o پi @sJddlmZddlZddlmZddlmZeddGdddeZdS) ) dataclassN)ForwardMetadata) ForwardBatchT)kw_onlyc @seZdZUeed<eed<eed<eed<ejed<ejed<edejdejd ed ejd df d d Ze dejdejde d dfddZ dS)BailingLinearMetadata num_prefillsnum_prefill_tokens num_decodes batch_sizehas_initial_states q_lengthsquery_start_locmamba_cache_indicesbsseq_lensreturnc Cs(t||||jdddt||dS)zTThis path is run during CUDA graph capture, i.e. decode only, so `num_prefills` is 0r)r r rr rrr r )rshapetorch ones_likediffr rrrrf/home/ubuntu/.local/lib/python3.10/site-packages/sglang/srt/layers/attention/linear/linear_metadata.pyprepare_decodesz$BailingLinearMetadata.prepare_decode forward_batchc Cs|jdur|j|||j|jdSt|j}|j}t|j|}|j}|dus(J|dk}|d|d}t|j|||||||dS)zEThis path cannot run with CUDA graph, as it contains extend requests.Nrr)r r rrrr r r ) extend_num_tokensrr rlenextend_seq_lensextend_prefix_lensrr) clsr rrrrr context_lens_tensorr rrr prepare_mixed%s0   z#BailingLinearMetadata.prepare_mixedN) __name__ __module__ __qualname__int__annotations__rTensor staticmethodr classmethodrr"rrrrr s:    r) dataclassesrr1sglang.srt.layers.attention.mamba.mamba2_metadatar,sglang.srt.model_executor.forward_batch_inforrrrrrs