+ "ݳjn^RIt^RIt^RIt^RIHt^RIHt^RIH t H t H t H t ]P!]4t]P!]P"!44^RIHt]!]!]P0!R^444tRR ltR R ltR R ltRRltR# ]d]P+R4RtLUi;i)N) is_compiling) out_dtype)_cpu_is_amx_tile_supported_cpu_is_vnni_supported _is_devicetorch_version_at_least) intmm_tritonzTWarning: Detected no triton, on systems without Triton certain kernels will not workTORCHAO_AUTOTUNER_ENABLEcxV^8dQhR\PR\PR\P/#)inputmat2returntorchTensor)formats"U/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torchao/kernel/intmm.py __annotate__r#s-: : u||: 5<<: ELL: c`\4'gRVP49dVPPR8Xde\ \ P PPP\ PVP4VP44#\ \ P PPP\ PW4#VPVP8Xg"QRVP RVP 24hRVPPVPP39pVP^,^,^8H;'dVP^,^8pVP^,^,^8H;'dVP^,^8pT;'dT'*pV'g V'd\ P!VP4P\ P4VP4P\ P44PVPP4#VP!4'gVP#4pVP!4'g0VP^,^,^8wdVP#4p\ \ P PPP\ PW4# \$dr\ P!TP\ P&4TP\ P&44P\ P4u#i;i)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 )dynamo_is_compiling__repr__devicetyperropsatenmmdefaultint32floatshapematmulrto is_contiguous contiguous Exceptionfloat32)r r device_cpuj_is_nonzero_multiple_of_8k_is_nonzero_multiple_of_8bad_dimensions_for_cublass&& r safe_int_mmr/#s  0@ @ <<   % !!))5;; tzz| **22EKKMM ;;%,, & =dkk]%PUP\P\~^ &4;;++U\\->->??J"'++a.1"4"9!S!S AQR@R"&**Q-!"3q"8!Q!Qtzz!}q?P"AA'A!.||EIIKNN5;;7u{{9STWW LL             ! ! Aa      **22EKKMM  ||EHHU]]3TWWU]]5KLOO KK   s*AL11A9N-,N-cxV^8dQhR\PR\PR\P/#)r abrr)rs"rrr`s-%,,5<<ELLrc\e6\'d*\PPP W4#\ W4#)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. )r AUTOTUNER_ENABLErrtorchao int_matmulr/)r1r2s&&rr6r6`s6$4$4yy  ++A11 q rcV^8dQhR\PR\PR\PR\P/#r r1r2scales1rr)rs"rrrqs< - - || -  -/4|| - \\ -rc>\4'g\4'd\R4'dVP\P 4^,P\P 4p\P!W4pVP^R\P R7^,pVPV4VPVP4V,#\P!W4pVPVP4V,#)a CPU-optimized path for scaled integer matrix multiplication. CPU prefers decomposed version to leverage the fusion capability of Inductor. It goes to u8s8 or s8s8 path based on ISA support for hardware. The selection is for performance only and both paths should work regardless of ISA support. Args: a (torch.Tensor): The first matrix to multiply (int8). b (torch.Tensor): The second matrix to multiply (int8). scales1 (torch.Tensor): The scaling factors, typically shape (M, 1). A scalar-like shape (1, 1) is also supported and will broadcast across all rows. Returns: torch.Tensor: The result of the scaled matrix multiplication. z 2.12.0.devT)dimkeepdimdtype) rrrr&rr"uint8_int_mmsumsub_r=)r1r2r9ccomps&&& r_int_scaled_matmul_cpurDqs( ' ( ( " $ $ "< 0 0TT%++  $ ( ( 5 MM! uuD ur_s  =7   8 $'%%'(+BII&@!DEF: z" -F#w NN^L sBB43B4