a dB@sddlZddlZddlmZmZddlmZmZmZddl Z ddl m Z ddl m Z ddl mZmZddlmZeeZdd ZeGd d d Zeed d dZGdddZdS)N) dataclassfield)AnyListOptional)fx)GraphCompileReason)deepcopy_to_fake_tensorfake_mode_from_tensors)NodecCspt|rd|jdSt|tr>dddd|DdSt|trdddd d|DdSt|SdS) NzT[]ztuple(, cSsg|] }t|qSargs_str.0xrrk/var/www/html/stable-diffusion-webui/venv/lib/python3.9/site-packages/torch/_dynamo/backends/distributed.py zargs_str..)zlist(cSsg|] }t|qSrrrrrrrr)torchZ is_tensorshape isinstancetuplejoinliststr)argsrrrrs   rc@sZeZdZUdZeed<eedZe e ed<eedZ e e j ed<eedZe ed<dS)Bucketrsize)default_factoryparamsnodes param_idsN)__name__ __module__ __qualname__r!int__annotations__rrr#rrr$rr r%rrrrr s  r )bucketscCsd}g}tt|D]R\}}t|jdkr|||j|jdf|jddD]}|dd|fqPqz(ddlm}td|||ddWnt ytdYn0dS) N)IndexzSize (b)z Param Namesr)tabulatez! DDPOptimizer bucket assignments Z simple_grid)headersZtablefmtzGPlease `pip install tabulate` in order to pretty-print ddp bucket sizes) enumeratereversedlenr#appendr!r.loginfo ImportError)r+r/rowsidxZbucketparamr.rrrpretty_print_buckets%s$   r:c@sFeZdZdZd eeedddZddZej e e j dd d Z dS) DDPOptimizeraNote [DDPOptimizer] DDPOptimizer applies when dynamo compiles models wrapped in DistributedDataParallel (DDP), breaking the dynamo graph into chunks to compile separately, with the breaks aligning to the boundaries of gradient-allreduce buckets chosen by DDP. Background/Motivation - DDP uses allreduce collectives to synchronize partial gradients computed on different workers - DDP groups gradient allreduces into 'buckets' to optimize communication efficiency of all-reduce - Parameters grouped into buckets are assumed to be adjacent in time, so they become ready at around the same time during backward and thus can share the same allreduce efficently - Allreduces must overlap with backward compute for optimal training performance - DDP schedules allreduces using 'hooks' fired from the c++ autograd engine in pytorch, which operates when individual grads become 'ready' - Dynamo+AOTAutograd produces a single fused graph that runs 'atomically' from the perspective of the autograd engine, such that all gradients become 'ready' at the same time. Hooks fire after the whole fused backward function executes, preventing any overlap of compute and communication Algorithm - DDPOptimizer starts off with an FX graph traced by dynamo which represents forward. It can traverse this graph in reverse order to determine the true order that gradients will become ready during backward. - Parameter sizes are counted in reverse order, up to a bucket size limit, at which point a new bucket is started and a graph break introduced - Each of the subgraphs is compiled by the compiler provided to dynamo by the user, and then fused back together into an outer module that is returned to the user Notes - It would be better to enforce (by adding an API to DDP) that the bucket splits chosen here are used by DDP, and that DDP does not need to detect or optimize bucket order by observing execution at runtime, as it does in eager. - If Dynamo can't capture a whole graph for the portion of the model wrapped by DDP, this algorithm will currently produce splits that do not necessarily align with the buckets used by DDP. This should result in performance degradation approaching the baseline case where graph-splits are not used, but not worse. - If the backend compiler fails to compile a single subgraph, it will execute eagerly despite the rest of the subgraphs being compiled - DDP has a 'parameters_and_buffers_to_ignore' field, which DDPOptimizer attempts to honor by reading markers left by DDP on individual parameters. In cases where other transformations, such as reparameterization, are also used, the ignore markers could be lost. If DDPOptimizer fails to ignore a parameter ignored by DDP, it is not catastrophic but could impact performance by choosing sub-optimal bucket splits. - DDPOptimizer always ignores all buffers, regardless of their ignore flag, since buffers do not require gradients, and therefore aren't allreduced by DDP. (They are broadcast during forward, but this is not covered by DDPOptimizer) Debugging - Generally, it is easiest to debug DDPOptimizer in a single process program, using pdb. - In many cases, the log messages are helpful (they show bucket size assignments)- just configure torch._dynamo.config.log_level to info or debug. - See `benchmarks/dynamo/distributed.py` for a simple harness that will run a toy model or a torchbench model in a single process (or with torchrun, in multiple processes) Args: bucket_bytes_cap (int): Controls the size of buckets, in bytes, used to determine graphbreaks. Should be set to match the equivalent parameter on the original DDP module. backend_compile_fn (callable): A dynamo compiler function, to be invoked to compile each subgraph. first_bucket_cap (int): Controls the size of the first bucket. Should match DDP's first bucket cap. DDP special-cases the first bucket size since it is sometimes optimal to start a small allreduce early. N)bucket_bytes_capfirst_bucket_capcCsP|dur||_ntjr&tjj|_n||_||_|j|jksFJd||_dS)NzQFirst bucket should be smaller/equal to other buckets to get comms warmed up ASAP)r=rZ distributedZ is_availableZ_DEFAULT_FIRST_BUCKET_BYTESr<backend_compile_fn)selfr<r>r=rrr__init__ws   zDDPOptimizer.__init__cCst|do|jS)N _ddp_ignored)hasattrrA)r?Z parameterrrr_ignore_parameterszDDPOptimizer._ignore_parameter)gmexample_inputscs&t|durtjjtg}t|jjD]V}|j dvrBq0|dj |j ksnt |dkr||dj |j kr||dt|j dkr ||j}|D]l\}}||}|jr||s|dj |7_ |dj|jd||djt|qnl|j dkrxt||j} | jrx|| sx|dj | 7_ |dj|j|djt| |dj|q0t |dkr|dj dkr|dj|djt |djdksJd|d=||_td |j d t |t |dkr|!||Sit"|D] \} } | jD]}| |<q6q(t#j$j%%|dfd d } d |jdd| jd} | &D]6\}}d|vrt |r| d|d|jd7} q| d7} t'| Gfdddtj#j(j)}|| |j!}|j*|| +t'dt,| jd| S)a1 Implements graph splitting, first determining a set of of buckets by counting parameter sizes in reverse graph order, then invoking the user/backend compiler to compile each subgraph. Finally, stiches compiled graphs into one graphmodule and returns its callable. N)output placeholderrr- call_module_Zget_attrz#Params should be empty if size is 0zDDPOptimizer used bucket cap z$ and produced the following buckets:cs|SNr)node) partition_maprrrz)DDPOptimizer.compile_fn..z ---orig graph---  z ---split graph--- . --- graph--- z --------------- cs8eZdZfddZddZeedfdd ZZS)z/DDPOptimizer.compile_fn..SubmodCompilercst|||_dSrJ)superr@compiler)r?modulerS __class__rrr@s z8DDPOptimizer.compile_fn..SubmodCompiler.__init__cSst|dksJdGdddtjj}d}|jjD],}|jdkr4t|jdt s4d}|jf|_q4| t dt tdtg|_|||||}|S) z Compile the submodule, using a wrapper to make sure its output is always a tuple, which is required by AotAutograd based compilers rz%We assume only args for these modulescs$eZdZfddZddZZS)zUDDPOptimizer.compile_fn..SubmodCompiler.compile_submod..WrapperModulecst||_||_dSrJ)rRr@submodunwrap_singleton_tuple)r?rWrXrUrrr@s z^DDPOptimizer.compile_fn..SubmodCompiler.compile_submod..WrapperModule.__init__cWs*|j|}|jr&t|ttfr&|dS|S)Nr)rWrXrrr)r?rrrrrforwards z]DDPOptimizer.compile_fn..SubmodCompiler.compile_submod..WrapperModule.forward)r&r'r(r@rY __classcell__rrrUr WrapperModules r[FrFTz{DDPOptimizer intentional graph-break (See Note [DDPOptimizer]). Set `torch._dynamo.config.optimize_ddp = False` to disable.)r2rnnModulegraphr$oprrr recompiler traceback FrameSummary__file__r;Zcompile_subgraph_reasonrS)r?Z input_modrkwargsr[rXZsnwrapperrrrcompile_submods&     z>DDPOptimizer.compile_fn..SubmodCompiler.compile_submod)nreturnc s||||\}}g}s(J|D]:}t|tjr\t|tjjs\||q,||q,t d|j d|j dt |t|tsJt|tsJ|j dkrn||j }rt|}n|}t d|j dt|j||||}|j|j d|j |_ |j|j |,||i|WdWdS1sb0Yn$t||j |j ||WdSWdn1s0YdS)Nz run_node r z got args rHrPrQZ compiled_)Z_set_current_nodeZfetch_args_kwargs_from_envrrTensor _subclassesZ FakeTensorr3Z from_tensorr4debugr_targetrrdictZ fetch_attrr rr^rfrTZdelete_submoduleZ add_submodulegetattr) r?rgrrdnew_argsargZreal_modZ curr_submodZcompiled_submod_real fake_moderrrun_node*s< $    >z8DDPOptimizer.compile_fn..SubmodCompiler.run_node) r&r'r(r@rfr rrsrZrrqrUrSubmodCompilers Brtz ---final graph--- )-r rrjZ fake_tensorZFakeTensorModer r1r^r$r_r!r<r2r=insertZ get_submodulerlZnamed_parametersZ get_parameterZ requires_gradrCZuntyped_storagenbytesr#r3r%idrnextendr+r4r5r:r>r0rZpassesZ split_moduleZ named_modulesrk interpreterZ Interpreterrunr`r)r?rDrEr+rKrlnamepr9Z maybe_paramr8bZsplit_gmZ debug_strrTrtZsubmod_compilerr)rrrLr compile_fns               y  zDDPOptimizer.compile_fn)N)r&r'r(__doc__r)rr@rCrZ GraphModulerrrir~rrrrr;:s@ r;)loggingraZ dataclassesrrtypingrrrrrZtorch._dynamo.output_graphrZtorch._dynamo.utilsr r Z torch.fx.noder getLoggerr&r4rr r:r;rrrrs