o i@sddlZddlZddlZddlmZddlmZddlm Z e e Z e ezddlmZWneyBe ddZYnweeeddZdejd ejd ejfd d Zd ejdejd ejfddZd ejdejdejd ejfddZdS)N) is_compiling) out_dtype)check_cpu_version) intmm_tritonzTWarning: Detected no triton, on systems without Triton certain kernels will not workTORCHAO_AUTOTUNER_ENABLEinputmat2returnc Csts d|vr+|jjdkrttjjjj tj | | Sttjjjj tj ||S|j|jks=Jd|jd|jd|jj|jjfv}|j dddkoV|j ddk}|j dddkof|j ddk}|oj| }|sp|rt |tj |tj |jjS|s|}|s|j dddkr|}z ttjjjj tj ||WStyt |tj|tjtj YSw)a Performs a safe integer matrix multiplication, considering different paths for torch.compile, cublas, and fallback cases. Args: input (torch.Tensor): The input tensor of shape [i, j]. mat2 (torch.Tensor): The matrix to multiply with, of shape [j, k]. Returns: torch.Tensor: The result of the matrix multiplication. Raises: AssertionError: If the tensors are not on the same device. FakeTensorcpuz3need both tensors to be on the same device but got z and r)dynamo_is_compiling__repr__devicetypertorchopsatenmmdefaultint32floatshapematmulr to is_contiguous contiguous Exceptionfloat32)rr device_cpuj_is_nonzero_multiple_of_8k_is_nonzero_multiple_of_8bad_dimensions_for_cublasr$H/home/ubuntu/.local/lib/python3.10/site-packages/torchao/kernel/intmm.py safe_int_mms>   $ r&abcCs&tdurtrtjj||St||S)a\ Performs integer matrix multiplication using intmm_triton if available and autotuner is enabled, otherwise falls back to safe_int_mm. Args: a (torch.Tensor): The first matrix to multiply. b (torch.Tensor): The second matrix to multiply. Returns: torch.Tensor: The result of the matrix multiplication. N)rAUTOTUNER_ENABLErrtorchao int_matmulr&)r'r(r$r$r%r+[s  r+scales1cCs|j\}}|j\}}||dks|dksJd|dks"J|s(J|||f}|dks7Jt|jrJt ||}| |j |St durYt rYtjj|||St||}||S)a Performs scaled integer matrix multiplication. Args: a (torch.Tensor): The first matrix to multiply. b (torch.Tensor): The second matrix to multiply. scales1 (torch.Tensor): The scaling factors for the rows of the result. Returns: torch.Tensor: The result of the scaled matrix multiplication. Raises: AssertionError: If the dimensions of the input tensors do not match the expected shapes. rr N)rsizenumelrexpanddimrrr_int_mmrdtyperr)rr*int_scaled_matmulr&)r'r(r,MKNcr$r$r%r4ls       r4)loggingosr torch._dynamorr!torch._higher_order_ops.out_dtyper torchao.utilsr getLogger__name__logger addHandler NullHandlertorchao.kernelr ImportErrorwarningboolintgetenvr)Tensorr&r+r4r$r$r$r%s8     =