+
    UV-jK  ã                   óŒ   € ^ RI t ^ RIt^ RIt^ RIt^RIHtHtHtHt ^RI	H
t
HtHtHtHt RtRt]P"                  ! R4      tR tR# )é    N)ÚDEFAULT_KV_GROUP_SIZEÚDEFAULT_KV_QUANT_SCHEMEÚDEFAULT_PREFILL_STEP_SIZEÚDEFAULT_QUANTIZED_KV_START)ÚDEFAULT_ENABLE_THINKINGÚget_server_max_tokensÚget_server_thinking_budgetÚget_server_thinking_end_tokenÚget_server_thinking_start_tokenz0.0.0.0i�  zmlx_vlm.serverc            	      ó  € \         P                  ! R R7      p V P                  R\        \        RR7       V P                  R\
        \        RR7       V P                  RRR	R
7       V P                  R\        RRR7       V P                  R\        RRR7       V P                  R\
        ^RR7       V P                  R\
        \        RR7       V P                  R\
        \        4       RR7       V P                  RR\        RR7       V P                  R\
        \        4       RR7       V P                  R\        \        4       RR7       V P                  RRR\        \        4       R R!7       V P                  R"\        RR#R7       V P                  R$\        RV\        R%R&7       V P                  R'\
        \        R(R7       V P                  R)\
        RR*R7       V P                  R+\
        \         R,R7       V P                  R-\        RR.R7       V P                  R/\        R. RWOR0R17       V P                  R2\
        RR3R7       V P                  R4\
        RR5R7       V P                  R6RR7R8R7       V P                  R9\        R:. RXOR;R17       V P#                  4       pVP$                  '       d   R<\&        P(                  R=&   VP*                  '       dM   VP*                  \&        P(                  R>&   VP,                  '       d   VP,                  \&        P(                  R?&   \        VP.                  4      \&        P(                  R@&   VP0                  '       d}   VP0                  \&        P(                  RA&   VP2                  e   VP2                  \&        P(                  RB&   VP4                  e'   \        VP4                  4      \&        P(                  RC&   VP6                  '       d'   \        VP6                  4      \&        P(                  RD&   \        VP8                  4      \&        P(                  RE&   VP:                  '       d   RFMRG\&        P(                  RH&   VP<                  e'   \        VP<                  4      \&        P(                  RI&   VP>                  e   VP>                  \&        P(                  RJ&   VP@                  e   VP@                  \&        P(                  RK&   VPB                  e'   \        VPB                  4      \&        P(                  RL&   \        VPD                  4      \&        P(                  RM&   VPF                  \&        P(                  RN&   VPH                  e'   \        VPH                  4      \&        P(                  RO&   \        VPJ                  4      \&        P(                  RP&   VPL                  e'   \        VPL                  4      \&        P(                  RQ&   \O        \P        VPR                  PU                  4       \P        PV                  4      p\P        PX                  ! VRRRS7       \Z        P]                  V4       \^        P`                  ! RTVPb                  VPd                  ^VPf                  R7RU7       R# )YzMLX VLM Http Server.)Údescriptionz--hostz*Host for the HTTP server (default:0.0.0.0))ÚtypeÚdefaultÚhelpz--portz(Port for the HTTP server (default: 8080)z--trust-remote-codeÚ
store_truez<Trust remote code when loading models from Hugging Face Hub.)Úactionr   z--modelNzMPre-load a model at startup (e.g. mlx-community/Qwen2.5-VL-3B-Instruct-4bit).z--adapter-pathz'Adapter weights to load with the model.z--vision-cache-sizez3Max number of cached vision features (default: 20).z--prefill-step-sizez/Tokens per prefill step (default: %(default)s).z--max-tokensz%Maximum number of tokens to generate.z--enable-thinkingzXEnable thinking mode by default for requests that do not set enable_thinking explicitly.)r   r   r   z--thinking-budgetzrDefault maximum number of tokens allowed inside a thinking block. Requests can override this with thinking_budget.z--thinking-start-tokenz`Default token that opens a thinking block. Requests can override this with thinking_start_token.z--thinking-end-tokenz--thinking-eos-tokenÚthinking_end_tokenz_Default token that closes a thinking block. Requests can override this with thinking_end_token.)Údestr   r   r   z	--kv-bitszCNumber of bits for KV cache quantization (e.g. 3.5 for TurboQuant).z--kv-quant-schemezKV cache quantization backend.)r   Úchoicesr   r   z--kv-group-sizez-Group size for uniform KV cache quantization.z--max-kv-sizez Maximum KV cache size in tokens.z--quantized-kv-startz#Start index for quantized KV cache.z--draft-modelzbSpeculative drafter path or HF id (e.g. z-lab/Qwen3.5-4B-DFlash, google/gemma-4-31B-it-assistant).z--draft-kindzrDrafter family -- 'dflash', 'eagle3', or 'mtp' (Gemma 4). Default: auto-detected from the drafter's HF model_type.)r   r   r   r   z--draft-block-sizez-Override the drafter's configured block size.z--top-logprobs-kzlServer-side cap for per-token top_logprobs (0-20, default 0 = disabled). Maps to the TOP_LOGPROBS_K env var.z--reloadFz#Enable auto-reload for development.z--log-levelÚINFOz&Set the logging level (default: INFO).ÚtrueÚMLX_TRUST_REMOTE_CODEÚMLX_VLM_PRELOAD_MODELÚMLX_VLM_PRELOAD_ADAPTERÚMLX_VLM_VISION_CACHE_SIZEÚMLX_VLM_DRAFT_MODELÚMLX_VLM_DRAFT_KINDÚMLX_VLM_DRAFT_BLOCK_SIZEÚPREFILL_STEP_SIZEÚMLX_VLM_MAX_TOKENSÚ1Ú0ÚMLX_VLM_ENABLE_THINKINGÚMLX_VLM_THINKING_BUDGETÚMLX_VLM_THINKING_START_TOKENÚMLX_VLM_THINKING_END_TOKENÚKV_BITSÚKV_GROUP_SIZEÚKV_QUANT_SCHEMEÚMAX_KV_SIZEÚQUANTIZED_KV_STARTÚTOP_LOGPROBS_Kz)%(asctime)s - %(levelname)s - %(message)s)ÚlevelÚformatzmlx_vlm.server:app)ÚhostÚportÚworkersÚreloadÚserver_header)ÚuniformÚ
turboquant)ÚdflashÚeagle3Úmtp)ÚDEBUGr   ÚWARNINGÚERRORÚCRITICAL)4ÚargparseÚArgumentParserÚadd_argumentÚstrÚDEFAULT_SERVER_HOSTÚintÚDEFAULT_SERVER_PORTr   r   r   r	   r   r
   Úfloatr   r   r   Ú
parse_argsÚtrust_remote_codeÚosÚenvironÚmodelÚadapter_pathÚvision_cache_sizeÚdraft_modelÚ
draft_kindÚdraft_block_sizeÚprefill_step_sizeÚ
max_tokensÚenable_thinkingÚthinking_budgetÚthinking_start_tokenr   Úkv_bitsÚkv_group_sizeÚkv_quant_schemeÚmax_kv_sizeÚquantized_kv_startÚtop_logprobs_kÚgetattrÚloggingÚ	log_levelÚupperr   ÚbasicConfigÚloggerÚsetLevelÚuvicornÚrunr/   r0   r2   )ÚparserÚargsr\   s      Úc/Volumes/fast/ai/experiments/ui-tars-smoke/.venv/lib/python3.14/site-packages/mlx_vlm/server/cli.pyÚmainrf      s÷  € Ü×$Ò$Ð1GÔH€FØ
×ÑØÜÜ#Ø9ð	 ô ð ×ÑØÜÜ#Ø7ð	 ô ð ×ÑØØØKð ô ð
 ×ÑØÜØØ\ð	 ô ð ×ÑØÜØØ6ð	 ô ð ×ÑØÜØØBð	 ô ð ×ÑØÜÜ)Ø>ð	 ô ð ×ÑØÜÜ%Ó'Ø4ð	 ô ð ×ÑØØÜ'ð*ð ô ð ×ÑØÜÜ*Ó,ð?ð ô ð ×ÑØ ÜÜ/Ó1ð.ð ô ð ×ÑØØØ!ÜÜ-Ó/ð,ð ô 
ð ×ÑØÜØØRð	 ô ð ×ÑØÜØ)Ü'Ø-ð ô ð ×ÑØÜÜ%Ø<ð	 ô ð ×ÑØÜØØ/ð	 ô ð ×ÑØÜÜ*Ø2ð	 ô ð ×ÑØÜØðOð ô ð ×ÑØÜØÚ+ðCð ô ð ×ÑØÜØØ<ð	 ô ð ×ÑØÜØð=ð ô ð ×ÑØØØØ2ð	 ô ð ×ÑØÜØÚAØ5ð ô ð ×ÑÓ€DØ××ÐØ.4Œ�
‰
Ð*Ñ+Ø‡z‡z€zØ.2¯j©jŒ�
‰
Ð*Ñ+Ø××ÐØ48×4EÑ4EŒB�J‰JÐ0Ñ1Ü.1°$×2HÑ2HÓ.I„B‡J�JÐ*Ñ+Ø××ÐØ,0×,<Ñ,<Œ�
‰
Ð(Ñ)Ø�?‰?Ò&Ø/3¯©ŒB�J‰JÐ+Ñ,Ø× Ñ Ò,Ü58¸×9NÑ9NÓ5OŒB�J‰JÐ1Ñ2Ø××ÐÜ*-¨d×.DÑ.DÓ*EŒ�
‰
Ð&Ñ'Ü'*¨4¯?©?Ó';„B‡J�JÐ#Ñ$Ø37×3G×3GÐ3G©CÈS„B‡J�JÐ(Ñ)Ø×ÑÒ'Ü03°D×4HÑ4HÓ0IŒ�
‰
Ð,Ñ-Ø× Ñ Ò,Ø59×5NÑ5NŒ�
‰
Ð1Ñ2Ø×ÑÒ*Ø37×3JÑ3JŒ�
‰
Ð/Ñ0Ø‡|�|ÒÜ # D§L¡LÓ 1Œ�
‰
�9ÑÜ"% d×&8Ñ&8Ó"9„B‡J�JˆÑØ$(×$8Ñ$8„B‡J�JÐ Ñ!Ø×ÑÒ#Ü$'¨×(8Ñ(8Ó$9Œ�
‰
�=Ñ!Ü'*¨4×+BÑ+BÓ'C„B‡J�JÐ#Ñ$Ø×ÑÒ&Ü'*¨4×+>Ñ+>Ó'?Œ�
‰
Ð#Ñ$äœ §¡×!5Ñ!5Ó!7¼¿¹ÓF€IÜ×ÒØØ:õô ‡O�O�IÔä‡K‚KØØ�Y‰YØ�Y‰YØØ�{‰{Ø÷ó    )r=   r[   rG   ra   Úgenerater   r   r   r   Ú
generationr   r   r	   r
   r   rA   rC   Ú	getLoggerr_   rf   © rg   re   Ú<module>rl      sJ   ðÛ Û Û 	ã ÷ó ÷õ ð  Ð ØÐ à	×	Ò	Ð+Ó	,€ôSrg   