xinference 1.3.0__py3-none-any.whl → 1.3.0.post2__py3-none-any.whl

xinference/_version.py

@@ -8,11 +8,11 @@ import json
 
 version_json = '''
 {
- "date": "2025-02-21T22:33:35+0800",
+ "date": "2025-02-22T23:10:02+0800",
  "dirty": false,
  "error": null,
- "full-revisionid": "bb778404b248ea91f4210e739e2b92b8a1be1455",
- "version": "1.3.0"
+ "full-revisionid": "378a47adad8506a13105b063322ccd7a04f7ea5d",
+ "version": "1.3.0.post2"
 }
 '''  # END VERSION_JSON
 

xinference/deploy/cmdline.py

@@ -821,12 +821,6 @@ def remove_cache(
     type=bool,
     help="Whether or not to allow for custom models defined on the Hub in their own modeling files.",
 )
-@click.option(
-    "--reasoning-content",
-    default=False,
-    type=bool,
-    help="Whether or not to enable reasoning content in model responses.",
-)
 @click.option(
     "--api-key",
     "-ak",
@@ -855,7 +849,6 @@ def model_launch(
     worker_ip: Optional[str],
     gpu_idx: Optional[str],
     trust_remote_code: bool,
-    reasoning_content: bool,
     api_key: Optional[str],
     model_path: Optional[str],
 ):
@@ -948,7 +941,6 @@ def model_launch(
         gpu_idx=_gpu_idx,
         trust_remote_code=trust_remote_code,
         model_path=model_path,
-        reasoning_content=reasoning_content,
         **kwargs,
     )
 

xinference/model/llm/reasoning_parsers/abs_reasoning_parsers.py

@@ -1,7 +1,7 @@
 from abc import ABC, abstractmethod
 from typing import Dict, Optional, Tuple, Type, Union
 
-from ....types import ChatCompletionChunkDelta, CompletionChoice, CompletionChunk
+from ....types import ChatCompletionChunkDelta, CompletionChoice
 
 
 class ReasoningParser(ABC):
@@ -26,7 +26,7 @@ class ReasoningParser(ABC):
         self,
         previous_text: str,
         current_text: str,
-        delta: Union[str, CompletionChunk],
+        delta: ChatCompletionChunkDelta,
     ) -> ChatCompletionChunkDelta:
         """Extract reasoning content from model output in a streaming fashion.
 

xinference/model/llm/reasoning_parsers/deepseek_r1_reasoning_parser.py

@@ -24,7 +24,7 @@ class DeepSeekR1ReasoningParser(ReasoningParser):
         previous_text: str,
         current_text: str,
         delta: ChatCompletionChunkDelta,
-    ) -> Optional[ChatCompletionChunkDelta]:
+    ) -> ChatCompletionChunkDelta:
         """Extract reasoning content from DeepSeek-R1 model output in a streaming fashion.
 
         Args:
@@ -122,7 +122,7 @@ class DeepSeekR1ReasoningParser(ReasoningParser):
         # Thus we assume the reasoning content is always at the start.
         # Ref https://huggingface.co/deepseek-ai/DeepSeek-R1/commit/8a58a132790c9935686eb97f042afa8013451c9f
         if self.reasoning_end_tag not in model_output:
-            return model_output, None
+            return model_output, ""
         else:
             # Add a start token if it's missing to keep compatibility.
             if self.reasoning_start_tag not in model_output:
@@ -136,5 +136,5 @@ class DeepSeekR1ReasoningParser(ReasoningParser):
             final_output = model_output[end_index:]
 
             if len(final_output) == 0:
-                return reasoning_content, None
+                return reasoning_content, ""
             return reasoning_content, final_output

xinference/model/llm/utils.py

@@ -374,19 +374,21 @@ class ChatModelMixin:
         current_text = ""
         async for chunk in chunks:
             if i == 0:
-                chunk = cls._get_first_chat_completion_chunk(chunk)
+                chat_chunk = cls._get_first_chat_completion_chunk(chunk)
             elif not chunk.get("choices"):
                 # usage
-                chunk = cls._get_final_chat_completion_chunk(chunk)
+                chat_chunk = cls._get_final_chat_completion_chunk(chunk)
             else:
-                chunk = cls._to_chat_completion_chunk(chunk)
+                chat_chunk = cls._to_chat_completion_chunk(chunk)
             if reasoning_parser is not None:
-                choices = chunk.get("choices")
+                choices = chat_chunk.get("choices")
+                if choices is None:
+                    continue
                 for choice in choices:
                     delta = choice.get("delta")
                     if not delta:
                         continue
-                    current_text = previous_text + delta.get("content")
+                    current_text = previous_text + delta.get("content", "")
                     choice[
                         "delta"
                     ] = reasoning_parser.extract_reasoning_content_streaming(
@@ -395,7 +397,7 @@ class ChatModelMixin:
                         delta=delta,
                     )
                     previous_text = current_text
-            yield chunk
+            yield chat_chunk
             i += 1
 
     @staticmethod
@@ -408,7 +410,7 @@ class ChatModelMixin:
             reasoning_content = None
 
             if reasoning_parser is not None:
-                reasoning_content, content = reasoning_parser.extract_reasoning_content(
+                reasoning_content, content = reasoning_parser.extract_reasoning_content(  # type: ignore
                     choice
                 )
 
@@ -429,7 +431,8 @@ class ChatModelMixin:
             "id": "chat" + completion["id"],
             "object": "chat.completion",
             "created": completion["created"],
-            "model": choices,
+            "model": completion["model"],
+            "choices": choices,  # type: ignore
             "usage": completion["usage"],
         }
 

{xinference-1.3.0.dist-info → xinference-1.3.0.post2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: xinference
-Version: 1.3.0
+Version: 1.3.0.post2
 Summary: Model Serving Made Easy
 Home-page: https://github.com/xorbitsai/inference
 Author: Qin Xuye

{xinference-1.3.0.dist-info → xinference-1.3.0.post2.dist-info}/RECORD

@@ -1,6 +1,6 @@
 xinference/__init__.py,sha256=nmTTrYbIpj964ZF6ojtgOM7E85JBOj1EyQbmYjbj1jw,915
 xinference/_compat.py,sha256=URSJQLXrcsTO9B_4x0wVDPijYQDhuVJmZ95npID560w,4197
-xinference/_version.py,sha256=WWhsA6HLqEmACoxpxfySEDjEnZMcVGekSJb9gJu4UIk,497
+xinference/_version.py,sha256=4EtXtCeHp6nyWCpJ7_0hRACaUqdifs7HuNYZPw5GSHM,503
 xinference/conftest.py,sha256=vETDpRBVIlWbWi7OTwf7og89U25KyYGyI7yPIB3O8N8,9564
 xinference/constants.py,sha256=mEW4HDzjXtDXN61Mt6TtJrJ4ljbB6VUkh97e3oDbNx4,3905
 xinference/device_utils.py,sha256=ELsqvnjvz9wYthTyQFzKSV4mZsaASz6hj_IsfMmfMWc,4447
@@ -34,7 +34,7 @@ xinference/core/supervisor.py,sha256=JgfEY7LkP9m48abukMXKBumEqeV9Q-IImSAJjk6YslE
 xinference/core/utils.py,sha256=TpUHNFcA4pEXq084Quz-ilIccuDMUAsdLrEJxj0Zn6I,10414
 xinference/core/worker.py,sha256=VpX9Mwa27FyNcgoXdFRe414iL4KySZNpMfPIOhrcVPE,51737
 xinference/deploy/__init__.py,sha256=h_JgzSqV5lP6vQ6XX_17kE4IY4BRnvKta_7VLQAL1ms,581
-xinference/deploy/cmdline.py,sha256=-RKCZaM2SlL8ZXfp_LQJ7tndn7ip9c9F-BKq0qINGkg,49565
+xinference/deploy/cmdline.py,sha256=fOuD4HixyI6VuuVO7XJ-0_C3JutDe1sv4MwVjZPluS8,49338
 xinference/deploy/local.py,sha256=sO3BcpEH9oCF87CxWVA4AXAYcfHGnrcop40ew5NOA2g,3979
 xinference/deploy/supervisor.py,sha256=68rB2Ey5KFeF6zto9YGbw3P8QLZmF_KSh1NwH_pNP4w,2986
 xinference/deploy/utils.py,sha256=jdL7i2WV6u_BZ8IiE1d3YktvCARcB3ntzMQ5rHGD5DM,6756
@@ -94,7 +94,7 @@ xinference/model/llm/llm_family_csghub.json,sha256=zMKWbihsxQNVB1u5iKJbZUkbOfQ4I
 xinference/model/llm/llm_family_modelscope.json,sha256=8RsMiGrbVdLQohnmdbVWIZgk4yYcFpnYZpkcqz-P8yE,289997
 xinference/model/llm/llm_family_openmind_hub.json,sha256=jl9pfbe5DztoxgEwKBxDk1Wd7TziTiJ48_Ie_lJdYjA,67872
 xinference/model/llm/memory.py,sha256=GLNmXBI-AtMbuaJfEf50fnhN4rdbOZjLyT6L_Vjqa5g,10206
-xinference/model/llm/utils.py,sha256=PMhIVa_KI6SBqPG1Usd3iMSUoIEMnN9U9lnW_iA82m0,29995
+xinference/model/llm/utils.py,sha256=ZzhXldHn2MHuMgyJkpwxFZlwS6mbdIL2_UMqKWKpkz4,30165
 xinference/model/llm/llama_cpp/__init__.py,sha256=h_JgzSqV5lP6vQ6XX_17kE4IY4BRnvKta_7VLQAL1ms,581
 xinference/model/llm/llama_cpp/core.py,sha256=3GSGk42c8Oy_jTqRv4nLC482V2tUis3V0LlohQy_I1U,11312
 xinference/model/llm/lmdeploy/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -102,8 +102,8 @@ xinference/model/llm/lmdeploy/core.py,sha256=WvSP3x6t-HBv6hKh1qWZatFAzlcZCyyKqvc
 xinference/model/llm/mlx/__init__.py,sha256=h_JgzSqV5lP6vQ6XX_17kE4IY4BRnvKta_7VLQAL1ms,581
 xinference/model/llm/mlx/core.py,sha256=WQN2iURiWSL_MY5hR0GkCYa08qr5wtOFx522_c2vW30,23130
 xinference/model/llm/reasoning_parsers/__init__.py,sha256=-sjSIQ4K6w-TEzx49kVaWeWC443fnZqODU91GCQ_JNo,581
-xinference/model/llm/reasoning_parsers/abs_reasoning_parsers.py,sha256=7jISylZEQDCw2-5kR4MsXkZH5xhkOlrBrJBAIAsnSZI,3040
-xinference/model/llm/reasoning_parsers/deepseek_r1_reasoning_parser.py,sha256=IpeXAJSApXAdcUj8lYC7jkBRkJbURCt9pBVa7fZHNKM,6359
+xinference/model/llm/reasoning_parsers/abs_reasoning_parsers.py,sha256=NWwx7Uj2K6SR1TZCfHeF1Q9Kpm9o6PNV8I16nyQNTh4,3020
+xinference/model/llm/reasoning_parsers/deepseek_r1_reasoning_parser.py,sha256=m_olMdPLJrozADlRuH_nLRrzqLgM2W1QY9G2fa7qzes,6345
 xinference/model/llm/sglang/__init__.py,sha256=-sjSIQ4K6w-TEzx49kVaWeWC443fnZqODU91GCQ_JNo,581
 xinference/model/llm/sglang/core.py,sha256=tR8PtdC_RBG00sBQp8lF-ow5xq9QkPlRbFPoHzzCDZk,20321
 xinference/model/llm/transformers/__init__.py,sha256=h_JgzSqV5lP6vQ6XX_17kE4IY4BRnvKta_7VLQAL1ms,581
@@ -15725,9 +15725,9 @@ xinference/web/ui/node_modules/yup/package.json,sha256=xRFSROB9NKxqSWHEVFvSTsPs9
 xinference/web/ui/node_modules/yup/node_modules/type-fest/package.json,sha256=JTv2zTTVgxQ2H82m1-6qEpdMv08lHjFx4Puf_MsbB_Q,1134
 xinference/web/ui/src/locales/en.json,sha256=OS25toIZHXcGYPtl5rZRQfJEDuAyrwBIYM6GGxOVe4Y,8641
 xinference/web/ui/src/locales/zh.json,sha256=mBhTne0sIYLe_xh5qyH45W_D50Hjp-DKdwcifYEHFnw,8386
-xinference-1.3.0.dist-info/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
-xinference-1.3.0.dist-info/METADATA,sha256=9Ad2j0ziQ8Xm7bxxRjr7URv5kS4k3m6xH2IcJSN345U,24121
-xinference-1.3.0.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
-xinference-1.3.0.dist-info/entry_points.txt,sha256=-lDyyzqWMFQF0Rgm7VxBNz0V-bMBMQLRR3pvQ-Y8XTY,226
-xinference-1.3.0.dist-info/top_level.txt,sha256=L1rQt7pl6m8tmKXpWVHzP-GtmzAxp663rXxGE7qnK00,11
-xinference-1.3.0.dist-info/RECORD,,
+xinference-1.3.0.post2.dist-info/LICENSE,sha256=QwcOLU5TJoTeUhuIXzhdCEEDDvorGiC6-3YTOl4TecE,11356
+xinference-1.3.0.post2.dist-info/METADATA,sha256=kFcBcCj6HdKTIrJASMPOHD69Lfa-TIliWblCYV6_1Bk,24127
+xinference-1.3.0.post2.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
+xinference-1.3.0.post2.dist-info/entry_points.txt,sha256=-lDyyzqWMFQF0Rgm7VxBNz0V-bMBMQLRR3pvQ-Y8XTY,226
+xinference-1.3.0.post2.dist-info/top_level.txt,sha256=L1rQt7pl6m8tmKXpWVHzP-GtmzAxp663rXxGE7qnK00,11
+xinference-1.3.0.post2.dist-info/RECORD,,