.`i <ddlmZddlmZddlZddlmZdZdZddgZ dej d e d e d e d ej f d Z dej e e ze ej zd e e fdZdej e e ze ej zdzde d e e fdZdej de de ee e e fd ej fdZdej de de ee e e fd ej fdZdejdej de de ee e e fd ej f dZdej dej d ej fdZdeej dee d eej dffdZdej e ej zd ej fd Zd!ej e ej zdej e e zdzd"e d ej fd#ZdS)$)Sequence)castNi)r)rr input_lengthpadding kernel_sizestridereturnc$|d|zz|z |zdzS)z6Calculate Conv1d output length using standard formula.rr)r r r r s {/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/vllm/model_executor/models/glmasr_utils.py_calculate_conv_output_lengthrs" 1w; & 4 ?! CC chunk_countsct|tjr|S|rQt|dtjr1t t tj|}d|DSd|DS)NrcPg|]#}t|$Sr)intitem.0cs r z)_as_list_chunk_counts.. s&555!AFFHH 555rc,g|]}t|Sr)rrs rrz)_as_list_chunk_counts..!s ) ) )qCFF ) ) )r) isinstancetorchTensortolistrlist)r tensor_countss r_as_list_chunk_countsr#s, --%""$$$6 <?ELAA6T%,/>> 55}5555 ) )L ) ) ))r num_chunksc0|dg|zSt|SNr)r#)rr$s r_normalize_chunk_countsr'$s%sZ  . ..r audio_lengths merge_factor conv_paramscN|D]\}}}t||||}||z |zdzSr&)r)r(r)r*r r r s r&_get_audio_output_lengths_from_lengthsr,-sK )4  $f5 7K   L (\ 9A ==rmaskcN|d}t|||S)N)sumr,)r-r)r*r(s r#_get_audio_output_lengths_from_maskr19s, HHRLLM 1|[  r audio_towerct|dr||\}}n|}|D]\}}}t||||}||z |zdzS)a Calculate the output lengths after audio processing. The output length accounts for: 1. Convolution layers (downsampling) 2. Merge factor (further downsampling during projection) Args: audio_tower: The audio encoder module audio_lengths: Input feature lengths [batch_size] merge_factor: Factor for merging adjacent features conv_params: List of (padding, kernel_size, stride) for each conv layer Returns: Output lengths after all processing [batch_size] _get_feat_extract_output_lengthsr)hasattrr4r) r2r(r)r*_conv_output_lengthsr r r s r#_get_audio_output_lengths_for_towerr8Ds.{>?? !,!M!M " "   ,,7   (G[&"?#Wk6##   , .< ?! CCraudio_featuresaudio_output_lengthsc|j\}}}|d}tj||||j|k}||d|S)Nrr/)shape unsqueezerarangeexpandtodeviceview)r9r:r$max_audio_tokens embed_dimaudio_features_masks r!_flatten_audio_features_by_lengthrFks/=.B+J )/99!<< %&&  , - -  ' ( (   - . 3 3B B BBrchunk_embeddings.cg}d}|D]=}||||z}|tj|d||z }>t|S)Nr)dim)appendrcattuple)rGrgrouped_embeddings current_idxcount audio_chunkss r_group_audio_embeddingsrQzspK' kE6I(IJ !!%)La"@"@"@AAAu # $ $$rct|trJ|r4t|dtjrtj|ntj|S|S)z>Convert mask to tensor, handling both list and tensor formats.r)rr!rrstacktensor)r-s r_normalize_to_tensorrUs^$  $"47EL99 $EK   d## Krfeature_attention_maskitem_idxcn|F||}t|tjr|dSt |St |}t |d|}|||z}t|tjr |||S|||}t |S)z1Extract attention mask for a specific audio item.Nr)rrrr=rUr#r0)rVrrWr-counts start_idxend_idx mask_slices r_extract_mask_for_itemr]s %h/ ,el ; ; %>>!$$ $#D)))#< 0 0FF9H9%&&I&**G(%,779%i&788' '(9:J  + ++r)collections.abcrtypingrrtorch.nnnnDEFAULT_MAX_AUDIO_LEN_SDEFAULT_MERGE_FACTORDEFAULT_CONV_PARAMSrrrr!r#r'rLr,r1Moduler8rFrQrUr]rrrrfsC%$$$$$ !),D,D),D;>DHKD \DDDD*,c*T%,-??* #Y****/,c*T%,-??$F// #Y//// >< > >eCcM*+ > \ > > > > ,eCcM*+ \ $D$D<$D$DeCcM*+ $D  \ $D$D$D$DN CL C, C \ C C C C %u|, %3- % 5<  % % % %u|d5<.@@U\,!L4 +==,,c*T1,, \ ,,,,,,r