^j,ddlZddlmZddlZddlmZddlmZmZm Z m Z ddl m Z dZ ddZdZdd Zd Zej$d ed fd Zej$d ed fdZej$d ed fdZy)N) Generator)global_decomposition_table) _rnn_helper gather_paramsgru_cell lstm_cell) while_loopcg}g}|D]M$tfd|Drj|j=|jOt|S)Nc3^K|]$}tjj|&yw)N)torch_C _overlaps).0 seen_tensortensors `/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/torch/export/_patches.py z,_clone_tensors_that_alias..s'& 8CEHH  v{ 3& s*-)anycloneappendtuple)tensorscloned_tensors seen_tensorsrs @r_clone_tensors_that_aliasr stNL(  #& GS& # \\^Ff%      '(   c*|d}|d|r|dnd}|r|dndt|dk(r|dnt|dk(r|dnd|djd}|djd}tjjj ||||rj dntf\tjjdgt|jdd|j|jd} fd } fd } tjdtj } t!| | | | ||g\} }}}|r|j d}||j#d|j#dffS) ay 1 layer fn for while loop LSTM Args: inp: Input tensor of shape (seq_len, batch, input_size) hidden: Tuple of (hx, cx) hidden states params: List of weight and bias tensors has_biases: Whether biases are included reverse: Whether to process sequence in reverse Returns: Tuple of (output, (final_hx, final_cx)) rNdtypedevicec,|jdkSNrsize)iouthxcxprecomputed_inputs rcond_fnz*one_layer_while_loop_lstm..cond_fnA$))!,,,rc *|j}tj|tj|jddz t |||d\}}|j }|j d||<|dz|||fS)Nrrmaxr) chunk_dim)itemr _check_is_sizer)rrsqueeze) idxr+r,r-r*hh_bias hh_weight hr_weightr.s rbody_fnz*one_layer_while_loop_lstm..body_fnDs HHJ Q Q$5$:$:1$=$AB a "b)WiST BiikAAQwR##rr$)len unsqueezer nn functionallinearflipremptyr)rshaper$r%rint64r r7)inphiddenparams has_biasesreverse ih_weightih_biasr,r- step_outputr/r<cnt_r+final_hxfinal_cxr9r:r;r.s @@@@rone_layer_while_loop_lstmrSsq Iq I%fQi4G%fQi4G[A%q F q8H6!9d   Q B   Q B++223 7K5<)..q1BS$= GY'%!Iw ++q! rxx| hhyy K- $ ,,q ,C!+3 R4"AsHhhhqk !!!$h&6&6q&9: ::rc t|dk7r tdt|||djd|djdk7}t t |d|d} t } t|| |||||||| \} } t t | } | tj| ddtj| ddfS)a LSTM implementation using while_loop for export compatibility. This is a drop-in replacement for the default LSTM decomposition that uses while_loop instead of Python loops, making it more suitable for torch.export. Args: input: Input tensor hx: Tuple of (h0, c0) hidden states params: List of weight and bias tensors has_biases: Whether biases are included num_layers: Number of LSTM layers dropout: Dropout probability train: Training mode bidirectional: Whether to use bidirectional LSTM batch_first: Whether batch dimension is first Returns: Tuple of (output, h_n, c_n) rzlstm expects two hidden statesrr) r>AssertionErrorrr)listziprSrr stack inputr,rIrJ num_layersdropouttrain bidirectional batch_firstrHlayer_fnr+ final_hiddenss rlstm_while_loop_implrb[s> 2w!|=>> 6:r!uzz!}1 1 /M NF #beRU# $F(H$   Cm,-M  M!,a0%++mA>NPQ2R RRrc~|d}|d|r|dnd}|r|dndtjjj||||rj dnt f\|j d}tjjdgt|jdd|j|jd}fd} fd} tjdtj } t| | | ||g\} } }|r| j d} | |j!dfS) ad 1 layer fn for while loop GRU Args: inp: Input tensor of shape (seq_len, batch, input_size) hidden: Hidden state tensor params: List of weight and bias tensors has_biases: Whether biases are included reverse: Whether to process sequence in reverse Returns: Tuple of (output, final_hidden) rrrNr r#c,|jdkSr'r()r*r+ cur_hiddenr.s rr/z)one_layer_while_loop_gru..cond_fnr0rc|j}tj|tj|jddz t ||dd}|j }|j d||<|dz||fS)Nrrr2)r5r r6r)rrr7)r8r+rer*r9r:r.s rr<z)one_layer_while_loop_gru..body_fns HHJ Q Q$5$:$:1$=$AB a *dD)W iik##A&AQwZ''rr=)r r@rArBrCrr?rDr)rrEr$r%rrFr r7)rGrHrIrJrKrLrMrerNr/r<rOrPr+ final_hiddenr9r:r.s @@@rone_layer_while_loop_grurhs>q Iq I%fQi4G%fQi4G++223 7K5<)..q1BS2Iw3GHIw!!!$J++q! z# $  K- ( ,,q ,C%gwk:8VWAsLhhqk  $$Q' ''rc t||d}t|jd} t} t || |||||||| \} } | t j | dfS)a GRU implementation using while_loop for export compatibility. This is a drop-in replacement for the default GRU decomposition that uses while_loop instead of Python loops, making it more suitable for torch.export. Args: input: Input tensor hx: Hidden state tensor params: List of weight and bias tensors has_biases: Whether biases are included num_layers: Number of GRU layers dropout: Dropout probability train: Training mode bidirectional: Whether to use bidirectional GRU batch_first: Whether batch dimension is first Returns: Tuple of (output, h_n) Fr)rrVunbindrhrr rXrYs rgru_while_loop_implrksn>6:u 5F "))A, F'H$   C  M1- --rreturn)NNNc#nKtd}|j|d}|jjtjj j d} |||<||jtjj j <d||||<n|j|d|2||jtjj j <y|jjtjj j dy#||||<n|j|d|2||jtjj j <w|jjtjj j dwxYww)a Generic context manager for registering while_loop-based RNN decompositions. Args: rnn_op: The aten operation to patch (e.g., torch.ops.aten.lstm.input) rnn_impl: The while_loop-based implementation function Note: This is an internal helper. Use register_lstm_while_loop_decomposition() or register_gru_while_loop_decomposition() instead. post_autogradN)rget py_kernelsr r DispatchKeyCompositeImplicitAutogradpop)rnn_oprnn_implregistryoriginal_decomporiginal_py_kernels r&_register_rnn_while_loop_decompositionrysr*/:Hll640O **.. 66X#LT%((..HHI   &.HV  LL &  )"   ehh22LL M    ! !%(("6"6"P"PRV W  &.HV  LL &  )"   ehh22LL M    ! !%(("6"6"P"PRV Ws AF5:D$B F5$BF22F5c#Kttjjjj t 5ddddy#1swYyxYww)a Context manager that temporarily registers the while_loop-based LSTM decomposition. The while_loop-based decomposition is more suitable for export and graph-based execution, as it avoids Python control flow that cannot be captured in the graph. This should support dynamic sequence lengths, however as while_loop does not support Autograd yet, an ExportedProgram created with this will not be trainable. Usage:: from torch.export._patches import register_lstm_while_loop_decomposition from torch.export import export with register_lstm_while_loop_decomposition(): # Export your model with LSTM ep = export(model, (x, h0, c0)) Note: This context manager temporarily modifies the global decomposition table and py_kernels registration. The original registrations are restored when exiting the context. N)ryr opsatenlstmrZrbrr®ister_lstm_while_loop_decompositionr'sC0 0 !!#7  =AA  A AAc#Kttjjjj t 5ddddy#1swYyxYww)a Context manager that temporarily registers the while_loop-based GRU decomposition. The while_loop-based decomposition is more suitable for export and graph-based execution, as it avoids Python control flow that cannot be captured in the graph. This should support dynamic sequence lengths, however as while_loop does not support Autograd yet, an ExportedProgram created with this will not be trainable. Usage:: from torch.export._patches import register_gru_while_loop_decomposition from torch.export import export with register_gru_while_loop_decomposition(): # Export your model with GRU ep = export(model, (x, h0)) Note: This context manager temporarily modifies the global decomposition table and py_kernels registration. The original registrations are restored when exiting the context. N)ryr r{r|grurZrkr~rr%register_gru_while_loop_decompositionrEsC0 0   "5  r)F) contextlibcollections.abcrr torch._decomprtorch._decomp.decompositionsrrrr"torch._higher_order_ops.while_loopr rrSrbrhrkcontextmanagerryrrr~rrrs% 4XX9 ! >;B1Sh4(n..b ,X ,X,X^  :J0K: y9I/Jr