o wiϜ @sUddlmZddlZddlZddlZddlmZmZddlm Z m Z ddl Z ddl m Zddl mZddlmZmZddlmZeeZe j Zed r]er]dd lmZmZmZmZdRd d ZdSddZ dTddZ!e j"e j#e j$e j%e j&e j'e j(e j)e j*e j+e j,d Z-ddZ. dUdVd$d%Z/d&d'Z0   dWdXd*d+Z1Gd,d-d-Z2Gd.d/d/e2Z3Gd0d1d1e2Z4Gd2d3d3e2Z5Gd4d5d5e2Z6Gd6d7d7e6Z7Gd8d9d9e2Z8Gd:d;d;e8Z9Gdd?d?eZ;e;ZdZdGdHZ?dIdJZ@dKdLZAd[dPdQZBdS)\) annotationsN)partialreduce)OptionalUnion)nn)is_torch_greater_or_equallogging)GeneralInterface2.5)DTensor Placement ReplicateShardc Cs|durdStdstdtjj}tt|}tjsyzIt t j d}t t j d}t t j d}dd d d d }| |}|d krOt t j ddrOd }tjj |||dtt|}|d krf||Wntyx} ztd| d} ~ ww|d kr|t t j d|d kr|nd} t|| } | dur| dkrddl} tt jd| _tt jd| _| } |dur|ntj}tj| j|f}| | |fS)z Sets up the device mesh and initilized the backend for tensor parallelism. This function is called when the model is loaded and the TP plan is set to 'auto'. NNNNr z3Tensor parallel is only supported for `torch>=2.5`.RANK LOCAL_RANK WORLD_SIZEncclgloocclhccl)cudacpuxpuhpurCCL_WORKER_COUNTr)backendrank world_sizezWe tried to initialize torch.distributed for you, but it failed. Make sure you init torch distributed in your script to use `tp_plan='auto'`.w)r OSErrortorch_C_get_acceleratortypegetattr distributedis_initializedintosenvirongetinit_process_group set_device Exceptioncurrent_devicedevicesysopendevnullstdoutstderrget_world_sizeinit_device_mesh)tp_plantp_size device_typer1r local_rankr backend_mapreindex tp_devicer3 device_map device_meshrDf/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/transformers/integrations/tensor_parallel.pyinitialize_tensor_parallelism(sP        rF total_sizer*blocksUnion[int, list[int]]return list[int]csxt|tr&t|}||dksJd|d|||fdd|DS||dks3Jd|||}|g|S)a Convert block count or proportions to block sizes. This function accepts - The number of blocks (int), in which case the block size is total_size//blocks; or - A list of block sizes (list[int]). In the second case, if sum(blocks) < total_size, the ratios between the block sizes will be preserved. For instance, if blocks is [2, 1, 1] and total_size is 1024, the returned block sizes are [512, 256, 256]. rz Cannot split z in proportional blocks: csg|]}|qSrDrD).0block part_sizerDrE qsz*_blocks_to_block_sizes..zPrepacked is not divisible by ) isinstancelistsum)rGrH total_blocks single_sizerDrNrE_blocks_to_block_sizes^s   rVparameter_namestrr:dict[str, str] Optional[str]cCsRtdd|}||vr||Sd|vr'|ddd|vr'||dddSdS)z Get the TP style for a parameter from the TP plan. The TP plan is a dictionary that maps parameter names to TP styles. The parameter name can be a generic name with wildcards (e.g. "*.weight") or a specific name (e.g. "layer_1.weight"). \d+*.rN)resubrsplit)rWr:generic_param_namerDrDrE_get_parameter_tp_planxs rc) BOOLU8I8I16F16BF16I32F32F64I64F8_E4M3cCs|}|j|}|}t|dd}g} d} |D]} | |} || } |d| }| t| | | |7} | | 7} q|}|dkrG|dtj}|dkrR|| df}n)|dksZ|dkrd|dd| df}n|dksl|d krs|d| f}ntd |d |t |S) u When weights are packed (gate_up_proj), we need to make sure each shard gets its correct share. So if you have: gate_proj ( 16, 5120, 8190) and up_proj ( 16, 5120, 8190) packed as gate_up_proj ( 16, 5120, 2 * 8190) And you shard along the last dimension, you need to interleave the gate and up values: Now, if we shard along the last dimension across TP_size (Tensor Parallelism size), we must interleave the values from gate and up projections correctly. Let's take TP_size = 4 for an example: Packed tensor `gate_up_proj` --------------------------------------------------------------- [ G0 G1 G2 G3 | G4 G5 G6 G7 | ... | U0 U1 U2 U3 | U4 U5 U6 U7 | ... ] ↑─────────────↑ ↑─────────────↑ ↑─────────────↑ ↑─────────────↑ Gate Slice 0 Gate Slice 1 Up Slice 0 Up Slice 1 Explanation: - The first half of the tensor (left of the center) holds the gate_proj values. - The second half (right of the center) holds the up_proj values. - For TP=4, we divide each half into 4 slices. In this example, we show two slices for brevity. - Each shard receives one slice from the gate part and the corresponding slice from the up part. For instance: • Shard 0 gets: [ Gate Slice 0, Up Slice 0 ] = [ G0, G1, G2, G3, U0, U1, U2, U3 ] • Shard 1 gets: [ Gate Slice 1, Up Slice 1 ] = [ G4, G5, G6, G7, U4, U5, U6, U7 ] • … and so on. This ensures that each shard receives an equal portion of both gate and up projections, maintaining consistency across tensor parallelism. r)rGrHrr^rn.NzUnsupported dim z", only dim 0, 1 or 2 are supported) shapesizerVrange get_dtypetor#float16 ValueErrorstr_to_torch_dtype)param empty_paramrCrdimslice_rGr block_sizestensors_slices block_offset block_sizeshard_block_sizestartstop slice_dtypetensorrDrDrEget_packed_weightss.    rpacked_parameter torch.Tensor sharded_dimr num_blockscCs|dkrtd|dkr|n||j}|j|}||}||}|jd|}|j|dd} |jg||||| R} t|} t|d} tt| j} | | | | | | <| | <| j| }||}|S)as Reorders a tensor that was reconstructed from sharded packed weights into its canonical packed format. For example, if a weight was packed (e.g., gate_proj and up_proj) and then sharded, DTensor.full_tensor() might produce an interleaved layout like [G0, U0, G1, U1, ...] along the sharded dimension. This function reorders it to [G0, G1, ..., U0, U1, ...]. This is an inverse operation to get_packed_weights. Args: reconstructed_tensor: The tensor reconstructed from DTensor (e.g., via .full_tensor().contiguous()). sharded_dim: The dimension index in the reconstructed_tensor that was originally sharded. world_size: The tensor parallel world size. num_packed_projs: The number of projections that were packed together (e.g., 2 for gate_up_proj). Returns: The reordered tensor in canonical packed format. rzNum blocks different from 2 is not supported yet. This is most likely a bug in your implementation as we only pack gate and up projections together.rNr^) rwndimrqviewlenrRrspermute reshape_as)rrr ractual_sharded_dimtotal_size_on_sharded_dimoriginal_block_size_on_dimshard_chunk_size prefix_shape suffix_shape tensor_view axis_ws_abs axis_npp_abs permute_ordertensor_permutedfinal_ordered_tensorrDrDrErepack_weightss8    rc Cs|}|dkr ||}||krtd|d||j}ttj|}||kr1td|d||j||}||} | |} tdg|} t| | | |<|t| S)a Generalized tensor sharding across a multi-dimensional device mesh. Args: param (torch.Tensor): The tensor to shard. empty_param (torch.Tensor): A tensor used for shape reference. device_mesh (torch.Tensor): Shape [d_0, ..., d_n] representing the mesh. rank (int): Global rank of the current process/device. dim (int): Dimension along which to shard the tensor. rzdim z* is out of bounds for tensor of dimension zRank z is out of bounds for mesh size N)r{rwrqroperatormulslicetuple) ryrzrCrr{ param_dim mesh_shaper shard_sizerend slice_indicesrDrDrEget_tensor_shards   rmodule nn.ModulecsJt|jdkr#dur|fdddur#|fdd|S)z Copy pasted from torch's function but we remove the communications (partitioning) as well as buffer registering that is similarly not efficient. rNcs ||SNrD)modinputs)rCinput_fnrDrEE z#distribute_module..cs ||SrrD)rroutputs)rC output_fnrDrErGr)r_forward_pre_hooksregister_forward_pre_hookregister_forward_hook)rrCrrrD)rCrrrEdistribute_module9s  rc@s>eZdZdZdZeddZeddZddZdd d Z dS)TensorParallelLayer9 General tensor parallel layer for transformers. TcCdSrrD input_layoutsdesired_input_layoutsrrrCrDrDrE_prepare_input_fnRz%TensorParallelLayer._prepare_input_fncCrrrDoutput_layoutsuse_local_outputrrrCrDrDrE_prepare_output_fnUrz&TensorParallelLayer._prepare_output_fncCstr)NotImplementedErrorselfryrz param_typeparam_casting_dtype to_contiguousrrCrDrDrEpartition_tensorXrz$TensorParallelLayer.partition_tensorrrrJcCs8|jrt||t|j|j|jt|j|j|jdSdSr) use_dtensorrrrrrrrrrrrCrDrDrEprepare_module_tp[sz%TensorParallelLayer.prepare_module_tpNrrrJr) __name__ __module__ __qualname____doc__r staticmethodrrrrrDrDrDrErKs  rcsDeZdZdZdddddfd d Zed d ZeddZZS)GatherParallelzi Simple class used to define the hooks to add to a layer when we just want to gather the outputs NT)rrrrOptional[Placement]rrboolcs2t|p tf|_||_tf|_||_dSr)super__init__rrrrr)rrrr __class__rDrErls   zGatherParallel.__init__cCs"|rt|dtr|d}|SNrrQr to_localrrDrDrErys z GatherParallel._prepare_input_fncCs tjj|dtjjjdd|S)NrF)opasync_op)r#r( all_reduceReduceOpSUMrrDrDrErsz!GatherParallel._prepare_output_fnrrrrrr) rrrrrrrr __classcell__rDrDrrErgs rc@s6eZdZdZed ddZed ddZd d d ZdS)IsolatedParallelz This class is used to isolate computation in a TP layer from the rest of the world. Parameters need to be LOCAL, so not dtensors NcCs|d}t|tr |}|SrrrrrrrC input_tensorrDrDrErs z"IsolatedParallel._prepare_input_fncCs|SrrDrrDrDrErsz#IsolatedParallel._prepare_output_fnrrrJcCs&t||t|jddt|jdddSr)rrrrrrDrDrErs   z"IsolatedParallel.prepare_module_tprr)rrrrrrrrrDrDrDrErs  rcsHeZdZdZdddfdd ZeddZedd Zd d ZZ S) ReplicateParallelz This class is used to replicate computation in a TP layer (used in SP regions when we don't use sequence parallelism for example) T)rrcs8ttf|_tf|_tf|_||_||_dSr)rrrrrrrr)rrrrrDrErs     zReplicateParallel.__init__cCs(|d}t|tstj|||dd}|S)NrF run_check)rQr from_localrrDrDrErs z#ReplicateParallel._prepare_input_fncCs|r|S|Sr)rrrDrDrErsz$ReplicateParallel._prepare_output_fncCs4|d|}|r |}tj||tgdd}|S)N.Fr)ru contiguousr rrrrDrDrErs z"ReplicateParallel.partition_tensor rrrrrrrrrrrDrDrrErs  rcsNeZdZdZddddddfd d Zed d ZddZeddZZ S)ColwiseParallelrNTrrrrrrrrrcsBt|p tf|_|ptdf|_tf|_||_||_dSNrp) rrrrrrrrrrrrrrrrDrErs   zColwiseParallel.__init__cCs>|d}t|tstj|||dd}||kr|j|dd}|S)NrFr placementsrrQr r redistributerrDrDrErs  z!ColwiseParallel._prepare_input_fnc Cs~|dkrt||||d}tdg} n tdg} t||||d}||}|r*|}|jr6tj||| dd}tj|| dS)NbiasrproFr requires_grad) rrrurrr rr Parameteris_floating_point rryrzrrrrrC parametershardrDrDrErs   z ColwiseParallel.partition_tensorcCs(|j|kr |j|dd}|r|S|S)NFr)rrrrrDrDrErs z"ColwiseParallel._prepare_output_fnr) rrrrrrrrrrrDrDrrErs  rc@eZdZddZdS)PackedColwiseParallelc CVt||||d}||}|r|}|jr"tj||tdgdd}tj|| dS)NroFrr rrurrr rrrrr rryrzrrrrrCrrDrDrEr z&PackedColwiseParallel.partition_tensorNrrrrrDrDrDrEr rcsXeZdZdZddddddfd d Zd d ZeddZeddZdddZ Z S)RowwiseParallela Partition a compatible nn.Module in a row-wise fashion. Currently supports nn.Linear and nn.Embedding. Users can compose it with ColwiseParallel to achieve the sharding of more complicated modules. (i.e. MLP, Attention) Keyword Args: input_layouts (Placement, optional): The DTensor layout of input tensor for the nn.Module, this is used to annotate the input tensor to become a DTensor. If not specified, we assume the input tensor to be sharded on the last dimension. output_layouts (Placement, optional): The DTensor layout of the output for the nn.Module, this is used to ensure the output of the nn.Module with the user desired layout. If not specified, the output tensor is replicated. use_local_output (bool, optional): Whether to use local :class:`torch.Tensor` instead of :class:`DTensor` for the module output, default: True. Returns: A :class:`ParallelStyle` object that represents Rowwise sharding of the nn.Module. NTrrrrrrcs8t|p tdf|_|ptf|_||_||_dSr)rrrrrrrrrrrDrEr"s  zRowwiseParallel.__init__c Csx|dkrt||||d}tdg} n tg} |dd}||}|r'|}|jr3tj||| dd}tj || dS)NrrpFrr) rrrrurrr rrrrrrDrDrEr0s   z RowwiseParallel.partition_tensorcCs`t|dr|jdur|j|_d|_|d}t|ts#tj|||dd}||kr.|j|dd}|S)NrrFrTr)hasattrr_biasrQr rrrrDrDrErBs z!RowwiseParallel._prepare_input_fncCs<|j|kr |j|dd}t|dr||j7}|r|S|S)NTrr)rrrrrrrDrDrErPs   z"RowwiseParallel._prepare_output_fnrrrJcCsd|_|jrGt|tjrtdf|_nt|tjrtf|_nt|tj r,tdf|_nt dt ||t |j |j|jt |j|j|jdSdS)NTrpzBRowwiseParallel currently only support nn.Linear and nn.Embedding!)_distribute_module_appliedrrQrLinearrr EmbeddingrrrrrrrrrrrrDrDrEr\s     z!RowwiseParallel.prepare_module_tprr) rrrrrrrrrrrrDrDrrErs   rc@r)PackedRowwiseParallelc Cr)NrpFrrrrrDrDrErtrz&PackedRowwiseParallel.partition_tensorNrrDrDrDrEr srr csLeZdZdZdddddfd d Zed d Zed dZddZZ S)SequenceParallelai SequenceParallel replicates a compatible ``nn.Module`` parameters and runs the sharded computation with input sharded on the sequence dimension. This currently supports ``nn.LayerNorm``, ``nn.Dropout``, and the `RMSNorm python implementation `__ This style implements the operation that is described in the paper `Reducing Activation Recomputation in Large Transformer Models `__ If the input passed in to this ``nn.Module`` is a :class:`torch.Tensor`, it assumes that the input is already sharded on the sequence dimension and converts the input to a :class:`DTensor` sharded on the sequence dimension. If the input passed in to this ``nn.Module`` is already a :class:`DTensor` but is not sharded on the sequence dimension, it would redistribute the input to be sharded on the sequence dimension. The output of the ``nn.Module`` will be sharded on the sequence dimension. Keyword Args: sequence_dim (int, optional): The sequence dimension of the input tensor for the ``nn.Module``, this is used to annotate the input tensor to become a DTensor that is sharded on the sequence dimension, default: 1. use_local_output (bool, optional): Whether to use local :class:`torch.Tensor` instead of :class:`DTensor` for the module output, default: False. Returns: A :class:`ParallelStyle` object that represents Sequence Parallel of the ``nn.Module``. Example:: >>> # xdoctest: +SKIP(failing) >>> from torch.distributed.tensor.parallel import parallelize_module, SequenceParallel >>> from torch.distributed.device_mesh import init_device_mesh >>> ... >>> m = Model(...) # m is a nn.Module that contains a "norm" nn.LayerNorm submodule >>> tp_mesh = init_device_mesh("cuda", (8,)) >>> >>> # By default, the input of the "norm" will be converted to DTensor that shards on the sequence dim >>> # and the output of "norm" will return a sharded on sequence dimension :class:`DTensor`. >>> >>> sharded_mod = parallelize_module(m, tp_mesh, {"norm": SequenceParallel()}), >>> ... .. note:: SequenceParallel style assumes ones initialization if there are weights in the nn.Module (i.e. ``nn.LayerNorm`` or ``RMSNorm``, and they by default have ones initialization). If you have custom inits for the weights on those modules, you need to broadcast the weights before/after parallelizing to ensure that they are replicated. r^F) sequence_dimrrr r*rrcsLttf|_tdf|_tf|_||_d|_t|f|_ ||_dS)Nr^T) rrrrrrrrrsequence_sharding)rr rrrrDrErs      zSequenceParallel.__init__cCs>|d}t|tstj|||dd}||kr|j|dd}|S)NrFrTrrrrDrDrErs  z"SequenceParallel._prepare_input_fncCs|jtfdd}|S)NTr)rrrrrDrDrErsz#SequenceParallel._prepare_output_fnc CsL|d}||}|r|}|jrtj||tgdd}tj||dS)N.Frr) rurrr rrrrrrrDrDrErs z!SequenceParallel.partition_tensor)r r*rrrrDrDrrEr s,  r c @sjeZdZedr1er1eeeedeededdeddee e dde e d Z dSiZ dS)ParallelInterfacer )r)rF)r) colwiserowwise colwise_rep rowwise_rep local_colwise local_rowwiselocalgatherlocal_packed_rowwisesequence_parallel replicateN)rrrr _torch_distributed_availablerrrrrr r r_global_mappingrDrDrDrEr s(   r ALL_PARALLEL_STYLESrr cCsd|vr |ddn|\}}t||}|s|S|dvr|S|dkr'tdg}n&|dkr:|dkr4tg}ntdg}n|dkrM|dkrHtdg}ntd g}tj|||d d S) z Converts a local variant of weights to a DTensor with corresponding placements. Shouldn't be done ever except of before saving the model. r]r^)rrrrrprrrroFr)rarcrrr r)rrWrCr:_rtp_stylerrDrDrEconvert_local_tensor_to_dtensors"      r state_dictdict[str, torch.Tensor]cCs>|D]\}}t|tjrt|tst||||||<q|S)z Replaces all tensors that were sharded with `local_*` strategy with DTensor to make determining their proper size possible. )itemsrQr#Tensorr r)rr:rCkeyvaluerDrDrE%replace_state_dict_local_with_dtensor s r%c sdur;t}z||Wnty/}ztd|dd|WYd}~nd}~ww_fdd_d|vrr|ddd }td d |} | | d } rtt| }| |||_fd d_dSdSdS)a Add hooks to the module holding the layer. Meaning: ``` class MyModel(nn.Module): def __init__(self): self.layer = nn.Linear(10, 10) ``` has state_dict like: ``` { "layer.weight": torch.Tensor, "layer.bias": torch.Tensor } ``` we add hooks to `MyModel` as well as `layer` to make sure that the tensors are correctly sharded and gathered. NTrying to prepare 0, but it's not supported. Corresponding module: z Fix it's TP plan: cdSNz TP Plan: __repr__rD)current_module_planrrDrEr5z5add_tensor_parallel_hooks_to_module..r]r^rr[r\Fcr(r)r*rD)r, module_to_tp_rDrErAr-) rrrprint _hf_tp_planr+rar_r`r- get_submodule) modelrr: layer_namer,rCtp_layerr?parent_layer_name generic_name module_planrD)r,rr.rE#add_tensor_parallel_hooks_to_modules.  r8c CsJd|vr |ddn|\}} |j} ||} t|}t|| } | dur5d} tdkr4td|dntdkrFtd|d| t | d d sXt || | || |d | _ zt | } | ||| ||||}Wn"ty}ztd |d | d| d|WYd}~nd}~wwt|tjjstjj||d}t| | ||S)a Main uses cases: - column / rowise parallelism, you just shard all the weights of the layer (weight and bias) - packed layers: you slice the weights, then shard like above - custom operation: - you want to add an all-gather at the end of a local layer. - you want to have a layer that is isolated from the rest of the world (because torch.DTensor does not work well with `.view` for instance) r]r^NrrzTensor parallel plan for z+ not found, using default 'replicate' plan.z: _is_hookedFTr&r'z" Fix it's TP plan, current layer: z : r)ra_tp_planr1r*rcdistget_rankloggerinfor'r8r9rrrr/rQr#rrrsetattr)r2ryrzrWr is_contiguousrrC param_namerr: module_to_tpr,r4r?rDrDrEshard_and_distribute_moduleDs<       rC expected_keys list[str]Optional[dict[str, str]]c Cs|durdSdd|D}t|}|}|D]B}d|vr#|dddn|}tdd|}||vr;||||qd|vrV|ddd}|vrV||||q qt|dkrftd |t|dkrytd d |dSdS) z Verify the TP plan of the model, log a warning if the layers that were not sharded and the rules that were not applied. NcSsh|] }tdd|qS)r[r\)r_r`)rLr#rDrDrE sz!verify_tp_plan..r]r^rr[r\z>The following TP rules were not applied on any of the layers: z'The following layers were not sharded: z, ) setrar_r`popdiscardrr=warningjoin) rDr: generic_keysunsharded_layers unused_rulesr#rArbparent_param_namerDrDrEverify_tp_planxs(       rQr)rGr*rHrIrJrK)rWrXr:rYrJrZ)r) rrrr*r r*rr*rJrrr)rrrWrXr:rYrJr )rr r:rYrJr )rDrEr:rF)C __future__rrr+r_ functoolsrrtypingrrr#torch.distributedr(r;rutilsr r utils.genericr get_loggerrr= is_availablertorch.distributed.tensorr rrrrFrVrcruint8int8int16rvbfloat16int32float32float64int64 float8_e4m3fnrxrrrrrrrrrrrr r r r__annotations__rr%r8rCrQrDrDrDrEsl         6 B A% #<dT  +4