.`ifddlmZmZejdZejdejfdZdS))tltritonctjd}||krdStj||z}d}|dkr|}ntj||zdz } || z }tj||z} |dz| z } tj|dk| d} tj||zdzdz } | | z } tj||z| tj||z| dS)a Fused kernel for Eagle prepare_input_padded. This kernel computes the token index to sample for each request, taking into account the number of draft tokens and the number of valid sampled tokens (which is one more than the number of accepted tokens). raxisN)r program_idloadwherestore)cu_num_draft_tokens_ptrvalid_sampled_tokens_count_ptrquery_start_loc_gpu_ptrtoken_indices_to_sample_ptrnum_rejected_tokens_gpu_ptrnum_reqsreq_idx cu_draft_currnum_draft_tokens cu_draft_prev valid_countnum_rejected_tokensq_last_tok_idxindex_to_samples m/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/vllm/v1/spec_decode/utils.py"eagle_prepare_inputs_padded_kernelrsm###G(G3g=>>M!||( 7' AA EFF (=8'87BCCK*Q.<(#3a#79LaPPW4w>BCCaGN$'::OH (7 2ODDDH (7 24GHHHHHBLOCK_SIZE_TOKENSc |tjd} | |krdStj|| z} | rjtj|| z} tjddtj} tj|| z| tj|| z| dStjd| }||k}|| |zz}tj||z|d}|dk||kz|z}tj|} | dkrmtjtj ||d}tjtj ||k|d}tj|| z|n/tj|| z} tj|| z| tj|| z| dS)a Fused kernel for Eagle prepare_next_token_ids_padded. This kernel computes the number of valid (1 + accepted) tokens for each request, and the corresponding "next" token id to sample from during speculative decoding. This is the "last accepted token" from the sampled tokens, or the backup token if no tokens were accepted or if the request is marked as discarded. rrN)dtype)maskother) rr r fulluint32r arangesummaxr )sampled_token_ids_ptrdiscard_request_mask_ptrbackup_next_token_ids_ptrnext_token_ids_ptrr vocab_sizenum_sampled_tokens_per_reqrstride_sampled_token_idsrr is_discarded backup_tokenr token_offs token_maskrow_ptr token_ids is_valid_masklast_valid_indexlast_valid_tokens r&eagle_prepare_next_token_padded_kernelr:1s(m###G(73g=>>L!Hw87BCC gb!29555  #g-|<<< /'9;GGGGGYq"344 "<< ''4L*LLGGj0zLLL #bY-CDzQ f]++ ?? "vbh}j"&M&MNN  "v'77AFF    H''13C D D D D7#r>s)((((((('I'I 'IT;H|;H;H;H ;H;H;Hr