Merge branch 'main' into release/3.0

Author: Jintao-Huang

docs/source/Instruction/支持的模型和数据集.md

@@ -344,6 +344,8 @@
 |[deepseek-ai/DeepSeek-V2-Chat](https://modelscope.cn/models/deepseek-ai/DeepSeek-V2-Chat)|deepseek_v2|deepseek|transformers>=4.39.3|-|[deepseek-ai/DeepSeek-V2-Chat](https://huggingface.co/deepseek-ai/DeepSeek-V2-Chat)|
 |[deepseek-ai/DeepSeek-V2.5](https://modelscope.cn/models/deepseek-ai/DeepSeek-V2.5)|deepseek_v2_5|deepseek_v2_5|transformers>=4.39.3|-|[deepseek-ai/DeepSeek-V2.5](https://huggingface.co/deepseek-ai/DeepSeek-V2.5)|
 |[deepseek-ai/DeepSeek-V2.5-1210](https://modelscope.cn/models/deepseek-ai/DeepSeek-V2.5-1210)|deepseek_v2_5|deepseek_v2_5|transformers>=4.39.3|-|[deepseek-ai/DeepSeek-V2.5-1210](https://huggingface.co/deepseek-ai/DeepSeek-V2.5-1210)|
+|[deepseek-ai/DeepSeek-V3-Base](https://modelscope.cn/models/deepseek-ai/DeepSeek-V3-Base)|deepseek_v2_5|deepseek_v2_5|transformers>=4.39.3|-|[deepseek-ai/DeepSeek-V3-Base](https://huggingface.co/deepseek-ai/DeepSeek-V3-Base)|
+|[deepseek-ai/DeepSeek-V3](https://modelscope.cn/models/deepseek-ai/DeepSeek-V3)|deepseek_v2_5|deepseek_v2_5|transformers>=4.39.3|-|[deepseek-ai/DeepSeek-V3](https://huggingface.co/deepseek-ai/DeepSeek-V3)|
 |[OpenBuddy/openbuddy-llama-65b-v8-bf16](https://modelscope.cn/models/OpenBuddy/openbuddy-llama-65b-v8-bf16)|openbuddy_llama|openbuddy|-|-|[OpenBuddy/openbuddy-llama-65b-v8-bf16](https://huggingface.co/OpenBuddy/openbuddy-llama-65b-v8-bf16)|
 |[OpenBuddy/openbuddy-llama2-13b-v8.1-fp16](https://modelscope.cn/models/OpenBuddy/openbuddy-llama2-13b-v8.1-fp16)|openbuddy_llama|openbuddy|-|-|[OpenBuddy/openbuddy-llama2-13b-v8.1-fp16](https://huggingface.co/OpenBuddy/openbuddy-llama2-13b-v8.1-fp16)|
 |[OpenBuddy/openbuddy-llama2-70b-v10.1-bf16](https://modelscope.cn/models/OpenBuddy/openbuddy-llama2-70b-v10.1-bf16)|openbuddy_llama|openbuddy|-|-|[OpenBuddy/openbuddy-llama2-70b-v10.1-bf16](https://huggingface.co/OpenBuddy/openbuddy-llama2-70b-v10.1-bf16)|

docs/source_en/Instruction/Supported-models-and-datasets.md

@@ -344,6 +344,8 @@ The table below introduces the models integrated with ms-swift:
 |[deepseek-ai/DeepSeek-V2-Chat](https://modelscope.cn/models/deepseek-ai/DeepSeek-V2-Chat)|deepseek_v2|deepseek|transformers>=4.39.3|-|[deepseek-ai/DeepSeek-V2-Chat](https://huggingface.co/deepseek-ai/DeepSeek-V2-Chat)|
 |[deepseek-ai/DeepSeek-V2.5](https://modelscope.cn/models/deepseek-ai/DeepSeek-V2.5)|deepseek_v2_5|deepseek_v2_5|transformers>=4.39.3|-|[deepseek-ai/DeepSeek-V2.5](https://huggingface.co/deepseek-ai/DeepSeek-V2.5)|
 |[deepseek-ai/DeepSeek-V2.5-1210](https://modelscope.cn/models/deepseek-ai/DeepSeek-V2.5-1210)|deepseek_v2_5|deepseek_v2_5|transformers>=4.39.3|-|[deepseek-ai/DeepSeek-V2.5-1210](https://huggingface.co/deepseek-ai/DeepSeek-V2.5-1210)|
+|[deepseek-ai/DeepSeek-V3-Base](https://modelscope.cn/models/deepseek-ai/DeepSeek-V3-Base)|deepseek_v2_5|deepseek_v2_5|transformers>=4.39.3|-|[deepseek-ai/DeepSeek-V3-Base](https://huggingface.co/deepseek-ai/DeepSeek-V3-Base)|
+|[deepseek-ai/DeepSeek-V3](https://modelscope.cn/models/deepseek-ai/DeepSeek-V3)|deepseek_v2_5|deepseek_v2_5|transformers>=4.39.3|-|[deepseek-ai/DeepSeek-V3](https://huggingface.co/deepseek-ai/DeepSeek-V3)|
 |[OpenBuddy/openbuddy-llama-65b-v8-bf16](https://modelscope.cn/models/OpenBuddy/openbuddy-llama-65b-v8-bf16)|openbuddy_llama|openbuddy|-|-|[OpenBuddy/openbuddy-llama-65b-v8-bf16](https://huggingface.co/OpenBuddy/openbuddy-llama-65b-v8-bf16)|
 |[OpenBuddy/openbuddy-llama2-13b-v8.1-fp16](https://modelscope.cn/models/OpenBuddy/openbuddy-llama2-13b-v8.1-fp16)|openbuddy_llama|openbuddy|-|-|[OpenBuddy/openbuddy-llama2-13b-v8.1-fp16](https://huggingface.co/OpenBuddy/openbuddy-llama2-13b-v8.1-fp16)|
 |[OpenBuddy/openbuddy-llama2-70b-v10.1-bf16](https://modelscope.cn/models/OpenBuddy/openbuddy-llama2-70b-v10.1-bf16)|openbuddy_llama|openbuddy|-|-|[OpenBuddy/openbuddy-llama2-70b-v10.1-bf16](https://huggingface.co/OpenBuddy/openbuddy-llama2-70b-v10.1-bf16)|

swift/llm/infer/infer_engine/infer_engine.py

@@ -244,12 +244,4 @@ def func(target, queue, args, kwargs):
 
     @staticmethod
     def safe_asyncio_run(coro):
-        try:
-            loop = asyncio.get_running_loop()
-        except RuntimeError:
-            loop = None
-        if loop:
-            result = InferEngine.thread_run(asyncio.run, args=(coro, ))
-        else:
-            result = asyncio.run(coro)
-        return result
+        return InferEngine.thread_run(asyncio.run, args=(coro, ))

swift/llm/model/model/deepseek.py

@@ -108,6 +108,8 @@ def _dtype_hook(module, input, output):
             ModelGroup([
                 Model('deepseek-ai/DeepSeek-V2.5', 'deepseek-ai/DeepSeek-V2.5'),
                 Model('deepseek-ai/DeepSeek-V2.5-1210', 'deepseek-ai/DeepSeek-V2.5-1210'),
+                Model('deepseek-ai/DeepSeek-V3-Base', 'deepseek-ai/DeepSeek-V3-Base'),
+                Model('deepseek-ai/DeepSeek-V3', 'deepseek-ai/DeepSeek-V3'),
             ]),
         ],
         TemplateType.deepseek_v2_5,

swift/llm/template/base.py

@@ -610,6 +610,7 @@ def _encode(self, inputs: StdTemplateInputs) -> Dict[str, Any]:
         if (self.template_meta.template_type == 'dummy' and self.use_chat_template and not self.is_training
                 and self.mode != 'seq_cls'):
             template_backend = 'jinja'
+            logger.info_once(f'Setting template_backend: {template_backend}')
         res_context_list, loss_scale_list, answer_len = (
             self._swift_encode(inputs) if template_backend == 'swift' else self._jinja_encode(inputs))
         encoded = {}

swift/ui/app.py

@@ -1,5 +1,6 @@
 # Copyright (c) Alibaba, Inc. and its affiliates.
 import os
+from copy import copy
 from dataclasses import fields
 from functools import partial
 from typing import List, Union
@@ -72,6 +73,7 @@ def run(self):
                     for f in fields(self.args):
                         if getattr(self.args, f.name):
                             LLMInfer.default_dict[f.name] = getattr(self.args, f.name)
+
                     LLMInfer.is_gradio_app = True
                     LLMInfer.is_multimodal = self.args.model_meta.is_multimodal
                     LLMInfer.build_ui(LLMInfer)
@@ -93,10 +95,20 @@ def run(self):
                         value = getattr(self.args, f.name)
                         if isinstance(value, list):
                             value = ' '.join([v or '' for v in value])
-                        LLMInfer.elements()[f.name].value = value
-                app.load(LLMInfer.deploy_model, list(LLMInfer.valid_elements().values()),
-                         [LLMInfer.element('runtime_tab'),
-                          LLMInfer.element('running_tasks')])
+                        LLMInfer.elements()[f.name].value = str(value)
+
+                args = copy(self.args)
+                args.port = find_free_port()
+
+                values = []
+                for key in LLMInfer.valid_elements():
+                    if key in args.__dict__:
+                        value = getattr(args, key)
+                    else:
+                        value = LLMInfer.element(key).value
+                    values.append(value)
+                _, running_task = LLMInfer.deploy_model(*values)
+                LLMInfer.element('running_tasks').value = running_task['value']
             else:
                 app.load(
                     partial(LLMTrain.update_input_model, arg_cls=RLHFArguments),

swift/ui/base.py

@@ -4,10 +4,11 @@
 import sys
 import time
 import typing
+from collections import OrderedDict
 from dataclasses import fields
 from datetime import datetime
 from functools import wraps
-from typing import Any, Dict, List, OrderedDict, Type
+from typing import Any, Dict, List, Type
 
 import gradio as gr
 import json
@@ -220,12 +221,12 @@ def elements(cls):
 
     @classmethod
     def valid_elements(cls):
+        valid_elements = OrderedDict()
         elements = cls.elements()
-        return {
-            key: value
-            for key, value in elements.items()
-            if isinstance(value, (Textbox, Dropdown, Slider, Checkbox)) and key != 'train_record'
-        }
+        for key, value in elements.items():
+            if isinstance(value, (Textbox, Dropdown, Slider, Checkbox)) and key != 'train_record':
+                valid_elements[key] = value
+        return valid_elements
 
     @classmethod
     def element_keys(cls):

swift/ui/llm_infer/llm_infer.py

@@ -1,6 +1,8 @@
 # Copyright (c) Alibaba, Inc. and its affiliates.
+import atexit
 import os
 import re
+import signal
 import sys
 import time
 from copy import deepcopy
@@ -298,13 +300,37 @@ def deploy_model(cls, *args):
                 cnt += 1
                 if cnt >= 60:
                     logger.warning_once(f'Deploy costing too much time, please check log file: {log_file}')
+            if cls.is_gradio_app:
+                cls.register_clean_hook()
             logger.info('Deploy done.')
         cls.deployed = True
         running_task = Runtime.refresh_tasks(log_file)
         if cls.is_gradio_app:
             cls.running_task = running_task['value']
         return gr.update(open=True), running_task
 
+    @classmethod
+    def clean_deployment(cls):
+        if not cls.is_gradio_app:
+            return
+
+        logger.info('Killing deployment')
+        _, args = Runtime.parse_info_from_cmdline(cls.running_task)
+        os.system(f'pkill -9 -f {args["log_file"]}')
+        logger.info('Done.')
+
+    @classmethod
+    def register_clean_hook(cls):
+        atexit.register(LLMInfer.clean_deployment)
+        signal.signal(signal.SIGINT, LLMInfer.signal_handler)
+        if os.name != 'nt':
+            signal.signal(signal.SIGTERM, LLMInfer.signal_handler)
+
+    @staticmethod
+    def signal_handler(*args, **kwargs):
+        LLMInfer.clean_deployment()
+        sys.exit(0)
+
     @classmethod
     def clear_session(cls):
         return '', [], gr.update(value=None), gr.update(value=None), gr.update(value=None), []

tests/test_align/test_template/test_template.py

@@ -5,6 +5,9 @@ def test_deepseek_v2_5():
     tokenizer = get_model_tokenizer('deepseek-ai/DeepSeek-V2.5-1210', load_model=False)[1]
     template = get_template(tokenizer.model_meta.template, tokenizer)
     inputs = TemplateInputs(messages=[{
+        'role': 'system',
+        'content': '000'
+    }, {
         'role': 'user',
         'content': 'aaa'
     }, {
@@ -17,8 +20,9 @@ def test_deepseek_v2_5():
     res = template.encode(inputs)
     template.print_inputs(res)
     template.template_backend = 'jinja'
-    res = template.encode(inputs)
+    res2 = template.encode(inputs)
     template.print_inputs(res)
+    assert res['input_ids'] == res2['input_ids']
 
 
 if __name__ == '__main__':