o پi@stdZddlmZmZddlmZddlmZddlm Z ddl m Z m Z m Z e eZGdddeZeejd <d S) z.LFM2 (Liquid Foundation Model 2) configuration)ListOptional)CONFIG_MAPPING) Lfm2Config)logging)Mamba2CacheParamsMamba2StateShapemamba2_state_dtypec@sdeZdZdZedeefddZedeefddZedefddZ ede e fd d Z d S) rz SGLang configuration for LFM2 models. Extends HuggingFace's Lfm2Config with hybrid model properties needed by SGLang. LFM2 uses a hybrid architecture mixing full attention and ShortConv layers. returncCddt|jDS)z0Return indices of attention layers for KV cache.cSsg|] \}}|dkr|qS)full_attention.0iltr r K/home/ubuntu/.local/lib/python3.10/site-packages/sglang/srt/configs/lfm2.py +sz7Lfm2Config.full_attention_layer_ids.. enumerate layer_typesselfr r rfull_attention_layer_ids(sz#Lfm2Config.full_attention_layer_idscCr )z3Return indices of conv layers for conv state cache.cSsg|] \}}|dvr|qS))conv short_convr rr r rr0sz/Lfm2Config.linear_layer_ids..rrr r rlinear_layer_ids-szLfm2Config.linear_layer_idscCsdS)zJReturn chunk size for Mamba2 backend. LFM2 doesn't use chunking, return 1.r rr r rmamba_chunk_size4szLfm2Config.mamba_chunk_sizec Cszddlm}|j}|s dS|j}t|j}z|}Wn ttfy'd}Ynwtj ||d||d|d}t ||t |dS)z Get cache params for HybridReqToTokenPool initialization. LFM2 uses ShortConv layers with a small fixed-size cache (kernel_size - 1). Unlike full Mamba2 models, LFM2 only uses the conv state, not SSM temporal state. r)get_attention_tp_sizeNr) tp_world_sizeintermediate_sizen_groups num_headshead_dim state_size conv_kernel)shapelayersdtype) sglang.srt.layers.dp_attentionrr hidden_sizeint conv_L_cacheAssertionError RuntimeErrorrcreaterr )rrconv_layer_idsr+r&tp_sizer'r r rmamba2_cache_params9s2    zLfm2Config.mamba2_cache_paramsN) __name__ __module__ __qualname____doc__propertyrr,rrrrrr3r r r rr srlfm2N)r7typingrr transformersrr HFLfm2Configtransformers.utilsrsglang.srt.configs.mamba_utilsrrr get_loggerr4logger_extra_contentr r r rs    H