README.md edited
Browse files
README.md
CHANGED
|
@@ -1,7 +1,344 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{\rtf1\ansi\ansicpg1252\cocoartf2639
|
| 2 |
+
\cocoatextscaling0\cocoaplatform0{\fonttbl\f0\froman\fcharset0 Times-Bold;\f1\froman\fcharset0 Times-Roman;\f2\fmodern\fcharset0 Courier;
|
| 3 |
+
\f3\fmodern\fcharset0 Courier-Bold;\f4\fnil\fcharset0 Menlo-Regular;\f5\fnil\fcharset0 AppleColorEmoji;
|
| 4 |
+
}
|
| 5 |
+
{\colortbl;\red255\green255\blue255;\red0\green0\blue0;\red0\green0\blue233;\red109\green109\blue109;
|
| 6 |
+
\red0\green0\blue0;}
|
| 7 |
+
{\*\expandedcolortbl;;\cssrgb\c0\c0\c0;\cssrgb\c0\c0\c93333;\cssrgb\c50196\c50196\c50196;
|
| 8 |
+
\csgray\c0;}
|
| 9 |
+
{\*\listtable{\list\listtemplateid1\listhybrid{\listlevel\levelnfc23\levelnfcn23\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{disc\}}{\leveltext\leveltemplateid1\'01\uc0\u8226 ;}{\levelnumbers;}\fi-360\li720\lin720 }{\listname ;}\listid1}
|
| 10 |
+
{\list\listtemplateid2\listhybrid{\listlevel\levelnfc23\levelnfcn23\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{disc\}}{\leveltext\leveltemplateid101\'01\uc0\u8226 ;}{\levelnumbers;}\fi-360\li720\lin720 }{\listname ;}\listid2}
|
| 11 |
+
{\list\listtemplateid3\listhybrid{\listlevel\levelnfc23\levelnfcn23\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{disc\}}{\leveltext\leveltemplateid201\'01\uc0\u8226 ;}{\levelnumbers;}\fi-360\li720\lin720 }{\listname ;}\listid3}
|
| 12 |
+
{\list\listtemplateid4\listhybrid{\listlevel\levelnfc23\levelnfcn23\leveljc0\leveljcn0\levelfollow0\levelstartat1\levelspace360\levelindent0{\*\levelmarker \{disc\}}{\leveltext\leveltemplateid301\'01\uc0\u8226 ;}{\levelnumbers;}\fi-360\li720\lin720 }{\listname ;}\listid4}}
|
| 13 |
+
{\*\listoverridetable{\listoverride\listid1\listoverridecount0\ls1}{\listoverride\listid2\listoverridecount0\ls2}{\listoverride\listid3\listoverridecount0\ls3}{\listoverride\listid4\listoverridecount0\ls4}}
|
| 14 |
+
\margl1440\margr1440\vieww24620\viewh8360\viewkind0
|
| 15 |
+
\deftab720
|
| 16 |
+
\pard\pardeftab720\sa321\partightenfactor0
|
| 17 |
|
| 18 |
+
\f0\b\fs48 \cf0 \expnd0\expndtw0\kerning0
|
| 19 |
+
\outl0\strokewidth0 \strokec2 Q2 Debugging LLMs (Friendli AI Take-Home)\
|
| 20 |
+
\pard\pardeftab720\sa240\partightenfactor0
|
| 21 |
+
|
| 22 |
+
\fs24 \cf0 Fixed Model Repo:
|
| 23 |
+
\f1\b0 \uc0\u8232 {\field{\*\fldinst{HYPERLINK "https://huggingface.co/syedahmedsoftware/broken-model-fixed"}}{\fldrslt \cf3 \ul \ulc3 \strokec3 https://huggingface.co/syedahmedsoftware/broken-model-fixed}}\
|
| 24 |
+
|
| 25 |
+
\f0\b Original Model:
|
| 26 |
+
\f1\b0 \uc0\u8232 {\field{\*\fldinst{HYPERLINK "https://huggingface.co/yunmorning/broken-model?utm_source=chatgpt.com"}}{\fldrslt \cf3 \ul \ulc3 \strokec3 https://huggingface.co/yunmorning/broken-model}}\
|
| 27 |
+
This repository contains the
|
| 28 |
+
\f0\b minimal, production-safe fixes
|
| 29 |
+
\f1\b0 required to make the original model usable behind an OpenAI-compatible
|
| 30 |
+
\f2\fs26 /chat/completions
|
| 31 |
+
\f1\fs24 API server.\uc0\u8232
|
| 32 |
+
\f0\b No model weights were modified.
|
| 33 |
+
\f1\b0 Only configuration metadata was corrected.\
|
| 34 |
+
\pard\pardeftab720\partightenfactor0
|
| 35 |
+
\cf0 \
|
| 36 |
+
\pard\pardeftab720\sa298\partightenfactor0
|
| 37 |
+
|
| 38 |
+
\f0\b\fs36 \cf0 Problem (a) \'97 Why inference failed\
|
| 39 |
+
\pard\pardeftab720\sa240\partightenfactor0
|
| 40 |
+
|
| 41 |
+
\f1\b0\fs24 \cf0 The original repository could not serve chat requests because
|
| 42 |
+
\f0\b two required metadata components were missing or incorrect
|
| 43 |
+
\f1\b0 .\
|
| 44 |
+
\pard\pardeftab720\sa280\partightenfactor0
|
| 45 |
+
|
| 46 |
+
\f0\b\fs28 \cf0 Root Causes\
|
| 47 |
+
\pard\pardeftab720\sa319\partightenfactor0
|
| 48 |
+
|
| 49 |
+
\fs24 \cf0 1) Missing
|
| 50 |
+
\f3\fs26 chat_template
|
| 51 |
+
\f0\fs24 (tokenizer_config.json)\
|
| 52 |
+
\pard\pardeftab720\sa240\partightenfactor0
|
| 53 |
+
|
| 54 |
+
\f1\b0 \cf0 Modern OpenAI-style runtimes rely on:\
|
| 55 |
+
python\
|
| 56 |
+
\pard\pardeftab720\partightenfactor0
|
| 57 |
+
|
| 58 |
+
\f2\fs26 \cf0 tokenizer.apply_chat_template(messages)\
|
| 59 |
+
\
|
| 60 |
+
\
|
| 61 |
+
\pard\pardeftab720\sa240\partightenfactor0
|
| 62 |
+
|
| 63 |
+
\f1\fs24 \cf0 The original model
|
| 64 |
+
\f0\b did not define a chat template
|
| 65 |
+
\f1\b0 , so:\
|
| 66 |
+
\pard\tx220\tx720\pardeftab720\li720\fi-720\sa240\partightenfactor0
|
| 67 |
+
\ls1\ilvl0\cf0 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext \uc0\u8226 }\expnd0\expndtw0\kerning0
|
| 68 |
+
\outl0\strokewidth0 \strokec2 Chat messages could not be rendered into prompts\
|
| 69 |
+
\ls1\ilvl0
|
| 70 |
+
\f2\fs26 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext \uc0\u8226 }\expnd0\expndtw0\kerning0
|
| 71 |
+
\outl0\strokewidth0 \strokec2 /chat/completions
|
| 72 |
+
\f1\fs24 servers had no deterministic formatting\
|
| 73 |
+
\ls1\ilvl0\kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext \uc0\u8226 }\expnd0\expndtw0\kerning0
|
| 74 |
+
\outl0\strokewidth0 \strokec2 Inference failed or produced undefined behavior\
|
| 75 |
+
\pard\pardeftab720\partightenfactor0
|
| 76 |
+
|
| 77 |
+
\f2\fs26 \cf0 \strokec2 \
|
| 78 |
+
\pard\pardeftab720\sa319\partightenfactor0
|
| 79 |
+
|
| 80 |
+
\f0\b\fs24 \cf0 2) Missing/incorrect
|
| 81 |
+
\f3\fs26 pad_token_id
|
| 82 |
+
\f0\fs24 (config.json)\
|
| 83 |
+
\pard\pardeftab720\sa240\partightenfactor0
|
| 84 |
+
|
| 85 |
+
\f1\b0 \cf0 Production inference uses:\
|
| 86 |
+
\pard\tx220\tx720\pardeftab720\li720\fi-720\sa240\partightenfactor0
|
| 87 |
+
\ls2\ilvl0\cf0 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext \uc0\u8226 }\expnd0\expndtw0\kerning0
|
| 88 |
+
\outl0\strokewidth0 \strokec2 batched decoding\
|
| 89 |
+
\ls2\ilvl0\kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext \uc0\u8226 }\expnd0\expndtw0\kerning0
|
| 90 |
+
\outl0\strokewidth0 \strokec2 attention masking\
|
| 91 |
+
\ls2\ilvl0\kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext \uc0\u8226 }\expnd0\expndtw0\kerning0
|
| 92 |
+
\outl0\strokewidth0 \strokec2 padded sequences\
|
| 93 |
+
\pard\pardeftab720\sa240\partightenfactor0
|
| 94 |
+
\cf0 \strokec2 The original
|
| 95 |
+
\f2\fs26 config.json
|
| 96 |
+
\f1\fs24 did not define a valid
|
| 97 |
+
\f2\fs26 pad_token_id
|
| 98 |
+
\f1\fs24 , making batching unsafe and causing runtime instability.\
|
| 99 |
+
\pard\pardeftab720\partightenfactor0
|
| 100 |
+
|
| 101 |
+
\f2\fs26 \cf0 \
|
| 102 |
+
\
|
| 103 |
+
\pard\pardeftab720\sa298\partightenfactor0
|
| 104 |
+
|
| 105 |
+
\f0\b\fs36 \cf0 Minimal Fixes Applied (No Weight Changes)\
|
| 106 |
+
|
| 107 |
+
\itap1\trowd \taflags0 \trgaph108\trleft-108 \trbrdrt\brdrnil \trbrdrl\brdrnil \trbrdrr\brdrnil
|
| 108 |
+
\clvertalc \clshdrawnil \clwWidth3432\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx2880
|
| 109 |
+
\clvertalc \clshdrawnil \clwWidth4607\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx5760
|
| 110 |
+
\clvertalc \clshdrawnil \clwWidth4118\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx8640
|
| 111 |
+
\pard\intbl\itap1\pardeftab720\qc\partightenfactor0
|
| 112 |
+
|
| 113 |
+
\fs24 \cf0 \strokec2 File\cell
|
| 114 |
+
\pard\intbl\itap1\pardeftab720\qc\partightenfactor0
|
| 115 |
+
\cf0 Change\cell
|
| 116 |
+
\pard\intbl\itap1\pardeftab720\qc\partightenfactor0
|
| 117 |
+
\cf0 Why\cell \row
|
| 118 |
+
|
| 119 |
+
\itap1\trowd \taflags0 \trgaph108\trleft-108 \trbrdrl\brdrnil \trbrdrr\brdrnil
|
| 120 |
+
\clvertalc \clshdrawnil \clwWidth3432\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx2880
|
| 121 |
+
\clvertalc \clshdrawnil \clwWidth4607\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx5760
|
| 122 |
+
\clvertalc \clshdrawnil \clwWidth4118\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx8640
|
| 123 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 124 |
+
|
| 125 |
+
\f2\b0\fs26 \cf0 tokenizer_config.json
|
| 126 |
+
\f1\fs24 \cell
|
| 127 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 128 |
+
\cf0 Added ChatML-style
|
| 129 |
+
\f2\fs26 chat_template
|
| 130 |
+
\f1\fs24 \cell
|
| 131 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 132 |
+
\cf0 Required for OpenAI-style chat formatting\cell \row
|
| 133 |
+
|
| 134 |
+
\itap1\trowd \taflags0 \trgaph108\trleft-108 \trbrdrl\brdrnil \trbrdrr\brdrnil
|
| 135 |
+
\clvertalc \clshdrawnil \clwWidth3432\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx2880
|
| 136 |
+
\clvertalc \clshdrawnil \clwWidth4607\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx5760
|
| 137 |
+
\clvertalc \clshdrawnil \clwWidth4118\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx8640
|
| 138 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 139 |
+
|
| 140 |
+
\f2\fs26 \cf0 config.json
|
| 141 |
+
\f1\fs24 \cell
|
| 142 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 143 |
+
\cf0 Set
|
| 144 |
+
\f2\fs26 pad_token_id
|
| 145 |
+
\f1\fs24 to tokenizer\'92s pad token\cell
|
| 146 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 147 |
+
\cf0 Enables safe batching and attention masks\cell \row
|
| 148 |
+
|
| 149 |
+
\itap1\trowd \taflags0 \trgaph108\trleft-108 \trbrdrl\brdrnil \trbrdrt\brdrnil \trbrdrr\brdrnil
|
| 150 |
+
\clvertalc \clshdrawnil \clwWidth3432\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx2880
|
| 151 |
+
\clvertalc \clshdrawnil \clwWidth4607\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx5760
|
| 152 |
+
\clvertalc \clshdrawnil \clwWidth4118\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx8640
|
| 153 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 154 |
+
|
| 155 |
+
\f2\fs26 \cf0 generation_config.json
|
| 156 |
+
\f1\fs24 \cell
|
| 157 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 158 |
+
\cf0 Normalized pad/eos fields (kept defaults)\cell
|
| 159 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 160 |
+
\cf0 Prevents decoding edge cases\cell \lastrow\row
|
| 161 |
+
\pard\pardeftab720\sa240\partightenfactor0
|
| 162 |
+
\cf0 \strokec2 \
|
| 163 |
+
No architecture, tokenizer vocab, or model weights were changed.\
|
| 164 |
+
\
|
| 165 |
+
\pard\pardeftab720\partightenfactor0
|
| 166 |
+
|
| 167 |
+
\f0\b \cf0 Verification (Remote)\
|
| 168 |
+
\
|
| 169 |
+
\
|
| 170 |
+
\pard\tx560\tx1120\tx1680\tx2240\tx2800\tx3360\tx3920\tx4480\tx5040\tx5600\tx6160\tx6720\pardirnatural\partightenfactor0
|
| 171 |
+
|
| 172 |
+
\f4\b0\fs22 \cf5 \kerning1\expnd0\expndtw0 \CocoaLigature0 \outl0\strokewidth0 python - <<'PY'\
|
| 173 |
+
from transformers import AutoTokenizer\
|
| 174 |
+
\
|
| 175 |
+
repo_id = "syedahmedsoftware/broken-model-fixed"\
|
| 176 |
+
tok = AutoTokenizer.from_pretrained(repo_id, use_fast=True)\
|
| 177 |
+
\
|
| 178 |
+
print("tokenizer:", tok.__class__.__name__)\
|
| 179 |
+
print("has_chat_template:", bool(getattr(tok, "chat_template", None)))\
|
| 180 |
+
print("pad_token_id:", tok.pad_token_id, "eos_token_id:", tok.eos_token_id)\
|
| 181 |
+
\
|
| 182 |
+
messages = [\
|
| 183 |
+
\{"role":"system","content":"You are helpful."\},\
|
| 184 |
+
\{"role":"user","content":"Say hello."\}\
|
| 185 |
+
]\
|
| 186 |
+
prompt = tok.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)\
|
| 187 |
+
print("
|
| 188 |
+
\f5 \uc0\u9989
|
| 189 |
+
\f4 REMOTE prompt renders. Preview:")\
|
| 190 |
+
print(prompt[:250])\
|
| 191 |
+
PY\
|
| 192 |
+
\
|
| 193 |
+
Expected output:\
|
| 194 |
+
\
|
| 195 |
+
tokenizer: Qwen2Tokenizer\
|
| 196 |
+
has_chat_template: True\
|
| 197 |
+
pad_token_id: 151643 eos_token_id: 151645\
|
| 198 |
+
|
| 199 |
+
\f5 \uc0\u9989
|
| 200 |
+
\f4 REMOTE prompt renders. Preview:\
|
| 201 |
+
<|im_start|>system\
|
| 202 |
+
You are helpful.<|im_end|>\
|
| 203 |
+
<|im_start|>user\
|
| 204 |
+
Say hello.<|im_end|>\
|
| 205 |
+
<|im_start|>assistant\
|
| 206 |
+
\
|
| 207 |
+
\
|
| 208 |
+
\
|
| 209 |
+
\pard\pardeftab720\sa298\partightenfactor0
|
| 210 |
+
|
| 211 |
+
\f0\b\fs36 \cf0 \expnd0\expndtw0\kerning0
|
| 212 |
+
\CocoaLigature1 \outl0\strokewidth0 \strokec2 Problem (b) \'97 Why
|
| 213 |
+
\f3\fs39 \strokec2 reasoning_effort
|
| 214 |
+
\f0\fs36 \strokec2 has no effect\
|
| 215 |
+
\pard\pardeftab720\sa280\partightenfactor0
|
| 216 |
+
|
| 217 |
+
\fs28 \cf0 Root Cause\
|
| 218 |
+
\pard\pardeftab720\sa240\partightenfactor0
|
| 219 |
+
|
| 220 |
+
\f2\b0\fs26 \cf0 reasoning_effort
|
| 221 |
+
\f1\fs24 is
|
| 222 |
+
\f0\b not a native Transformers generation parameter
|
| 223 |
+
\f1\b0 .\uc0\u8232 Unless the
|
| 224 |
+
\f0\b serving runtime
|
| 225 |
+
\f1\b0 explicitly maps it to real compute policies, it is silently ignored.\
|
| 226 |
+
The base model has
|
| 227 |
+
\f0\b no internal mechanism
|
| 228 |
+
\f1\b0 to interpret \'93effort.\'94\
|
| 229 |
+
\pard\pardeftab720\partightenfactor0
|
| 230 |
+
\cf0 \
|
| 231 |
+
\pard\pardeftab720\sa298\partightenfactor0
|
| 232 |
+
|
| 233 |
+
\f0\b\fs36 \cf0 What is required to make
|
| 234 |
+
\f3\fs39 \strokec2 reasoning_effort
|
| 235 |
+
\f0\fs36 \strokec2 meaningful\
|
| 236 |
+
\pard\pardeftab720\sa280\partightenfactor0
|
| 237 |
+
|
| 238 |
+
\fs28 \cf0 1) Runtime orchestration (required)\
|
| 239 |
+
\pard\pardeftab720\sa240\partightenfactor0
|
| 240 |
+
|
| 241 |
+
\f1\b0\fs24 \cf0 The server must map
|
| 242 |
+
\f2\fs26 reasoning_effort
|
| 243 |
+
\f1\fs24 to actual strategies, for example:\
|
| 244 |
+
|
| 245 |
+
\itap1\trowd \taflags0 \trgaph108\trleft-108 \trbrdrt\brdrnil \trbrdrl\brdrnil \trbrdrr\brdrnil
|
| 246 |
+
\clvertalc \clshdrawnil \clwWidth1246\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx4320
|
| 247 |
+
\clvertalc \clshdrawnil \clwWidth3074\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx8640
|
| 248 |
+
\pard\intbl\itap1\pardeftab720\qc\partightenfactor0
|
| 249 |
+
|
| 250 |
+
\f0\b \cf0 \strokec2 Effort Level\cell
|
| 251 |
+
\pard\intbl\itap1\pardeftab720\qc\partightenfactor0
|
| 252 |
+
\cf0 Runtime Policy\cell \row
|
| 253 |
+
|
| 254 |
+
\itap1\trowd \taflags0 \trgaph108\trleft-108 \trbrdrl\brdrnil \trbrdrr\brdrnil
|
| 255 |
+
\clvertalc \clshdrawnil \clwWidth1246\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx4320
|
| 256 |
+
\clvertalc \clshdrawnil \clwWidth3074\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx8640
|
| 257 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 258 |
+
|
| 259 |
+
\f1\b0 \cf0 low\cell
|
| 260 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 261 |
+
\cf0 single pass, greedy\cell \row
|
| 262 |
+
|
| 263 |
+
\itap1\trowd \taflags0 \trgaph108\trleft-108 \trbrdrl\brdrnil \trbrdrr\brdrnil
|
| 264 |
+
\clvertalc \clshdrawnil \clwWidth1246\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx4320
|
| 265 |
+
\clvertalc \clshdrawnil \clwWidth3074\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx8640
|
| 266 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 267 |
+
\cf0 medium\cell
|
| 268 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 269 |
+
\cf0 temperature + nucleus sampling\cell \row
|
| 270 |
+
|
| 271 |
+
\itap1\trowd \taflags0 \trgaph108\trleft-108 \trbrdrl\brdrnil \trbrdrr\brdrnil
|
| 272 |
+
\clvertalc \clshdrawnil \clwWidth1246\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx4320
|
| 273 |
+
\clvertalc \clshdrawnil \clwWidth3074\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx8640
|
| 274 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 275 |
+
\cf0 high\cell
|
| 276 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 277 |
+
\cf0 multi-sample + rerank\cell \row
|
| 278 |
+
|
| 279 |
+
\itap1\trowd \taflags0 \trgaph108\trleft-108 \trbrdrl\brdrnil \trbrdrt\brdrnil \trbrdrr\brdrnil
|
| 280 |
+
\clvertalc \clshdrawnil \clwWidth1246\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx4320
|
| 281 |
+
\clvertalc \clshdrawnil \clwWidth3074\clftsWidth3 \clmart10 \clmarl10 \clmarb10 \clmarr10 \clbrdrt\brdrnil \clbrdrl\brdrnil \clbrdrb\brdrnil \clbrdrr\brdrnil \clpadt20 \clpadl20 \clpadb20 \clpadr20 \gaph\cellx8640
|
| 282 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 283 |
+
\cf0 very_high\cell
|
| 284 |
+
\pard\intbl\itap1\pardeftab720\partightenfactor0
|
| 285 |
+
\cf0 tree-of-thought / verifier loop\cell \lastrow\row
|
| 286 |
+
\pard\pardeftab720\sa240\partightenfactor0
|
| 287 |
+
\cf0 \strokec2 The runtime must log which policy is applied.\
|
| 288 |
+
\pard\pardeftab720\sa280\partightenfactor0
|
| 289 |
+
|
| 290 |
+
\f0\b\fs28 \cf0 2) Architectural support (one of)\
|
| 291 |
+
\pard\tx220\tx720\pardeftab720\li720\fi-720\sa240\partightenfactor0
|
| 292 |
+
\ls3\ilvl0
|
| 293 |
+
\f1\b0\fs24 \cf0 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext \uc0\u8226 }\expnd0\expndtw0\kerning0
|
| 294 |
+
\outl0\strokewidth0 \strokec2 Multi-pass verifier loop\
|
| 295 |
+
\ls3\ilvl0\kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext \uc0\u8226 }\expnd0\expndtw0\kerning0
|
| 296 |
+
\outl0\strokewidth0 \strokec2 Tree search / self-consistency\
|
| 297 |
+
\ls3\ilvl0\kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext \uc0\u8226 }\expnd0\expndtw0\kerning0
|
| 298 |
+
\outl0\strokewidth0 \strokec2 Reflection / critique agent\
|
| 299 |
+
\ls3\ilvl0\kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext \uc0\u8226 }\expnd0\expndtw0\kerning0
|
| 300 |
+
\outl0\strokewidth0 \strokec2 Budgeted decoding controller\
|
| 301 |
+
\ls3\ilvl0\kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext \uc0\u8226 }\expnd0\expndtw0\kerning0
|
| 302 |
+
\outl0\strokewidth0 \strokec2 Control-token trained model\
|
| 303 |
+
\pard\pardeftab720\sa240\partightenfactor0
|
| 304 |
+
\cf0 \strokec2 Without these,
|
| 305 |
+
\f2\fs26 reasoning_effort
|
| 306 |
+
\f1\fs24 remains a
|
| 307 |
+
\f0\b no-op
|
| 308 |
+
\f1\b0 .\
|
| 309 |
+
\pard\pardeftab720\partightenfactor0
|
| 310 |
+
\cf0 \
|
| 311 |
+
\pard\pardeftab720\sa298\partightenfactor0
|
| 312 |
+
|
| 313 |
+
\f0\b\fs36 \cf0 Final Notes\
|
| 314 |
+
\pard\pardeftab720\sa240\partightenfactor0
|
| 315 |
+
|
| 316 |
+
\f1\b0\fs24 \cf0 These fixes:\
|
| 317 |
+
\pard\tx220\tx720\pardeftab720\li720\fi-720\sa240\partightenfactor0
|
| 318 |
+
\ls4\ilvl0\cf0 \kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext \uc0\u8226 }\expnd0\expndtw0\kerning0
|
| 319 |
+
\outl0\strokewidth0 \strokec2 Restore deterministic chat formatting\
|
| 320 |
+
\ls4\ilvl0\kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext \uc0\u8226 }\expnd0\expndtw0\kerning0
|
| 321 |
+
\outl0\strokewidth0 \strokec2 Enable production-safe batching\
|
| 322 |
+
\ls4\ilvl0\kerning1\expnd0\expndtw0 \outl0\strokewidth0 {\listtext \uc0\u8226 }\expnd0\expndtw0\kerning0
|
| 323 |
+
\outl0\strokewidth0 \strokec2 Make the model compatible with OpenAI-style
|
| 324 |
+
\f2\fs26 /chat/completions
|
| 325 |
+
\f1\fs24 servers\
|
| 326 |
+
\pard\pardeftab720\sa240\partightenfactor0
|
| 327 |
+
\cf0 \strokec2 The model is now deployable in real inference environments.\
|
| 328 |
+
\pard\tx560\tx1120\tx1680\tx2240\tx2800\tx3360\tx3920\tx4480\tx5040\tx5600\tx6160\tx6720\pardirnatural\partightenfactor0
|
| 329 |
+
|
| 330 |
+
\f4\fs22 \cf5 \kerning1\expnd0\expndtw0 \CocoaLigature0 \outl0\strokewidth0 \
|
| 331 |
+
\
|
| 332 |
+
\pard\pardeftab720\partightenfactor0
|
| 333 |
+
|
| 334 |
+
\f0\b\fs24 \cf0 \expnd0\expndtw0\kerning0
|
| 335 |
+
\CocoaLigature1 \outl0\strokewidth0 \strokec2 \
|
| 336 |
+
|
| 337 |
+
\f1\b0 \
|
| 338 |
+
\pard\pardeftab720\partightenfactor0
|
| 339 |
+
|
| 340 |
+
\f2\fs26 \cf0 \
|
| 341 |
+
\
|
| 342 |
+
\
|
| 343 |
+
\
|
| 344 |
+
}
|