^j++ddlmZddlmZmZddlmZmZddlZddl Zddl m Z ddl m Z ddl mZmZmZer dd lmZddlZGd d eZGd d eZy)) annotations)ABCabstractmethod)Any TYPE_CHECKINGN)ir)V) FixedLayoutFlexibleLayoutLayout)SequenceceZdZdZ d ddZdddZeddZeddZddZ ddZ dd Z dd Z dd Z dd Zdd ZdddZeddZd dZed!d"dZy)# KernelInputsz Class to store and provide access to input nodes for kernels. This class takes in a tuple of input nodes and provides methods to access information about these nodes, such as their device type and device. Ncn||_d|_||ni|_||_t |dkDsJdy)z Initialize with a tuple of input nodes. Args: input_nodes: A tuple of input nodes to store out_dtype: Optional output dtype to store Nrz Expected at least one input node) _input_nodes _device_name_scalars _out_dtypelen)self input_nodesscalars out_dtypes h/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/torch/_inductor/kernel_inputs.py__init__zKernelInputs.__init__sB((,#*#6B #;!#G%GG#c| |jSt|jt|k(s)Jdt|jdt||Dcgc]}|j|c}Scc}w)a7 Return the stored input nodes, optionally reordered. Args: reorder: Optional sequence of indices to reorder the nodes. For example, (2, 0, 1) would return nodes in that order. Returns: The tuple of input nodes, optionally reordered zReorder length mismatch:  vs )rr)rreorderis rnodeszKernelInputs.nodes.sz ?$$ $4$$%W5 'D,=,=(>'?tCL> R 5/66!!!$666sA6c,t|jS)zh Get the number of input nodes. Returns: The number of input nodes )rrrs rcountzKernelInputs.count@s4$$%%rcFtj|jdS)z| Get the device type of the first node. Returns: The device type (e.g., 'cuda', 'cpu') r)rget_device_typerr$s r device_typezKernelInputs.device_typeJs !!$"3"3A"677rc<|jdjS)zn Get the device of the first node. Returns: The device of the first node r)r get_devicer$s rdevicezKernelInputs.deviceUs  #..00rc|jO|j}|jdk(r0tjj |}|j |_|jS)zu Get the device name information. Returns: A tuple of (gpu_name, vendor, model) cuda)rr+r(torchr-get_device_properties gcnArchName)rr+device_propertiess r device_namezKernelInputs.device_name^sX    $[[]F6)$)JJ$D$DV$L!$5$A$A!   rc:td|jDS)z Get the symbolic shapes of all input nodes. Returns: A tuple of shape tuples for each input node c3<K|]}|jywN)get_size.0nodes r z/KernelInputs.shapes_symbolic..ssCT]]_Ctuplerr$s rshapes_symboliczKernelInputs.shapes_symboliclsC1B1BCCCrc:td|jDS)z Get the size hints for shapes of all input nodes. Returns: A tuple of shape tuples with integer hints for each input node c3K|];}tjjj|j =ywr5)r graphsizevarsoptimization_hintsr6r7s rr:z-KernelInputs.shapes_hinted..|s3  GG   / /  @ AAr<r$s r shapes_hintedzKernelInputs.shapes_hintedu$ ))   rc:td|jDS)z Get the symbolic strides of all input nodes. Returns: A tuple of stride tuples for each input node c3<K|]}|jywr5) get_strider7s rr:z0KernelInputs.strides_symbolic..sE4T__&Er;r<r$s rstrides_symboliczKernelInputs.strides_symbolicsE43D3DEEErc:td|jDS)z Get the size hints for strides of all input nodes. Returns: A tuple of stride tuples with integer hints for each input node c3K|];}tjjj|j =ywr5)r rArBrCrIr7s rr:z.KernelInputs.strides_hinted..s4  GG   / /0A B rDr<r$s rstrides_hintedzKernelInputs.strides_hintedrFrc:td|jDS)zx Get the dtypes of all input nodes. Returns: A tuple of dtypes for each input node c3<K|]}|jywr5) get_dtyper7s rr:z&KernelInputs.dtypes..sD$T^^%Dr;r<r$s rdtypeszKernelInputs.dtypessD$2C2CDDDrc<|j|jS)z Get the dtype of a specific input node. Args: idx: Index of the node to get the dtype from (default: 0) Returns: The dtype of the specified input node )rrP)ridxs rdtypezKernelInputs.dtypes  %//11rcy) Get the output dtype, whether passed in or inferred from the nodes Returns: The output dtype Nr$s rrzKernelInputs.out_dtypercR||jvs Jd|d|j|S)z Get the scalar value for a given name. Args: name: Name of the scalar to get Returns: The scalar value zScalar z not found, but required)r)rnames r get_scalarzKernelInputs.get_scalars4t}}$Nv5M&NN$}}T""rcy)z Abstract method to handle output layout generation. Args: out_dtype: Optional output dtype. If not provided, infer from inputs flexible: If True, return FlexibleLayout, otherwise FixedLayout NrW)rflexibles r output_layoutzKernelInputs.output_layoutrXr)NN)r list[Any]rdict[str, float | int] | Nonertorch.dtype | Noner5)r zSequence[int] | Nonereturnr_rbint)rbz str | None)rbz torch.device)rbztuple[tuple[Any, ...], ...])rbztuple[tuple[int, ...], ...])rbz%tuple[tuple[sympy.Integer, ...], ...])rbztuple[torch.dtype, ...])r)rSrdrb torch.dtyperbre)rZstrrbz float | intTr]boolrbr )__name__ __module__ __qualname____doc__rr"propertyr%r(r+r2r>rErJrMrQrTrrr[r^rWrrrrs26(, HH/H& H&7$&&881 !D  F  E 2   #  rrcveZdZdZ d d fd Z d dZd dZd ddZddZddZ ddZ xZ S)MMKernelInputsz Specialized KernelInputs for matrix multiplication operations. Provides additional methods to access M, N, K dimensions. cVt||||t|jdk\sJd||}}|dkr|t|z }|dkr|t|z }d|cxkrt|ks nJd|d|cxkrt|ks nJd|||_||_y)z Initialize with a tuple of input nodes. By default, we assume the last 2 input nodes are mat1 and mat2, but the caller can adjust when necessary zExpected at least 2 input nodesrzInvalid mat1_idx: zInvalid mat2_idx: N)superrrr _mat1_idx _mat2_idx) rrrrmat1_idxmat2_idxm1_idxm2_idx __class__s rrzMMKernelInputs.__init__s gy94$$%*M,MM*"8 a< c+& &F a< c+& &FF-S--N1CH:/NN-F-S--N1CH:/NN-!!rcl|j|j}|j|j}|jd}|jd}|jd}|jd}tj j j|||||fS)a Get the symbolic M, N, K dimensions for matrix multiplication. Handles both 2D (MM) and 3D (BMM) tensors. M is extracted from the second-to-last dimension of the first operand (mat1). N is extracted from the last dimension of the second operand (mat2). K is extracted from the last dimension of the first operand (mat1). Returns: A tuple of (M, N, K) dimensions )r"rurvr6r rArB check_equals)rmat1mat2mknk0s r mnk_symboliczMMKernelInputs.mnk_symbolicszz|DNN+zz|DNN+ MMOB  MMOB  MMOB ]]_R  %%a,1ayrct|j |jS|jdjS)rVr)rmat1mat2rPr$s rrzMMKernelInputs.out_dtype s2 ?? &?? "}}q!++--rc|j\}}|j}|j^}}}|j^}} } t||D cgc]/\} } tj j j| | 1c} } } g| || } |rt|j|| St|j|| Scc} } w)z Handle output layout generation for matrix multiplication. Args: out_dtype: Optional output dtype. If not provided, infer from inputs flexible: If True, return FlexibleLayout, otherwise FixedLayout ) rrr6zipr rArBcheck_equals_and_simplifyr r+r )rr]rrrb1rk1b2k2rabsizes rr^zMMKernelInputs.output_layouts]]_ dNN$ ]]_ Q]]_ RJMbRT+ V$!QQWW   7 71 = VzAzqz !$++-DA At{{}i> > Ws4Cc^|j}||j||jfS)zj Get the mat1 and mat2 nodes. Returns: A tuple of (mat1, mat2) nodes )r"rurv)rr"s rrzMMKernelInputs.mat1mat2,s, T^^$eDNN&;;;rc|j}||j}||j}|d}|d}|d}|d}||k(s Jd|d||||fS)a  Get the hinted M, N, K dimensions for matrix multiplication. Handles both 2D (MM) and 3D (BMM) tensors. Uses shapes_hinted from the base class to get integer hints for dimensions. Returns: A tuple of (M, N, K) dimensions as integers r}r~zK dimensions don't match: r)rErurv)r hinted_shapes mat1_shape mat2_shaperrrk_checks r mnk_hintedzMMKernelInputs.mnk_hinted6s**, "4>>2 "4>>2 rN rN rNR.G|J9!D JJ|1ayrch|j}||j}t|dk\r|dSy)z Get the hinted batch size for batched matrix multiplication. Returns 1 for non-batched (2D) operations. Returns: The batch size as an integer r )rErur)rrrs r batch_hintedzMMKernelInputs.batch_hintedNs9**, "4>>2 z?a b> !r)NNr}r~) rr_rr`rrarwrdrxrd)rbz2tuple[sympy.Integer, sympy.Integer, sympy.Integer]rfrhri)rbztuple[Any, Any])rbztuple[int, int, int]rc) rkrlrmrnrrrr^rrr __classcell__)r{s@rrqrqsr26(, ""/"& "  "  "B ;4 .?(<0 rrq) __future__rabcrrtypingrrr.torch._inductor.configtorch._inductorrtorch._inductor.virtualizedr r r r collections.abcrsympyrrqrWrrrsH"#% )33(u 3u pO\Or