fix(model): Fix reasoning output bug (#2393)

Fix reasoning output bug

configs/dbgpt-local-glm.toml

@@ -34,4 +34,4 @@ provider = "hf"
 # If not provided, the model will be downloaded from the Hugging Face model hub
 # uncomment the following line to specify the model path in the local file system
 # path = "the-model-path-in-the-local-file-system"
-path = "models/BAAI/glm-4-9b-chat-hf"
+path = "models/BAAI/bge-large-zh-v1.5"

docs/docs/quickstart.md

@@ -285,8 +285,23 @@ uv run dbgpt start webserver --config configs/dbgpt-local-vllm.toml
 ```
 
   </TabItem>
-      <TabItem value="llama_cpp" label="LLAMA_CPP(local)">
+  <TabItem value="llama_cpp" label="LLAMA_CPP(local)">
 
+If you has a Nvidia GPU, you can enable the CUDA support by setting the environment variable `CMAKE_ARGS="-DGGML_CUDA=ON"`.
+
+```bash
+# Use uv to install dependencies needed for llama-cpp
+# Install core dependencies and select desired extensions
+CMAKE_ARGS="-DGGML_CUDA=ON" uv sync --all-packages \
+--extra "base" \
+--extra "llama_cpp" \
+--extra "rag" \
+--extra "storage_chromadb" \
+--extra "quant_bnb" \
+--extra "dbgpts"
+```
+
+Otherwise, run the following command to install dependencies without CUDA support.
 ```bash
 # Use uv to install dependencies needed for llama-cpp
 # Install core dependencies and select desired extensions

packages/dbgpt-accelerator/dbgpt-acc-auto/README.md

@@ -0,0 +1,4 @@
+# DB-GPT Accelerator Module
+
+Building across multiple platforms and hardware is complex, and the DB-GPT Accelerator aims to provide compatibility handling for this, offering as consistent an interface as possible for other core models.
+

packages/dbgpt-accelerator/dbgpt-acc-auto/pyproject.toml

@@ -1,5 +1,5 @@
 [project]
-name = "dbgpt-accelerator"
+name = "dbgpt-acc-auto"
 version = "0.7.0"
 description = "Add your description here"
 authors = [
@@ -16,21 +16,6 @@ Documentation = "http://docs.dbgpt.cn/docs/overview"
 Repository = "https://github.com/eosphoros-ai/DB-GPT.git"
 Issues = "https://github.com/eosphoros-ai/DB-GPT/issues"
 
-[build-system]
-requires = ["hatchling"]
-build-backend = "hatchling.build"
-
-[tool.hatch.build.targets.wheel]
-packages = ["src/dbgpt_accelerator"]
-exclude = [
-    "src/dbgpt_accelerator/**/tests",
-    "src/dbgpt_accelerator/**/tests/*",
-    "src/dbgpt_accelerator/tests",
-    "src/dbgpt_accelerator/tests/*",
-    "src/dbgpt_accelerator/**/examples",
-    "src/dbgpt_accelerator/**/examples/*"
-]
-
 [project.optional-dependencies]
 # Auto install dependencies
 auto = [
@@ -76,10 +61,10 @@ vllm = [
     # Just support GPU version on Linux
     "vllm>=0.7.0; sys_platform == 'linux'",
 ]
-#vllm_pascal = [
+# vllm_pascal = [
 #    # https://github.com/sasha0552/pascal-pkgs-ci
 #    "vllm-pascal==0.7.2; sys_platform == 'linux'"
-#]
+# ]
 quant_bnb = [
     "bitsandbytes>=0.39.0; sys_platform == 'win32' or sys_platform == 'linux'",
     "accelerate"
@@ -103,6 +88,10 @@ quant_gptq = [
     "optimum",
     "auto-gptq",
 ]
+flash_attn = [
+    # "torch>=2.2.1",
+    "dbgpt-acc-flash-attn"
+]
 
 [dependency-groups]
 auto = [

packages/dbgpt-accelerator/dbgpt-acc-flash-attn/README.md

@@ -0,0 +1,3 @@
+# DB-GPT-Accelerator for Flash Attention
+
+Wrapper for the Flash Attention module in the DB-GPT-Accelerator.

packages/dbgpt-accelerator/dbgpt-acc-flash-attn/pyproject.toml

@@ -0,0 +1,24 @@
+# Install the flash-attn package for uv
+# https://github.com/astral-sh/uv/issues/2252#issuecomment-2624150395
+[project]
+name = "dbgpt-acc-flash-attn"
+version = "0.1.0"
+description = "Add your description here"
+readme = "README.md"
+requires-python = ">=3.10"
+dependencies = []
+
+[dependency-groups]
+build = [
+    "setuptools>=75.8.0",
+]
+direct = [
+    "torch>=2.2.1",
+]
+main = [
+    "flash-attn>=2.5.8",
+]
+
+[tool.uv]
+default-groups = ["build", "direct", "main"]
+no-build-isolation-package = ["flash-attn"]

packages/dbgpt-app/pyproject.toml

@@ -10,7 +10,7 @@ readme = "README.md"
 requires-python = ">= 3.10"
 
 dependencies = [
-    "dbgpt-accelerator",
+    "dbgpt-acc-auto",
     "dbgpt",
     "dbgpt-ext",
     "dbgpt-serve",

packages/dbgpt-core/src/dbgpt/core/interface/parameter.py

@@ -102,6 +102,15 @@ class LLMDeployModelParameters(BaseDeployModelParameters, RegisterParameters):
             )
         },
     )
+    reasoning_model: Optional[bool] = field(
+        default=None,
+        metadata={
+            "help": _(
+                "Whether the model is a reasoning model. If None, it is "
+                "automatically determined from model."
+            )
+        },
+    )
 
     @property
     def real_provider_model_name(self) -> str:
@@ -202,8 +211,10 @@ class BitsandbytesQuantization(BaseHFQuantization):
         real_cls = cls
         if load_in_8bits:
             real_cls = BitsandbytesQuantization8bits
+            data["type"] = BitsandbytesQuantization8bits.__type__
         if load_in_4bits:
             real_cls = BitsandbytesQuantization4bits
+            data["type"] = BitsandbytesQuantization4bits.__type__
         real_data = prepare_data_func(real_cls, data)
         return real_cls(**real_data)
 

packages/dbgpt-core/src/dbgpt/model/adapter/base.py

@@ -251,6 +251,27 @@ class LLMModelAdapter(ABC):
         """Load the model and tokenizer according to the given parameters"""
         raise NotImplementedError
 
+    def is_reasoning_model(
+        self,
+        deploy_model_params: LLMDeployModelParameters,
+        lower_model_name_or_path: Optional[str] = None,
+    ) -> bool:
+        """Whether the model is a reasoning model"""
+        if (
+            deploy_model_params.reasoning_model is not None
+            and deploy_model_params.reasoning_model
+        ):
+            return True
+        return (
+            lower_model_name_or_path
+            and "deepseek" in lower_model_name_or_path
+            and (
+                "r1" in lower_model_name_or_path
+                or "reasoning" in lower_model_name_or_path
+                or "reasoner" in lower_model_name_or_path
+            )
+        )
+
     def support_async(self) -> bool:
         """Whether the loaded model supports asynchronous calls"""
         return False

packages/dbgpt-core/src/dbgpt/model/adapter/hf_adapter.py

@@ -88,6 +88,15 @@ class HFLLMDeployModelParameters(LLMDeployModelParameters):
             "valid_values": ["auto", "float16", "bfloat16", "float", "float32"],
         },
     )
+    attn_implementation: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": _(
+                "The attention implementation, only valid in multi-GPU configuration"
+            ),
+            "valid_values": ["flash_attention_2"],
+        },
+    )
 
     @property
     def real_model_path(self) -> Optional[str]:

packages/dbgpt-core/src/dbgpt/model/adapter/llama_cpp_adapter.py

@@ -301,15 +301,15 @@ class LlamaServerParameters(LLMDeployModelParameters):
                 config_dict[fd.name] = curr_config[fd.name]
 
         if (
-            "device" in config_dict
-            and config_dict["device"] == "cuda"
+            self.real_device
+            and self.real_device == "cuda"
             and ("n_gpu_layers" not in config_dict or not config_dict["n_gpu_layers"])
         ):
             # Set n_gpu_layers to a large number to use all layers
             logger.info("Set n_gpu_layers to a large number to use all layers")
             config_dict["n_gpu_layers"] = 1000000000
         config_dict["model_alias"] = self.name
-        config_dict["model_file"] = self.path
+        config_dict["model_file"] = self._resolve_root_path(self.path)
         model_file = config_dict.get("model_file")
         model_url = config_dict.get("model_url")
         model_hf_repo = config_dict.get("model_hf_repo")

packages/dbgpt-core/src/dbgpt/model/adapter/loader.py

@@ -143,6 +143,8 @@ def huggingface_loader(
     if "device_map" in kwargs and "low_cpu_mem_usage" not in kwargs:
         # Must set low_cpu_mem_usage to True when device_map is set
         kwargs["low_cpu_mem_usage"] = True
+    if model_params.attn_implementation:
+        kwargs["attn_implementation"] = model_params.attn_implementation
 
     model, tokenizer = _hf_try_load_default_quantization_model(
         model_path, llm_adapter, device, num_gpus, model_params, kwargs

packages/dbgpt-core/src/dbgpt/model/adapter/vllm_adapter.py

@@ -62,6 +62,9 @@ class VLLMDeployModelParameters(LLMDeployModelParameters):
         model = data.get("path", None)
         if not model:
             model = data.get("name", None)
+        else:
+            # Path is specified, so we use it as the model
+            model = self._resolve_root_path(model)
         if not model:
             raise ValueError(
                 "Model is required, please specify the model path or name."

packages/dbgpt-core/src/dbgpt/model/cluster/worker/default_worker.py

@@ -421,6 +421,10 @@ class DefaultModelWorker(ModelWorker):
             span_params["messages"] = list(
                 map(lambda m: m.dict(), span_params["messages"])
             )
+        if self.llm_adapter.is_reasoning_model(
+            self._model_params, self.model_name.lower()
+        ):
+            params["is_reasoning_model"] = True
 
         metadata = {
             "is_async_func": self.support_async(),

packages/dbgpt-core/src/dbgpt/model/cluster/worker/manager.py

@@ -613,39 +613,41 @@ class LocalWorkerManager(WorkerManager):
         )
 
     async def _start_all_worker(
-        self, apply_req: WorkerApplyRequest
+        self, apply_req: WorkerApplyRequest, parallel_num: int = 1
     ) -> WorkerApplyOutput:
         from httpx import TimeoutException, TransportError
 
         # TODO avoid start twice
         start_time = time.time()
         logger.info(f"Begin start all worker, apply_req: {apply_req}")
+        semaphore = asyncio.Semaphore(parallel_num)
 
         async def _start_worker(worker_run_data: WorkerRunData):
             _start_time = time.time()
             info = worker_run_data._to_print_key()
             out = WorkerApplyOutput("")
             try:
-                await self.run_blocking_func(
-                    worker_run_data.worker.start,
-                    worker_run_data.command_args,
-                )
-                worker_run_data.stop_event.clear()
-                if worker_run_data.worker_params.register and self.register_func:
-                    # Register worker to controller
-                    await self.register_func(worker_run_data)
-                    if (
-                        worker_run_data.worker_params.send_heartbeat
-                        and self.send_heartbeat_func
-                    ):
-                        asyncio.create_task(
-                            _async_heartbeat_sender(
-                                worker_run_data,
-                                worker_run_data.worker_params.heartbeat_interval,
-                                self.send_heartbeat_func,
+                async with semaphore:
+                    await self.run_blocking_func(
+                        worker_run_data.worker.start,
+                        worker_run_data.command_args,
+                    )
+                    worker_run_data.stop_event.clear()
+                    if worker_run_data.worker_params.register and self.register_func:
+                        # Register worker to controller
+                        await self.register_func(worker_run_data)
+                        if (
+                            worker_run_data.worker_params.send_heartbeat
+                            and self.send_heartbeat_func
+                        ):
+                            asyncio.create_task(
+                                _async_heartbeat_sender(
+                                    worker_run_data,
+                                    worker_run_data.worker_params.heartbeat_interval,
+                                    self.send_heartbeat_func,
+                                )
                             )
-                        )
-                out.message = f"{info} start successfully"
+                    out.message = f"{info} start successfully"
             except TimeoutException:
                 out.success = False
                 out.message = (

packages/dbgpt-core/src/dbgpt/model/llm/llama_cpp/llama_cpp.py

@@ -13,7 +13,11 @@ from dbgpt.core import ModelOutput
 from dbgpt.model.adapter.llama_cpp_py_adapter import LlamaCppModelParameters
 from dbgpt.model.utils.llm_utils import parse_model_request
 
-from ...utils.parse_utils import ParsedChatMessage, parse_chat_message
+from ...utils.parse_utils import (
+    _DEFAULT_THINK_START_TOKEN,
+    ParsedChatMessage,
+    parse_chat_message,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -113,6 +117,8 @@ class LlamaCppModel:
         messages = request.to_common_messages()
         repetition_penalty = float(params.get("repetition_penalty", 1.1))
         top_k = int(params.get("top_k", -1))  # -1 means disable
+        think_start_token = params.get("think_start_token", _DEFAULT_THINK_START_TOKEN)
+        is_reasoning_model = params.get("is_reasoning_model", False)
         # Handle truncation
         completion_chunks = self.model.create_chat_completion(
             messages=messages,
@@ -129,6 +135,7 @@ class LlamaCppModel:
         usage = None
         msg = ParsedChatMessage()
         finish_reason: Optional[str] = None
+        is_first = True
         for r in completion_chunks:
             if not r.get("choices"):
                 continue
@@ -136,11 +143,16 @@ class LlamaCppModel:
             if delta.get("content") is not None:
                 content = delta["content"]
                 text += content
-                msg, _ = parse_chat_message(
-                    content,
-                    extract_reasoning=True,
-                    is_streaming=True,
-                    streaming_state=msg,
+                if (
+                    is_reasoning_model
+                    and not text.startswith(think_start_token)
+                    and is_first
+                ):
+                    text = think_start_token + "\n" + text
+                    is_first = False
+                msg = parse_chat_message(
+                    text,
+                    extract_reasoning=is_reasoning_model,
                 )
                 finish_reason = delta.get("finish_reason")
             if text:

packages/dbgpt-core/src/dbgpt/model/llm/llama_cpp/llama_cpp_server.py

@@ -17,7 +17,11 @@ from typing import Dict, Optional
 
 from dbgpt.core import ModelOutput
 
-from ...utils.parse_utils import ParsedChatMessage, parse_chat_message
+from ...utils.parse_utils import (
+    _DEFAULT_THINK_START_TOKEN,
+    ParsedChatMessage,
+    parse_chat_message,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -78,7 +82,10 @@ def chat_generate_stream(
 ):
     req = _build_chat_completion_request(params, stream=True)
     text = ""
+    think_start_token = params.get("think_start_token", _DEFAULT_THINK_START_TOKEN)
+    is_reasoning_model = params.get("is_reasoning_model", False)
     msg = ParsedChatMessage()
+    is_first = True
     for r in model.stream_chat_completion(req):
         if len(r.choices) == 0:
             continue
@@ -86,13 +93,17 @@ def chat_generate_stream(
         if r.choices[0] is not None and r.choices[0].delta is None:
             continue
         content = r.choices[0].delta.content
+        if content is None:
+            continue
+
+        text += content
+        if is_reasoning_model and not text.startswith(think_start_token) and is_first:
+            text = think_start_token + "\n" + text
+            is_first = False
+
+        msg = parse_chat_message(text, extract_reasoning=is_reasoning_model)
         finish_reason = _parse_finish_reason(r.choices[0].finish_reason)
 
-        if content is not None:
-            text += content
-            msg, _ = parse_chat_message(
-                content, extract_reasoning=True, is_streaming=True, streaming_state=msg
-            )
         yield ModelOutput.build(
             msg.content,
             msg.reasoning_content,

packages/dbgpt-core/src/dbgpt/model/llm/llm_out/hf_chat_llm.py

@@ -6,7 +6,11 @@ from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStream
 
 from dbgpt.core import ModelOutput
 
-from ...utils.parse_utils import ParsedChatMessage, parse_chat_message
+from ...utils.parse_utils import (
+    _DEFAULT_THINK_START_TOKEN,
+    ParsedChatMessage,
+    parse_chat_message,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -27,6 +31,8 @@ def huggingface_chat_generate_stream(
     stop_token_ids = params.get("stop_token_ids", [])
     do_sample = params.get("do_sample", True)
     custom_stop_words = params.get("custom_stop_words", [])
+    think_start_token = params.get("think_start_token", _DEFAULT_THINK_START_TOKEN)
+    is_reasoning_model = params.get("is_reasoning_model", False)
 
     input_ids = tokenizer(prompt).input_ids
     # input_ids = input_ids.to(device)
@@ -65,15 +71,23 @@ def huggingface_chat_generate_stream(
     text = ""
     usage = None
     msg = ParsedChatMessage()
+    is_first = True
     for new_text in streamer:
         text += new_text
-        msg, _ = parse_chat_message(
-            new_text, extract_reasoning=True, is_streaming=True, streaming_state=msg
-        )
         if custom_stop_words:
             for stop_word in custom_stop_words:
                 if text.endswith(stop_word):
                     text = text[: -len(stop_word)]
+
+        if (
+            prompt.rstrip().endswith(think_start_token)
+            and is_reasoning_model
+            and is_first
+        ):
+            text = think_start_token + "\n" + text
+            is_first = False
+
+        msg = parse_chat_message(text, extract_reasoning=is_reasoning_model)
         yield ModelOutput.build(
             msg.content,
             msg.reasoning_content,

packages/dbgpt-core/src/dbgpt/model/llm/llm_out/vllm_llm.py

@@ -31,6 +31,7 @@ async def generate_stream(
     best_of = params.get("best_of", None)
     stop_str = params.get("stop", None)
     think_start_token = params.get("think_start_token", _DEFAULT_THINK_START_TOKEN)
+    is_reasoning_model = params.get("is_reasoning_model", False)
     # think_end_token = params.get("think_end_token", _DEFAULT_THINK_END_TOKEN)
 
     stop_token_ids = params.get("stop_token_ids", None) or []
@@ -104,11 +105,11 @@ async def generate_stream(
         )
         if text_outputs:
             # Tempora
-            if prompt.rstrip().endswith(think_start_token):
+            if prompt.rstrip().endswith(think_start_token) and is_reasoning_model:
                 text_outputs = think_start_token + "\n" + text_outputs
             msg = parse_chat_message(
                 text_outputs,
-                extract_reasoning=True,
+                extract_reasoning=is_reasoning_model,
             )
             yield ModelOutput.build(
                 msg.content,

packages/dbgpt-core/src/dbgpt/model/proxy/llms/ollama.py

@@ -9,13 +9,13 @@ from dbgpt.core.awel.flow import (
     ResourceCategory,
     auto_register_resource,
 )
+from dbgpt.core.interface.parameter import LLMDeployModelParameters
 from dbgpt.model.proxy.base import (
     AsyncGenerateStreamFunction,
     GenerateStreamFunction,
     ProxyLLMClient,
     register_proxy_model_adapter,
 )
-from dbgpt.model.proxy.llms.chatgpt import OpenAICompatibleDeployModelParameters
 from dbgpt.model.proxy.llms.proxy_model import ProxyModel, parse_model_request
 from dbgpt.util.i18n_utils import _
 
@@ -31,7 +31,7 @@ logger = logging.getLogger(__name__)
     show_in_ui=False,
 )
 @dataclass
-class OllamaDeployModelParameters(OpenAICompatibleDeployModelParameters):
+class OllamaDeployModelParameters(LLMDeployModelParameters):
     """Deploy model parameters for Ollama."""
 
     provider: str = "proxy/ollama"

packages/dbgpt-core/src/dbgpt/model/utils/parse_utils.py

@@ -128,13 +128,13 @@ def process_streaming_chunk(
             if end_marker in remaining_chunk:
                 end_idx = remaining_chunk.find(end_marker)
                 # Output reasoning content event
-                if end_idx > 0:
-                    reasoning_part = remaining_chunk[:end_idx]
-                    events.append(
-                        StreamingEvent(type="reasoning_content", content=reasoning_part)
-                    )
-                    # Append reasoning content instead of replacing
-                    msg.reasoning_content += reasoning_part
+                # if end_idx > 0:
+                reasoning_part = remaining_chunk[:end_idx]
+                events.append(
+                    StreamingEvent(type="reasoning_content", content=reasoning_part)
+                )
+                # Append reasoning content instead of replacing
+                msg.reasoning_content += reasoning_part
 
                 # Output reasoning end event
                 events.append(StreamingEvent(type="reasoning_end", content=""))
@@ -220,6 +220,49 @@ def process_streaming_chunk(
                 remaining_chunk = ""
             continue
 
+        # Check for reasoning end markers without matching start markers
+        # This is the special case to handle
+        found_end_marker = False
+        for pattern in reasoning_patterns:
+            start_marker = pattern["start"]
+            end_marker = pattern["end"]
+            if end_marker in remaining_chunk and not state["in_reasoning"]:
+                end_idx = remaining_chunk.find(end_marker)
+                start_idx = 0
+                if start_marker in remaining_chunk:
+                    start_idx = remaining_chunk.find(start_marker) + len(start_marker)
+
+                # This is content that should be treated as reasoning but didn't have a
+                # start tag
+                # if end_idx > 0:
+                reasoning_part = remaining_chunk[start_idx:end_idx]
+                # Clear regular content
+                reasoning_part = msg.content + reasoning_part
+                msg.content = ""
+
+                # First, emit a reasoning_start event
+                events.append(StreamingEvent(type="reasoning_start", content=""))
+
+                # Then emit the content as reasoning content
+                events.append(
+                    StreamingEvent(type="reasoning_content", content=reasoning_part)
+                )
+
+                # Add to reasoning content
+                msg.reasoning_content += reasoning_part
+
+                # Emit the reasoning_end event
+                events.append(StreamingEvent(type="reasoning_end", content=""))
+                # Move past the end marker
+                remaining_chunk = remaining_chunk[end_idx + len(end_marker) :]
+                found_end_marker = True
+                state["reasoning_pattern"] = None
+                break
+
+        # If we found an end marker, continue to the next iteration
+        if found_end_marker:
+            continue
+
         # Check for reasoning start markers
         reasoning_start_found = False
         for pattern in reasoning_patterns:
@@ -228,10 +271,10 @@ def process_streaming_chunk(
                 start_idx = remaining_chunk.find(start_marker)
 
                 # Output regular content before the marker
-                if start_idx > 0:
-                    content_part = remaining_chunk[:start_idx]
-                    events.append(StreamingEvent(type="content", content=content_part))
-                    msg.content += content_part
+                # if start_idx > 0:
+                content_part = remaining_chunk[:start_idx]
+                events.append(StreamingEvent(type="content", content=content_part))
+                msg.content += content_part
 
                 # Output reasoning start event
                 events.append(StreamingEvent(type="reasoning_start", content=""))
@@ -257,12 +300,10 @@ def process_streaming_chunk(
                     start_idx = remaining_chunk.find(start_marker)
 
                     # Output regular content before the marker
-                    if start_idx > 0:
-                        content_part = remaining_chunk[:start_idx]
-                        events.append(
-                            StreamingEvent(type="content", content=content_part)
-                        )
-                        msg.content += content_part
+                    # if start_idx > 0:
+                    content_part = remaining_chunk[:start_idx]
+                    events.append(StreamingEvent(type="content", content=content_part))
+                    msg.content += content_part
 
                     # Output tool call start event
                     events.append(StreamingEvent(type="tool_call_start", content=""))
@@ -355,6 +396,7 @@ def parse_chat_message(
     reasoning_content = ""
     content = input_text
 
+    # First check for the normal case with proper start and end markers
     for pattern in reasoning_patterns:
         start_marker = pattern["start"]
         end_marker = pattern["end"]
@@ -372,6 +414,44 @@ def parse_chat_message(
                     content = content[:start_idx] + content[end_idx + len(end_marker) :]
                 break
 
+    # If no reasoning content was found with the standard pattern, check for the
+    # special case
+    # where content starts with reasoning but has no start marker
+    if not reasoning_content:
+        for pattern in reasoning_patterns:
+            start_marker = pattern["start"]
+            end_marker = pattern["end"]
+
+            if end_marker in content:
+                # Check if this is at the beginning of the content or
+                # if there's no matching start marker before it
+                end_idx = content.find(end_marker)
+                start_marker = pattern["start"]
+                start_idx = content.find(start_marker)
+
+                # If no start marker or end marker appears before start marker
+                if start_idx == -1 or end_idx < start_idx:
+                    # This is our special case - treat the content up to the end marker
+                    # as reasoning
+                    reasoning_content = string_strip(content[:end_idx])
+
+                    # Remove reasoning part from original content
+                    if extract_reasoning:
+                        content = content[end_idx + len(end_marker) :]
+                    break
+            elif start_marker in content:
+                # If there's a start marker but no end marker, treat the content
+                # as reasoning content
+                start_idx = content.find(start_marker)
+                reasoning_content = string_strip(
+                    content[start_idx + len(start_marker) :]
+                )
+
+                # Remove reasoning part from original content
+                if extract_reasoning:
+                    content = ""
+                break
+
     # Parse tool calls
     tool_calls_text = ""
 

packages/dbgpt-core/src/dbgpt/model/utils/tests/test_parse_utils.py

@@ -1,5 +1,3 @@
-import pytest
-
 from ..parse_utils import parse_chat_message
 
 
@@ -357,8 +355,8 @@ def test_streaming_mode_without_tool_calls():
 
     # Verify final message should include tool call markers
     assert (
-        "I will search for data.<｜tool▁calls▁begin｜>Tool call content<｜tool▁calls▁end｜>Search complete."  # noqa
-        == msg.content
+        "I will search for data.<｜tool▁calls▁begin｜>Tool call content"
+        "<｜tool▁calls▁end｜>Search complete." == msg.content  # noqa
     )
     assert len(msg.tool_calls) == 0
 
@@ -449,48 +447,6 @@ def test_incomplete_markers():
     assert "Tool content" in msg2.streaming_state.get("tool_call_text", "")
 
 
-def test_multiple_special_sections():
-    """Test handling multiple special sections"""
-    input_text = """<think>Reasoning content 1</think>Regular content 1
-<｜tool▁calls▁begin｜>Tool call content<｜tool▁calls▁end｜>
-Regular content 2<think>Reasoning content 2</think>End"""
-
-    result = parse_chat_message(input_text, extract_tool_calls=True)
-
-    # Verify only first reasoning content is extracted
-    assert "Reasoning content 1" == result.reasoning_content
-    assert (
-        "Regular content 1\n\nRegular content 2<think>Reasoning content 2</think>End"
-        == result.content
-    )
-
-    # Use streaming processing to handle multiple reasoning parts
-    chunks = [
-        "<think>Reasoning content 1</think>Regular content 1\n",
-        "<｜tool▁calls▁begin｜>Tool call content<｜tool▁calls▁end｜>\n",
-        "Regular content 2<think>Reasoning content 2</think>End",
-    ]
-
-    msg = None
-    all_events = []
-
-    for chunk in chunks:
-        msg, events = parse_chat_message(
-            chunk, is_streaming=True, streaming_state=msg, extract_tool_calls=True
-        )
-        all_events.extend(events)
-
-    # Verify event sequence contains two reasoning sections
-    reasoning_start_counts = sum(1 for e in all_events if e.type == "reasoning_start")
-    reasoning_end_counts = sum(1 for e in all_events if e.type == "reasoning_end")
-
-    assert reasoning_start_counts == 2
-    assert reasoning_end_counts == 2
-
-    # In streaming mode, reasoning content should accumulate
-    assert "Reasoning content 1Reasoning content 2" == msg.reasoning_content
-
-
 def test_custom_streaming_patterns():
     """Test custom streaming pattern markers"""
     custom_reasoning = [{"start": "{{thinking}}", "end": "{{/thinking}}"}]
@@ -530,41 +486,239 @@ def test_custom_streaming_patterns():
     assert "tool_call_end" in event_types
 
 
-def test_alternating_reasoning_and_tool_calls():
-    """Test alternating between reasoning and tool calls in a single message"""
-    # Use streaming to capture all sections
+def test_missing_start_token_non_streaming():
+    """Test parsing messages with missing start token but having end token
+    (non-streaming mode)
+    """
+    input_text = """Model reasoning content without start token.
+</think>
+This is the regular content part."""
+
+    result = parse_chat_message(input_text, extract_reasoning=True)
+
+    assert "This is the regular content part." == result.content
+    assert "Model reasoning content without start token." == result.reasoning_content
+
+
+def test_missing_start_token_streaming():
+    """Test parsing messages with missing start token but having end token
+    (streaming mode)
+    """
     chunks = [
-        "<think>First reasoning block</think>Content 1\n",
-        "<｜tool▁calls▁begin｜>Tool call 1<｜tool▁calls▁end｜>\n",
-        "Content 2<think>Second reasoning block</think>\n",
-        "<｜tool▁calls▁begin｜>Tool call 2<｜tool▁calls▁end｜>\n",
-        "Final content",
+        "Model reasoning content ",
+        "without start token.</think>",
+        "This is the regular content part.",
     ]
 
     msg = None
     all_events = []
 
     for chunk in chunks:
-        msg, events = parse_chat_message(
-            chunk, is_streaming=True, streaming_state=msg, extract_tool_calls=True
-        )
+        msg, events = parse_chat_message(chunk, is_streaming=True, streaming_state=msg)
         all_events.extend(events)
 
-    # Verify content is parsed correctly - note the double newlines
-    assert "Content 1\n\nContent 2\n\nFinal content" == msg.content
-    assert "First reasoning blockSecond reasoning block" == msg.reasoning_content
+    # Verify final message - match the expected result in test
+    assert "This is the regular content part." == msg.content
+    assert "Model reasoning content without start token." == msg.reasoning_content
 
-    # Count events by type
-    event_counts = {}
-    for e in all_events:
-        event_counts[e.type] = event_counts.get(e.type, 0) + 1
-
-    assert event_counts.get("reasoning_start", 0) == 2
-    assert event_counts.get("reasoning_end", 0) == 2
-    assert event_counts.get("tool_call_start", 0) == 2
-    assert event_counts.get("tool_call_end", 0) == 2
+    # Verify event sequence contains correct reasoning events
+    event_types = [e.type for e in all_events]
+    assert "reasoning_start" in event_types
+    assert "reasoning_content" in event_types
+    assert "reasoning_end" in event_types
 
 
-if __name__ == "__main__":
-    # Run tests
-    pytest.main(["-v", "test_parse_utils.py"])
+def test_missing_start_token_deepseek_chinese():
+    """Test the DeepSeek example with Chinese content missing start token"""
+    input_text = """您好！我是由中国的深度求索（DeepSeek）公司开发的智能助手\
+DeepSeek-R1。有关模型和产品的详细内容请参考官方文档。
+</think>
+您好！我是由中国的深度求索（DeepSeek）公司开发的智能助手DeepSeek-R1。有关模型\
+和产品的详细内容请参考官方文档。"""
+
+    result = parse_chat_message(input_text, extract_reasoning=True)
+
+    assert (
+        "您好！我是由中国的深度求索（DeepSeek）公司开发的智能助手DeepSeek-R1。"
+        "有关模型和产品的详细内容请参考官方文档。" == result.content
+    )
+    assert (
+        "您好！我是由中国的深度求索（DeepSeek）公司开发的智能助手DeepSeek-R1。"
+        "有关模型和产品的详细内容请参考官方文档。" == result.reasoning_content
+    )
+
+
+def test_multiple_missing_start_tokens():
+    """Test multiple occurrences of missing start tokens in the same message"""
+    input_text = """First reasoning section.
+</think>
+Some regular content.
+Second reasoning section.
+</reasoning>
+More regular content."""
+
+    result = parse_chat_message(input_text, extract_reasoning=True)
+
+    # Note: In non-streaming mode, only the first matching reasoning content is
+    # extracted
+    assert (
+        "Some regular content.\nSecond reasoning section.\n</reasoning>\nMore regular "
+        "content." == result.content
+    )
+    assert "First reasoning section." == result.reasoning_content
+
+    # Use streaming to capture all sections
+    chunks = [
+        "First reasoning section.\n</think>\n",
+        "Some regular content.\n",
+        "Second reasoning section.\n</reasoning>\n",
+        "More regular content.",
+    ]
+
+    msg = None
+    all_events = []
+
+    for chunk in chunks:
+        msg, events = parse_chat_message(chunk, is_streaming=True, streaming_state=msg)
+        all_events.extend(events)
+
+    # In streaming mode, reasoning content should match the expected format
+    assert (
+        "First reasoning section.\n\nSome regular content.\nSecond reasoning section.\n"
+        == msg.reasoning_content
+    )
+
+    # Verify event sequence contains two reasoning sections
+    reasoning_start_counts = sum(1 for e in all_events if e.type == "reasoning_start")
+    reasoning_end_counts = sum(1 for e in all_events if e.type == "reasoning_end")
+
+    assert reasoning_start_counts == 2
+    assert reasoning_end_counts == 2
+
+
+def test_missing_start_token_with_tools():
+    """Test missing reasoning start token with tool calls"""
+    input_text = """Analyzing user request to query weather information.
+</think>
+I'll look up the weather data for you.
+
+<｜tool▁calls▁begin｜>
+<｜tool▁call▁begin｜>function<｜tool▁sep｜>get_weather
+```json
+{
+    "location": "Beijing",
+    "date": "2023-05-20"
+}
+```
+<｜tool▁call▁end｜>
+<｜tool▁calls▁end｜>"""
+
+    result = parse_chat_message(
+        input_text, extract_reasoning=True, extract_tool_calls=True
+    )
+
+    assert "I'll look up the weather data for you." in result.content
+    assert (
+        "Analyzing user request to query weather information."
+        == result.reasoning_content
+    )
+    assert len(result.tool_calls) == 1
+    assert result.tool_calls[0]["name"] == "get_weather"
+    assert result.tool_calls[0]["arguments"]["location"] == "Beijing"
+
+
+def test_mixed_language_missing_start_token():
+    """Test mixed Chinese and English content with missing start token"""
+    input_text = """这是一段中英文混合的思考内容 with both languages mixed together.
+</think>
+Here's the regular content with 中文 mixed in."""
+
+    result = parse_chat_message(input_text, extract_reasoning=True)
+
+    assert "Here's the regular content with 中文 mixed in." == result.content
+    assert (
+        "这是一段中英文混合的思考内容 with both languages mixed together."
+        == result.reasoning_content
+    )
+
+
+def test_streaming_mixed_language_missing_start():
+    """Test streaming mixed language content with missing start token"""
+    chunks = [
+        "Analysis 分析: The user needs ",
+        "information about 关于天气的信息。</reasoning>",
+        "I'll provide weather information 我将提供天气信息。",
+    ]
+
+    msg = None
+    all_events = []
+
+    for chunk in chunks:
+        msg, events = parse_chat_message(chunk, is_streaming=True, streaming_state=msg)
+        all_events.extend(events)
+
+    # Verify final message includes mixed language content properly parsed
+    assert "I'll provide weather information 我将提供天气信息。" == msg.content
+    assert (
+        "Analysis 分析: The user needs information about 关于天气的信息。"
+        == msg.reasoning_content
+    )
+
+    # Verify events sequence
+    reasoning_events = [e for e in all_events if e.type.startswith("reasoning_")]
+    assert len(reasoning_events) >= 3  # At least start, content, and end events
+
+
+def test_chinese_pattern_missing_start():
+    """Test Chinese pattern with missing start token"""
+    input_text = """这里是模型的思考内容，但是没有开始标记。
+</思考>
+这是正常的响应内容。"""
+
+    result = parse_chat_message(input_text, extract_reasoning=True)
+
+    assert "这是正常的响应内容。" == result.content
+    assert "这里是模型的思考内容，但是没有开始标记。" == result.reasoning_content
+
+
+#
+# def test_multiple_special_sections():
+#     """Test handling multiple special sections"""
+#     input_text = """<think>Reasoning content 1</think>Regular content 1
+# <｜tool▁calls▁begin｜>Tool call content<｜tool▁calls▁end｜>
+# Regular content 2<think>Reasoning content 2</think>End"""
+#
+#     result = parse_chat_message(input_text, extract_tool_calls=True)
+#
+#     # Verify only first reasoning content is extracted
+#     assert "Reasoning content 1" == result.reasoning_content
+#     assert (
+#         "Regular content 1\n\nRegular content 2<think>Reasoning content 2</think>End"
+#         == result.content
+#     )
+#
+#     # Use streaming processing to handle multiple reasoning parts
+#     chunks = [
+#         "<think>Reasoning content 1</think>Regular content 1\n",
+#         "<｜tool▁calls▁begin｜>Tool call content<｜tool▁calls▁end｜>\n",
+#         "Regular content 2<think>Reasoning content 2</think>End",
+#     ]
+#
+#     msg = None
+#     all_events = []
+#
+#     for chunk in chunks:
+#         msg, events = parse_chat_message(
+#             chunk, is_streaming=True, streaming_state=msg, extract_tool_calls=True
+#         )
+#         all_events.extend(events)
+#
+#     # Verify event sequence contains two reasoning sections
+#     reasoning_start_counts = sum(1 for e in all_events if e.type == "reasoning_start")
+#     reasoning_end_counts = sum(1 for e in all_events if e.type == "reasoning_end")
+#
+#     assert reasoning_start_counts == 2
+#     assert reasoning_end_counts == 2
+#
+#     # In streaming mode, reasoning content should match the expected format
+#     assert "Reasoning content 1Reasoning content 2" == msg.reasoning_content

packages/dbgpt-core/src/dbgpt/rag/embedding/embeddings.py

@@ -113,7 +113,7 @@ class HFEmbeddingDeployModelParameters(EmbeddingDeployModelParameters):
     @property
     def real_provider_model_name(self) -> str:
         """Get the real provider model name."""
-        return self.path or self.name
+        return self.real_model_path or self.name
 
     @property
     def real_model_path(self) -> Optional[str]:

packages/dbgpt-ext/src/dbgpt_ext/datasource/rdbms/conn_clickhouse.py

@@ -40,8 +40,7 @@ class ClickhouseParameters(BaseDatasourceParameters):
     user: str = field(metadata={"help": _("Database user to connect")})
     database: str = field(metadata={"help": _("Database name")})
     engine: str = field(
-        default="MergeTree", 
-        metadata={"help": _("Storage engine, e.g., MergeTree")}
+        default="MergeTree", metadata={"help": _("Storage engine, e.g., MergeTree")}
     )
     password: str = field(
         default="${env:DBGPT_DB_PASSWORD}",

pyproject.toml

@@ -13,16 +13,22 @@ readme = "README.md"
 requires-python = ">= 3.10"
 
 [tool.uv.sources]
-dbgpt-accelerator = { workspace = true }
 dbgpt = { workspace = true }
 dbgpt-client = { workspace = true }
 dbgpt-ext = { workspace = true }
 dbgpt-serve = { workspace = true }
 dbgpt-app = { workspace = true }
+dbgpt-acc-auto = { workspace = true }
+dbgpt-acc-flash-attn = { workspace = true }
 
 [tool.uv.workspace]
 members = [
-  "packages/dbgpt-*"
+  "packages/dbgpt-app", 
+  "packages/dbgpt-client",
+  "packages/dbgpt-core", 
+  "packages/dbgpt-ext", 
+  "packages/dbgpt-serve", 
+  "packages/dbgpt-accelerator/*"
 ]
 
 [tool.uv]
@@ -65,4 +71,4 @@ select = ["E", "F", "I"]
 
 [tool.ruff.isort]
 # Specify the local modules (first-party)
-known-first-party = ["dbgpt", "dbgpt_accelerator", "dbgpt_client", "dbgpt_ext", "dbgpt_serve", "dbgpt_app"]
+known-first-party = ["dbgpt", "dbgpt_acc_auto", "dbgpt_client", "dbgpt_ext", "dbgpt_serve", "dbgpt_app"]