PhddlZddlmZddlmZddlZddlmZddlm Z m Z ddl m Z erddl mZGd d eZGd d eZd efdZddd efdZGddeZGddeZGddeZej,dgdggZej,dgdggdgdggdgdgggZej,gdgdgdgZddd efd Zy)!N)IntEnum) TYPE_CHECKING)ir)get_dtype_size sympy_product)V)BaseSchedulerNodeceZdZdZdZdZy) NCCL_COLLrrN)__name__ __module__ __qualname__ ALL_REDUCE ALL_GATHERREDUCE_SCATTERhC:\Users\daisl\Desktop\realtime-object-detection\venv\Lib\site-packages\torch/_inductor/comm_analysis.pyr r sJJNrr ceZdZdZdZdZy)NVIDIA_GPU_TYPErrr N)rrrVOLTAAMPEREHOPPERrrrrrs E F Frrreturnc0tjjjtjjj}d|vrt j Sd|vrt jSd|vrt jSt jS)NV100A100H100) torchutils collect_env get_gpu_inforunrrrr)gpu_infos r get_gpu_typer'sv{{&&33EKK4K4K4O4OPH $$$ 8 %%% 8 %%%%%%rsnoder ct|jtjtjfrt j St|jtjtjfrt jSt|jtjtjfrt jStd|j)NzUnsupported collective type: ) isinstancenoder AllReduceAllReduceCoalescedr rAllGatherIntoTensorAllGatherIntoTensorCoalescedrReduceScatterTensorReduceScatterTensorCoalescedr Exception)r(s rget_collective_typer3)s%**r||R-B-BCD###  R++R-L-LM ###  R++R-L-LM '''7 |DEErceZdZdZdZdZy)NCCL_HWrrr N)rrrNVLINKPCINETrrrr5r5=s F C Crr5ceZdZdZdZy) NCCL_ALGOrrN)rrrTREERINGrrrr:r:Cs D Drr:ceZdZdZy) NCCL_PROTOrN)rrrLLrrrr>r>Hs  Brr>g333333@gffffff@g333333??g@g@)C@rAgffffff4@)gU@g6@g3@c tjjjt |j j j}|j j j}|t|z}|dz dz dz }d}|j j\}}}tj||z }|} | dkrytj} tj } t#|} t$j&j(j*} t$j&j(j,}t/}|dkr|dz nd}|dk(r|nd}t0||j3}|dk(r| n|}d}||z}t5|||dkDs| t6j8k(rdndz}| t6j8k(r d| dz z}n'| t6j:t6j<fvr| dz }d| zz }||z}|d z }t>j@}|dk(r|nt>jB}| t6j8k(r|dkDrd|z}n*d}n'| t6j:t6j<fvr|dz }tD| | j3}tF|| | j3}tFt>jB| | j3}d } |dkDrd} tI|| }||z |z||zzz }|d z}!||z }"|"|!zS) a9 Returns estimated NCCL collective runtime in nanoseconds (ns). The following heuristics are copied from https://github.com/NVIDIA/nccl/blob/master/src/graph/tuning.cc. We aim to estimate the runtime as accurately as possible. Assumptions: - only ring algorithm (NCCL_ALGO_RING) is used - only Low-Latency protocol (NCCL_PROTO_LL) is used, i.e. Simple or LL128 is not used - 8 gpus per node # TODO: Need to find a way to get accurate "gpus per node" and "# nodes" info. - collective is one of: allreduce, reducescatter, allgather irrr g?gUUUUUU?r@geAgg@@)%r graphsizevars size_hintrr+layoutsizedtyper constant_argsmathceilr:r<r>r?r3r! _inductorconfig intra_node_bw inter_node_bwr'llMaxBwsitemminr rrrr5r6r8baseLathwLatmax)#r( tensor_numel tensor_dtypetensor_storage_size_bytestensor_storage_size_GBnum_gpus_per_node_ group_sizenNodesnRanks nccl_algo nccl_protocollbwIntrabwInter compCapIndexindex2index1llMaxBwbw nChannelsbusBwnstepsratio bandwidthbandwidth_GB_per_nsintraHwhw nInterStepslatencyintraLatinterLat netOverhead latency_ns transport_nss# r estimate_nccl_collective_runtimerys77##--mEJJL!Q;VaZAF#q[\aFvv&++-GaKWBI NE  !ty/C/C'C9) U E  y###fqj! )**I,@,@A A!6\V #E I#c/nnGaKW[[B y### A:f*KK )**I,@,@A Aqj i ,113GW~i(499;HW[[!),Z8==?HK z 8[)H $0;3IIIG3J*,??L * $$r)rKenumrtypingrr!rr"rr virtualizedr torch._inductor.schedulerr r rr'r3r5r:r>tensorrTrUrQfloatryrrrrs1  0; g &o & F2 Fy F( g      %,,      E E E E E E  , 5<<    0e%,?e%Ee%r