Wj*ddlZddlmZddlZddlmZddlmZmZm Z m Z ddl m Z ddZ dZdd Zd Zejd ed fd Zejd ed fdZejd ed fdZdS)N) Generator)global_decomposition_table) _rnn_helper gather_paramsgru_cell lstm_cell) while_loopFcb|d}|d|r|dnd}|r|dndt|dkr|dnt|dkr|dnd|dd}|dd}tjj||||rdntjdgt|j ddR|j |j d} fd } fd } tj dtj } t| | | | ||g\} }}}|r|d}||d|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)) rNdtypedevicec6|dkSNrsize)iouthxcxprecomputed_inputs Z/home/longshao/multi-rider-rag/.venv/lib/python3.11/site-packages/torch/export/_patches.pycond_fnz*one_layer_while_loop_lstm..cond_fn.$))!,,,,c T|}tj|tj|ddz t |||d\}}|}|d||<|dz|||fS)Nrr maxr ) chunk_dim)itemtorch_check_is_sizerrclonesqueeze) idxrrrrhh_bias hh_weight hr_weightrs rbody_fnz*one_layer_while_loop_lstm..body_fn1s HHJJ Q Q$5$:$:1$=$=$ABBBB a "b)WiST   BiikkAAQwR##rr)len unsqueezer%nn functionallinearflipemptyrtupleshaperrtensorint64r r()inphiddenparams has_biasesreverse ih_weightih_biasrr step_outputrr-cnt_rfinal_hxfinal_cxr*r+r,rs @@@@rone_layer_while_loop_lstmrF sq Iq I%/fQii4G%/fQii4G[[A%%q F q8H8H6!99d   Q  B   Q  B+223 7KK5<S)..q111BS+q!! rx|  hy K----- $ $ $ $ $ $ $ $ ,q , , ,C!+3 R4""AsHhhhqkk !!!$$h&6&6q&9&9: ::rc t|dkrtdt|||dd|ddk}t t |d|d} t } t|| |||||||| \} } t t | } | tj | ddtj | ddfS)a LSTM implementation using while_loop for export compatibility. This is a drop-in replacement for the default LSTM decomposition that uses while_loop instead of Python loops, making it more suitable for torch.export. Args: input: Input tensor hx: Tuple of (h0, c0) hidden states params: List of weight and bias tensors has_biases: Whether biases are included num_layers: Number of LSTM layers dropout: Dropout probability train: Training mode bidirectional: Whether to use bidirectional LSTM batch_first: Whether batch dimension is first Returns: Tuple of (output, h_n, c_n) r zlstm expects two hidden statesrr ) r/AssertionErrorrrlistziprFrr%stack inputrr<r= num_layersdropouttrain bidirectional batch_firstr;layer_fnr final_hiddenss rlstm_while_loop_implrUIs> 2ww!||=>>> 6:r!uzz!}}1 1 /M N NF #beRU## $ $F(H$    Cm,--M  M!,a00%+mA>NPQ2R2R RRrc |d|d|r|dnd|r|dnd tjj||rdn|d}tjdgt|j ddR|j |j d}fd} fd}tj dtj } t||| ||g\} } } |r| d} | | dfS) ad 1 layer fn for while loop GRU Args: inp: Input tensor of shape (seq_len, batch, input_size) hidden: Hidden state tensor params: List of weight and bias tensors has_biases: Whether biases are included reverse: Whether to process sequence in reverse Returns: Tuple of (output, final_hidden) rr r Nr rc6|dkSrr)rr cur_hiddenrs rrz)one_layer_while_loop_gru..cond_fnrrcH|}tj|tj|ddz t ||}|}|d||<|dz||fS)Nrr r!)r$r%r&rrr'r() r)rrXrr*r+r@r?rs rr-z)one_layer_while_loop_gru..body_fns HHJJ Q Q$5$:$:1$=$=$ABBBB a *i)W  iikk##A&&AQwZ''rr.)r%r1r2r3r4r0r5rr6r7rrr8r9r r()r:r;r<r=r>rXrArr-rBrCr final_hiddenr*r+r@r?rs @@@@@rone_layer_while_loop_grur[}sq Iq I%/fQii4G%/fQii4G+223 7KK5<S)..q111BS!!!$$J+q!! z# $ $ K----- ( ( ( ( ( ( ( ( ( ,q , , ,C%gwk:8VWWAsLhhqkk  $$Q'' ''rc t||d}t|d} 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)rrIunbindr[rr%rKrLs rgru_while_loop_implr^s|>6:u 5 5F "))A,,  F'H$    C  M1-- --rreturn)NNNc#Ktd}||d}|jtjjjd} |||<||jtjjj<dV||||<n||d| ||jtjjj<dS|jtjjjddS#||||<n||d|||jtjjj<w|jtjjjdwxYw)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%_C DispatchKeyCompositeImplicitAutogradpop)rnn_oprnn_implregistryoriginal_decomporiginal_py_kernels r&_register_rnn_while_loop_decompositionrmsq*/:Hll6400O *.. 6X#LT%(.HI   &.HV   LL & & &  )"  eh2L M M M   ! !%("6"PRV W W W W W  &.HV   LL & & &  )"  eh2L M M   ! !%("6"PRV W W W Ws 'C..A0Ec#Kttjjjjt 5dVddddS#1swxYwYdS)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)rmr%opsatenlstmrMrUrr®ister_lstm_while_loop_decompositionrss0 0 !#7   AA Ac#Kttjjjjt 5dVddddS#1swxYwYdS)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)rmr%rorpgrurMr^rrrr%register_gru_while_loop_decompositionrw2s0 0  "5   rt)F) contextlibcollections.abcrr% torch._decomprtorch._decomp.decompositionsrrrr"torch._higher_order_ops.while_loopr rFrUr[r^contextmanagerrmrsrwrrrrr~se%%%%%% 444444XXXXXXXXXXXX999999<;<;<;<;~1S1S1Sh3(3(3(3(l......b ,X ,X,X,X,X^  :J0K: y9I/Jr