o i>@s<ddlmZddlmZddlmZddlZddlmZddlm Z ddl m Z ddl m Z dd lmZdd lmZdd lmZmZdd lmZdd lmZmZddlmZmZddlmZddlm Z m!Z!ddl"m#Z#m$Z$m%Z%m&Z&m'Z'Gdddej(Z)Gdddej(Z*e Gdddej(Z+Gdddej(e e!Z,dS))Iterable)islice)AnyN)nn) LlamaConfig)support_torch_compile) get_pp_group)ReLUSquaredActivation)RMSNorm)ColumnParallelLinearRowParallelLinear)LogitsProcessor)ParallelLMHeadVocabParallelEmbedding)default_weight_loadermaybe_remap_kv_scale_name)IntermediateTensors) SupportsLoRA SupportsPP)AutoWeightsLoaderPPMissingLayeris_pp_missing_parameter'make_empty_intermediate_tensors_factory make_layerscsfeZdZdZ    ddededed edBd ed ed ed dffdd Zde j d e j fddZ Z S)ArceeMLPzQFeed-forward layer for Arcee using ReLU^2 activation (no gating as in LLaMA).NFT hidden_sizeintermediate_size hidden_act quant_configbiasprefixreduce_resultsreturncsdtt|||||dd|_t||||||dd|_|dkr,td|dt|_dS)Nz.up_proj) input_size output_sizer!r r"z .down_proj)r%r&r!r r#r"relu2zUnsupported activation: z$. Only 'relu2' is supported for AFM.) super__init__r up_projr down_proj ValueErrorr act_fn)selfrrrr r!r"r# __class__V/home/ubuntu/vllm_env/lib/python3.10/site-packages/vllm/model_executor/models/arcee.pyr)1s*   zArceeMLP.__init__xcCs*||\}}||}||\}}|SN)r*r-r+)r.r3_r1r1r2forwardVs zArceeMLP.forward)NFrT) __name__ __module__ __qualname____doc__intstrrboolr)torchTensorr6 __classcell__r1r1r/r2r-s0 %rc sveZdZdZ   ddededBdedBdeddf fd d Zd ej d ej d ej dBde ej ej ffddZ Z S)ArceeDecoderLayerzLTransformer decoder block for Arcee, with self-attention and ReLU^2 MLP.Nrconfig cache_configr r"r$c st|j|_t|dd}t|ddpt|dd}|}t|dr%|j}ddlm}|||j|jt|d |j||||||d t|d d d |_ t |j|j |j |t|dd|dd|_ t|j|jd|_t|j|jd|_dS)Nmax_position_embeddingsi attention_biasFr!qkv_biasr)LlamaAttentionnum_key_value_headsz .self_attn attn_typedecoder) rBr num_heads num_kv_headsrDr r! bias_o_projrCr"rImlp_biasz.mlp)rrrr r!r"eps)r(r)rgetattrhasattrrF vllm.model_executor.models.llamarGnum_attention_heads self_attnrrrmlpr rms_norm_epsinput_layernormpost_attention_layernorm) r.rBrCr r"rDrErMrGr/r1r2r)asN       zArceeDecoderLayer.__init__ positions hidden_statesresidualcCsX|dur |}||}n|||\}}|j||d}|||\}}||}||fS)N)rZr[)rXrUrYrV)r.rZr[r\r1r1r2r6s  zArceeDecoderLayer.forward)NNr) r7r8r9r:rrr<r)r>r?tupler6r@r1r1r/r2rA]s27rAcseZdZdZdeddedeejddffdd Z d e j de j fd d Z dd e j dBd e j de dBde j dBde j e Bee j ee j fBf ddZdeeee j fdeefddZZS) ArceeModelzeThe transformer model backbone for Arcee (embedding layer + stacked decoder blocks + final norm).r)r" layer_typer"r_r$Ncst|jj|j|j|_|_j|_tj s$j r/tj r/t |jj d|_nt|_tjfdd|dd\|_|_|_tj rYtj jd|_nt|_t|_tddgj |_dS) N)r cs|dS)N)rBrCr r"r1r"rCrBr_r r1r2s z%ArceeModel.__init__..z.layersr`rOr[r\)r(r) model_config hf_configrCr rB vocab_sizer is_first_ranktie_word_embeddings is_last_rankrr embed_tokensrrnum_hidden_layers start_layer end_layerlayersr rWnormr]aux_hidden_state_layersrmake_empty_intermediate_tensors)r. vllm_configr"r_r/rar2r)s<     zArceeModel.__init__ input_idscCs ||Sr4)rir.rrr1r1r2embed_input_idss zArceeModel.embed_input_idsrZintermediate_tensors inputs_embedsc Cstjr|dur |n||}d}n|dusJd|d}|d}g}tt|j|j|jD]\}} ||jvr?| ||| |||\}}q/tj sSt ||dS| ||\}} t |dkre||fS|S)NzAIntermediateTensors must be provided for non-first pipeline ranksr[r\)r[r\r)rrfrt enumeraterrmrkrlroappendrhrrnlen) r.rrrZrurvr[r\aux_hidden_statesidxlayerr5r1r1r2r6s8   zArceeModel.forwardweightscCsgd}t|}t}|D]\}}d|vrqd|vs d|vr!q|jdurO|j|}rO||}t|dt} |dkr@|n|d}| ||||qd|vsWd |vrct ||} | duraq| }d } |D]:\} } }| |vrqqg| | | }| d r||vrd } nt ||rd } n||}|j } | |||||d } | rq| d r||vrqt ||rq||}t|dt} | ||||q|S) z:Load weights, mapping q/k/v projections to fused qkv_proj.)) .qkv_projz.q_projq)r~z.k_projk)r~z.v_projvzrotary_emb.inv_freqzrotary_emb.cos_cachedzrotary_emb.sin_cachedN weight_loaderrscale zero_pointFz.biasT)dictnamed_parameterssetr get_cache_scalerQrdimaddrreplaceendswithrr)r.r}stacked_params_mapping params_dict loaded_paramsname loaded_weight scale_nameparamr remapped_namemapped param_name weight_nameshard_idr1r1r2 load_weightssf               zArceeModel.load_weightsr4)r7r8r9r:rAr<typerModuler)r>r?rtrr]listr6rrrr@r1r1r/r2r^s26 ,+r^c seZdZdZdgdiZdddeddffd d Z  dd ejdBd ejd e dBdejdBdeje Bf ddZ dejdejdBfddZ d ejdejfddZ de eeejfdeefddZZS)ArceeForCausalLMzKArcee Model for causal language modeling, integrated with vLLM runtime.qkv_proj)q_projk_projv_projrr`r"r$Ncst|jj}||_t||dd|_tjrIt |j |j |j t |dd|dd|_|jr:|j|jj|_t |dd}t|j |d |_nt|_|jj|_dS) Nz.model)rqr" lm_head_biasFz.lm_head)r r!r" logit_scaleg?)r)r(r)rcrdrBr^modelrrhrrerr rQlm_headrg tie_weightsrir logits_processorrrp)r.rqr"rBrr/r1r2r)is*    zArceeForCausalLM.__init__rrrZrurvcCs|j||||d}|S)N)rrrZrurv)r)r.rrrZrurv model_outputr1r1r2r6szArceeForCausalLM.forwardr[cCs||j|}|Sr4)rr)r.r[logitsr1r1r2compute_logitsszArceeForCausalLM.compute_logitscCs |j|Sr4)rrtrsr1r1r2rts z ArceeForCausalLM.embed_input_idsr}cCs(t||jjr dgnddgd}||S)z[Load weights into the model (delegates to inner model and handles tied embeddings).zlm_head.N gate_proj) skip_prefixes skip_substrs)rrBrgr)r.r}loaderr1r1r2rs  zArceeForCausalLM.load_weights)NN)r7r8r9r:packed_modules_mappingr<r)r>r?rr6rrtrr]rrr@r1r1r/r2r_s*& ,r)-collections.abcr itertoolsrtypingrr>r transformersrvllm.compilation.decoratorsrvllm.distributedr%vllm.model_executor.layers.activationr $vllm.model_executor.layers.layernormr !vllm.model_executor.layers.linearr r +vllm.model_executor.layers.logits_processorr 3vllm.model_executor.layers.vocab_parallel_embeddingrr-vllm.model_executor.model_loader.weight_utilsrr vllm.sequencer interfacesrrutilsrrrrrrrrAr^rr1r1r1r2s.           0O3