o پi@sddlmZddlZddlmZddlmZddlmZddddddefdej d ej d eej d eej d eej d eej deej dee de fddZ ddddefdej dej d ej d eej dee deej deej de fddZdS))OptionalN)causal_conv1d_fwd)causal_conv1d_update) PAD_SLOT_IDsiluxweightbiasquery_start_loc cache_indiceshas_initial_state conv_states activation pad_slot_idc KsZ|dvrtd|ddkr|}|dur|nd}t||||||||dv| |S)a x: (batch, dim, seqlen) or (dim,cu_seq_len) for varlen sequences are concatenated from left to right for varlen weight: (dim, width) bias: (dim,) query_start_loc: (batch + 1) int32 The cumulative sequence lengths of the sequences in the batch, used to index into sequence. prepended by 0. for example: query_start_loc = torch.Tensor([0,10,16,17]), x.shape=(dim,17) cache_indices: (batch) int32 indicates the corresponding state index, like so: conv_state = conv_states[cache_indices[batch_id]] has_initial_state: (batch) bool indicates whether should the kernel take the current state as initial state for the calculations conv_states: (...,dim,width - 1) itype updated inplace if provided activation: either None or "silu" or "swish" pad_slot_id: int if cache_indices is passed, lets the kernel identify padded entries that will not be processed, for example: cache_indices = [pad_slot_id, 1, 20, pad_slot_id] in this case, the kernel will not process entries at indices 0 and 3 out: (batch, dim, seqlen) Nrswishz'activation must be None, silu, or swishrNrr)NotImplementedErrorstride contiguousr) rr r r r r rrrkwargsrc/home/ubuntu/.local/lib/python3.10/site-packages/sglang/srt/layers/attention/mamba/causal_conv1d.pycausal_conv1d_fns") r conv_state cache_seqlensconv_state_indicesc Cs`|dvr td||dv}|dk} | r|d}t||||||||| r.|d}|S)a x: (batch, dim) or (batch, dim, seqlen) conv_state: (batch, dim, state_len), where state_len >= width - 1 weight: (dim, width) bias: (dim,) cache_seqlens: (batch,), dtype int32. If not None, the conv_state is treated as a circular buffer. The conv_state will be updated by copying x to the conv_state starting at the index @cache_seqlens % state_len. conv_state_indices: (batch,), dtype int32 If not None, the conv_state is a larger tensor along the batch dim, and we are selecting the batch coords specified by conv_state_indices. Useful for a continuous batching scenario. pad_slot_id: int if cache_indices is passed, lets the kernel identify padded entries that will not be processed, for example: cache_indices = [pad_slot_id, 1 ,20 ,pad_slot_id] in this case, the kernel will not process entries at indices 0 and 3 out: (batch, dim) or (batch, dim, seqlen) rz1activation must be None, silu, or swish, actual: rr)rdim unsqueezecausal_conv1d_update_kernelsqueeze) rrr r rrrractivation_valr!rrrrMs*    r) typingrtorch sgl_kernelrrr"causal_conv1d_tritonrTensorstrintrrrrrsj     A