)`i VddlZddlZddlmZddlZddlmZddlm Z ej dZ dej fdZ defd Zdej fd Zd ej d ed edefdZdej dej ddfdZddZdefdZdefdZdeedejdejdej fdZddZddZdS)N)Sequence)env)gen_nvshmem_modulec6tj}d}ddg}|D](}|D]}||z }|r|}n |n)|td|t j|tjt}|S)Nzlibnvshmem_host.sozlibnvshmem_host.so.3z=Could not find libnvshmem_host.so or libnvshmem_host.so.3 in )mode) jit_envget_nvshmem_lib_dirsexistsFileNotFoundErrorctypesCDLL RTLD_GLOBALrbuild_and_load)lib_dirslib_path lib_nameslib_dirlib_namecandidate_pathmodules k/home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/flashinfer/comm/nvshmem.pyget_nvshmem_moduler s+--HH%'=>I!  H$x/N$$&& )    E  VH V V    Kv12222  ! ! 0 0 2 2F Mreturncdt}t||SN)alloc_empty_unique_idrnvshmem_get_unique_id)uids r get_unique_idr!&s,  ! !C..s333 JrcBtSr)rnvshmem_unique_id_sizerrunique_id_sizer%,s    6 6 8 88rc\tjttjdS)Ncpu)dtypedevice)torchzerosr%uint8r$rrrr0s! ;~''u{5 I I IIrr rank world_sizect|||}tj|Sr)r nvshmem_initr*cuda synchronize)r r-r.statuss rinitr44s8  ! ! . .sD* E EF J MrdestsourcecFt||Sr)rnvshmem_alltoall)r5r6s ralltoallr9:s    0 0v > >>rctjtdSr)r*r1r2rnvshmem_finalizer$rrfinalizer<>s4 J))+++++rcBtSr)r nvshmem_my_per$rrmy_per?C    - - / //rcBtSr)r nvshmem_n_pesr$rrn_pesrCGr@rshaper(r)cpt|||}tj|S)aAllocates memory using NVSHMEM collective malloc operation. This is a collective operation that requires participation by all PEs (Processing Elements). All participants must call this function with the same parameters. Note: This tensor should be explicitly deleted (del tensor) to ensure proper ordering of nvshmem_free operations rather than relying on garbage collection. Args: shape: The shape of the tensor to allocate. dtype: The data type of the tensor. device: The device to allocate the tensor on. Returns: A tensor allocated using NVSHMEM collective malloc. Reference: https://docs.nvidia.com/nvshmem/api/gen/api/memory.html#nvshmem-malloc-nvshmem-free-nvshmem-align )rnvshmem_mallocr* from_dlpack)rDr(r)outputs rmallocrIKs22 ! ! 0 0v F FF  V $ $$rcFtdSr)rnvshmem_barrier_allr$rr barrier_allrLhs ,,.....rcFtdSr)r%nvshmem_barrier_all_on_current_streamr$rrbarrier_all_on_current_streamrOls >>@@@@@r)rN)r functoolstypingrr*jitrr jit.commrcacherTensorr!intr%rr4r9r<r?rCr(r)rIrLrOr$rrrWs  )))))) 4u| 99999Ju|JJJJel#33 ?5<??$????,,,, 0s00000s0000% C=% ;% L% \ %%%%:////AAAAAAr