fix: modify qwq reasoning output (#90)

* merge

* merge

* add Mistral-Small-3.1-24B-Instruct-2503

* modify qwq-32b deploy

Author: 11zhouxuan

docs/en/best_deployment_practices.md

@@ -4,6 +4,11 @@ This document provides examples of best practices for deploying models using EMD
 
 ## Famous Models
 
+### Mistral Small Series
+```
+emd deploy --model-id Mistral-Small-3.1-24B-Instruct-2503 --instance-type g5.12xlarge --engine-type vllm --service-type sagemaker_realtime
+```
+
 ### Gemma 3 Series
 
 ```

docs/en/supported_models.md

@@ -44,6 +44,7 @@
 | gemma-3-4b-it                        | gemma3                   | vlm         | vllm                    | g5.2xlarge,g5.4xlarge,g5.8xlarge,g5.16xlarge                                                                                                                                                      | sagemaker_realtime,sagemaker_async,ecs | ❎                     |
 | gemma-3-12b-it                       | gemma3                   | vlm         | vllm                    | g5.2xlarge,g5.4xlarge,g5.8xlarge,g5.16xlarge                                                                                                                                                      | sagemaker_realtime,sagemaker_async,ecs | ❎                     |
 | gemma-3-27b-it                       | gemma3                   | vlm         | vllm                    | g5.12xlarge,g5.24xlarge,g5.48xlarge                                                                                                                                                               | sagemaker_realtime,sagemaker_async,ecs | ❎                     |
+| Mistral-Small-3.1-24B-Instruct-2503  | mistral                  | vlm         | vllm                    | g5.12xlarge,g5.24xlarge,g5.48xlarge                                                                                                                                                               | sagemaker_realtime,sagemaker_async,ecs | ❎                     |
 | txt2video-LTX                        | comfyui                  | video       | comfyui                 | g5.4xlarge,g5.8xlarge,g6e.2xlarge                                                                                                                                                                 | sagemaker_async                        | ❎                     |
 | whisper                              | whisper                  | whisper     | huggingface             | g5.xlarge,g5.2xlarge,g5.4xlarge,g5.8xlarge,g5.16xlarge                                                                                                                                            | sagemaker_async                        | ❎                     |
 | bce-embedding-base_v1                | bce                      | embedding   | vllm                    | g4dn.2xlarge,g5.xlarge,g5.2xlarge,g5.4xlarge,g5.8xlarge,g5.16xlarge                                                                                                                               | sagemaker_realtime,ecs                 | ✅                     |

src/emd/models/engines.py

@@ -77,6 +77,17 @@ class KtransformersEngine(OpenAICompitableEngine):
     }
 )
 
+
+vllm_mistral_small_engine082 = VllmEngine(
+    **{
+    **vllm_engine064.model_dump(),
+    "engine_dockerfile_config": {"VERSION":"v0.8.2"},
+    "dockerfile_name":"Dockerfile",
+    "default_cli_args": " --tokenizer-mode mistral --config-format mistral --load-format mistral  --limit-mm-per-prompt 'image=4' --max-model-len 16384",
+    "environment_variables": ""
+    }
+)
+
 vllm_deepseek_r1_distill_qwen_engine071 = VllmEngine(**{
             **vllm_engine064.model_dump(),
             "engine_dockerfile_config": {"VERSION":"v0.7.1"},
@@ -124,6 +135,13 @@ class KtransformersEngine(OpenAICompitableEngine):
             "default_cli_args": " --chat-template emd/models/chat_templates/qwq_32b_add_prefill_chat_template.jinja --max_model_len 16000  --max_num_seq 10 --disable-log-stats --enable-auto-tool-choice --tool-call-parser hermes"
 })
 
+vllm_qwq_engine082 = VllmEngine(**{
+            **vllm_qwen25vl72b_engine073.model_dump(),
+            "engine_dockerfile_config": {"VERSION":"v0.8.2"},
+            "environment_variables": "export VLLM_ATTENTION_BACKEND=FLASHINFER && export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True",
+            "default_cli_args": " --chat-template emd/models/chat_templates/qwq_32b_add_prefill_chat_template.jinja --max_model_len 16000  --max_num_seq 10 --disable-log-stats --enable-auto-tool-choice --tool-call-parser hermes  --enable-reasoning --reasoning-parser deepseek_r1"
+})
+
 
 vllm_internvl2d5_76b_engine064 = VllmEngine(**{
              **vllm_engine064.model_dump(),

src/emd/models/llms/qwen.py

@@ -7,7 +7,8 @@
     tgi_qwen2d5_on_inf2,
     tgi_qwen2d5_72b_on_inf2,
     vllm_qwen2d5_72b_engine064,
-    vllm_qwq_engine073
+    vllm_qwq_engine073,
+    vllm_qwq_engine082
 )
 from ..services import (
     sagemaker_service,
@@ -471,7 +472,7 @@
 Model.register(
     dict(
         model_id = "QwQ-32B",
-        supported_engines=[vllm_qwq_engine073],
+        supported_engines=[vllm_qwq_engine082],
         supported_instances=[
             g5d12xlarge_instance,
             g5d24xlarge_instance,

src/emd/models/model_series.py

@@ -97,6 +97,12 @@
     reference_link="https://blog.google/technology/developers/gemma-3/"
 )
 
+MISTRAL_SERIES = ModelSeries(
+    model_series_name=ModelSeriesType.MISTRAL,
+    description="LLMs and VLMs provided by MISTRAL AI.",
+    reference_link="https://huggingface.co/mistralai"
+)
+
 DEEPSEEK_REASONING_MODEL = ModelSeries(
     model_series_name=ModelSeriesType.DEEPSEEK_REASONING_MODEL,
     description="DeepSeek-R1-Zero and DeepSeek-R1 are innovative reasoning models, with the former showcasing strong performance through reinforcement learning alone, while the latter enhances reasoning capabilities by incorporating cold-start data, achieving results comparable to OpenAI-o1 and setting new benchmarks with its distilled versions.",

src/emd/models/services.py

@@ -91,6 +91,7 @@
         "ServiceType":"service_type",
         "EngineType":"engine_type",
         "Region": "region",
+        "DesiredCapacity": "desired_capacity",
         "ContainerCpu": "container_cpu",
         "ContainerMemory": "container_memory",
         "ContainerGpu":"instance_gpu_num"

src/emd/models/utils/constants.py

@@ -214,6 +214,7 @@ def get_service_quota_code(cls, instance_type: str):
 
 class ModelSeriesType(ConstantBase):
     GEMMA3 = "gemma3"
+    MISTRAL = "mistral"
     QWEN2D5 = "qwen2.5"
     GLM4 = "glm4"
     INTERLM2d5 = "internlm2.5"

src/emd/models/vlms/__init__.py

@@ -1,3 +1,4 @@
 from . import qwen
 from . import internvl
 from . import gemma3
+from . import mistral

src/emd/models/vlms/mistral.py

@@ -0,0 +1,54 @@
+from ..engines import vllm_mistral_small_engine082
+from .. import Model
+from ..frameworks import fastapi_framework
+from ..services import (
+    sagemaker_service,
+    sagemaker_async_service,
+    ecs_service,
+    local_service
+)
+from emd.models.utils.constants import ModelType
+from ..model_series import MISTRAL_SERIES
+from ..instances import (
+    g5d2xlarge_instance,
+    g5d4xlarge_instance,
+    g5d8xlarge_instance,
+    g5d12xlarge_instance,
+    g5d16xlarge_instance,
+    g5d24xlarge_instance,
+    g5d48xlarge_instance,
+    g6e2xlarge_instance,
+    local_instance
+)
+from ..utils.constants import ModelFilesDownloadSource
+
+
+Model.register(
+    dict(
+        model_id = "Mistral-Small-3.1-24B-Instruct-2503",
+        supported_engines=[vllm_mistral_small_engine082],
+        supported_instances=[
+            g5d12xlarge_instance,
+            g5d24xlarge_instance,
+            g5d48xlarge_instance,
+            local_instance
+        ],
+        supported_services=[
+            sagemaker_service,
+            sagemaker_async_service,
+            ecs_service,
+            local_service
+        ],
+        supported_frameworks=[
+            fastapi_framework
+        ],
+        huggingface_model_id="unsloth/Mistral-Small-3.1-24B-Instruct-2503",
+        # require_huggingface_token=False,
+        modelscope_model_id="mistralai/Mistral-Small-3.1-24B-Instruct-2503",
+        # model_files_download_source=ModelFilesDownloadSource.MODELSCOPE,
+        application_scenario="vision llms for image understanding",
+        description="The latest series of mistral small",
+        model_type=ModelType.VLM,
+        model_series=MISTRAL_SERIES,
+    )
+)

src/pipeline/backend/backend.py

@@ -134,9 +134,10 @@ def start_server(self, server_start_command):
         logger.info(f"Starting {self.engine_type} server with command: {server_start_command}")
         t = threading.Thread(target=os.system,args=(server_start_command,),daemon=True)
         t.start()
-        t2 = threading.Thread(target=self.check_model_serve_ready,args=(t, "127.0.0.1", self.server_port),daemon=True)
-        t2.start()
-        t2.join()
+        self.check_model_serve_ready(t, "127.0.0.1", self.server_port)
+        logger.info(f"Server started successfully.")
+        # t2.start()
+        # t2.join()
         return