`i@ UddlZddlmZmZmZddlZddlmZddl m Z ddl m Z ddl mZddlmZddlmZdd lmZmZmZmZmZmZmZdd lmZmZmZmZm Z m!Z!dd l"m#Z#m$Z$d gZ%e&e'e(d <dedede)fdZ*de&ede)fdZ+de&ede&ede&efdZ,dej-defdZ.de defdZ/dede defdZ0de'dede fdZ1de'dede de fdZ2de'dej-de fdZ3de'd efd!Z4d"Z5d#Z6de'd$ed%e&ede&efd&Z7d'edefd(Z8de'd)ede&e fd*Z9dedefd+Z:dej-de&efd,Z;de'd-ed.ede&efd/Zd1ed2ed3ed4efd5Z?dS)6N)AnyCallablecast)_get_device_module) ShardMetadata) ShardedTensor)DTensor)%compute_local_shape_and_global_offset)BytesStorageMetadataChunkStorageMetadata MetadataIndexSTATE_DICT_TYPE STORAGE_TYPESTensorPropertiesTensorStorageMetadata) LoadItemTypeReadItemSavePlanTensorWriteData WriteItem WriteItemType)"_check_shard_metadata_pair_overlap+_shards_get_overlap_region_wrt_saved_tensor create_read_items_for_chunk_list__all__plan other_planreturncJ|j|jkrdSt|jt|jkrdSt|j|jD]\}}|j|jkrdS|j}|j}|j|jks |j|jks|j|jkrdS|j}|j}|r|r|s|rdS|rU|rS|j |j krdS|j }|j } |r| r|s| rdS|r%| r#|j | j ks|j | j krdSdS)a Compare the two Save plans and return True if they are equal. Args: plan (SavePlan): First SavePlan to compare. other_plan (SavePlan): Second SavePlan to compare. Returns: True if the two plans are equal, False otherwise. FT) usablelenitemsziptypeindexfqnoffset tensor_datasizechunkoffsetssizes) rr plan_itemother_plan_itemplan_metadata_indexother_plan_metadata_indexr)other_tensor_datar+ other_chunks /home/jaya/work/projects/VOICE-AGENT/VIET/agent-env/lib/python3.11/site-packages/torch/distributed/checkpoint/planner_helpers.py_compare_save_plansr5's {j'''u 4:#j.////u'*$*j6F&G&G)!)!" ? >_1 1 155'o$3$9! #'@'D D D")-F-MMM"(,E,KKK55 + +7   1   1 55  !, !#4#999uu %E+1K k 5 [ uu ! !M[%888{k&777 55 4 delta_plansc4td|DS)z Check if any delta plan is usable, indicating the plan has changed. Args: delta_plans (List[SavePlan]): A list of delta plans to check. Returns: True if any delta plan is usable, False otherwise. c3(K|] }|o|jVdSN)r!).0 delta_plans r4 z(_contains_usable_plan..qs*NNJz/j/NNNNNNr6)any)r7s r4_contains_usable_planr?hs! NN+NNN N NNr6 cached_planscg}t||D]9\}}|r|js||$||:|S)ac Merge a list of delta plans into a single plan. Args: cached_plans (List[SavePlan]): A list of cached plans. delta_plans (List[SavePlan]): A list of delta plans to merge. It can contain empty plans Returns: A single merged plan. If a delta plan is not usable, use the cached plan. Otherwise, use the delta plan. )r$r!append)r@r7 merged_plans cached_planr<s r4_merge_delta_local_plansrEtsmL#&|[#A#A,, Z  ,j/ ,    , , , ,    + + + + r6tensorc ttjdgt|z|S)Nrr,r-)r torchSizer"r*)rFs r4_create_chunk_from_tensorrKsG  A3V[[]]!3!3344FKKMM   r6shard_mdcttj|jtj|jSNrH)r rIrJ shard_offsets shard_sizes)rLs r4_chunk_for_shardrQs9  8122j-..   r6sharded_tensorc|j}t|j|j|j|j|j}tt|||j S)N)dtypelayout requires_grad memory_format pin_memoryr+ propertiesr*) metadatatensor_propertiesrrTrUrVrWrXrrQr*)rRrLshard_propertiesrZs r4_sharded_tensor_metadatar^s&..00B!$&&4&4#. J x((  $ $ & & +   r6r'c t|j|j|j\}}t j|t j|}}t t||tj tt||tj ||S)NrHrYr&r%r))r shape device_mesh placementsrIrJrrrSHARDrr rcreate_from_tensorto_localr*)r'rFr-r,s r4_create_write_items_for_dtensorrgs: f(&*;NE7Z&& 7(;(;7E C))  #&(:6??;L;LMM        r6ctj|j}tt ||t jt||S)Nr`)rIrJrOrrrrdr^)r'rRrLr,s r4_create_write_item_for_shardrisMj/00G C))  ,^XFF   r6c ntjdgt|z}t t ||t jtt||tj ||S)NrrHrYr`) rIrJr"r*rrrTENSORrr rre)r'rFr,s r4_create_write_item_for_tensorrlsj!s6;;==11122G C))  !#&wfkkmmLLL':6BB      r6bytescRtt|tjS)N)r&r%)rrrBYTE_IO)r'rms r4_create_write_item_for_bytesiorps* C   "   r6c ttj|tj|f|tj|ftj|fSN)r% dest_index dest_offsets storage_indexstorage_offsetslengths)rrrorIrJrs dest_offsetrustorage_offsetlengths r4_create_read_item_for_byteior|sU   !Z//# N#455 F9%%    r6c ttj|tj||tj|tj|Srr)rrrkrIrJrsrtrurvrws r4_create_read_item_for_tensorrsN   Z --# ?33 7##    r6 checkpoint_md local_chunksc g}t|D]\}}t|jD]\}}t||sg}g} g} t||D]F\} } } }|| | | | |G|t t ||j|| t ||j||| Č|S)aW Create a list of ``ReadItem`` based on the checkpoint and local chunks. This applies the resharding algorithm and computes the reads needed to satisfy ``local_chunks`` with a checkpoint described by ``checkpoint_md``. Args: fqn (str) : The state_dict FQN to pass to ``ReadItem``. checkpoint_md (TensorStorageMetadata): metadata for a given tensor from a checkpoint. local_chunks (List[ChunkStorageMetadata]): Local chunks that needs to be loaded. Returns: A list of ``ReadItem`` that will satisfy all input chunks. ) saved_shard current_shardr~) enumeratechunksrrrBrrr,)r'rr read_itemsidxshard storage_idx storage_mdrvrtrw_dimoffset_for_saved_tensoroffset_for_current_tensorr{s r4rrsD*J -- U'01E'F'F   #K5eZHH  OLG =&e ' '') &&'>???##$=>>>v&&&&   ,,S%-EE!-"/Z5G"U"U$3#     ' 8 r6 state_dictc.g}|D]\ttr$|t >tt r:|fdjDttj r$|t|tt|S)Nc3:K|]}t|VdSr:)ri)r;rLr'objs r4r=z5_create_default_metadata_only_plan..8sC-S#x@@r6)r# isinstancer rBrgrextendr[shards_metadatarITensorrlrpr)rrequestsr'rs @@r4"_create_default_metadata_only_planr2s&H$$&& F FS c7 # # F OO;CEE F F F F ] + + F OO #  >    U\ * * F OO9#sCC D D D D OO:3DD E E E E H  r6objectc6tdrSttr!fdDStt jrtgStgS)N__create_write_items__c<g|]}t|jS)rir[)r;rr'rs r4 z'_create_write_items..Hs7    )fen E E   r6) hasattrrrr local_shardsrIrrlrp)r'rs``r4_create_write_itemsrCsv/00 =,,S&999 FM * *=     ,,..     FEL ) )=-c6::;;.sF;;<.bs0   16 U^ , ,   r6zMUnsupported Type, expecting one of [Tensor, DTensor, ShardedTensor] ,but got ) rrrrrrIrrK ValueErrorr%)rFrs r4_create_chunk_listr]sv.//  3355 FM * *    :@:M:M:O:O    FEL ) ) 1&99:  'V  ' '   r6mdrc@t|ts[ t|}n9#t$r,}td|ddt |z|d}~wwxYwt |||St t|dt|ddgS)Nz Invalid checkpoint metadata for z, z(expected BytesStorageMetadata but found rrx)rr rrr%rr|r)r'rrrexs r4_create_read_itemsrps b. / /  -c22LL   :3:::GT"XXGGH   0RFFF )(--+C00       s' A'AAcvdtfd}dtfd}dtjfd}t ||||dS)zP Initializes meta tensor if the meta tensor is DTensor or torch.Tensor. valuect|dd}|tjdkrtjj}ttjt| }tj | |}tj ||j|j||}|S|S)Ndevicemetar)rbrcrastride)getattrrIrdistdistributed_c10d_get_pg_default_devicer%rrcurrent_device empty_likerfr from_localrbrcr*r)rr device_typenew_local_tensordtensors r4 dtensor_funcz&_init_state_dict..dtensor_funcs$// U\&)) ) )/FFHHMK 0==LLNNF %/0@0@PPP ( !- +jjll||~~ GNLr6ct|dd}|tjdkr tdt |d|S)NrrzFound unsupported type z for meta device loading.)rrIr RuntimeErrorr%)rrs r4sharded_tensor_funcz-_init_state_dict..sharded_tensor_funcsU$// U\&)) ) )P$u++PPP Lr6c@t|dd}|tjdkrttjj}ttjt| }tj ||}|S|S)Nrrr) rrIrrrrr%rrrr)rrrrFs r4 tensor_funcz%_init_state_dict..tensor_funcs$// U\&)) ) )/FFHHMK 0==LLNNF%eF;;;FMLr6N)r rrIr_iterate_state_dict)rrrrs r4_init_state_dictrs G*3 5<     r6 iter_objectrrrctt|tr |St|tr |St|tjr |St|t t tttj fs||St|tr1| D]\}}t|||<|St|ttfr6fd|D}t|trt|}|SdS)a$ Iterate through the state dict, applying the given functions to each tensor type and update the state dict in place. Args: iter_object (Any): the target state_dict. sharded_tensor_func (Callable): the function to apply to ShardedTensor dtensor_func (Callable): the function to apply to DTensor tensor_func (Callable): the function to apply to Tensor # TODO: let state_dict_util._iterate_state_dict() to support in place option so we don't need to have two versions of _iterate_state_dict. Nc4g|]}t|Sr)r)r;vrrrs r4rz'_iterate_state_dict..s8    <1Dk R R   r6)rr rrIrintfloatstrrmioBytesIOdictr#rlisttuple)rrrrkeyrrets ``` r4rrsk(+w''|K((( K / /"";/// K . .{;''';eS% DEE   K & & %++--  JC2|%8+  K   K$ / /           k5 ) ) **C r6)@rtypingrrrrItorch.distributed distributedr torch._utilsr!torch.distributed._shard.metadatar'torch.distributed._shard.sharded_tensorrtorch.distributed.tensorr torch.distributed.tensor._utilsr r[r r rrrrrplannerrrrrrr reshardingrrrrr__annotations__boolr5r?rErrKrQr^rgrirlrpr|rrrrrrrrrrrr6r4rs &&&&&&&&&& ++++++;;;;;;AAAAAA,,,,,,QQQQQQ 99c999>h>H>>>>>B OtH~ O$ O O O Ox.h (^2el7K }1E!-:(g)( +7D s EL Y    C      4 4(4+,4 (^ 4444n?x" =S =# =$y/ = = = =w3Gu|5I0J& C ]  h    .4c3h4C4444n,,,", ,,,,,,r6