;i L S r SSKrSSKrSSKrSSKrSSKrSSKrSSKrSSKJ r SSK
J r SSK J
r
Jr SSKJr SSKJrJrJrJrJrJrJr SSKrSS KJrJ r J r J r J r J r J!r!J"r"J#r# SS
K$J%r%J&r& SS K'J(r(J)r)J*r*J+r+J,r,J-r-J.r.J/r/ SS K0J1r1J2r2 \( a SS
K3J4r4 \Rj " \65 r7 " S S5 r8 " S S\Rr 5 r: " S S\Rv \:S9r<\
" S S\85 5 r= " S S\<5 r>\
" S S\85 5 r?\
" S S \85 5 r@\
" S S \85 5 rA\S rB\S rC\S! rD\S" rE\
" S# S$\85 5 rF\
" S% S&\85 5 rG\
" S' S(\85 5 rH\
" S) S*\G\H5 5 rI\S+ rJ\
" S, S-\85 5 rK " S. S/\L\Rv 5 rM " S0 S1\L\Rv 5 rN " S2 S3\L\Rv 5 rO " S4 S5\L\Rv 5 rP " S6 S7\L\<5 rQ " S8 S9\<5 rR " S: S;\L\<5 rS " S< S=\<5 rT " S> S?\Rv 5 rU\
" S@ SA5 5 rV\
" SB SC5 5 rW\
" SD SE5 5 rX\
" SF SG\85 5 rY\
" SH SI\85 5 rZ\
" SJ SK5 5 r[\
" SL SM5 5 r\\
" SN SO5 5 r]\
" SP SQ5 5 r^\
" SR SS5 5 r_\
" ST SU5 5 r`0 raSV\L4SW jrb\b" SX5 SY 5 rc\b" SZ5 S[ 5 rd\b" S\5 S] 5 re\b" S^5 S_ 5 rf\
" S` Sa5 5 rgSb rhg)cz1
General namespace and dataclass related classes
N)Iterable)contextmanager) dataclassfield) timedelta)
TYPE_CHECKINGAnyCallableLiteralOptionalUnionget_args ) !BETA_CP_AVAILABLE_PYTORCH_VERSION!BETA_TP_AVAILABLE_PYTORCH_VERSION&BETA_TP_AVAILABLE_TRANSFORMERS_VERSIONFSDP2_PYTORCH_VERSIONFSDP_AUTO_WRAP_POLICYFSDP_BACKWARD_PREFETCHFSDP_SHARDING_STRATEGY(MITA_PROFILING_AVAILABLE_PYTORCH_VERSION'XPU_PROFILING_AVAILABLE_PYTORCH_VERSION)parse_flag_from_env str_to_bool)is_cuda_availableis_hpu_availableis_mlu_availableis_msamp_availableis_musa_availableis_npu_available is_transformer_engine_availableis_xpu_available)compare_versionsis_torch_version)Float8LinearConfigc $ \ rS rSrSrS rS rSrg)
KwargsHandlerC zH
Internal mixin that implements a `to_kwargs()` method for a dataclass.
c B [ R " U R 5 $ N)copydeepcopy__dict__selfs l/home/dmtnaga/Documents/work/airagagent/rag_env/lib/python3.13/site-packages/accelerate/utils/dataclasses.pyto_dictKwargsHandler.to_dictH s }}T]]++ c SSK Jn U" 5 U R 5 R 5 nSSS5 U R 5 nUR 5 VVs0 s H u pEWU U:w d M XE_M snn$ ! , ( d f NN= fs snnf )zf
Returns a dictionary containing the attributes with values different from the default of this class.
r )clear_environmentN) environmentr5 __class__r1 items)r/ r5 default_dict this_dictkvs r0 to_kwargsKwargsHandler.to_kwargsK so
3
>>+335L !LLN !* !2K!2l1o6J!2KK !
Ls A5B+B5
B N)__name__
__module__ __qualname____firstlineno____doc__r1 r= __static_attributes__r? r3 r0 r' r' C s ,
Lr3 r' c \ rS rSrSrS rSrg)EnumWithContainsX z\A metaclass that adds the ability to check if `self` contains an item with the `in` operatorc 6 U " U5 g! [ a gf = f)NFT)
ValueError)clsitems r0 __contains__ EnumWithContains.__contains__[ s' I s
r? N)r@ rA rB rC rD rM rE r? r3 r0 rG rG X s
br3 rG c . \ rS rSrSrS r\S 5 rSrg)BaseEnumc zDAn enum class that can get the value of an item with `str(Enum.key)`c U R $ r* )valuer. s r0 __str__BaseEnum.__str__f s zzr3 c 4 [ [ [ U 5 5 $ )z.Method to list all the possible items in `cls`)listmapstr)rK s r0 rW
BaseEnum.listi s CSM""r3 r? N) r@ rA rB rC rD rT classmethodrW rE r? r3 r0 rP rP c s J # #r3 rP ) metaclassc < \ rS rSr% SrSr\\S' Sr\ \ \S' Sr
g)AutocastKwargso a
Use this object in your [`Accelerator`] to customize how `torch.autocast` behaves. Please refer to the
documentation of this [context manager](https://pytorch.org/docs/stable/amp.html#torch.autocast) for more
information on each argument.
Example:
```python
from accelerate import Accelerator
from accelerate.utils import AutocastKwargs
kwargs = AutocastKwargs(cache_enabled=True)
accelerator = Accelerator(kwargs_handlers=[kwargs])
```
TenabledN
cache_enabledr? ) r@ rA rB rC rD r` bool__annotations__ra r rE r? r3 r0 r^ r^ o s! GT$(M8D>(r3 r^ c , \ rS rSrSrSrSrSrSrSr Sr
g )
DDPCommunicationHookType a
Represents a type of communication hook used in DDP.
Values:
- **NO** -- no communication hook
- **FP16** -- DDP communication hook to compress the gradients in FP16
- **BF16** -- DDP communication hook to compress the gradients in BF16
- **POWER_SGD** -- DDP communication hook to use PowerSGD
- **BATCHED_POWER_SGD** -- DDP communication hook to use batched PowerSGD
nofp16bf16 power_sgdbatched_power_sgdr? N) r@ rA rB rC rD NOFP16BF16 POWER_SGDBATCHED_POWER_SGDrE r? r3 r0 re re s#
B D DI+r3 re c < ^ \ rS rSr% SrSr\\S' Sr\ \S' Sr
\\S' S r \ \S
' S r \ \S ' S r
\ \S ' S r\ \S
' \R r\\S' \R r\\R \R( \R* 4 \S' \" \S9r\\S' SU 4S jjrS rSrU =r $ ) DistributedDataParallelKwargs a
Use this object in your [`Accelerator`] to customize how your model is wrapped in a
`torch.nn.parallel.DistributedDataParallel`. Please refer to the documentation of this
[wrapper](https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html) for more
information on each argument.
`gradient_as_bucket_view` is only available in PyTorch 1.7.0 and later versions.
`static_graph` is only available in PyTorch 1.11.0 and later versions.
Example:
```python
from accelerate import Accelerator
from accelerate.utils import DistributedDataParallelKwargs
kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
accelerator = Accelerator(kwargs_handlers=[kwargs])
```
r dimTbroadcast_buffers
bucket_cap_mbFfind_unused_parameterscheck_reductiongradient_as_bucket_view static_graph comm_hook comm_wrapper)default_factorycomm_state_optionc > [ TU ] 5 R 5 VVs0 s H u p#X!; d M X#_M snn$ s snnf r* )superr1 r8 )r/ ignore_keysr; r< r7 s r0 r1 %DistributedDataParallelKwargs.to_dict s8 !& !2!8!8!:S!:a>R!:SSSs ::c SSK JnJn [ R UR
[ R UR [ R UR [ R UR 0n[ R UR [ R UR 0nUR U R 5 nUR U R 5 nU( a U( a U" U5 nU( a] U R [ R [ R 4; a UR " S0 U R" D6OS nUR% UUS9 g g )Nr )
default_hooks
powerSGD_hook)statehookr* )+torch.distributed.algorithms.ddp_comm_hooksr r re rm fp16_compress_hookrn bf16_compress_hookro rp batched_powerSGD_hookfp16_compress_wrapperbf16_compress_wrappergetr| r}
PowerSGDStater register_comm_hook) r/ modelr r hook_map wrapper_mapr wrapperr s r0 r 0DistributedDataParallelKwargs.register_comm_hook s!
% ) )=+K+K $ ) )=+K+K $ . .
0K0K $ 6 6
8[8[ >
% ) )=+N+N $ ) )=+N+NA
$,<<#?&1ood6G6G&H G4=D >>,66,>> ++KD4J4JK
$ $
%
r3 r? ))r| r} r ) r@ rA rB rC rD rt intrc ru rb rw rx ry rz r{ re rl r| r} r rm rn r dictr r1 r rE
__classcell__r7 s @r0 rr rr s 2 CL "t"M3#(D( !OT!$)T) L$ *B*E*EI'E
!## ' ## %% %% ' $
$D 9t9T% %r3 rr c ` \ rS rSr% SrSr\\S' Sr\\S' Sr \\S' S r
\ \S
' S r \
\S ' S
rg)GradScalerKwargs a
Use this object in your [`Accelerator`] to customize the behavior of mixed precision, specifically how the
`torch.amp.GradScaler` or `torch.cuda.amp.GradScaler` used is created. Please refer to the documentation of this
[scaler](https://pytorch.org/docs/stable/amp.html?highlight=gradscaler) for more information on each argument.
`torch.cuda.amp.GradScaler` is only available in PyTorch 1.5.0 and later versions, and `torch.amp.GradScaler` is
only available in PyTorch 2.4.0 and later versions.
Example:
```python
from accelerate import Accelerator
from accelerate.utils import GradScalerKwargs
kwargs = GradScalerKwargs(backoff_factor=0.25)
accelerator = Accelerator(kwargs_handlers=[kwargs])
```
g @
init_scaleg @
growth_factorg ?backoff_factori growth_intervalTr` r? N)r@ rA rB rC rD r floatrc r r r r r` rb rE r? r3 r0 r r s; . J M5 NE OS GTr3 r c \ \ rS rSr% SrSr\\ \S' Sr \\ \S' Sr
\\ \S' S r S r
g)
InitProcessGroupKwargsi am
Use this object in your [`Accelerator`] to customize the initialization of the distributed processes. Please refer
to the documentation of this
[method](https://pytorch.org/docs/stable/distributed.html#torch.distributed.init_process_group) for more
information on each argument.
Note: If `timeout` is set to `None`, the default will be based upon how `backend` is set.
```python
from datetime import timedelta
from accelerate import Accelerator
from accelerate.utils import InitProcessGroupKwargs
kwargs = InitProcessGroupKwargs(timeout=timedelta(seconds=800))
accelerator = Accelerator(kwargs_handlers=[kwargs])
```
ncclbackendN init_methodtimeoutc d U R c# U R S:w a SOSn[ US9U l g g )Nr i iX )seconds)r r r )r/ r s r0
__post_init__$InitProcessGroupKwargs.__post_init__' s/ << "llf 4d#G$W5DL r3 )r )r@ rA rB rC rD r r rY rc r r r r rE r? r3 r0 r r s9 $ $GXc]#!%K# %#'GXi
'6r3 r )MSAMPTEO1O2)HYBRIDE4M3E5M2)max most_recentc B \ rS rSr% SrSr\S \S' Sr\\ \S' Sr
g)AORecipeKwargsi5 a
Use this object in your [`Accelerator`] to customize the initialization of the recipe for FP8 mixed precision
training with `torchao` FP8.
Args:
config (`torchao.float8.Float8LinearConfig`, *optional*, default to `None`):
The configuration for the FP8 training. In general, the default config should be sufficient.
module_filter_func (`Callable`, *optional*, default to `None`):
Optional function that must take in a module and layer name, and returns a boolean indicating whether the
module should be converted to FP8. Defaults to `accelerate.utils.ao.filter_linear_layers`. See it for an
example.
Nr% configmodule_filter_funcr? ) r@ rA rB rC rD r r rc r r
rE r? r3 r0 r r 5 s) .2FH) *1-1*1r3 r c \ rS rSr% SrSr\\ \S' Sr \\
\S' Sr \\
\S' Sr \
\S' Sr\\
\S' Sr\\S ' Sr\\\\4 \S
' Sr\\ \S ' S rS
rg)TERecipeKwargsiH a
Use this object in your [`Accelerator`] to customize the initialization of the recipe for FP8 mixed precision
training with `transformer-engine`.
For more information on the args, please refer to the API
[documentation](https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/api/common.html).
```python
from accelerate import Accelerator
from accelerate.utils import TERecipeKwargs
kwargs = TERecipeKwargs(fp8_format="HYBRID")
accelerator = Accelerator(mixed_precision="fp8", kwargs_handlers=[kwargs])
```
Args:
use_autocast_during_eval (`bool`, *optional*, default to `False`):
Whether to use FP8 autocast during eval mode. Generally better metrics are found when this is `False`.
margin (`int`, *optional*, default to 0):
The margin to use for the gradient scaling.
interval (`int`, *optional*, default to 1):
The interval to use for how often the scaling factor is recomputed.
fp8_format (`str`, *optional*, default to "HYBRID"):
The format to use for the FP8 recipe. Must be one of `HYBRID`, `E4M3` or `E5M2`. (Generally `HYBRID` for
training, `E4M3` or `E5M2` for evaluation)
amax_history_len (`int`, *optional*, default to 1024):
The length of the history to use for the scaling factor computation
amax_compute_algo (`str`, *optional*, default to "most_recent"):
The algorithm to use for the scaling factor computation. Must be one of `max` or `most_recent`.
override_linear_precision (`tuple` of three `bool`, *optional*, default to `(False, False, False)`):
Whether or not to execute `fprop`, `dgrad`, and `wgrad` GEMMS in higher precision.
Nuse_autocast_during_evalmargininterval
fp8_formatamax_history_lenamax_compute_algooverride_linear_precisionuse_mxfp8_block_scalingc Sn[ 5 ( d [ S5 eU R c [ US- 5 U l U R c1 [ [ R R US- S5 5 U l U R c1 [ [ R R US- S5 5 U l U R c( [ R R US- S 5 U l
U R R 5 U l
U R [ [ 5 ; a+ [ S
S R [ [ 5 5 S 35 eU R c( [ R R US
- S5 U l U R R# 5 U l U R [ [$ 5 ; a* [ SS R [ [$ 5 5 35 eU R& c1 [ [ R R US- S5 5 U l U R( c3 [ US- 5 n[ US- 5 n[ US- 5 nX#U4U l U R* c [ US- 5 U l g g )NACCELERATE_FP8_zQTransformerEngine is not available. Please install it or use a different backend.USE_AUTOCAST_DURING_EVALMARGINr INTERVALr FORMATr z `fp8_format` must be one of or .AMAX_COMPUTE_ALGOr z#`amax_compute_algo` must be one of AMAX_HISTORY_LEN OVERRIDE_FPROPOVERRIDE_DGRADOVERRIDE_WGRADUSE_MXFP8_BLOCK_SCALING)r! ImportErrorr r r r osenvironr r r upperr FP8FormatrJ joinr lowerAmaxComputeAlgorithmr r r )r/
env_prefixfpropdgradwgrads r0 r TERecipeKwargs.__post_init__x s &
.00 qr r ( ( 0,?
Mg@g,hD ) ;; bjjnnZ(-BA FGDK ==
zJ/F J KDM ?? " jjnnZ(-BH MDO ////1 ??(9"5 5 ;FKKQZH[<\;]]^ _` ` ! ! )%'ZZ^^JAT4TVc%dD "!%!7!7!=!=!? ! !2F)G G B6;;xXlOmCnBo pq q ($'
zDV7VX\(]$^D ! ) ) 1'
5E(EFE'
5E(EFE'
5E(EFE.3E-BD * ' ' /+>zLe?e+fD ( 0r3 )r r r r r r r r )r@ rA rB rC rD r r rb rc r r r r r r r r r tupler r rE r? r3 r0 r r H s #J 04 htn3 FHSM "Hhsm" J &*hsm*.2+29= uT4%5 6=.2Xd^2 gr3 r c . \ rS rSr% SrSr\\S' S rSr g)MSAMPRecipeKwargsi z
Use this object in your [`Accelerator`] to customize the initialization of the recipe for FP8 mixed precision
training with `ms-amp`.
N opt_levelc SnU R c( [ R R US- S5 U l U R [ [
5 ; a* [
SSR [ [
5 5 35 eg )Nr OPT_LEVELr z`opt_level` must be one of r )r r r r r OptLevelrJ r )r/ r s r0 r MSAMPRecipeKwargs.__post_init__ se &
>> ! ZZ^^J,Dd KDN >>(!3 3 :6;;xPXGY;Z:[ \] ] 4r3 )r )
r@ rA rB rC rD r r rc r rE r? r3 r0 r r s
Ix ^r3 r c < ^ \ rS rSr% SrSr\\S' U 4S jrSr U =r
$ )FP8RecipeKwargsi z}
Deprecated. Please use one of the proper FP8 recipe kwargs classes such as `TERecipeKwargs` or `MSAMPRecipeKwargs`
instead.
Nr c z > Sn[ R " S[ 5 [ 5 ( a SOSnU R c( [
R R US- U5 U l U R R 5 U l U R [ [ 5 ; a [ S5 e[ TU ]5 5 g )Nr zFP8RecipeKwargs is deprecated and will be removed in Accelerate v2.0.0. Please use one of the proper FP8 recipe kwargs classes such as TERecipeKwargs or MSAMPRecipeKwargs instead.msampteBACKENDzO`backend` must be 'MSAMP' or 'TE' (TransformerEngine) to use `FP8RecipeKwargs`.)warningswarn
FutureWarningr r r r r r r BackendrJ r r )r/ r default_backendr7 s r0 r FP8RecipeKwargs.__post_init__ s &
z
&8%9%9't << ::>>*y*@/RDL||))+ <
`torch.profiler` is only available in PyTorch 1.8.1 and later versions.
Example:
```python
from accelerate import Accelerator
from accelerate.utils import ProfileKwargs
kwargs = ProfileKwargs(activities=["cpu", "cuda"])
accelerator = Accelerator(kwargs_handlers=[kwargs])
```
Args:
activities (`List[str]`, *optional*, default to `None`):
The list of activity groups to use in profiling. Must be one of `"cpu"`, `"xpu"`, `"mtia"`, "hpu" or
`"cuda"`.
schedule_option (`Dict[str, int]`, *optional*, default to `None`):
The schedule option to use for the profiler. Available keys are `wait`, `warmup`, `active`, `repeat` and
`skip_first`. The profiler will skip the first `skip_first` steps, then wait for `wait` steps, then do the
warmup for the next `warmup` steps, then do the active recording for the next `active` steps and then
repeat the cycle starting with `wait` steps. The optional number of cycles is specified with the `repeat`
parameter, the zero value means that the cycles will continue until the profiling is finished.
on_trace_ready (`Callable`, *optional*, default to `None`):
Callable that is called at each step when schedule returns `ProfilerAction.RECORD_AND_SAVE` during the
profiling.
record_shapes (`bool`, *optional*, default to `False`):
Save information about operator’s input shapes.
profile_memory (`bool`, *optional*, default to `False`):
Track tensor memory allocation/deallocation
with_stack (`bool`, *optional*, default to `False`):
Record source information (file and line number) for the ops.
with_flops (`bool`, *optional*, default to `False`):
Use formula to estimate the FLOPS of specific operators
with_modules (`bool`, *optional*, default to `False`):
Record module hierarchy (including function names) corresponding to the callstack of the op.
output_trace_dir (`str`, *optional*, default to `None`):
Exports the collected trace in Chrome JSON format. Chrome use 'chrome://tracing' view json file. Defaults
to None, which means profiling does not store json files.
N
activitiesschedule_optionon_trace_readyF
record_shapesprofile_memory
with_stack
with_flops with_modulesoutput_trace_diractivityreturnc [ R R R [ R R R S.n[ 5 ( a' [ R R R US' [ S[ 5 ( aJ [ R R 5 ( a' [ R R R US' [ S[ 5 ( aJ [ R R 5 ( a' [ R R R US' X; a [ SU S[! U5 S35 eX! $ ) zGet the profiler activity from the string.
Args:
activity (str): The profiler activity name.
Returns:
torch.profiler.ProfilerActivity: The profiler activity.
)r r r >=r r zInvalid profiler activity: . Must be one of r )torchprofilerProfilerActivityCPUCUDAr HPUr$ r r is_availableXPUr r MTIArJ rW )r/ r profiler_activity_maps r0 _get_profiler_activity$ProfileKwargs._get_profiler_activity s >>2266NN3388M
+0>>+J+J+N+N !% ( D"I J Jyy%%''/4~~/N/N/R/R%e, D"J K Kzz&&((050O0O0T0T%f- 0 :8*DUVZ[pVqUrrs tu u$..r3 c SnU R b* U R Vs/ s H o R U5 PM nnSnU R b* [ R R
" S0 U R D6n[ R R
UUU R U R U R U R U R U R S9$ s snf )zt
Build a profiler object with the current configuration.
Returns:
torch.profiler.profile: The profiler object.
N)r scheduler r r r r r r? )
r r r r
r r profiler r r r r r )r/ r r r s r0 buildProfileKwargs.build s 8<
?? &PTP_P_`P_H55h?P_J`6: + ~~..F1E1EFH~~%%! .. ,, .. ** &
as Cr? ) r@ rA rB rC rD r r rW r rc r r rY r r r
r rb r r r r r r
r r r r rE r? r3 r0 r r s /b 48J ./0704OXd38n-4)-NHX&- M4 ND J J L$ &*hsm* //? /ENNDcDc /<
u~~--
r3 r c L \ rS rSrSrSrSrSrSrSr Sr
S r S
r S r
S rS
rSrSrSrg)DistributedTypei: ay
Represents a type of distributed environment.
Values:
- **NO** -- Not a distributed environment, just a single process.
- **MULTI_CPU** -- Distributed on multiple CPU nodes.
- **MULTI_GPU** -- Distributed on multiple GPUs.
- **MULTI_MLU** -- Distributed on multiple MLUs.
- **MULTI_SDAA** -- Distributed on multiple SDAAs.
- **MULTI_MUSA** -- Distributed on multiple MUSAs.
- **MULTI_NPU** -- Distributed on multiple NPUs.
- **MULTI_XPU** -- Distributed on multiple XPUs.
- **MULTI_HPU** -- Distributed on multiple HPUs.
- **DEEPSPEED** -- Using DeepSpeed.
- **XLA** -- Using TorchXLA.
rl MULTI_CPU MULTI_GPU MULTI_NPU MULTI_MLU
MULTI_SDAA
MULTI_MUSA MULTI_XPU DEEPSPEEDFSDPXLA MEGATRON_LM MULTI_HPUr? N)r@ rA rB rC rD rl r r r r r! r" r# r$ r% r& r' r( rE r? r3 r0 r r : sJ &
BIIII J JII D
C KIr3 r c $ \ rS rSrSrSrSrSrSrg)SageMakerDistributedTypei] a
Represents a type of distributed environment.
Values:
- **NO** -- Not a distributed environment, just a single process.
- **DATA_PARALLEL** -- using sagemaker distributed data parallelism.
- **MODEL_PARALLEL** -- using sagemaker distributed model parallelism.
rl
DATA_PARALLELMODEL_PARALLELr? N) r@ rA rB rC rD rl r+ r, rE r? r3 r0 r* r* ] s
B#M%Nr3 r* c ( \ rS rSrSrSrSrSrSrSr g) FP8BackendTypein zy
Represents the backend used for FP8.
Values:
- **TE** -- using TransformerEngine.
- **MSAMP** -- using msamp.
rl r r AOr? N)
r@ rA rB rC rD rl r r r/ rE r? r3 r0 r. r. n s
B
B E
Br3 r. c \ rS rSrSrSrSrSrg)ComputeEnvironmenti z
Represents a type of the compute environment.
Values:
- **LOCAL_MACHINE** -- private/custom cluster hardware.
- **AMAZON_SAGEMAKER** -- Amazon SageMaker as compute environment.
LOCAL_MACHINEAMAZON_SAGEMAKERr? N)r@ rA rB rC rD r2 r3 rE r? r3 r0 r1 r1 s $M)r3 r1 c X \ rS rSrSrSrSrSrSrSr Sr
S r S
r S r
S rS
rSrSrSrSrSrSrg)
DynamoBackendi a
Represents a dynamo backend (see https://pytorch.org/docs/stable/torch.compiler.html).
Values:
- **NO** -- Do not use torch dynamo.
- **EAGER** -- Uses PyTorch to run the extracted GraphModule. This is quite useful in debugging TorchDynamo
issues.
- **AOT_EAGER** -- Uses AotAutograd with no compiler, i.e, just using PyTorch eager for the AotAutograd's
extracted forward and backward graphs. This is useful for debugging, and unlikely to give speedups.
- **INDUCTOR** -- Uses TorchInductor backend with AotAutograd and cudagraphs by leveraging codegened Triton
kernels. [Read
more](https://dev-discuss.pytorch.org/t/torchinductor-a-pytorch-native-compiler-with-define-by-run-ir-and-symbolic-shapes/747)
- **AOT_TS_NVFUSER** -- nvFuser with AotAutograd/TorchScript. [Read
more](https://dev-discuss.pytorch.org/t/tracing-with-primitives-update-1-nvfuser-and-its-primitives/593)
- **NVPRIMS_NVFUSER** -- nvFuser with PrimTorch. [Read
more](https://dev-discuss.pytorch.org/t/tracing-with-primitives-update-1-nvfuser-and-its-primitives/593)
- **CUDAGRAPHS** -- cudagraphs with AotAutograd. [Read more](https://github.com/pytorch/torchdynamo/pull/757)
- **OFI** -- Uses Torchscript optimize_for_inference. Inference only. [Read
more](https://pytorch.org/docs/stable/generated/torch.jit.optimize_for_inference.html)
- **FX2TRT** -- Uses Nvidia TensorRT for inference optimizations. Inference only. [Read
more](https://github.com/pytorch/TensorRT/blob/master/docsrc/tutorials/getting_started_with_fx_path.rst)
- **ONNXRT** -- Uses ONNXRT for inference on CPU/GPU. Inference only. [Read more](https://onnxruntime.ai/)
- **TENSORRT** -- Uses ONNXRT to run TensorRT for inference optimizations. [Read
more](https://github.com/onnx/onnx-tensorrt)
- **AOT_TORCHXLA_TRACE_ONCE** -- Uses Pytorch/XLA with TorchDynamo optimization, for training. [Read
more](https://github.com/pytorch/xla/blob/r2.0/docs/dynamo.md)
- **TORCHXLA_TRACE_ONCE** -- Uses Pytorch/XLA with TorchDynamo optimization, for inference. [Read
more](https://github.com/pytorch/xla/blob/r2.0/docs/dynamo.md)
- **IPEX** -- Uses IPEX for inference on CPU. Inference only. [Read
more](https://github.com/intel/intel-extension-for-pytorch).
- **TVM** -- Uses Apache TVM for inference optimizations. [Read more](https://tvm.apache.org/)
- **HPU_BACKEND** -- Uses HPU backend for inference optimizations.
rl EAGER AOT_EAGERINDUCTORAOT_TS_NVFUSERNVPRIMS_NVFUSER
CUDAGRAPHSOFIFX2TRTONNXRTTENSORRTAOT_TORCHXLA_TRACE_ONCETORCHXLA_TRACE_ONCEIPEXTVM HPU_BACKENDr? N)r@ rA rB rC rD rl r6 r7 r8 r9 r: r; r< r= r> r? r@ rA rB rC rD rE r? r3 r0 r5 r5 s\ "J
B EIH%N'O J
C
F
FH 7/ D
C Kr3 r5 c @ \ rS rSrSrSrSrSrSrSr Sr
S r S
r S r
S rS
rg)
LoggerTypei a= Represents a type of supported experiment tracker
Values:
- **ALL** -- all available trackers in the environment that are supported
- **TENSORBOARD** -- TensorBoard as an experiment tracker
- **WANDB** -- wandb as an experiment tracker
- **TRACKIO** -- trackio as an experiment tracker
- **COMETML** -- comet_ml as an experiment tracker
- **MLFLOW** -- mlflow as an experiment tracker
- **CLEARML** -- clearml as an experiment tracker
- **DVCLIVE** -- dvclive as an experiment tracker
- **SWANLAB** -- swanlab as an experiment tracker
allaim tensorboardwandbtrackiocomet_mlmlflowclearmldvcliveswanlabr? N)r@ rA rB rC rD ALLAIM TENSORBOARDWANDBTRACKIOCOMETMLMLFLOWCLEARMLDVCLIVESWANLABrE r? r3 r0 rF rF s;
C
C K EGG
FGGGr3 rF c ( \ rS rSrSrSrSrSrSrSr g)
PrecisionTypei zRepresents a type of precision used on floating point values
Values:
- **NO** -- using full precision (FP32)
- **FP16** -- using half precision
- **BF16** -- using brain floating point precision
rg fp8rh ri r? N)
r@ rA rB rC rD rl FP8rm rn rE r? r3 r0 r\ r\ s
B
C D Dr3 r\ c < \ rS rSrSrSrSrSrSrSr Sr
S r S
r S r
S rg
)RNGTypei r
r mlusdaamusanpuxlar r generatorr? N)r@ rA rB rC TORCHr MLUSDAAMUSANPUr& r r GENERATORrE r? r3 r0 r` r` s4 E D
C D D
C
C
C
CIr3 r` c $ \ rS rSrSrSrSrSrSrg) CustomDtypei z\
An enum that contains multiple custom dtypes that can be used for `infer_auto_device_map`.
r] int4int2r? N) r@ rA rB rC rD r^ INT4INT2rE r? r3 r0 rn rn s C D Dr3 rn c R \ rS rSr% \R
\S' \R \S' Srg)TensorInformationi shapedtyper? N) r@ rA rB rC r
Sizerc rv rE r? r3 r0 rt rt s :: ;;r3 rt c \ rS rSr% Sr\" SSS0S9r\\S' \" SSS 0S9r \\S
' \" S SS 0S9r
\\S
' \" SSS0S9r \\S' \" SSS0S9r \
\S' \" SSS0S9r\\S' \" SSS0S9r\\S' Srg)DataLoaderConfigurationi a
Configuration for dataloader-related items when calling `accelerator.prepare`.
Args:
split_batches (`bool`, defaults to `False`):
Whether or not the accelerator should split the batches yielded by the dataloaders across the devices. If
`True`, the actual batch size used will be the same on any kind of distributed processes, but it must be a
round multiple of `num_processes` you are using. If `False`, actual batch size used will be the one set in
your script multiplied by the number of processes.
dispatch_batches (`bool`, defaults to `None`):
If set to `True`, the dataloader prepared by the Accelerator is only iterated through on the main process
and then the batches are split and broadcast to each process. Will default to `True` for `DataLoader` whose
underlying dataset is an `IterableDataset`, `False` otherwise.
even_batches (`bool`, defaults to `True`):
If set to `True`, in cases where the total batch size across all processes does not exactly divide the
dataset, samples at the start of the dataset will be duplicated so the batch can be divided equally among
all workers.
use_seedable_sampler (`bool`, defaults to `False`):
Whether or not use a fully seedable random sampler ([`data_loader.SeedableRandomSampler`]). Ensures
training results are fully reproducible using a different sampling technique. While seed-to-seed results
may differ, on average the differences are negligible when using multiple different seeds to compare.
Should also be ran with [`~utils.set_seed`] for the best results.
data_seed (`int`, defaults to `None`):
The seed to use for the underlying generator when using `use_seedable_sampler`. If `None`, the generator
will use the current default seed from torch.
non_blocking (`bool`, defaults to `False`):
If set to `True`, the dataloader prepared by the Accelerator will utilize non-blocking host-to-device
transfers, allowing for better overlap between dataloader communication and computation. Recommended that
the prepared dataloader has `pin_memory` set to `True` to work properly.
use_stateful_dataloader (`bool`, defaults to `False`):
If set to `True`, the dataloader prepared by the Accelerator will be backed by
[torchdata.StatefulDataLoader](https://github.com/pytorch/data/tree/main/torchdata/stateful_dataloader).
This requires `torchdata` version 0.8.0 or higher that supports StatefulDataLoader to be installed.
Fhelpau Whether or not the accelerator should split the batches yielded by the dataloaders across the devices. If `True` the actual batch size used will be the same on any kind of distributed processes, but it must be a round multiple of the `num_processes` you are using. If `False`, actual batch size used will be the one set in your script multiplied by the number of processes.defaultmetadata
split_batchesNa If set to `True`, the dataloader prepared by the Accelerator is only iterated through on the main process and then the batches are split and broadcast to each process. Will default to `True` for `DataLoader` whose underlying dataset is an `IterableDataset`, `False` otherwise.dispatch_batchesTzIf set to `True`, in cases where the total batch size across all processes does not exactly divide the dataset, samples at the start of the dataset will be duplicated so the batch can be divided equally among all workers. even_batchesar Whether or not use a fully seedable random sampler ([`data_loader.SeedableRandomSampler`]).Ensures training results are fully reproducible using a different sampling technique. While seed-to-seed results may differ, on average the differences are negligible when usingmultiple different seeds to compare. Should also be ran with [`~utils.set_seed`] for the best results.use_seedable_samplerzThe seed to use for the underlying generator when using `use_seedable_sampler`. If `None`, the generator will use the current default seed from torch. data_seeda If set to `True`, the dataloader prepared by the Accelerator will utilize non-blocking host-to-device transfers, allowing for better overlap between dataloader communication and computation. Recommended that the prepared dataloader has `pin_memory` set to `True` to work properly. non_blockinga If set to `True`, the dataloader prepared by the Accelerator will be backed by [torchdata.StatefulDataLoader](https://github.com/pytorch/data/tree/main/torchdata/stateful_dataloader). This requires `torchdata` version 0.8.0 or higher that supports StatefulDataLoader to be installed.use_stateful_dataloaderr? )r@ rA rB rC rD r r~ rb rc r r r r r r r rE r? r3 r0 ry ry s !F E
M4 # N
d
L$ "' u
"$ =
Is T
L$ %* [
%T r3 ry c \ rS rSr% Sr\" SSS0S9r\\S' \" SSS0S9r \\S ' \" S
SS 0S9r
\ \S ' \" SSS
0S9r \
\S' \" SSS0S9r\
\S' \" S
SS0S9r\ \S' SS\\ 4S jjrS rSrg)ProjectConfigurationiq aD
Configuration for the Accelerator object based on inner-project needs.
Args:
project_dir (`str`, defaults to `None`):
A path to a directory for storing data.
logging_dir (`str`, defaults to `None`):
A path to a directory for storing logs of locally-compatible loggers. If None, defaults to `project_dir`.
automatic_checkpoint_naming (`bool`, defaults to `False`):
Whether saved states should be automatically iteratively named.
total_limit (`int`, defaults to `None`):
The maximum number of total saved states to keep.
iteration (`int`, defaults to `0`):
The current save iteration.
save_on_each_node (`bool`, defaults to `False`):
When doing multi-node distributed training, whether to save models and checkpoints on each node, or only on
the main one.
Nrz z'A path to a directory for storing data.r{ project_dirziA path to a directory for storing logs of locally-compatible loggers. If None, defaults to `project_dir`. logging_dirFz?Whether saved states should be automatically iteratively named.automatic_checkpoint_namingz1The maximum number of total saved states to keep. total_limitr zThe current save iteration. iterationzxWhen doing multi-node distributed training, whether to save models and checkpoints on each node, or only on the main onesave_on_each_nodec 8 Xl U R c Xl gg)zISets `self.project_dir` and `self.logging_dir` to the appropriate values.N)r r )r/ r s r0 set_directories$ProjectConfiguration.set_directories s & # * $r3 c : U R U R 5 g r* )r r r. s r0 r "ProjectConfiguration.__post_init__ s T - -.r3 )r r r* )r@ rA rB rC rD r r rY rc r r rb r r r r r r r rE r? r3 r0 r r q s & TV=f4ghKh
K ).[\)
MNK
78Is
$ (
t +8C= + /r3 r c \ rS rSr% Sr\" SSS0S9r\\S' \" SSS 0S9r \
\S
' \" SSS 0S9r \
\S ' \" S
SS0S9r \
\S' Sr
g)GradientAccumulationPlugini a
A plugin to configure gradient accumulation behavior. You can only pass one of `gradient_accumulation_plugin` or
`gradient_accumulation_steps` to [`Accelerator`]. Passing both raises an error.
Parameters:
num_steps (`int`):
The number of steps to accumulate gradients for.
adjust_scheduler (`bool`, *optional*, defaults to `True`):
Whether to adjust the scheduler steps to account for the number of steps being accumulated. Should be
`True` if the used scheduler was not adjusted for gradient accumulation.
sync_with_dataloader (`bool`, *optional*, defaults to `True`):
Whether to synchronize setting the gradients when at the end of the dataloader.
sync_each_batch (`bool`, *optional*):
Whether to synchronize setting the gradients at each data batch. Setting to `True` may reduce memory
requirements when using gradient accumulation with distributed training, at expense of speed.
Example:
```python
from accelerate.utils import GradientAccumulationPlugin
gradient_accumulation_plugin = GradientAccumulationPlugin(num_steps=2)
accelerator = Accelerator(gradient_accumulation_plugin=gradient_accumulation_plugin)
```
Nrz z0The number of steps to accumulate gradients for.r{ num_stepsTzWhether to adjust the scheduler steps to account for the number of steps being accumulated. Should be `True` if the used scheduler was not adjusted for gradient accumulation.adjust_schedulerzWhether to synchronize setting the gradients when at the end of the dataloader. Should only be set to `False` if you know what you're doing.sync_with_dataloaderFzWhether to synchronize setting the gradients at each data batch. Setting to `True` may reduce memory requirements when using gradient accumulation with distributed training, at expense of speed.sync_each_batchr? )r@ rA rB rC rD r r r rc r rb r r rE r? r3 r0 r r s 4 LMIs # E
d "' c
"$ " Y
OT r3 r c d ^ \ rS rSr% Sr\" SSS\ V Vs/ s H oR R 5 PM snn 30S9r \\
S' \" SSS0S9r \ \
S ' \" SSS
0S9r
\\
S ' \" SSS 0S9r\\
S
' \" SSS0S9r\\
S' \" SSS0S9r\\
S' \" SSS0S9r\\
S' S rS rU4S jrSrU=r$ s snn f )TorchDynamoPlugini a
This plugin is used to compile a model with PyTorch 2.0
Args:
backend (`DynamoBackend`, defaults to `None`):
A valid Dynamo backend. See https://pytorch.org/docs/stable/torch.compiler.html for more details.
mode (`str`, defaults to `None`):
Possible options are 'default', 'reduce-overhead' or 'max-autotune'.
fullgraph (`bool`, defaults to `None`):
Whether it is ok to break model into several subgraphs.
dynamic (`bool`, defaults to `None`):
Whether to use dynamic shape for tracing.
options (`Any`, defaults to `None`):
A dictionary of options to pass to the backend.
disable (`bool`, defaults to `False`):
Turn torch.compile() into a no-op for testing
use_regional_compilation (`bool`, defaults to `None`):
Use it to reduce the cold start compilation time of torch.compile() by targeting repeated blocks of the
same class and compiling them sequentially to hit the compiler's cache. For example, in `GPT2LMHeadModel`,
the repeated block/class is `GPT2Block`, and can be accessed as `model.transformer.h[0]`. The rest of the
model (e.g model.lm_head) is compiled separately.
Nrz zPossible options are r{ r zCPossible options are 'default', 'reduce-overhead' or 'max-autotune'modez6Whether it is ok to break model into several subgraphs fullgraphz(Whether to use dynamic shape for tracingdynamicz/A dictionary of options to pass to the backend.optionsFz-Turn torch.compile() into a no-op for testingdisablean Use it to reduce the cold start compilation time of torch.compile() by targeting repeated blocks of the same class and compiling them sequentially to hit the compiler's cache. For example, in `GPT2LMHeadModel`, the repeated block/class is `GPT2Block`, and can be accessed as `model.transformer.h[0]`. The rest of the model (e.g model.lm_head) is compiled separately.use_regional_compilationc SnU R c( [ R R US- S5 U l [ U R R 5 5 U l U R c( [ R R US- S5 U l U R c4 [ [ R R US- S5 5 S:H U l U R c4 [ [ R R US - S5 5 S:H U l U R cZ [ R R US
- S 5 b5 [ [ R R US
- S5 5 S:H U l
g g g ) NACCELERATE_DYNAMO_r rg MODEr|
USE_FULLGRAPHFalser USE_REGIONAL_COMPILATION USE_DYNAMIC) r r r r r5 r r r r r r r/ prefixs r0 r TorchDynamoPlugin.__post_init__ s& % << ::>>&9*> ! ( 8PRY)Z [_` `DN ( ( 0BJJNN64N+NPW XY]^^
) << BJJNN6M3I4$P$\&rzz~~f}6Lg'VW[\\DL %] r3 c [ R " U R 5 nUS R R 5 US' U$ )Nr )r+ r, r- rS r )r/
dynamo_configs r0 r1 TorchDynamoPlugin.to_dict2 s8
dmm4
#0#;#A#A#G#G#I
i r3 c H > [ TU ] 5 nUR SS 5 U$ )Nr )r r= pop)r/ kwargsr7 s r0 r= TorchDynamoPlugin.to_kwargs7 s$ "$
-t4
r3 )r r r r r )r@ rA rB rC rD r r5 rS r r rc r rY r rb r r r r r r r1 r= rE r ).0br7 s 00@r0 r r s . # 1M2ZMq77==?M2Z1[\] G] _`D# RSIt $&:d1efGTfKLGS IJGT
&+ q
& d ]$
q 3[s #B,r c d \ rS rSr% Sr\" SSS0S9r\\S' \" SSS0S9r \
\S ' \" SSS
0S9r \ \S ' \" SSS 0S9r
\
\S
' \" SSS0S9r\\S' \" SSS0S9r\\S' \" SSS0S9r\\S' \" SSS0S9r\\S' \" SSS0S9r\\S' \" SSS0S9r\\S' \" SSS0S9r\\S' \" SSS 0S9r\\S ' \" SSS 0S9r\\S ' \" SSS 0S9r\\S! \S"' S# r S6S$ jr S% r S7S& jr S8S' jr S( r!S) r"S* r#\$S9S+ j5 r%S, r&S- r'S9S.\4S/ jjr(S0 r)S1\4S2 jr*\+S3 5 r,\,RZ S4 5 r,S5r.g):DeepSpeedPlugini= a
This plugin is used to integrate DeepSpeed.
Args:
hf_ds_config (`Any`, defaults to `None`):
Path to DeepSpeed config file or dict or an object of class `accelerate.utils.deepspeed.HfDeepSpeedConfig`.
gradient_accumulation_steps (`int`, defaults to `None`):
Number of steps to accumulate gradients before updating optimizer states. If not set, will use the value
from the `Accelerator` directly.
gradient_clipping (`float`, defaults to `None`):
Enable gradient clipping with value.
zero_stage (`int`, defaults to `None`):
Possible options are 0, 1, 2, 3. Default will be taken from environment variable.
is_train_batch_min (`bool`, defaults to `True`):
If both train & eval dataloaders are specified, this will decide the `train_batch_size`.
offload_optimizer_device (`str`, defaults to `None`):
Possible options are none|cpu|nvme. Only applicable with ZeRO Stages 2 and 3.
offload_param_device (`str`, defaults to `None`):
Possible options are none|cpu|nvme. Only applicable with ZeRO Stage 3.
offload_optimizer_nvme_path (`str`, defaults to `None`):
Possible options are /nvme|/local_nvme. Only applicable with ZeRO Stage 3.
offload_param_nvme_path (`str`, defaults to `None`):
Possible options are /nvme|/local_nvme. Only applicable with ZeRO Stage 3.
zero3_init_flag (`bool`, defaults to `None`):
Flag to indicate whether to save 16-bit model. Only applicable with ZeRO Stage-3.
zero3_save_16bit_model (`bool`, defaults to `None`):
Flag to indicate whether to save 16-bit model. Only applicable with ZeRO Stage-3.
transformer_moe_cls_names (`str`, defaults to `None`):
Comma-separated list of Transformers MoE layer class names (case-sensitive). For example,
`MixtralSparseMoeBlock`, `Qwen2MoeSparseMoeBlock`, `JetMoEAttention`, `JetMoEBlock`, etc.
enable_msamp (`bool`, defaults to `None`):
Flag to indicate whether to enable MS-AMP backend for FP8 training.
msasmp_opt_level (`Optional[Literal["O1", "O2"]]`, defaults to `None`):
Optimization level for MS-AMP (defaults to 'O1'). Only applicable if `enable_msamp` is True. Should be one
of ['O1' or 'O2'].
Nrz zkpath to DeepSpeed config file or dict or an object of class `accelerate.utils.deepspeed.HfDeepSpeedConfig`.r{ hf_ds_configzNumber of steps to accumulate gradients before updating optimizer states. If not set, will use the value from the `Accelerator` directly.gradient_accumulation_stepsz#Enable gradient clipping with valuegradient_clippingzMPossible options are 0,1,2,3; Default will be taken from environment variable
zero_stageTzUIf both train & eval dataloaders are specified, this will decide the train_batch_sizeis_train_batch_minzMPossible options are none|cpu|nvme. Only applicable with ZeRO Stages 2 and 3.offload_optimizer_devicezFPossible options are none|cpu|nvme. Only applicable with ZeRO Stage 3.offload_param_devicezJPossible options are /nvme|/local_nvme. Only applicable with ZeRO Stage 3.offload_optimizer_nvme_pathoffload_param_nvme_pathz{Flag to indicate whether to enable `deepspeed.zero.Init` for constructing massive models.Only applicable with ZeRO Stage-3.zero3_init_flagzQFlag to indicate whether to save 16-bit model. Only applicable with ZeRO Stage-3.zero3_save_16bit_modelzcomma-separated list of transformers MoE layer class names (case-sensitive), e.g : `MixtralSparseMoeBlock`, `Qwen2MoeSparseMoeBlock`, `JetMoEAttention,JetMoEBlock` ...transformer_moe_cls_nameszCFlag to indicate whether to enable MS-AMP backend for FP8 training. enable_msampz}Optimization level for MS-AMP (defaults to 'O1'). Only applicable if `enable_msamp` is True. Should be one of ['O1' or 'O2'].r msamp_opt_levelc SSK Jn U R cG [ R R SS5 nUR
5 ( a [ U5 OUU l U R c8 [ R R SS5 nUS:X a UO
[ U5 U l U R c. [ [ R R SS5 5 U l
U R c% [ R R SS 5 U l U R c% [ R R S
S 5 U l U R c% [ R R S S 5 U l
U R c% [ R R S S 5 U l U R c6 [ R R S
S5 R! 5 S:H U l U R" c( [ R R SS 5 S:H U l U R$ c% [ R R SS5 U l U R& c% [ R R SS 5 U l [) U R& [* 5 ( dK [) U R& [, 5 ( a U R& S :w d [) U R& U5 ( Gaw [) U R& U5 ( d U" U R& 5 U l SU R& R. ; a SU R& R. S' SU R& R. ; a [1 S5 eU R3 5 SSSSSS S S S .nUR5 5 VVs0 s H u pV[7 X5 c M U[7 X5 _M! nnnUR9 5 H nU R: " U40 UDS S!0D6 M U R& R= 5 UR5 5 H9 u pU R& R? U 5 n
U
c M% U
S:w d M- [A XU
5 M; OSSU R U R U R U R S":X a U R OS S#.U R U R S":X a U R OS S#.U R S$.S%.n U R ( a U R U S' U" U 5 U l U R& R. U l! [ S&5 U RB S'' U RD cR [G [ R R S([- U R& RI 5 5 5 5 S:H U l" U RD ( a< U R& RI 5 ( d [J RL " S)5 S!U l" U RO S!5 U R" ( ad S*U RB ; aS U R S+:X a [Q S,5 eU R$ S-; a [1 S.5 eS/U R$ S0.U RB S*' g g g s snnf )1Nr )HfDeepSpeedConfig&ACCELERATE_GRADIENT_ACCUMULATION_STEPSauto ACCELERATE_GRADIENT_CLIPPING ACCELERATE_DEEPSPEED_ZERO_STAGE -ACCELERATE_DEEPSPEED_OFFLOAD_OPTIMIZER_DEVICEnone)ACCELERATE_DEEPSPEED_OFFLOAD_PARAM_DEVICE0ACCELERATE_DEEPSPEED_OFFLOAD_OPTIMIZER_NVME_PATH,ACCELERATE_DEEPSPEED_OFFLOAD_PARAM_NVME_PATH+ACCELERATE_DEEPSPEED_ZERO3_SAVE_16BIT_MODELfalsetrueACCELERATE_FP8_BACKENDr ACCELERATE_FP8_OPT_LEVELr ACCELERATE_DEEPSPEED_CONFIG_FILEr zero_optimizationzDPlease specify the ZeRO optimization config in the DeepSpeed config.r zzero_optimization.stagez*zero_optimization.offload_optimizer.devicez&zero_optimization.offload_param.devicez)zero_optimization.offload_param.nvme_pathz-zero_optimization.offload_optimizer.nvme_pathz;zero_optimization.stage3_gather_16bit_weights_on_model_save)r r r r r r r r
must_matchFnvme)device nvme_path)stageoffload_optimizer
offload_param)stage3_gather_16bit_weights_on_model_save)train_batch_size train_micro_batch_size_per_gpur r infsteps_per_print ACCELERATE_DEEPSPEED_ZERO3_INITzSDeepSpeed Zero3 Init flag is only applicable for ZeRO Stage 3. Setting it to False.r zSMS-AMP is not supported for ZeRO Stage 3. Please use ZeRO Stage 0, 1, or 2 instead.r zGInvalid optimization level for MS-AMP. Please use one of ['O1' or'O2'].Tr` r )) deepspeedr r r r r isdigitr r r r r r r r r r r r r
isinstancer rY r rJ _deepspeed_config_checksr8 getattrkeys
fill_matchset_stage_and_offload get_valuesetattrdeepspeed_configr r is_zero3r r
_set_selectedNotImplementedError) r/ r gasr plugin_to_config_mappingr; r< r keyrS config_valuer s r0 r DeepSpeedPlugin.__post_init__ sd 0 + + 3**..!I6RC;>;;==s3xcD , ! ! ) "
/Mv V :Kv:U%6[`ar[sD " ?? " !"**..1RTU"V WDO ( ( 0,.JJNN;jlr,sD ) $ $ ,(*
7bdj(kD % + + 3/1zz~~BF0D , ' ' /+-::>>:hjp+qD ( & & .
LgV\\^bhh
' $ "
/G NRY YD '#%::>>2Ld#SD $ "
/QSY ZD t(($ / /4,,c22t7H7HF7R$++->?? d / /1BCC$5d6G6G$H!,D4E4E4L4LLJK!!(()FG"$*;*;*B*BB !ghh ) ) +/L%8 7,X(P+V/^*g ( $ 7O6T6T6Vw6VdaZabfZj)a))6VFw {{}@v@%@ % 3 3 5 7 < < >
# 0 0 : :5 A + 0FD|4 ? %+28/3/O/O !__"&"?"?@D@]@]ag@gD < ::f '$!'!4 } %''2m 4)) F# +"7 ,,!!E+ axx }TUV\ViUj"kl - #8 r3 c J U R R U5 nUc gUS:H $ )NFr r r )r/ r vals r0 is_autoDeepSpeedPlugin.is_autoA s* ))+6 ;&= r3 c 8 U R R X5 $ r* r )r/ r r| s r0 r DeepSpeedPlugin.get_valueH s **;@@r3 c d Uc / OUnUc U R nUR 5 HP u pg[ U[ 5 ( a U R " S
X- S- UUUS.UD6 M8 U R
" X- U4SU0UD6 MR [
U5 S: a' US:X a SR U5 n[ SU S 35 egg) z=Process the DeepSpeed config with the values from the kwargs.Nr )r r r r r r
zSPlease correct the following DeepSpeed config values that mismatch kwargs values:
zF
The easiest method is to set these DeepSpeed config values to 'auto'.r? ) r r8 r r deepspeed_config_processr lenr rJ ) r/ r r r r r r rS mismatches_msgs r0 r (DeepSpeedPlugin.deepspeed_config_processK s %-R:
>**F ,,.JC%&&-- !<#- ) )
jZZZSYZ ) z?Q 6R< !YYz 2N + ,,su
$0 r3 c U R nUS:H US; S.nUS:X a SU; a SSS.US' OUS; a
SU; a SS0US' US:X a) U R ( a S U; a SU R S
.US ' US :w aY US:X a SOSn[ UR U0 5 R SS 5 5 R 5 S
:X a [
SU SU S35 eS H nXR; d M
SS0X%' M U R " SSS0UD6 U R " SSS0UD6 g )Nrh )ri r] ) fp16.enabled bf16.enabledT)r` auto_castri r` r] r r rg r r z*`--mixed_precision` arg cannot be set to `z` when `z&` is set in the DeepSpeed config file.)rh ri Fr )r )r )r r r rY r r rJ r )r/ mixed_precision ds_configr
diff_dtyperv s r0 set_mixed_precision#DeepSpeedPlugin.set_mixed_precisionb sP )) +v 5 + >
f $Y&044$H &!
/Y&%.$5 &! e #(9(9i' #!%!5!5& '"
d "#2f#<&J9==R044YHIOOQU[[ @ @QQYZdYe fL M &E%$-u#5 &
C5CFC C5CFCr3 c SSK Jn [ R " U R 5 nU R
( a U" 5 ( d [
S5 eSU; d US S:X a SUS' SU; d US S:X a SUS' UR SS 5 S:X a US [ SS S
5 ( a S S K J
nJ n OS S K J
nJ n U" 5 U" U5 U l
g )
Nr )is_transformers_availablezoWhen `zero3_init_flag` is set, it requires Transformers to be installed. Please run `pip install transformers`.r r r r transformers# U H oT; v M g 7fr* r? )r name env_variable_names_to_ignores r0 ;DeepSpeedPlugin._deepspeed_config_checks.. s h@g33@gs z_When using `deepspeed_config_file`, the following accelerate config variables will be ignored: am .
Please specify them appropriately in the DeepSpeed config file.
If you are using an accelerate config file, remove others config variables mentioned in the above specified list.
The easiest method is to create a new config following the questionnaire via `accelerate config`.
It will only ask for the necessary config variables when using `deepspeed_config_file`.)replacer r r r splitanyrJ )r/ r4 'deepspeed_fields_from_accelerate_configr5 s @r0 r (DeepSpeedPlugin._deepspeed_config_checks s
(
$ [w(
ZvRVLL + 3 3L" E K K MZv(
$ 35**..A^`b2c2i2ijm2n/ h@gh h h q sO rP Pj j
i
(
s 9B c U R c% [ R R SS 5 U l U R b [ SSS5 ( a [ S5 eSSKJn U R R S5 n/ nU H1 n[ X5 nUc [ S U S
35 eUR U5 M3 U" X5 g g ) N(ACCELERATE_DEEPSPEED_MOE_LAYER_CLS_NAMESr r z0.14.0zPDeepSpeed version must be >= 0.14.0 to use MOE support. Please update DeepSpeed.r )set_z3_leaf_modulesr2 z1Could not find a transformer layer class called 'z' to wrap in the model.) r r r r r# r deepspeed.utilsr? r9 get_module_class_from_namer r )r/ r r? class_namestransformer_moe_cls layer_classtransformer_clss r0 set_moe_leaf_modules$DeepSpeedPlugin.set_moe_leaf_modules s ) ) 1-/ZZ^^>sCK"$ * "$ ) )6<( !d !" 4 __
r3 r c \ rS rSr% Sr\" SSS0S9r\\S' \" SSS0S9r \
\ S 4 \S
' \" SSS 0S9r \
\ S \
4 \S ' \" SSS
0S9r\\
\ S4 \S' \" SSS0S9r\\
\\ SS4 \S' \" SSS0S9r\\
\\S 4 \S' \" SSS0S9r\
\
SS4 \S' \" SSS0S9r\\
\\R2 R4 \ 4 \S ' \" SSS 0S9r\
\ S 4 \S ' \" SSS 0S9r \\
S! \S"' \" SSS#0S9r \\
S$ \S%' \" S&SS'0S9r \
\S(' \" SSS)0S9r \\
\S*' \" SSS+0S9r \\\R2 R4 /S4 \S,' \" SSS-0S9r!\\
\S.' \" SSS/0S9r"\
\S0' \" SSS10S9r#\
\S2' \" SSS30S9r$\
\S4' \" SSS50S9r%\\&\ \S6' \" SSS70S9r'\\ \S8' S9 r(SAS: jr)S; r*SBS< jr+S= r,S> r-S? r.S@r/g)C FullyShardedDataParallelPlugini a
This plugin is used to enable fully sharded data parallelism.
Args:
fsdp_version (`int`, defaults to `1`):
The version of FSDP to use. Defaults to 1. If set to 2, launcher expects the config to be converted to
FSDP2 format.
sharding_strategy (`Union[str, torch.distributed.fsdp.ShardingStrategy]`, defaults to `'FULL_SHARD'`):
Sharding strategy to use. Should be either a `str` or an instance of
`torch.distributed.fsdp.fully_sharded_data_parallel.ShardingStrategy`. Is deprecated in favor of
`reshard_after_forward`.
reshard_after_forward (`Union[str, torch.distributed.fsdp.ShardingStrategy, bool]`, defaults to `'FULL_SHARD'` for `fsdp_version=1` and `True` for `fsdp_version=2`):
Sharding strategy to use. Should be a bool if `fsdp_version` is set to 2 else a `str` or an instance of
`torch.distributed.fsdp.fully_sharded_data_parallel.ShardingStrategy`.
backward_prefetch (`Union[str, torch.distributed.fsdp.BackwardPrefetch]`, defaults to `'NO_PREFETCH'`):
Backward prefetch strategy to use. Should be either a `str` or an instance of
`torch.distributed.fsdp.fully_sharded_data_parallel.BackwardPrefetch`.
mixed_precision_policy (`Optional[Union[dict, str, torch.distributed.fsdp.MixedPrecision, torch.distributed.fsdp.MixedPrecisionPolicy]]`, defaults to `None`):
A config to enable mixed precision training with FullyShardedDataParallel. If passing in a `dict`, it
should have the following keys: `param_dtype`, `reduce_dtype`, and `buffer_dtype`, can be an instance of
`torch.distributed.fsdp.MixedPrecisionPolicy` if `fsdp_version` is set to 2. If passing in a `str`, it
should be one of the following values: fp8, fp16, bf16, fp32, and used to set `param_dtype`,
`reduce_dtype`, and `buffer_dtype`.
auto_wrap_policy (`Optional(Union[Callable, Literal["transformer_based_wrap", "size_based_wrap", "no_wrap"]]), defaults to `NO_WRAP`):
A callable or string specifying a policy to recursively wrap layers with FSDP. If a string, it must be one
of `transformer_based_wrap`, `size_based_wrap`, or `no_wrap`. See
`torch.distributed.fsdp.wrap.size_based_wrap_policy` for a direction on what it should look like.
cpu_offload (`Union[bool, torch.distributed.fsdp.CPUOffload, torch.distributed.fsdp.CPUOffloadPolicy]`, defaults to `False`):
Whether to offload parameters to CPU. Should be either a `bool` or an instance of
`torch.distributed.fsdp.fully_sharded_data_parallel.CPUOffload` or
`torch.distributed.fsdp.fully_sharded_data_parallel.CPUOffloadPolicy` if `fsdp_version` is set to 2.
ignored_modules (`Optional[Union[Iterable[torch.nn.Module], str]]`, defaults to `None`):
A list of modules to ignore when wrapping with FSDP. When passing a string, will match the modules by name
using regex fullmatch. If `fsdp_version` is set to 2, the modules are converted to parameters and used.
state_dict_type (`Union[str, torch.distributed.fsdp.StateDictType]`, defaults to `'FULL_STATE_DICT'`):
State dict type to use. If a string, it must be one of `full_state_dict`, `local_state_dict`, or
`sharded_state_dict`.
state_dict_config (`Optional[Union[torch.distributed.fsdp.FullStateDictConfig, torch.distributed.fsdp.ShardedStateDictConfig]`, defaults to `None`):
State dict config to use. Is determined based on the `state_dict_type` if not passed in.
optim_state_dict_config (`Optional[Union[torch.distributed.fsdp.FullOptimStateDictConfig, torch.distributed.fsdp.ShardedOptimStateDictConfig]`, defaults to `None`):
Optim state dict config to use. Is determined based on the `state_dict_type` if not passed in.
limit_all_gathers (`bool`, defaults to `True`):
Whether to have FSDP explicitly synchronizes the CPU thread to prevent too many in-flight all-gathers. This
bool only affects the sharded strategies that schedule all-gathers. Enabling this can help lower the number
of CUDA malloc retries.
use_orig_params (`bool`, defaults to `False`):
Whether to use the original parameters for the optimizer.
param_init_fn (`Optional[Callable[[torch.nn.Module], None]`, defaults to `None`):
A `Callable[torch.nn.Module] -> None` that specifies how modules that are currently on the meta device
should be initialized onto an actual device. Only applicable when `sync_module_states` is `True`. By
default is a `lambda` which calls `to_empty` on the module.
sync_module_states (`bool`, defaults to `False`):
Whether each individually wrapped FSDP unit should broadcast module parameters from rank 0 to ensure they
are the same across all ranks after initialization. Defaults to `False` unless `cpu_ram_efficient_loading`
is `True`, then will be forcibly enabled.
forward_prefetch (`bool`, defaults to `False`):
Whether to have FSDP explicitly prefetches the next upcoming all-gather while executing in the forward
pass. only use with Static graphs.
activation_checkpointing (`bool`, defaults to `False`):
A technique to reduce memory usage by clearing activations of certain layers and recomputing them during a
backward pass. Effectively, this trades extra computation time for reduced memory usage.
cpu_ram_efficient_loading (`bool`, defaults to `None`):
If True, only the first process loads the pretrained model checkoint while all other processes have empty
weights. Only applicable for Transformers. When using this, `sync_module_states` needs to be `True`.
transformer_cls_names_to_wrap (`Optional[List[str]]`, defaults to `None`):
A list of transformer layer class names to wrap. Only applicable when `auto_wrap_policy` is
`transformer_based_wrap`.
min_num_params (`Optional[int]`, defaults to `None`):
The minimum number of parameters a module must have to be wrapped. Only applicable when `auto_wrap_policy`
is `size_based_wrap`.
Nrz ztThe version of FSDP to use. Defaults to 1. If set to 2, launcher expects the config to be converted to FSDP2 format.r{ fsdp_versionzSharding strategy to use. Should be either a `str` or an instance of `torch.distributed.fsdp.fully_sharded_data_parallel.ShardingStrategy`. Defaults to 'FULL_SHARD'. Is deprecated in favor of `reshard_after_forward` z'torch.distributed.fsdp.ShardingStrategysharding_strategyzSharding strategy to use. Should be a bool if `fsdp_version` is set to 2 else a `str` or an instance of `torch.distributed.fsdp.fully_sharded_data_parallel.ShardingStrategy`. Defaults to 'FULL_SHARD'reshard_after_forwardzBackward prefetch strategy to use. Should be either a `str` or an instance of `torch.distributed.fsdp.fully_sharded_data_parallel.BackwardPrefetch`. Defaults to 'NO_PREFETCH'. This becomes obsolete in FSDP2.z'torch.distributed.fsdp.BackwardPrefetchbackward_prefetcha A config to enable mixed precision training with FullyShardedDataParallel. If passing in a `dict`, it should have the following keys: `param_dtype`, `reduce_dtype`, and `buffer_dtype`.Can also be an instance of `torch.distributed.fsdp.MixedPrecisionPolicy` if `fsdp_version` is set to 2.z%torch.distributed.fsdp.MixedPrecisionz+torch.distributed.fsdp.MixedPrecisionPolicymixed_precision_policya$ A callable or string specifying a policy to recursively wrap layers with FSDP. If a string, it must be one of `transformer_based_wrap`, `size_based_wrap`, or `no_wrap`. Defaults to `NO_WRAP`. See `torch.distributed.fsdp.wrap.size_based_wrap_policy` for a direction on what it should look like)transformer_based_wrapsize_based_wrapno_wrapauto_wrap_policya
Whether to offload parameters to CPU. Should be either a `bool` or an instance of `torch.distributed.fsdp.fully_sharded_data_parallel.CPUOffload` or `torch.distributed.fsdp.fully_sharded_data_parallel.CPUOffloadPolicy` if `fsdp_version` is set to 2. Defaults to `False`z!torch.distributed.fsdp.CPUOffloadz'torch.distributed.fsdp.CPUOffloadPolicy cpu_offloadz4A list of modules to ignore when wrapping with FSDP.ignored_moduleszState dict type to use. If a string, it must be one of `full_state_dict`, `local_state_dict`, or `sharded_state_dict`. Defaults to `FULL_STATE_DICT`z$torch.distributed.fsdp.StateDictTypestate_dict_typezXState dict config to use. Is determined based on the `state_dict_type` if not passed in.)z*torch.distributed.fsdp.FullStateDictConfigz-torch.distributed.fsdp.ShardedStateDictConfigstate_dict_configz^Optim state dict config to use. Is determined based on the `state_dict_type` if not passed in.)z/torch.distributed.fsdp.FullOptimStateDictConfigz2torch.distributed.fsdp.ShardedOptimStateDictConfigoptim_state_dict_configTzWhether to have FSDP explicitly synchronizes the CPU thread to prevent too many in-flight all-gathers. This bool only affects the sharded strategies that schedule all-gathers. Enabling this can help lower the number of CUDA malloc retries.limit_all_gathersznWhether to use the original parameters for the optimizer. Defaults to `False`. This becomes obsolete in FSDP2.use_orig_paramsa A Callable[torch.nn.Module] -> None that specifies how modules that are currently on the meta device should be initialized onto an actual device. Only applicable when `sync_module_states` is `True`. By default is a `lambda` which calls `to_empty` on the module.
param_init_fna Whether each individually wrapped FSDP unit should broadcast module parameters from rank 0 to ensure they are the same across all ranks after initialization. Defaults to `False` unless `cpu_ram_efficient_loading` is `True`, then will be forcibly enabled. This becomes obsolete in FSDP2.sync_module_stateszWhether to have FSDP explicitly prefetches the next upcoming all-gather while executing in the forward pass. only use with Static graphs. Defaults to `False`forward_prefetchzA technique to reduce memory usage by clearing activations of certain layers and recomputing them during a backward pass. Effectively, this trades extra computation time for reduced memory usage. Defaults to `False`activation_checkpointingu If True, only the first process loads the pretrained model checkoint while all other processes have empty weights. Only applicable for 🤗 Transformers. When using this, `sync_module_states` needs to be `True`. Defaults to `False`.cpu_ram_efficient_loadingzuA list of transformer layer class names to wrap. Only applicable when `auto_wrap_policy` is `transformer_based_wrap`. transformer_cls_names_to_wrapzThe minimum number of parameters a module must have to be wrapped. Only applicable when `auto_wrap_policy` is `size_based_wrap`.min_num_paramsc ^ SSK JnJn [ 5 nSnU R c1 [ [ R R US- S5 5 U l U R S:X a' [ S[ 5 ( d [ S[ 35 eU R b UR S 5 U R S
:X Ga, U R c( [ R R US - S 5 U l [ U R [ 5 ( a U R R! 5 [" ; a6 [" R$ " U R R! 5 5 S
- U l [ U R [
5 ( d U R R' 5 ( a! U" [ U R 5 5 U l O!X R R! 5 U l U R( cv U R ci [ R R US
- U R S:X a SOS 5 nU R S:X a [+ UR- 5 SS9U l OXPl [ U R( [ 5 ( Ga U R S:X a( [+ U R( R- 5 SS9U l OU R( R! 5 [" ; a6 [" R$ " U R( R! 5 5 S
- U l [ U R( [
5 ( d U R( R' 5 ( a! U" [ U R( 5 5 U l O!X R( R! 5 U l U R S:X a8 [ U R( [. 5 ( d [1 SU R( S35 eU R S
:X a8 [ U R( [. 5 ( a [1 SU R( S35 eU R2 c4 [+ [ R R US- S5 5 S
:H U l U R5 5 U R7 5 U R8 c( [ R R US- S 5 U l [ U R8 [ 5 ( a% U R8 R! 5 S:X a S U l U R8 Gb [ U R8 U5 ( d [ U R8 [ 5 ( aX U R8 R! 5 [: ; a6 [: R$ " U R8 R! 5 5 S
- U l [ U R8 [
5 ( d U R8 R' 5 ( a! U" [ U R8 5 5 U l O!XR8 R! 5 U l U R S:X a% U R8 b UR S5 S U l U R= 5 U R> c( [ R R US- S5 U l [ U R> [ 5 ( Ga U R> R! 5 [@ ; a [1 SU R> S [@ 35 eSS K!J"nJ#n U R> R! 5 S :X a{ Xpl U RH c( [ R R US - S 5 U l$ [ U RH [ 5 ( a U RH RK S 5 U l$ OU R> R! 5 S!:X a X`l U RL c2 [ [ R R US"- S5 5 U l& Os[ U RL [
5 ( d. [1 S#U RL S$[O U RL 5 35 eO%U R> R! 5 S:X a S U l U RP cD U R S
:X a4 [+ [ R R US%- S5 5 S
:H U l( U R S:X a% U RP b UR S&5 S U l( U RR cD U R S
:X a4 [+ [ R R US'- S5 5 S
:H U l) U R S:X a% U RR b UR S(5 S U l) U RT cD U R S
:X a4 [+ [ R R US)- S5 5 S
:H U l* U R S:X a U RT b [1 S*5 eU RV c4 [+ [ R R US+- S5 5 S
:H U l+ U RX c( [ R R US,- S 5 U l, U RZ c4 [+ [ R R US-- S5 5 S
:H U l- U R S
:X a? U RZ ( a. U RR ( d [\ R^ " S.5 SU l) U RZ [/ [+ [ R R US-- S5 5 5 :w aE US-- n[\ R^ " S/U S035 [ U RZ 5 [ R U' [ U R` [ 5 ( a U Rc U R` SS19 O:[ U R` [d 5 ( a U Rc U R` 5 U R` b U Rg 5 U RR ( Ga* [i 5 ( a [j Rl Ro 5 m O[q 5 ( a [j Rr Ro 5 m O[u 5 ( a [j Rv Ro 5 m O[y 5 ( a [j Rz Ro 5 m Og[} 5 ( a [j R~ Ro 5 m O9[ 5 ( a [j R Ro 5 m O [ S25 eU 4S3 jU lC [ S4S55 ( a5 U R S:X a% U RX b UR S65 S U l, U( a% [ R S7R U5 5 g g )8Nr )BackwardPrefetchShardingStrategyFSDP_VERSION1r r zFSDP2 requires PyTorch >= zzsharding_strategy is deprecated in favor of reshard_after_forward. This will be removed in a future version of Accelerate.r SHARDING_STRATEGY
FULL_SHARDRESHARD_AFTER_FORWARDr T)to_boolz reshard_after_forward set to z:. This is not supported with FSDP2, please set to a `bool`z. This is not supported with FSDP1, please set to a `str` or an instance of `torch.distributed.fsdp.fully_sharded_data_parallel.ShardingStrategy`OFFLOAD_PARAMSr BACKWARD_PREFETCH NO_PREFETCHzObackward_prefetch is not supported in FSDP2. Setting backward prefetch to None.AUTO_WRAP_POLICYNO_WRAPzInvalid auto wrap policy: r size_based_auto_wrap_policy transformer_auto_wrap_policyTRANSFORMER_BASED_WRAPTRANSFORMER_CLS_TO_WRAPr2 SIZE_BASED_WRAPMIN_NUM_PARAMSz)`min_num_params` must be an integer. Got z of type USE_ORIG_PARAMSzSuse_orig_params is obsolete in FSDP2, as FSDP2 always uses the original parameters.SYNC_MODULE_STATESzhsync_module_states is obsolete in FSDP2, as it is not needed anymore.Setting sync_module_states to None.FORWARD_PREFETCHzUforward_prefetch is not yet implemented in FSDP2, set to None or use `fsdp_version=1`ACTIVATION_CHECKPOINTINGIGNORED_MODULESCPU_RAM_EFFICIENT_LOADINGzosync_module_states cannot be False since efficient cpu ram loading enabled. Setting sync_module_states to True.zrThe `cpu_ram_efficient_loading` flag for `FullyShardedDataParallelPlugin` does not match the environment variable zD. Setting environment variable to match `cpu_ram_efficient_loading`.)overridezmThere are currently no available devices found, must be one of 'XPU', 'CUDA', 'MLU', 'NPU', 'MUSA', or 'HPU'.c $ > U R TSS9$ )NF)r recurse)to_empty)xr s r0 >FullyShardedDataParallelPlugin.__post_init__.. s 1::VU:+Sr3 r z2.7.0zoFSDP2 ignored_params/ignored_modules is not available for torch version < 2.7.0Setting ignored_modules to None.z7Multiple deprecation warnings due to FSDP2 conversion:
)Gtorch.distributed.fsdprt ru setr\ r r r r r$ r r r] addr rY r r indexr r^ r r rb rJ re set_cpu_offloadvalidate_cpu_offloadr_ r set_state_dict_typerd r torch.distributed.fsdp.wrapr r rq r9 rr typerk rm rn ro rf rp r r r` r r validate_mixed_precision_policyr r
rd current_devicer ra r rc r r r" r r r RuntimeErrorrl loggerwarningr )
r/ rt ru _fsdp2_warningsr r^ r r env_varr s
@r0 r ,FullyShardedDataParallelPlugin.__post_init__ s M %
$ #BJJNN: 3I3$O PD !#D*?@@!$>?T>U"VWW ! ! -
J
!%%-)+
EX8XZf)g&$00#66))//15KK-C-I-I$J`J`JfJfJh-ilm-mD* d 4 4c::d>T>T>\>\>^>^-=c$BXBX>Y-ZD*-=>T>T>Z>Z>\-]D* % % -$2H2H2P$&JJNN 44 ++q0l% ! A%-89N9T9T9V`d-e*-B* d00# 6 6 A%-89S9S9Y9Y9[ei-j* --3359OO1G1M1MdNhNhNnNnNp1qtu1uD. d 8 8#>>$B\B\BdBdBfBf1A#dF`F`Ba1bD.1AB\B\BbBbBd1eD. !*T5O5OQU*V*V /0J0J/K LF G
!j1K1KT&R&R /0J0J/K L] ^
# *2::>>*GW:WY`+a bfg gD !!# ! ! )%'ZZ^^JAT4TVZ%[D " d,,c 2 2t7M7M7S7S7UYf7f%)D " ! ! -jAWAWYi6j6j$00#664;Q;Q;W;W;Y]s;s)?)E)EdF\F\FbFbFd)ehi)i&$00#66$:P:P:X:X:Z:Z)9#d>T>T:U)V&)9:P:P:V:V:X)Y& !d&<&<&H q r%)D " " ($&JJNN:@R3RT]$^D ! d++S 1 1$$**,4II 01F1F0GGXYnXop
$$**,0HH(D%55=9;
UnHnpt9uD6 d @ @#FF9=9[9[9a9abe9fD6&&,,.2CC(C%&&.*-bjjnnZJZ=Z\].^*_D'#D$7$7==$CDDWDWCXXabfgkgzgzb{a|} > &&,,.);(,% 'D,=,=,B#.rzz~~jK\>\^e/f#gkl#lD !d&:&:&F u v#'D " " *t/@/@A/E&1"**..NbAbdk2l&mqr&rD # !d&=&=&I 6
'+D # (T->->!-C$/
zL^?^`g0h$imn$nD ! !d&;&;&G tu u ( ( 0BJJNN:8R+RT[ \]abb
) '#%::>>*?P2PRV#WD ) ) 1BJJNN:8S+SU\ ]^bcc
* !d&D&DTMdMd MM6
'+D # ) )T
z4O'OQXY Z.
!#>>G MM E FM EN NU U
#&d&D&D"EBJJw d113 7 7
$ $T%@%@4 $ P
33T
:
: $ $T%@%@ A & & 2 0 0 2 " " " !! 113!## 113"$$ 224"$$ 224!## 113!## 113" D
"TD C ) )d.?.?1.DI]I]Ii 3
$(D NNUZZ[jk l r3 c T SSK JnJnJnJnJn Ub Xl U R c7 [ R R SU R S:X a SOS5 U l [ U R [ 5 ( aa U R R 5 ( a! U" [ U R 5 5 U l O!X`R R 5 U l U R UR :X a3 U R" c U" SSS 9U l U R$ c U" SSS 9U l OJU R UR& :X a0 U R" c U" SS
9U l U R$ c U" SS
9U l U R S :X a& U R UR( :X a [+ S 5 egg)
z9
Set the state dict config based on the `StateDictType`.
r )FullOptimStateDictConfigFullStateDictConfigShardedOptimStateDictConfigShardedStateDictConfig
StateDictTypeNFSDP_STATE_DICT_TYPEr FULL_STATE_DICTSHARDED_STATE_DICTT)offload_to_cpu
rank0_only)r r zxFSDP2 does not support LOCAL_STATE_DICT. Please set `fsdp_state_dict_type` to `SHARDED_STATE_DICT` or `FULL_STATE_DICT`.)2torch.distributed.fsdp.fully_sharded_data_parallelr r r r r rg r r r r\ r rY r r r r rh ri r LOCAL_STATE_DICTrJ )r/ rg r r r r r s r0 r 2FullyShardedDataParallelPlugin.set_state_dict_type s
'#%::>>&%)%6%6!%;!AU$D d**C 0 0##++--'4S9M9M5N'O$'45I5I5O5O5Q'R$ =#@#@ @%%-)\>\&\ b
'] !r3 c " SSK JnJn [ USS5 nUb [ U5 O/ nU R
U:X a U R c XPl [ 5 nU R H1 n[ X5 nUc [ SU S35 eUR U5 M3 [ R " U R
US9U l gU R
U:X aG U R S: a/ [ R " U R
U R S9U l gSU l gg) z|
Given `model`, creates an `auto_wrap_policy` based on the passed in policy and if we can use the
`transformer_cls_to_wrap`
r r _no_split_modulesNz+Could not find the transformer layer class z in the model.)transformer_layer_cls)rr )r r r r rW rd rq r rA rJ r functoolspartialrr ) r/ r r r no_split_modules%default_transformer_cls_names_to_wraptransformer_cls_to_wraprD rE s r0 set_auto_wrap_policy3FullyShardedDataParallelPlugin.set_auto_wrap_policy s
#5*=tDJZJf5E0Fln- $@ @1195Z2&)e # # A A " UR
US5 nUc( [ SU S[ UR 5 5 35 eOZ[ U[ R 5 ( a; XR 5 ; a( [ SU S[ UR 5 5 35 eU( a [ R OUnU R S:X a SSK
Jn OU R S:X a SS K
Jn U( d
U R c+ XUS
.nU R S:X a XhS ' OXXS ' W" S0 UD6U l g[ U R [" 5 ( a S
S/U R S:X a S /OS /- n U V
s/ s H oU R ; d M U
PM n n
U R R% 5 V
V s/ s H u pXR 5 ; d M U
PM n
n
n U ( d U
( a5 [ SU R SU S[ UR 5 5 35 eW" S0 U R D6U l ggs sn
f s sn n
f )z(Sets the mixed precision policy for FSDP)r] rh ri fp32NzInvalid mixed precision: r r r MixedPrecisionr MixedPrecisionPolicy) param_dtype reduce_dtype buffer_dtype output_dtyper r z Invalid mixed precision policy: z . Must be a `dict` with keys z.Values must be one of r? )r
bfloat16float16float32r rY r rJ rW r rv valuesr\ r r r r` r r8 )r/ r buffer_autocastr mixed_precision_mappingrv buffer_typer
dtype_args
valid_keysr; missing_keysr< invalid_valuess r0 r 2FullyShardedDataParallelPlugin.set_mixed_precision sY >>MMNNMM #
os + ++//FE} //@@QRVWnWsWsWuRvQwx
5
5/QoQoQq:q +O+<*-2>**8*F:*FD '
33T
:
:' 8$($5$5$: @PJ (2Zz!d>Y>Y5YAzLZ " 9 9 ? ? A AdaQNlNlNnEn A
~ 6t7R7R6S T22< >--12I2P2P2R-S,TV
+9*W4;V;V*WD ' ;
[ s 2I0 I0.I5 I5c U R S:X a SSKJn OSSKJn [ U R
U5 ( d# U R S:X a SOSn[
SU S35 eg )
zb
Validates the mixed precision policy, abstracted away to not bring in the imports if not needed.
r r r r z-`torch.distributed.fsdp.MixedPrecisionPolicy`z'`torch.distributed.fsdp.MixedPrecision`z.mixed_precision_policy must be an instance of r N)r\ r r r r r` rJ )r/ r
required_types r0 r >FullyShardedDataParallelPlugin.validate_mixed_precision_policy6 sg ! U =$55~FF $$) @>
Mm_\] ^_ _
Gr3 c " U R S:X a SSKJnJn OSSKJn [ U R [ 5 ( aQ U R S:X a+ U R ( d
W" 5 U l g W" 5 U l g W" U R S9U l g g )Nr r )CPUOffloadPolicy
OffloadPolicy
CPUOffload)offload_params)r\ r r r r r re rb )r/ r r r s r0 r .FullyShardedDataParallelPlugin.set_cpu_offloadG sp ! N N 9 d&& - - A%'''4D$'7'9D$#-T=M=M#N .r3 c L U R S:X a SSKJn OSSKJn U R S:X a3 [ U R
W5 ( d [
SU R
35 eU R S:X a4 [ U R
W5 ( d [
SU R
35 eg g )Nr r )r r zZ`cpu_offload` must be an instance of `torch.distributed.fsdp.OffloadPolicy` in FSDP2, got r zW`cpu_offload` must be an instance of `torch.distributed.fsdp.CPUOffload` in FSDP1, got )r\ r r r r re rJ )r/ r r s r0 r 3FullyShardedDataParallelPlugin.validate_cpu_offloadV s ! < 9 !*T5E5E}*U*U lmqm}m}l~
!*T5E5Ez*R*R ijnjzjzi{|
+S !r3 )ro rd r_ re rp rn r\ rf rr r` ri rl r^ r] rh rg rm rq rk r* )FF)0r@ rA rB rC rD r r\ r rc r] r
rY r^ rb r_ r r` r rd r
r re rf r r
nnModulerg rh ri rj rk rl rm rn ro rp rq rW rr r r r r r r r rE r? r3 r0 r[ r[ s FP K
L# PU o
PuS"K KL Z_ ^
Z5&OQU!UV Z_ f
Zxc+T&T UV v
H
3 9
;
N
huXw7m/n%n op d
+1 3 HMPQHOXeHUXX__$=s$B CD
KP k
KOU3 FFG tu x
=
t
X
B
$ N
t ', E
'OXd^ BG B
BM8Hehhoo%6%< => */ t
* # o
d &+ <
& d ', D
' t :? L
:!8DI#6 %* W
%NHSM JmX+Z -D2Xh`"
O
r3 r[ c R \ rS rSr% Sr\" SSS0S9r\\S' \" SS 9r \
S
\S ' S r g)
TorchTensorParallelPluginif zH
This plugin is used to enable tensor parallelism using PyTorch >= 2.0.
r rz z@tensor parallel size will be used in the device mesh preparationr{ tp_sizeN)r| z torch.distributed.DeviceMeshtorch_device_meshr? ) r@ rA rB rC rD r r r rc r r rE r? r3 r0 r r f sA \]GS CHPTBUx >?Ur3 r c B \ rS rSr% Sr\" SSS0S9r\\ \ S' S r
S r g)
TorchContextParallelConfigiu zH
This class holds the configuration for context parallelism in PyTorch.
Nrz zqCommunication strategy for context parallelism. Can be one of 'allgather' or 'alltoall'. Defaults to 'allgather'.r{ cp_comm_strategyc
[ S[ 5 ( d [ S[ S35 eU R c% [ R
R
SS5 U l U R S; a [ SU R S35 eg ) Nr z1Context parallelism is only available in PyTorch 9 and later versions. Please upgrade your PyTorch version.#PARALLELISM_CONFIG_CP_COMM_STRATEGY allgather)r alltoallzInvalid cp_comm_strategy: z+. Must be one of 'allgather' or 'alltoall'.)r$ r rJ r r r r r. s r0 r (TorchContextParallelConfig.__post_init__ s &GHH CDeCf g7 7
($&JJNN3XZe$fD ! (A A ,T-B-B,CCno
Br3 )r ) r@ rA rB rC rD r r r rY rc r rE r? r3 r0 r r u s5 ', H
'hsm r3 r c . \ rS rSr% SrSr\\S' S rSr g)TorchTensorParallelConfigi zU
Use this object in your [`Accelerator`] to customize your torch tensor parallelism.
Fenable_async_tpc [ S[ 5 ( d [ S[ S35 e[ SS[ 5 ( d [ S[ 35 eU R
( a [ R " S5 g g )Nr z6Torch tensor parallelism is only available in PyTorch r r z TP requires transformers >= zJAsync tensor parallelism is currently not supported, ignoring this option.)r$ r rJ r# r r r r r. s r0 r 'TorchTensorParallelConfig.__post_init__ ss &GHH HIjHk l7 7
6\]] ;' \" S?SS@0S9r(\\SA' \" SSSB0S9r)\\SC' \" SSSD0S9r*\\SE' \" SSSF0S9r+\\SG' \" SSSH0S9r,\\SI' \" S?SSJ0S9r-\\SK' \" SLSSM0S9r.\\SN' \" SOSSP0S9r/\\SQ' \" S?SSR0S9r0\\SS' \" SSST0S9r1\$\2 \SU' \" SSSV0S9r3\$\4\\24 \SW' \" SSSX0S9r5\$\% \SY' \" SSSZ0S9r6\$\% \S[' \" SSS\0S9r7\$\% \S]' \" SSS^0S9r8\$\% \S_' \" SSS`0S9r9\$\% \Sa' \" SSSb0S9r:\$\4\\24 \Sc' Sd r;SlSe jrSh r?Si r@Sj rASkrBg)mMegatronLMPlugini an
Plugin for Megatron-LM to enable tensor, pipeline, sequence and data parallelism. Also to enable selective
activation recomputation and optimized fused kernels.
Args:
tp_degree (`int`, defaults to `None`):
Tensor parallelism degree.
pp_degree (`int`, defaults to `None`):
Pipeline parallelism degree.
num_micro_batches (`int`, defaults to `None`):
Number of micro-batches.
gradient_clipping (`float`, defaults to `None`):
Gradient clipping value based on global L2 Norm (0 to disable).
sequence_parallelism (`bool`, defaults to `None`):
Enable sequence parallelism.
recompute_activations (`bool`, defaults to `None`):
Enable selective activation recomputation.
use_distributed_optimizr (`bool`, defaults to `None`):
Enable distributed optimizer.
pipeline_model_parallel_split_rank (`int`, defaults to `None`):
Rank where encoder and decoder should be split.
num_layers_per_virtual_pipeline_stage (`int`, defaults to `None`):
Number of layers per virtual pipeline stage.
is_train_batch_min (`str`, defaults to `True`):
If both tran & eval dataloaders are specified, this will decide the `micro_batch_size`.
train_iters (`int`, defaults to `None`):
Total number of samples to train over all training runs. Note that either train-iters or train-samples
should be provided when using `MegatronLMDummyScheduler`.
train_samples (`int`, defaults to `None`):
Total number of samples to train over all training runs. Note that either train-iters or train-samples
should be provided when using `MegatronLMDummyScheduler`.
weight_decay_incr_style (`str`, defaults to `'constant'`):
Weight decay increment function. choices=["constant", "linear", "cosine"].
start_weight_decay (`float`, defaults to `None`):
Initial weight decay coefficient for L2 regularization.
end_weight_decay (`float`, defaults to `None`):
End of run weight decay coefficient for L2 regularization.
lr_decay_style (`str`, defaults to `'linear'`):
Learning rate decay function. choices=['constant', 'linear', 'cosine'].
lr_decay_iters (`int`, defaults to `None`):
Number of iterations for learning rate decay. If None defaults to `train_iters`.
lr_decay_samples (`int`, defaults to `None`):
Number of samples for learning rate decay. If None defaults to `train_samples`.
lr_warmup_iters (`int`, defaults to `None`):
Number of iterations to linearly warmup learning rate over.
lr_warmup_samples (`int`, defaults to `None`):
Number of samples to linearly warmup learning rate over.
lr_warmup_fraction (`float`, defaults to `None`):
Fraction of lr-warmup-(iters/samples) to linearly warmup learning rate over.
min_lr (`float`, defaults to `0`):
Minimum value for learning rate. The scheduler clip values below this threshold.
consumed_samples (`List`, defaults to `None`):
Number of samples consumed in the same order as the dataloaders to `accelerator.prepare` call.
no_wd_decay_cond (`Optional`, defaults to `None`):
Condition to disable weight decay.
scale_lr_cond (`Optional`, defaults to `None`):
Condition to scale learning rate.
lr_mult (`float`, defaults to `1.0`):
Learning rate multiplier.
megatron_dataset_flag (`bool`, defaults to `False`):
Whether the format of dataset follows Megatron-LM Indexed/Cached/MemoryMapped format.
seq_length (`int`, defaults to `None`):
Maximum sequence length to process.
encoder_seq_length (`int`, defaults to `None`):
Maximum sequence length to process for the encoder.
decoder_seq_length (`int`, defaults to `None`):
Maximum sequence length to process for the decoder.
tensorboard_dir (`str`, defaults to `None`):
Path to save tensorboard logs.
set_all_logging_options (`bool`, defaults to `False`):
Whether to set all logging options.
eval_iters (`int`, defaults to `100`):
Number of iterations to run for evaluation validation/test for.
eval_interval (`int`, defaults to `1000`):
Interval between running evaluation on validation set.
return_logits (`bool`, defaults to `False`):
Whether to return logits from the model.
custom_train_step_class (`Optional`, defaults to `None`):
Custom train step class.
custom_train_step_kwargs (`Optional`, defaults to `None`):
Custom train step kwargs.
custom_model_provider_function (`Optional`, defaults to `None`):
Custom model provider function.
custom_prepare_model_function (`Optional`, defaults to `None`):
Custom prepare model function.
custom_megatron_datasets_provider_function (`Optional`, defaults to `None`):
Custom megatron train_valid_test datasets provider function.
custom_get_batch_function (`Optional`, defaults to `None`):
Custom get batch function.
custom_loss_function (`Optional`, defaults to `None`):
Custom loss function.
other_megatron_args (`Optional`, defaults to `None`):
Other Megatron-LM arguments. Please refer Megatron-LM.
Nrz ztensor parallelism degree.r{ tp_degreez pipeline parallelism degree. pp_degreeznumber of micro-batches.num_micro_batchesz>gradient clipping value based on global L2 Norm (0 to disable)r zenable sequence parallelismsequence_parallelismz)enable selective activation recomputationrecompute_activationsz enable distributed optimizeruse_distributed_optimizerz/Rank where encoder and decoder should be split."pipeline_model_parallel_split_rankz,Number of layers per virtual pipeline stage.%num_layers_per_virtual_pipeline_stageTzUIf both train & eval dataloaders are specified, this will decide the micro_batch_sizer zTotal number of iterations to train over all training runs. Note that either train-iters or train-samples should be provided when using `MegatronLMDummyScheduler` train_iterszTotal number of samples to train over all training runs. Note that either train-iters or train-samples should be provided when using `MegatronLMDummyScheduler`
train_samplesconstantzKWeight decay increment function. choices=["constant", "linear", "cosine"]. weight_decay_incr_stylez7Initial weight decay coefficient for L2 regularization.start_weight_decayz:End of run weight decay coefficient for L2 regularization.end_weight_decaylinearzGLearning rate decay function. choices=['constant', 'linear', 'cosine'].lr_decay_stylezPNumber of iterations for learning rate decay. If None defaults to `train_iters`.lr_decay_iterszONumber of samples for learning rate decay. If None defaults to `train_samples`.lr_decay_samplesz;number of iterations to linearly warmup learning rate over.lr_warmup_itersz8number of samples to linearly warmup learning rate over.lr_warmup_sampleszLfraction of lr-warmup-(iters/samples) to linearly warmup learning rate over.lr_warmup_fractionr zPMinimum value for learning rate. The scheduler clip values below this threshold.min_lrz^Number of samples consumed in the same order as the dataloaders to `accelerator.prepare` call.consumed_samplesz"Condition to disable weight decay.no_wd_decay_condz!Condition to scale learning rate.
scale_lr_cond ?zLearning rate multiplier.lr_multFzUWhether the format of dataset follows Megatron-LM Indexed/Cached/MemoryMapped format.megatron_dataset_flagz#Maximum sequence length to process.
seq_lengthz3Maximum sequence length to process for the encoder.encoder_seq_lengthz3Maximum sequence length to process for the decoder.decoder_seq_lengthz Path to save tensorboard logs.tensorboard_dirz#Whether to set all logging options.set_all_logging_optionsd z?Number of iterations to run for evaluation validation/test for.
eval_itersi z6Interval between running evaluation on validation set.
eval_intervalz(Whether to return logits from the model.
return_logitszCustom train step class.custom_train_step_classzCustom train step kwargs.custom_train_step_kwargsz Custom model provider function. custom_model_provider_functionz Custom prepare model function. custom_prepare_model_functionz> ! 0Da!H IDN >> ! 0Da!H IDN ! ! )%( AT8TVW)X%YD " ! ! )%*2::>>&CV:VX[+\%]D " % % -)4RZZ^^FMdDdfm5n)ost)tD & ) ) 1BJJNN64O+OQX YZ^__
* $ $ ,(3BJJNN6LbCbdk4l(mqr(rD % >>A !?!? #DM #DM ,4(()Q.%%,,aV4T**+q0%%,,Q/)
($..)
*DNN)
1$2Y2Y)
4T5_5_ )
)
()G)G
)
!:!:)
//)
!7!7)
5 5)
5 5)
T / /)
t||)
$T%?%? )
$// )
T / /!)
%$ % %EPD ) )*A B +?C?S?SD ) )*; <++446 # # / ) ) 0 01I1I J 0r3 c UR R R 5 n[ R 5 H nXC; d M
[ U " XU5 g [ SU S35 e)Nz0Accelerate Megatron-LM integration not supports z1 model. You can add your own model config parser.)r
model_typer !MODEL_CONFIGS_TO_MEGATRON_PARSERSr rJ )r/ r
batch_datamodel_config_typer; s r0 set_network_size_args&MegatronLMPlugin.set_network_size_args sj !LL 3 3 9 9 ;;@@BJ.1*=d:V C >?P>Q R8
8
r3 c US:X a SU R S' g US:X a0 SU R S' SU l U R U R S' g g )Nrh Tri r- r0 )r6 r0 )r/ r s r0 r $MegatronLMPlugin.set_mixed_precision sR f $48D ) )& 1
&48D ) )& 1 #DM8<
D ) )* 5 'r3 c X l Xl X!- U R - U l U R U R S' U R U R S' U R U R S' g )Ndata_parallel_sizemicro_batch_sizeglobal_batch_size)rD rE r rF r6 )r/ rE dp_degrees r0 set_training_args"MegatronLMPlugin.set_training_args si "+ 0 !*!=@V@V!V >B>U>U%%&:;<@D ) )+ 68Z%%n5r3 c U R cK UR U R S - U l U R b S U l [ R
" S5 U R cK UR U R S - U l U R b [ R
" S5 SU l U R U R S' U R U R S' U R U R S' U R U R S' U R U R S ' U R U R S
' U R U R S ' U R U R S ' U R U R S
' U R U R S' U R U R S' U R U R S' g )NrD zXIgnoring `train_samples` as `train_iters` based on scheduler is being used for training.z`Ignoring `lr_warmup_samples` as `lr_warmup_iters` based on scheduler is being used for training.r r r r r
r r r r r r r r )r total_num_stepsr6 r r r r warmup_num_stepsr
r r r r r r r r )r/ schedulers r0 set_scheduler_args#MegatronLMPlugin.set_scheduler_args
s # ( 8 8D:M:M%%&67<@B>U>U%%&:;:>:M:M%%&67CGC_C_%%&?@>B>U>U%%&:;<@:;9!8H#5 FKXYF.0B 5:675 x 1 05120(8, 5:QR5$sCx.1
7Kr
FT[ >: Mr3 r r; c ^ U 4S jnU$ )Nc Z >^ [ R " T 5 U 4S j5 nT [ T' U$ )Nc > T" U 0 UD6$ r* r? )argsr funcs r0 r \add_model_config_to_megatron_parser..add_model_config_parser_helper..wrapperD
s (( (r3 )r wrapsr< )rt r r; s ` r0 add_model_config_parser_helperKadd_model_config_to_megatron_parser..add_model_config_parser_helperC
s0 )
) 9=)*5r3 r? )r; rw s ` r0 #add_model_config_to_megatron_parserry B
s *)r3 z
megatron-bertc SnUR R nUR R nUR R nUR R nUR R
nUR R n Sn
SUR R R 5 ; a Sn
U R b5 U R b [ R " S5 U R U l
O@U R b U R U l
O!Ub US R S U l
OXpl
U R U R S' X0R S ' X@R S
' XPR S ' X`R S ' XpR S
' XR S' XR S' UR R U R S' XR S' g )NbertFmaskedlmTzOBoth `seq_length` and `encoder_seq_length` are set. Using `encoder_seq_length`. input_idsr r model_type_name
num_layers hidden_sizenum_attention_headsmax_position_embeddingspretraining_flagorig_vocab_sizemodel_return_dict
num_labels)r num_hidden_layersr r r r
vocab_sizer7 r@ r r r r r ru r6 return_dict) megatron_lm_pluginr r= r~ r r r r r r r s r0 parse_bert_configr N
s O //J,,**K ,,:: #ll B B ((Jll--O U__--3355 $$0 0 0 < MMk l(:(M(M% . . :(:(M(M% (2;(?(E(Ea(H%(?%@R@]@]// =ET//0AB@J// =AL//
>I\//0EFMd//0IJFV//0BCET//0ABGL||G_G_//0CD@J// =r3 gpt2c x SnUR R nUR R nUR R nUR R nUR R
nSn U R b5 U R b [ R " S5 U R U l O@U R b U R U l O!Ub US R S U l OXpl U R U R S' U R U R S' SU R S ' X0R S
' X@R S ' XPR S ' X`R S
' XpR S' XR S' XR S' UR R U R S' g )NgptTzOBoth `seq_length` and `decoder_seq_length` are set. Using `decoder_seq_length`.r} r r r GPT2BPETokenizertokenizer_typer~ r r r r r r r )r n_layern_embdn_head n_positionsr r r r r ru r6 r r
r r r= r~ r r r r r r s
r0 parse_gpt2_configr p
s O %%J,,%%K ,,-- #ll 6 6ll--O$$0 0 0 < MMk l(:(M(M% . . :(:(M(M% (2;(?(E(Ea(H%(?%@R@]@]// =CUCcCc//@DV//0@AET//0AB@J// =AL//
>I\//0EFMd//0IJFV//0BCET//0ABGL||G_G_//0CDr3 t5c H SnUR R nUR R nUR R n[ UR S5 ( a UR R
OSnUR R nSn U R c! Ub US R S U l OXpl U R c! Ub US R S U l OXpl U R U R S' U R U R S ' X0R S
' X@R S ' XPR S ' X`R S
' XpR S' XR S' XR S' UR R U R S' g )Nr r r Tr} r labelsr r r~ r r r r r r r ) r r d_model num_headshasattrr r r ru r r6 r r s
r0 parse_t5_configr
s O ((J,,&&K ,,00:A%,,P]:^:^ell 6 6dhll--O,,4 !4>{4K4Q4QRS4T 14K 1,,4 !4>x4H4N4Nq4Q 14K 1HZHmHm//0DEHZHmHm//0DEET//0AB@J// =AL//
>I\//0EFMd//0IJFV//0BCET//0ABGL||G_G_//0CDr3 llamac P SnUR R nSnUR R nUR R nUR R nUR R
n [
UR SS 5 n
U R cJ U
b Xl O@U R b U R U l O!Ub US R S U l OXl U R U R S' SU R S' X0R S ' X@R S
' XPR S ' X`R S ' XpR S
' XR S' XR S' U R U R S' UR R U R S' g )Nr Tmax_sequence_lengthr} r r Llama2Tokenizerr r~ r r r r r r r r )
r r r r r r r r r ru r r6 r ) r r r= r~ r r r r r r r s r0 parse_llama_configr
s O //J,,**K ,,::ll--O #ll B B ',>,Q,Q )
#,6{,C,I,I!,L ),C )CUCcCc//@DU//0@AET//0AB@J// =FV//0BCAL//
>I\//0EFET//0ABMd//0IJ@R@]@]// =GL||G_G_//0CDr3 c < \ rS rSr% Sr\" SSS0S9r\\S' \" SSS 0S9r \
\S
' \" SSS 0S9r \\S ' \" S
SS0S9r \
\S' \" SSS0S9r\\S' \" SSS0S9r\
\S' \" SSS0S9r\R$ \S' \" SSS0S9r\\
\S' \" SSS0S9r\\
\S' S rS rg) BnbQuantizationConfigi
a
A plugin to enable BitsAndBytes 4bit and 8bit quantization
Args:
load_in_8bit (`bool`, defaults to `False`):
Enable 8bit quantization.
llm_int8_threshold (`float`, defaults to `6.0`):
Value of the outliner threshold. Only relevant when `load_in_8bit=True`.
load_in_4_bit (`bool`, defaults to `False`):
Enable 4bit quantization.
bnb_4bit_quant_type (`str`, defaults to `fp4`):
Set the quantization data type in the `bnb.nn.Linear4Bit` layers. Options are {'fp4','np4'}.
bnb_4bit_use_double_quant (`bool`, defaults to `False`):
Enable nested quantization where the quantization constants from the first quantization are quantized
again.
bnb_4bit_compute_dtype (`bool`, defaults to `fp16`):
This sets the computational type which might be different than the input time. For example, inputs might be
fp32, but computation can be set to bf16 for speedups. Options are {'fp32','fp16','bf16'}.
torch_dtype (`torch.dtype`, defaults to `None`):
This sets the dtype of the remaining non quantized layers. `bitsandbytes` library suggests to set the value
to `torch.float16` for 8 bit model and use the same dtype as the compute dtype for 4 bit model.
skip_modules (`List[str]`, defaults to `None`):
An explicit list of the modules that we don't quantize. The dtype of these modules will be `torch_dtype`.
keep_in_fp32_modules (`List`, defaults to `None`):
An explicit list of the modules that we don't quantize. We keep them in `torch.float32`.
Frz zenable 8bit quantization.r{ load_in_8bit @zEvalue of the outliner threshold. only relevant when load_in_8bit=Truellm_int8_thresholdzenable 4bit quantization. load_in_4bitfp4z\set the quantization data type in the `bnb.nn.Linear4Bit` layers. Options are {'fp4','nf4'}.bnb_4bit_quant_typezlenable nested quantization where the quantization constants from the first quantization are quantized again.bnb_4bit_use_double_quantrh zThis sets the computational type which might be different than the input time. For example, inputs might be fp32, but computation can be set to bf16 for speedups. Options are {'fp32','fp16','bf16'}.bnb_4bit_compute_dtypeNzthis sets the dtype of the remaining non quantized layers. `bitsandbytes` library suggests to set the valueto `torch.float16` for 8 bit model and use the same dtype as the compute dtype for 4 bit model torch_dtypezian explicit list of the modules that we don't quantize. The dtype of these modules will be `torch_dtype`. skip_moduleszXan explicit list of the modules that we don't quantize. We keep them in `torch.float32`.keep_in_fp32_modulesc [ U R [ 5 ( d [ S5 e[ U R [ 5 ( d [ S5 eU R ( a U R ( a [ S5 eU R ( d U R ( d [ S5 e[ U R
[ [ 45 ( d [ S5 e[ U R [ 5 ( d [ S5 eU R S; a [ SU R 35 e[ U R [ 5 ( d [ S 5 e[ U R [ 5 ( a U R S
:X a [ R U l OU R S :X a [ R U l OrU R S :X a [ R U l OL[ S
U R 35 e[ U R [ R 5 ( d [ S5 eU R" b* [ U R" [$ 5 ( d [ S5 eU R& b* [ U R& [$ 5 ( d [ S5 eU R ( a [( R* U l U R ( a [ R. U l U R ( a& U R
S:w a [0 R2 " S5 [ U R4 [ 5 ( a U R4 S
:X a [ R U l OdU R4 S :X a [ R U l O>U R4 S :X a [ R U l O[ SU R4 35 eU R ( a" U R4 c [ R U l U R ( a U R4 c U R U l [ U R4 [ R 5 ( d [ S5 eg)zn
Safety checker that arguments are correct - also replaces some NoneType arguments with their default values.
z load_in_8bit must be a booleanz load_in_4bit must be a booleanz0load_in_4bit and load_in_8bit can't be both Truez1load_in_4bit and load_in_8bit can't be both Falsez,llm_int8_threshold must be a float or an intz$bnb_4bit_quant_type must be a string)r nf4z7bnb_4bit_quant_type must be in ['fp4','nf4'] but found z+bnb_4bit_use_double_quant must be a booleanr rh ri zCbnb_4bit_compute_dtype must be in ['fp32','fp16','bf16'] but found z8bnb_4bit_compute_dtype must be a string or a torch.dtypeNz&skip_modules must be a list of stringsz/keep_in_fp_32_modules must be a list of stringsr z >$++T22 => > !2!2 OP P ):): PQ Q$11C<@@ KL L$22C88 CD D
%
%^
; VW[WoWoVp qr r$88$?? JK K d113 7 7**f4.3mm+,,6.3mm+,,6.3nn+ YZ^ZuZuYvw D 7 7EE WX X (D> #[\`\l\l[m!noo !1!1!9 $}}D !1!1!9 # : :D $**EKK88 @A A 9r3 )r r r )r@ rA rB rC rD r r rb rc r r r r rY r r r r
rv r rW r r rE r? r3 r0 r r
s9 6 u@[7\]L$] %ab!
u@[7\]L$] $ r
', C
' t #( i
#C % n
K $
L$s) ',tu'$s)
HBr3 r c [ U R 5 5 nU R R U:X a U R $ [ U5 S:X a gU H n[ X15 nUc M Us $ g)z
Gets a class from a module by its name.
Args:
module (`torch.nn.Module`): The module to get the class from.
name (`str`): The name of the class.
r N)rW childrenr7 r@ r
rA )moduler4 modules_children child_module module_classs r0 rA rA l sg FOO -.
D(
! # ,L5lIL'## -r3 )irD rg r+ enumr loggingr r collections.abcr
contextlibr dataclassesr r datetimer typingr r r
r r r
r r
constantsr r r r r r r r r r6 r r r r r r r r r r! r" versionsr# r$ torchao.float8r% getLoggerr@ r r' EnumMetarG EnumrP r^ re rr r r r r r r r r r r r r rY r r* r. r1 r5 rF r\ r` rn rt ry r r r r r[ r r r r r< ry r r r r r rA r? r3 r0 r sd $ % ( S S S
: 9 1
8 $L L*t}} #tyy$4 # )] ) )*,x ,( RM R Rj } > 6] 6 6< -
: , - 34 2] 2 2$ Kg] Kg Kg\
^
^
^ n&7 2 >? q
M q
q
h c499 F&sDII &"S$)) " *dii * 5 C 5 p 8
C
h
$))
[ [ [| ;/ ;/ ;/| 0 0 0f U
U Up @
@
@
F a a a H V V V 4 h h h* UM UM UMp %' ! *C * %_5 K 6 KB %V, ` - `> %T*` +`< %W- ` . `B WB WB WBt$r3