.`iu+ d dl Zd dlZd dlmZ d dlmZ d dlZerd dl m
Z
d dl m Z d dl
mZ d dlmZ d dlmZ d d lmZ ee Z G d
d ej Z ej d
G d d Zdedede edz edz f fdZ ej G d d Z G d de Z dS ) N)Sequence)
TYPE_CHECKING)
TokenizerLike)ChatCompletionRequest) DeltaMessage)ResponsesRequest) init_logger)ReasoningParserc e Zd ZdZdZdS )Olmo3ReasoningState N)__name__
__module__ __qualname__ REASONINGCONTENT y/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/vllm/reasoning/olmo3_reasoning_parser.pyr r s IGGGr r T)frozenc * e Zd ZU eed< eed< d ZdS )Indicesstartendc | j | j z
S N)r r selfs r __len__zIndices.__len__& s x$*$$r N)r r r int__annotations__r r r r r r ! s7 JJJ HHH% % % % %r r abreturnc T t | t | k r| |dfn|| df\ } }}| |v rmt dt | }t | | | | t | z }|r||fn||fS t t | dz
dd D ]b}| | d |d| k rIt t | |z
t | }t d| }|r||fn||fc S ct t | dz
dd D ]b}|| d | d| k rIt d| }t t | |z
t | }|r||fn||fc S cdS )a.
Find the longest overlap where the end of string a matches the start
of string b.
Args:
a: First string
b: Second string
Returns:
Tuple of IndicesTuples representing the overlapping portions in each
string, or a tuple of None if no overlap exists
FTr r
N)NN)lenr indexrange)r# r$ swapind_aind_bis r string_overlapr/ * s #&a&&3q66//!Q 1d|JAq$ Avv3q66""
AGGAJJQ$788!%9u~~E5>9 3q66A:q"
%
% > > aRSS6QrrU??CFFQJA//EAqMME%)=E5>>u~ = = = 3q66A:q"
%
% > > aRSS6QrrU??AqMMECFFQJA//E%)=E5>>u~ = = =
:r c e Zd ZU dZeed< dZeed< dZeed< ej Z
eed< de d z fd
Z d Z
d ede d z fd
Zd S )Olmo3ReasoningBuffer think_start think_end bufferstater% Nc | j | j }|dk r`t j | _ | j d | | j |t
| j z d c}| _ |dk rt | S | j | j }|dk r`t j
| _ | j d | | j |t
| j z d c}| _ |dk rt | S | j t j k r | j dc}| _ t | S | j t j
k r | j dc}| _ t | S d S )Nr )content) reasoningr6 ) r7 findr3 r r r8 r( r rfindr5 r )r start_think_idxpretext
end_think_idx text_buffers r process_bufferz#Olmo3ReasoningBuffer.process_buffer` s +**4+;<< a ,6DJ ,_ ,- Oc$2B.C.C C E EF
!GT[ "" $G4444 ))$.99
A ,4DJ N]N+ MC,?,? ? A AB
!GT[ q $g6666 :,6 6 6 R
+666 6 :,4 4 4 R
444 4 tr c * t | j S r )r( r7 r s r r z Olmo3ReasoningBuffer.__len__ s 4; r
delta_textc | xj |z
c_ d }t || j \ }}t || j \ }}|d uo$t | t | j k }|d uo$t | t | j k }|r%| j | j v r|s| }n?|r#| j | j v r| }n|s|rd S | }|S r )r7 r/ r3 r5 r( rB )r rD
delta_message_overlap_think_startoverlap_think_endpartial_overlap_startpartial_overlap_ends r add_textz Olmo3ReasoningBuffer.add_text sG z! .2
"0
DE >>>. t 3 . . . .` / 3 / <$+> / / / / / / r r1 c e Zd ZdZd fd Zdee defdZde e de e fdZ
d e d
e e
z dee d z e d z f fd Zd
e de de dee dee dee ded z fdZ xZS )Olmo3ReasoningParsera
Reasoning parser for Olmo 3 model
Olmo3ReasoningParser
This class implements a reasoning parser specifically designed for the
Olmo 3 family of models. Olmo 3 models do not use special tokens to
indicate reasoning; rather, reasoning trace is wrapped in `` and
``, which are tokenized using standard vocabulary entries.
Because of this, the parser operates in string space, accumulating the
characters in a buffer until it sees `` or ``. tokens
to switch modes.
Key Features:
- For non-stream output, Recognizes and extracts reasoning (text
bracketed by `` and ``) and content (everything
after the first ``).
- For stream process, it uses a buffer to accumulate delta text,
and output progressive delta messages as soon as thinking starts
or ends.
- For reliability, some Olmo 3 models may hardcode the first
`` token is the input text (similar to Deepseek R1,
or reasoning-only Qwen models). To support such variants, the
parser can optionally work in cases where the first ``
token is missing from generation.
tokenizerr c t j |g|R i | d| _ d| _ d| j d| j dz }t j |t j | _ t | j | j | _ d S )Nr2 r4 z^(?:z)?(?P.*?)z(?P.*)$)r3 r5 )
super__init__r3 r5 recompileDOTALLreasoning_regexr1 r7 )r rP argskwargsreasoning_expr __class__s r rS z Olmo3ReasoningParser.__init__ s 4T444V444%$
;D$ : : :.222
3 "z.") D D * (DN
r input_idsr% c H | j | }| j |v S r )model_tokenizerdecoder5 )r r\ texts r is_reasoning_endz%Olmo3ReasoningParser.is_reasoning_end s% #**955~%%r c g S r r )r r\ s r extract_content_idsz(Olmo3ReasoningParser.extract_content_ids s
r model_outputrequestNc | j | }|r2| d pd}| d pd}||fS d|fS )a1 Extract the reasoning content & content sections, respectively.
If the sequence doesn't match what we expect, i.e., the model generates
something else, all content is considered non-reasoning content.
Args:
model_output (str): Output of the model to be parsed.
request (ChatCompletionRequest | ResponsesRequest): Request being
processed.
Returns:
tuple[Optional[str], Optional[str]]: Tuple pair containing the
reasoning content and non-reasoning content.
r; Nr: )rW matchgroup)r rd re re_matchr; r: s r extract_reasoningz&Olmo3ReasoningParser.extract_reasoning sf & '--l;; & {33;tI nnY//74G g% % \!!r
previous_text current_textrD previous_token_idscurrent_token_idsdelta_token_idsc | j | }|1| j j | j j v r| j }|S )z5Extract content using token ID sequence state machine)r7 rL r5 rB )r rk rl rD rm rn ro rF s r extract_reasoning_streamingz0Olmo3ReasoningParser.extract_reasoning_streaming sN ,,Z88
T[%:dk>P%P%P !K 6 6 8 8M r )rP r )r r r __doc__rS r r! boolra listrc rM r r tuplerj r rq
__classcell__)r[ s @r rO rO sH 6
&&(3- &D & & & &T#Y 49 "" ')99"
sTz3:% & " " " "8
%SM $C=
"#
r rO ) dataclassesdtenumcollections.abcr typingr regexrT vllm.tokenizersr 0vllm.entrypoints.openai.chat_completion.protocolr 'vllm.entrypoints.openai.engine.protocolr *vllm.entrypoints.openai.responses.protocolr vllm.loggerr vllm.reasoningr
r loggerEnumr dataclassr rM ru r/ r1 rO r r r r s $ $ $ $ $ $ .------ $ # # # # # * * * * * * X $)
T% % % % % % % %'c 'c 'eGdNGdN,J&K ' ' ' 'T n n n n n n n n bk k k k k ? k k k k k r