DŽgQddlmZmZmZmZddlZddlmcmZ ddlm Z m Z ddl m Z mZddlmZmZgdZ ddee ee ee dffd eed eed e fd Ze ej,d Zddddddd e fdZej2fde dedeee fdeee fd e f dZ d de dee dee deedeedeed e fdZde de d e fdZy)!)ListOptionalTupleUnionN)SymIntTensor) _add_docstr_nested)_device_dtype)to_padded_tensoras_nested_tensor nested_tensornested_tensor_from_jaggednarrow masked_selectts.dtypedevicereturnc~t|ttfxrtd|D}t|ts |s t d|rt|ts t|}t|tr|j dkr tdt|tr9|jr-||jk(r|j||Std|tj}|tjk(rt|tr|jjdj||}tj|Dcgc]}|j c}}tj"||gtj$|St|tsJtj&||d|dS|tj(k(rt|tr| |j*}|jj-d d j||}|j d } |j d } tj.d | | zd z| |tj0} d d lm} | || | | Sd d lm} t|tsJ| |d||\}}|Std|cc}w)a& Constructs a nested tensor preserving autograd history from a tensor or a list / tuple of tensors. If a nested tensor is passed, it will be returned directly unless the device / dtype / layout differ. Note that converting device / dtype will result in a copy, while converting layout is not currently supported by this function. If a non-nested tensor is passed, it is treated as a batch of constituents of consistent size. A copy will be incurred if the passed device / dtype differ from those of the input OR if the input is non-contiguous. Otherwise, the input's storage will be used directly. If a tensor list is provided, tensors in the list are always copied during construction of the nested tensor. Args: ts (Tensor or List[Tensor] or Tuple[Tensor]): a tensor to treat as a nested tensor OR a list / tuple of tensors with the same ndim Keyword arguments: dtype (:class:`torch.dtype`, optional): the desired type of returned nested tensor. Default: if None, same :class:`torch.dtype` as leftmost tensor in the list. device (:class:`torch.device`, optional): the desired device of returned nested tensor. Default: if None, same :class:`torch.device` as leftmost tensor in the list layout (:class:`torch.layout`, optional): the desired layout of returned nested tensor. Only strided and jagged layouts are supported. Default: if None, the strided layout. Example:: >>> a = torch.arange(3, dtype=torch.float, requires_grad=True) >>> b = torch.arange(5, dtype=torch.float, requires_grad=True) >>> nt = torch.nested.as_nested_tensor([a, b]) >>> nt.is_leaf False >>> fake_grad = torch.nested.nested_tensor([torch.ones_like(a), torch.zeros_like(b)]) >>> nt.backward(fake_grad) >>> a.grad tensor([1., 1., 1.]) >>> b.grad tensor([0., 0., 0., 0., 0.]) >>> c = torch.randn(3, 5, requires_grad=True) >>> nt2 = torch.nested.as_nested_tensor(c) c3<K|]}t|tyw)N) isinstancer).0ts ]/home/mcse/projects/flask_80/flask-venv/lib/python3.12/site-packages/torch/nested/__init__.py z#as_nested_tensor..Hs:]UV:a;P:]szXas_nested_tensor(): Expected first argument to be a tensor or a list / tuple of tensors z>as_nested_tensor(): Expected tensor argument to have dim() > 1rrzMas_nested_tensor(): Converting between nested tensor layouts is not supportedNrnested_view_from_values_offsets) min_seqlen max_seqlenjagged_from_listoffsetsrr4Specified layout is unsupported for nested tensors: )rlisttupleallr TypeErrordim RuntimeError is_nestedlayouttotorchstrided contiguousviewtensorshape_nested_view_from_buffer*_nested_compute_contiguous_strides_offsets_nested_tensor_from_tensor_listjaggedrflattenarangeint64$torch.nested._internal.nested_tensorr#r')rrrr2is_tensor_listbufferr nested_sizesvalues batch_sizeseq_lenr)r#r'nt_s rrrs|b T5M2]s:]Z\:]7]N b& !. f  jT2 "X"f"&&(Q,[\\"f",, RYY 55e54 4_a a~  b& !]]_))"-00e0LF <<"(=Q(=>L11QAA,OQ Q b$' ''88UD&RVW W 5<<  b& !~]]_,,Q255V55QF!JhhqkGll1j7&:Q&>*0 EG ]2G  Nb$' ''$RfEREBIQRXQYZ[[C)>sJ:a to_padded_tensor(input, padding, output_size=None, out=None) -> Tensor Returns a new (non-nested) Tensor by padding the :attr:`input` nested tensor. The leading entries will be filled with the nested data, while the trailing entries will be padded. .. warning:: :func:`to_padded_tensor` always copies the underlying data, since the nested and the non-nested tensors differ in memory layout. Args: padding (float): The padding value for the trailing entries. Keyword args: output_size (Tuple[int]): The size of the output tensor. If given, it must be large enough to contain all nested data; else, will infer by taking the max size of each nested sub-tensor along each dimension. out (Tensor, optional): the output tensor. Example:: >>> nt = torch.nested.nested_tensor([torch.randn((2, 5)), torch.randn((3, 4))]) nested_tensor([ tensor([[ 1.6862, -1.1282, 1.1031, 0.0464, -1.3276], [-1.9967, -1.0054, 1.8972, 0.9174, -1.4995]]), tensor([[-1.8546, -0.7194, -0.2918, -0.1846], [ 0.2773, 0.8793, -0.5183, -0.6447], [ 1.8009, 1.8468, -0.9832, -1.5272]]) ]) >>> pt_infer = torch.nested.to_padded_tensor(nt, 0.0) tensor([[[ 1.6862, -1.1282, 1.1031, 0.0464, -1.3276], [-1.9967, -1.0054, 1.8972, 0.9174, -1.4995], [ 0.0000, 0.0000, 0.0000, 0.0000, 0.0000]], [[-1.8546, -0.7194, -0.2918, -0.1846, 0.0000], [ 0.2773, 0.8793, -0.5183, -0.6447, 0.0000], [ 1.8009, 1.8468, -0.9832, -1.5272, 0.0000]]]) >>> pt_large = torch.nested.to_padded_tensor(nt, 1.0, (2, 4, 6)) tensor([[[ 1.6862, -1.1282, 1.1031, 0.0464, -1.3276, 1.0000], [-1.9967, -1.0054, 1.8972, 0.9174, -1.4995, 1.0000], [ 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000], [ 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000]], [[-1.8546, -0.7194, -0.2918, -0.1846, 1.0000, 1.0000], [ 0.2773, 0.8793, -0.5183, -0.6447, 1.0000, 1.0000], [ 1.8009, 1.8468, -0.9832, -1.5272, 1.0000, 1.0000], [ 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000]]]) >>> pt_small = torch.nested.to_padded_tensor(nt, 2.0, (2, 2, 2)) RuntimeError: Value in output_size is less than NestedTensor padded size. Truncation is not supported. F)rr2r requires_grad pin_memoryc|tj}|tjk(rtj|||||S|tjk(r|Dcgc])}t |t r|ntj|+}}ddlm }tj5||d||\} } ddd j||r| j} | Std|cc}w#1swYAxYw)a Constructs a nested tensor with no autograd history (also known as a "leaf tensor", see :ref:`Autograd mechanics `) from :attr:`tensor_list` a list of tensors. Args: tensor_list (List[array_like]): a list of tensors, or anything that can be passed to torch.tensor, where each element of the list has the same dimensionality. Keyword arguments: dtype (:class:`torch.dtype`, optional): the desired type of returned nested tensor. Default: if None, same :class:`torch.dtype` as leftmost tensor in the list. layout (:class:`torch.layout`, optional): the desired layout of returned nested tensor. Only strided and jagged layouts are supported. Default: if None, the strided layout. device (:class:`torch.device`, optional): the desired device of returned nested tensor. Default: if None, same :class:`torch.device` as leftmost tensor in the list requires_grad (bool, optional): If autograd should record operations on the returned nested tensor. Default: ``False``. pin_memory (bool, optional): If set, returned nested tensor would be allocated in the pinned memory. Works only for CPU tensors. Default: ``False``. Example:: >>> a = torch.arange(3, dtype=torch.float, requires_grad=True) >>> b = torch.arange(5, dtype=torch.float, requires_grad=True) >>> nt = torch.nested.nested_tensor([a, b], requires_grad=True) >>> nt.is_leaf True N)rrrJrKrr&r(r*)r4r5r rr=rr as_tensorrAr'no_gradrequires_grad_rKr0) tensor_listrr2rrJrKrlist_of_tensorsr'rHrIs rrrs:~ $$ '! # # 5<< WbcRS 1f 515??1;MMccI ]]_ `$_d6Y^_EB ` -( B QRXQYZ[[d ` `s.C,!C11C:r8r/startlengthcBt|tttfs t dt|tttfs t d|t j k(rlt|tst|tr t dtt j|t j j|||}|S|t jk(r|dk7r t dddl m }t|ttfr1t j|g|jt j }t|ttfr1t j|g|jt j }||||\}}}|St d |) a Constructs a nested tensor (which might be a view) from :attr:`tensor`, a strided tensor. This follows similar semantics to torch.Tensor.narrow, where in the :attr:`dim`-th dimension the new nested tensor shows only the elements in the interval `[start, start+length)`. As nested representations allow for a different `start` and `length` at each 'row' of that dimension, :attr:`start` and :attr:`length` can also be tensors of shape `tensor.shape[0]`. There's some differences depending on the layout you use for the nested tensor. If using strided layout, torch.narrow will do a copy of the narrowed data into a contiguous NT with strided layout, while jagged layout narrow() will create a non-contiguous view of your original strided tensor. This particular representation is really useful for representing kv-caches in Transformer models, as specialized SDPA kernels can deal with format easily, resulting in performance improvements. Args: tensor (:class:`torch.Tensor`): a strided tensor, which will be used as the underlying data for the nested tensor if using the jagged layout or will be copied for the strided layout. dim (int): the dimension where narrow will be applied. Only `dim=1` is supported for the jagged layout, while strided supports all dim start (Union[int, :class:`torch.Tensor`]): starting element for the narrow operation length (Union[int, :class:`torch.Tensor`]): number of elements taken during the narrow op Keyword arguments: layout (:class:`torch.layout`, optional): the desired layout of returned nested tensor. Only strided and jagged layouts are supported. Default: if None, the strided layout. Example:: >>> starts = torch.tensor([0, 1, 2, 3, 4], dtype=torch.int64) >>> lengths = torch.tensor([3, 2, 2, 1, 5], dtype=torch.int64) >>> narrow_base = torch.randn(5, 10, 20) >>> nt_narrowed = torch.nested.narrow(narrow_base, 1, starts, lengths, layout=torch.jagged) >>> nt_narrowed.is_contiguous() False z$start must be an integer or a tensorz%length must be an integer or a tensorz@start and length must be integers for the strided layout NT impl)r2r!z!jagged layout only supports dim=1r)jagged_from_tensor_and_lengthsrz3Specified layout is unsupported for nested narrow: )rintrrr0r4r5runbindrr=rArUr8rr@)r8r/rRrSr2rHrUrIs rrrsFH ec662 3ABB fsFF3 4BCC  eV $ 66(Bab b ell625== I P PQTV[]c d" I! 5<<  !8BC CW ec6] +LL%ekkRE fsFm ,\\6(6== TF1&%HAq IPQWPXYZZrEr)lengths jagged_dimr$r%cddlm}|r td|4| tdtj|j dd}d}|d}ddlm}||||||| S) aH Constructs a jagged layout nested tensor from the given jagged components. The jagged layout consists of a required values buffer with the jagged dimension packed into a single dimension. The offsets / lengths metadata determines how this dimension is split into batch elements and are expected to be allocated on the same device as the values buffer. Expected metadata formats: * offsets: Indices within the packed dimension splitting it into heterogeneously-sized batch elements. Example: [0, 2, 3, 6] indicates that a packed jagged dim of size 6 should be conceptually split into batch elements of length [2, 1, 3]. Note that both the beginning and ending offsets are required for kernel convenience (i.e. shape batch_size + 1). * lengths: Lengths of the individual batch elements; shape == batch_size. Example: [2, 1, 3] indicates that a packed jagged dim of size 6 should be conceptually split into batch elements of length [2, 1, 3]. Note that it can be useful to provide both offsets and lengths. This describes a nested tensor with "holes", where the offsets indicate the start position of each batch item and the length specifies the total number of elements (see example below). The returned jagged layout nested tensor will be a view of the input values tensor. Args: values (:class:`torch.Tensor`): The underlying buffer in the shape of (sum_B(*), D_1, ..., D_N). The jagged dimension is packed into a single dimension, with the offsets / lengths metadata used to distinguish batch elements. offsets (optional :class:`torch.Tensor`): Offsets into the jagged dimension of shape B + 1. lengths (optional :class:`torch.Tensor`): Lengths of the batch elements of shape B. jagged_dim (optional int): Indicates which dimension in values is the packed jagged dimension. If None, this is set to dim=1 (i.e. the dimension immediately following the batch dimension). Default: None min_seqlen (optional int): If set, uses the specified value as the cached minimum sequence length for the returned nested tensor. This can be a useful alternative to computing this value on-demand, possibly avoiding a GPU -> CPU sync. Default: None max_seqlen (optional int): If set, uses the specified value as the cached maximum sequence length for the returned nested tensor. This can be a useful alternative to computing this value on-demand, possibly avoiding a GPU -> CPU sync. Default: None Example:: >>> values = torch.randn(12, 5) >>> offsets = torch.tensor([0, 3, 5, 6, 10, 12]) >>> nt = nested_tensor_from_jagged(values, offsets) >>> # 3D shape with the middle dimension jagged >>> nt.shape torch.Size([5, j2, 5]) >>> # Length of each item in the batch: >>> offsets.diff() tensor([3, 2, 1, 4, 2]) >>> values = torch.randn(6, 5) >>> offsets = torch.tensor([0, 2, 3, 6]) >>> lengths = torch.tensor([1, 1, 2]) >>> # NT with holes >>> nt = nested_tensor_from_jagged(values, offsets, lengths) >>> a, b, c = nt.unbind() >>> # Batch item 1 consists of indices [0, 1) >>> torch.equal(a, values[0:1, :]) True >>> # Batch item 2 consists of indices [2, 3) >>> torch.equal(b, values[2:3, :]) True >>> # Batch item 3 consists of indices [3, 5) >>> torch.equal(c, values[3:5, :]) True r) is_fx_tracingztorch.nested.nested_tensor_from_jagged does not support tracing with fx.symbolic_trace. Use fx.wrap to wrap the function that calls nested_tensor_from_jagged.NzLnested_tensor_from_jagged(): At least one of offsets or lengths is required.r!rr!)'nested_view_from_values_offsets_lengths) ragged_idxr$r%)torch.fx._symbolic_tracer\r0FpadcumsumrAr^)rEr)rYrZr$r%r\r^s rrr<sR7 U   ?^  eeGNN1-v6GG \ 2ZJ[e ggrXmaskc|jtjk7rtd|j|jtjk7rtd|j|j |}|j |j }|j|jdz jd}ddl m }||tj|jdd S) a Constructs a nested tensor given a strided tensor input and a strided mask, the resulting jagged layout nested tensor will have values retain values where the mask is equal to True. The dimensionality of the mask is preserved and is represented with the offsets, this is unlike :func:`masked_select` where the output is collapsed to a 1D tensor. Args: tensor (:class:`torch.Tensor`): a strided tensor from which the jagged layout nested tensor is constructed from. mask (:class:`torch.Tensor`): a strided mask tensor which is applied to the tensor input Example:: >>> tensor = torch.randn(3, 3) >>> mask = torch.tensor([[False, False, True], [True, False, True], [False, False, True]]) >>> nt = torch.nested.masked_select(tensor, mask) >>> nt.shape torch.Size([3, j4]) >>> # Length of each item in the batch: >>> nt.offsets().diff() tensor([1, 2, 1]) >>> tensor = torch.randn(6, 5) >>> mask = torch.tensor([False]) >>> nt = torch.nested.masked_select(tensor, mask) >>> nt.shape torch.Size([6, j5]) >>> # Length of each item in the batch: >>> nt.offsets().diff() tensor([0, 0, 0, 0, 0, 0]) zrws//  ): "#  n\ fd6lE&#+$667n\ E?n\ V n\  n\h ##25n)-T$V[hm5\rx5\pdicpcp?6??E#v+,>?cSYkHZ?u{?H!% $ $ $ $ ag ag f agf ag ag  ag  ag agF2&2262rX