: List[TextBlock | ToolCallBlock | ThinkingBlock] = []- message = response.output.choices[0].message+ if response.choices:+ choice = response.choices[0]+ reasoning = getattr(choice.message, "reasoning_content", None)+ if isinstance(reasoning, str) and reasoning:+ content_blocks.append(ThinkingBlock(thinking=reasoning))- # The content and tool calls in the message- content, tool_calls = message.get("content"), message.get("tool_calls")-- if isinstance(content, str):- content_blocks.append(TextBlock(text=content))-- elif isinstance(content, list):- content_blocks.append(- TextBlock(- text="".join(- [- _["text"]- for _ in content- if isinstance(_, dict)- and isinstance(_.get("text"), str)- ],- ),- ),- )-- if isinstance(tool_calls, list):- for tool_call in tool_calls:- function = tool_call.get("function") or {}- name = function.get("name") or ""- input_ = function.get("arguments") or "{}"+ if choice.message.content:+ content_blocks.append(TextBlock(text=choice.message.content))+ for tool_call in choice.message.tool_calls or []:content_blocks.append(ToolCallBlock(- id=tool_call.get("id", uuid.uuid4().hex),- name=name,- input=input_,+ id=tool_call.id,+ name=tool_call.function.name,+ input=tool_call.function.arguments,),)- # Usage informationusage = Noneif response.usage:u = response.usage- ptd = u.get("prompt_tokens_details", None)- if isinstance(ptd, dict):- cache_creation = ptd.get("cache_creation_input_tokens", 0)- cache_read = ptd.get("cached_tokens", 0)+ ptd = getattr(u, "prompt_tokens_details", None)+ if ptd and hasattr(ptd, "cached_tokens"):+ cache_read = ptd.cached_tokens or 0else:- cache_creation = 0- cache_read = u.get("cached_tokens", 0)+ cache_read = 0usage = ChatUsage(- input_tokens=u.input_tokens,- output_tokens=u.output_tokens,+ input_tokens=u.prompt_tokens,+ output_tokens=u.completion_tokens,time=(datetime.now() - start_datetime).total_seconds(),- cache_creation_input_tokens=cache_creation,cache_input_tokens=cache_read,)- return ChatResponse(- id=getattr(response, "request_id", uuid.uuid4().hex),- content=content_blocks,- is_last=True,- usage=usage,- )+ resp_kwargs: dict[str, Any] = {+ "content": content_blocks,+ "is_last": True,+ "usage": usage,+ }+ response_id = getattr(response, "id", None)+ if response_id:+ resp_kwargs["id"] = response_id++ return ChatResponse(**resp_kwargs)def _format_tools(self,@@ -515,12 +454,11 @@ class DashScopeChatModel(ChatModelBase):) -> tuple[list[dict] | None, str | dict | None]:"""Validate and format tools and tool_choice for DashScope.- DashScope only supports "auto" and "none" modes; "required" is- converted to "auto" with a warning. When ``tool_choice.tools``- is specified the schemas list is filtered to only those tools.- When ``tool_choice.mode`` is a specific tool name (str) the- model is forced to call exactly that tool without needing to- filter the list, preserving prompt-cache efficiency.+ DashScope supports "auto", "none", and "required" modes in+ OpenAI-compatible format. When ``tool_choice.tools`` is specified+ the schemas list is filtered to only those tools. When+ ``tool_choice.mode`` is a specific tool name (str) the model is+ forced to call exactly that tool.Args:tools (`list[dict] | None`, optional):@@ -559,7 +497,6 @@ class DashScopeChatModel(ChatModelBase):mode = tool_choice.modeif mode not in _TOOL_CHOICE_LITERAL_MODES:- # mode is a specific tool name — force call itreturn fmt_tools, {"type": "function","function": {"name": mode},@@ -567,9 +504,10 @@ class DashScopeChatModel(ChatModelBase):if mode == "required":warnings.warn(- f"'{mode}' is not supported by DashScope API. "+ f"'{mode}' is not fully supported by DashScope API. ""It will be converted to 'auto'.",DeprecationWarning,+ stacklevel=2,)return fmt_tools, "auto"diff --git a/tests/formatter_dashscope_test.py b/tests/formatter_dashscope_test.pyindex 31235f139..870b79847 100644--- a/tests/formatter_dashscope_test.py+++ b/tests/formatter_dashscope_test.py@@ -46,7 +46,7 @@ class TestDashScopeFormatter(IsolatedAsyncioTestCase):media_type="audio/mpeg",)self.image_url = str(_img_src.url)- audio_url = str(_aud_src.url)+ self.audio_url = str(_aud_src.url)# --- Base64 fixtures ---self.image_b64 = "ZmFrZSBpbWFnZSBkYXRh"@@ -94,7 +94,7 @@ class TestDashScopeFormatter(IsolatedAsyncioTestCase):),DataBlock(source=URLSource(- url=audio_url,+ url=self.audio_url,media_type="audio/mpeg",),),@@ -152,44 +152,78 @@ class TestDashScopeFormatter(IsolatedAsyncioTestCase):]# ---------------------------------------------------------------- # Ground truth: DashScopeChatFormatter- # - System/user/assistant content is always a list of dicts.+ # Ground truth: DashScopeChatFormatter (OpenAI-compatible format)+ # - Content blocks use {"type": "text", "text": ...} format.+ # - Images use {"type": "image_url", "image_url": {"url": ...}}.+ # - Audio uses {"type": "input_audio", "input_audio": {...}}.# - Tool-result content is a plain string.# ---------------------------------------------------------------self.gt_chat = [{"role": "system",- "content": [{"text": "You're a helpful assistant."}],+ "content": [+ {"type": "text", "text": "You're a helpful assistant."},+ ],},{"role": "user","content": [- {"text": "What is the capital of France?"},- {"image": self.image_url},+ {+ "type": "text",+ "text": "What is the capital of France?",+ },+ {+ "type": "image_url",+ "image_url": {"url": self.image_url},+ },],},{"role": "assistant",- "content": [{"text": "The capital of France is Paris."}],+ "content": [+ {+ "type": "text",+ "text": "The capital of France is Paris.",+ },+ ],},{"role": "user","content": [- {"text": "What is the capital of Germany?"},- {"audio": audio_url},+ {+ "type": "text",+ "text": "What is the capital of Germany?",+ },+ {+ "type": "input_audio",+ "input_audio": {+ "data": self.audio_url,+ "format": "mpeg",+ },+ },],},{"role": "assistant",- "content": [{"text": "The capital of Germany is Berlin."}],+ "content": [+ {+ "type": "text",+ "text": "The capital of Germany is Berlin.",+ },+ ],},{"role": "user",- "content": [{"text": "What is the capital of Japan?"}],+ "content": [+ {+ "type": "text",+ "text": "What is the capital of Japan?",+ },+ ],},{"role": "assistant",- "content": [],+ "content": None,"tool_calls": [{"id": "call_1",@@ -209,7 +243,12 @@ class TestDashScopeFormatter(IsolatedAsyncioTestCase):},{"role": "assistant",- "content": [{"text": "The capital of Japan is Tokyo."}],+ "content": [+ {+ "type": "text",+ "text": "The capital of Japan is Tokyo.",+ },+ ],},]@@ -222,18 +261,11 @@ class TestDashScopeFormatter(IsolatedAsyncioTestCase):DashScopeMultiAgentFormatter().conversation_history_prompt)- # The trailing assistant message ("The capital of Japan is Tokyo.")- # is always converted to a user <history> message by the multiagent- # formatter. Whether it carries the full history_prompt prefix- # depends on whether an "agent_message" group appeared before it:- # * is_first=False (conversation preceded the tool sequence):- # only <history>...</history>- # * is_first=True (no prior agent_message, e.g. tools-only):- # full history_prompt + <history>...</history>_gt_trailing_asst_nonfirst = {"role": "user","content": [{+ "type": "text","text": ("<history>\n""assistant: The capital of Japan is Tokyo.\n"@@ -246,6 +278,7 @@ class TestDashScopeFormatter(IsolatedAsyncioTestCase):"role": "user","content": [{+ "type": "text","text": (_hist_prompt + "<history>\n""assistant: The capital of Japan is Tokyo.\n"@@ -257,7 +290,7 @@ class TestDashScopeFormatter(IsolatedAsyncioTestCase):self._gt_tool_call = {"role": "assistant",- "content": [],+ "content": None,"tool_calls": [{"id": "call_1",@@ -285,32 +318,32 @@ class TestDashScopeFormatter(IsolatedAsyncioTestCase):"role": "user","content": [{+ "type": "text","text": (_hist_prompt + "<history>\n"- "user: What is the capital of France?"- ),- },- {"image": self.image_url},- {- "text": (+ "user: What is the capital of France?\n""assistant: The capital of France is Paris.\n"- "user: What is the capital of Germany?"- ),- },- {"audio": audio_url},- {- "text": (+ "user: What is the capital of Germany?\n""assistant: The capital of Germany is Berlin.\n""user: What is the capital of Japan?\n""</history>"),},+ {+ "type": "image_url",+ "image_url": {"url": self.image_url},+ },+ {+ "type": "input_audio",+ "input_audio": {+ "data": self.audio_url,+ "format": "mpeg",+ },+ },],},self._gt_tool_call,self._gt_tool_result,- # Trailing assistant message: is_first=False because conversation- # preceded the tool sequence._gt_trailing_asst_nonfirst,]@@ -373,8 +406,11 @@ class TestDashScopeFormatter(IsolatedAsyncioTestCase):{"role": "user","content": [- {"text": "What's in this image?"},- {"image": self.image_data_uri},+ {"type": "text", "text": "What's in this image?"},+ {+ "type": "image_url",+ "image_url": {"url": self.image_data_uri},+ },],},],@@ -440,7 +476,10 @@ class TestDashScopeFormatter(IsolatedAsyncioTestCase):# The promoted multimodal user messageself.assertEqual(res[2]["role"], "user")image_blocks = [- b for b in res[2]["content"] if b.get("image") == self.image_url+ b+ for b in res[2]["content"]+ if b.get("type") == "image_url"+ and b.get("image_url", {}).get("url") == self.image_url]self.assertEqual(len(image_blocks), 1)diff --git a/tests/model_dashscope_test.py b/tests/model_dashscope_test.pyindex f126ce75f..a74a6e11d 100644--- a/tests/model_dashscope_test.py+++ b/tests/model_dashscope_test.py@@ -7,7 +7,6 @@ Formatter tests have been moved to tests/formatter_dashscope_test.py.import jsonfrom typing import Anyfrom datetime import datetime-from http import HTTPStatusimport unittestfrom unittest import IsolatedAsyncioTestCasefrom unittest.mock import MagicMock@@ -28,7 +27,6 @@ def _make_model() -> Any:credential=DashScopeCredential(api_key="test"),model="qwen3-max",stream=False,- multimodality=True,max_retries=3,context_size=int(1500),parameters=DashScopeChatModel.Parameters(@@ -56,29 +54,32 @@ class TestDashScopeModelParsing(IsolatedAsyncioTestCase):self,content: Any = None,tool_calls: Any = None,+ reasoning_content: Any = None,) -> "MagicMock":- """Build a minimal DashScope GenerationResponse mock."""- message = {}- if content is not None:- message["content"] = content- if tool_calls is not None:- message["tool_calls"] = tool_calls+ """Build a minimal OpenAI ChatCompletion-style mock."""+ message = MagicMock()+ message.content = content+ message.tool_calls = tool_calls+ message.reasoning_content = reasoning_content- msg_mock = MagicMock()- msg_mock.get = lambda key, default=None: message.get(key, default)+ choice = MagicMock()+ choice.message = message++ usage = MagicMock()+ usage.prompt_tokens = 10+ usage.completion_tokens = 5+ usage.prompt_tokens_details = Noneresp = MagicMock()- resp.status_code = HTTPStatus.OK- resp.output.choices[0].message = msg_mock- resp.request_id = "req-1"- resp.usage.input_tokens = 10- resp.usage.output_tokens = 5+ resp.id = "req-1"+ resp.choices = [choice]+ resp.usage = usagereturn respasync def test_parse_text_response(self) -> None:"""Parsing a text response creates a TextBlock."""resp = self._mock_response(content="Hello!")- result = await self.model._parse_dashscope_generation_response(+ result = self.model._parse_completion_response(self.start,resp,)@@ -88,17 +89,13 @@ class TestDashScopeModelParsing(IsolatedAsyncioTestCase):async def test_parse_tool_call_response(self) -> None:"""Parsing a tool-call response creates a ToolCallBlock."""- tool_calls = [- {- "id": "call-1",- "function": {- "name": "get_weather",- "arguments": '{"city":"Beijing"}',- },- },- ]- resp = self._mock_response(tool_calls=tool_calls)- result = await self.model._parse_dashscope_generation_response(+ tc_mock = MagicMock()+ tc_mock.id = "call-1"+ tc_mock.function.name = "get_weather"+ tc_mock.function.arguments = '{"city":"Beijing"}'++ resp = self._mock_response(tool_calls=[tc_mock])+ result = self.model._parse_completion_response(self.start,resp,)@@ -108,16 +105,6 @@ class TestDashScopeModelParsing(IsolatedAsyncioTestCase):self.assertEqual(tcs[0].name, "get_weather")self.assertEqual(json.loads(tcs[0].input)["city"], "Beijing")- async def test_parse_response_with_status_error(self) -> None:- """Non-OK status raises RuntimeError."""- resp = self._mock_response(content="text")- resp.status_code = HTTPStatus.BAD_REQUEST- with self.assertRaises(RuntimeError):- await self.model._parse_dashscope_generation_response(- self.start,- resp,- )-# ---------------------------------------------------------------------------# Shared _format_tools fixtures