o hF?@s.ddlZddlZddlmZmZddlZddlmZddlm Z m Z gdZ e e Z dZGdddZGd d d eZeed ejZdZGd d d ZGdddZddZ  ddeedfdeeeefdedeeedfdeeeefdeeeeeff ddZdeefddZdS)N)AnyOptional map_aggregate) tree_flattentree_unflatten)TensorChunkSpecsplit_args_kwargs_into_chunks merge_chunksFc@seZdZdZddZdS)_CustomReducera$ Custom reducer class that can be used to specify a custom operation that reduces losses of multiple microbatches into one value. Example: >>> # xdoctest: +SKIP >>> sum_reducer = _CustomReducer( >>> torch.tensor(0.0), >>> lambda a, b: a + b >>> ) cCs||_||_dSN) init_value reduce_fn)selfr rrm/home/www/facesmatcher.com/frenv_anti/lib/python3.10/site-packages/torch/distributed/pipelining/microbatch.py__init__)s z_CustomReducer.__init__N)__name__ __module__ __qualname____doc__rrrrrr s r c@ eZdZdS) _LossReducerNrrrrrrrr.rgc@sfeZdZUdZddZeed<ddZddZe d e ed ffd d Z e d e e effd dZdS)rz2 Class used to specify chunking of inputs cCs ||_dSr  split_dim)rrrrrr>s zTensorChunkSpec.__init__rcCs |jjd|jjd|jdS)N.()) __class__rrrrrrr__repr__CszTensorChunkSpec.__repr__cCsd|jdS)NzTensorChunkSpec(rrr!rrr__str__HszTensorChunkSpec.__str__ chunk_dims.cCt|dd}|S)a A helper for creating a tuple of `TensorChunkSpec` from a tuple of chunk dimensions (int's). Example: >>> # xdoctest: +SKIP >>> # There are three positional arguments to the model, and >>> # we are chunking them along dimension 0, 0 and 1, respectively >>> args_chunk_spec = TensorChunkSpec.from_tuple((0, 0, 1)) cSt|Sr rdimrrrZz,TensorChunkSpec.from_tuple..r)r$args_chunk_specrrr from_tupleKs  zTensorChunkSpec.from_tuplecCr%)a\ A helper for creating a dictionary of `TensorChunkSpec` from a dictionary of chunk dimensions (int's). Example: >>> # xdoctest: +SKIP >>> # Chunk dimension 0 for the "id" argument, 1 for the "mask" argument >>> kwargs_chunk_spec = TensorChunkSpec.from_dict({"id": 0, "mask": 1}) cSr&r r'r(rrrr*lr+z+TensorChunkSpec.from_dict..r)r$kwargs_chunk_specrrr from_dict^s  zTensorChunkSpec.from_dictN)rrrrrint__annotations__r"r# staticmethodtupler-dictstrr/rrrrr9s    rc@r) _ReplicateNrrrrrr6rrr6c! si}g}|}d}t|t|ks"Jdt|dt||D]\}}t|\} } || ||} | dus?Jt| \} } t| t| krWtd|d| g}t| | D]\}}|tuslt |t j su||g|q^t |t rt |t j sJ|d| |j}||kr|rtd|d |d |d |}ntd |d |d|dt |||j}trg}d}|D]2}t |}|| |j}tdddg|j}t||||j<|||<|||| |j7}q||n||d}q^td||||<q&g}t|D]!i}|D]\}}fdd|D}|||<q||qg}|D]+}i}t|t|ksLJt||D]\\}}} t|| ||<qS||q=|S)aW Given a dictionary of args, and a dictionary of chunking specs, shard the args according to the chunking specs. Args: args_dict: Dictionary of args args_chunk_spec: Dictionary of chunking specs num_chunks: Number of chunks to shard the args into Returns: args_split: List of sharded args Tzargs_dict.keys() = z args_chunk_spec.keys() = NzArgument value z9 did not have the same number of values as as chunk spec z is not a tensorz%Tensor size on chunking dimension is z', downsizing the number of chunks from z to rzArg z% on chunking dimension has a size of z$, smaller than the number of chunks z. PiPPy cannot reduce the number of chunks because other arguments have bigger chunk-dimension sizes. Please adjust your num_chunks setting.rFzUnrecognized chunk spec: csg|]}|qSrr).0Zv_flat chunk_idxrr sz'_shard_dict_of_args..)lenlistkeysitemsrappend ValueErrorzipr6 isinstancetorchZTensorrsizerloggerwarning RuntimeError tensor_split_debug_mask_minibatchesZ zeros_likeslicendim TypeErrorranger)!Z args_dictr,Z num_chunksZargs_sharded_replicatedZ arg_specsreal_num_chunksZ first_tensorZarg_keyargZflatspec chunk_specZchunk_spec_flat_Zsharded_arg_flatvZchunk_vZv_split_dim_sizeZ chunk_tensorsZexpanded_chunksZ split_dim_idxZ chunk_tensornew_valZ upper_idx slice_indicesZ chunks_flat chunk_argskeyZarg_single_chunk args_splitchunkZper_chunk_argsZarg_specrr8r_shard_dict_of_argsvs              rZargs.kwargschunksr,r.returnc Cs|duri}|durttft|}|durt|tt}ttt|tt||}t|}t|||}t||krOt|}ttt|tt||}t|t|kretdt|dt|dd|D}||fS)a Given a sequence of args and kwargs, split them into a number of chunks according to their respective chunking specs. Args: args: Tuple of args kwargs: Dict of kwargs chunks: Number of chunks to split the args and kwargs into args_chunk_spec: chunking specs for args, in same shape as args kwargs_chunk_spec: chunking specs for kwargs, in same shape as kwargs Returns: args_split: List of sharded args kwargs_split: List of sharded kwargs Nz;args and kwargs are split into different number of chunks: z, cs*g|]tfddttDqS)c3s|]}|VqdSr r)r7irVrr Wsz;split_args_kwargs_into_chunks...)r3rMr;)r7rr`rr:Vsz1split_args_kwargs_into_chunks..)rDEFAULT_CHUNK_DIMr;r4fromkeysrZ enumeraterG) r[r\r]r,r.Zargs_split_dictrNZ kwargs_splitrXrrrr sH8     r cs0|dur t|\}}nt|d\}}ttgt|}g|D]}t|\}}t|t|kr:td|d||q g}t|D]\} t| trчfddttD} t r| dj } | ddD] } | j | kssJqjt j t j | dd it| | jd } g}d}t| t| ksJt| | D])\}}||| j}tdddg|j}t|||| j<||}|||}qn| }|t j|| jd qFt| tr| j}ttD] }| ||}q||qFd}tdtD] }||ks Jq||qFt||S) z Given a list of chunks, merge them into a single value according to the chunk spec. Args: chunks: list of chunks chunk_spec: Chunking spec for the chunks Returns: value: Merged value NrzChunk z did not match chunk spec csg|]}|qSrr)r7r9Zarg_idxZchunks_flattenedrrr:s z merge_chunks..Zdevicemeta)sectionsr)r()rrrbr;r@r?rdrBrMrIshaperCrHemptyrrArDrJrKcatr r rr)r]rQZspec_flattenedZ flatten_specZ chunk0_flatrYZchunk_flattenedrRZargs_flattenedrOZpartial_valuesZ overall_shapevalZ meta_chunksZ values_to_catZchunk_start_idxZ partial_valueZ meta_chunkZ chunk_end_idxrUZslicedZ reduced_valr9valuerrerr ^sd-              r )NN) loggingoperatortypingrrrCZ torch.fx.noderZtorch.utils._pytreerr__all__ getLoggerrrErIr rZtensoraddZ sum_reducerrbrr6rZr3r4r5r0r<r r rrrrsH  9  i