bidZddlZddlZddlZddlZddlZddlmZddl m Z ddl m Z ddl mZmZmZmZmZddlmZdd lmZmZmZerddlZej4eZGd d ee Zerdd lm Z Gd dejBZ"eGddZ#eGdde#Z$eGdde#Z%eGdde#Z&eGdde#Z'y)z Adapted from https://github.com/huggingface/transformers/blob/52cb4034ada381fe1ffe8d428a1076e5411a8026/src/transformers/utils/quantization_config.py N) dataclass)Enum)partial)AnyDictListOptionalUnion)version)is_torch_availableis_torchao_availableloggingceZdZdZdZdZdZy)QuantizationMethod bitsandbytesgguftorchaoquantoN)__name__ __module__ __qualname__BITS_AND_BYTESGGUFTORCHAOQUANTOc/home/cdr/jupyterlab/.venv/lib/python3.12/site-packages/diffusers/quantizers/quantization_config.pyrr,s#N DG Frr) MappingTypeceZdZfdZxZS)TorchAoJSONEncodercZt|tr |jSt||SN) isinstancer namesuperdefault)selfobj __class__s rr(zTorchAoJSONEncoder.default7s%#{+xx7?3' 'r)rrrr( __classcell__)r+s@rr"r"6s  ( (rr"ceZdZUdZeed<gZeddZde e e jffdZ dee effdZdZd Zdd ede fd Zd Zy )QuantizationConfigMixinz- Mixin class for quantization config quant_methodc |di|}g}|jD]0\}}t||st||||j|2|D]}|j |d|r||fS|S)a Instantiates a [`QuantizationConfigMixin`] from a Python dictionary of parameters. Args: config_dict (`Dict[str, Any]`): Dictionary that will be used to instantiate the configuration object. return_unused_kwargs (`bool`, *optional*, defaults to `False`): Whether or not to return a list of unused keyword arguments. Used for `from_pretrained` method in `PreTrainedModel`. kwargs (`Dict[str, Any]`): Additional parameters from which to initialize the configuration object. Returns: [`QuantizationConfigMixin`]: The configuration object instantiated from those parameters. Nr)itemshasattrsetattrappendpop)cls config_dictreturn_unused_kwargskwargsconfig to_removekeyvalues r from_dictz!QuantizationConfigMixin.from_dictFs$#{#  ,,. &JCvs#U+  % & "C JJsD ! " 6> !Mrjson_file_pathct|dd5}|j}tj|dddz}|j |dddy#1swYyxYw) a Save this instance to a JSON file. Args: json_file_path (`str` or `os.PathLike`): Path to the JSON file in which this configuration instance's parameters will be saved. use_diff (`bool`, *optional*, defaults to `True`): If set to `True`, only the difference between the config instance and the default `QuantizationConfig()` is serialized to JSON file. wzutf-8)encodingr Tindent sort_keys N)opento_dictjsondumpswrite)r)r?writerr7 json_strings r to_json_filez$QuantizationConfigMixin.to_json_filegsU.# 8 &F,,.K**[dKdRK LL %  & & &s =AAreturnc@tj|jS) Serializes this instance to a Python dictionary. Returns: `Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance. )copydeepcopy__dict__r)s rrHzQuantizationConfigMixin.to_dictxs }}T]]++rc#Ktj|jjD] \}}||f yw)zTallows `dict(obj)` for situations where obj may be a dict or QuantizationConfigMixinN)rRrSrTr1)r)attrr=s r__iter__z QuantizationConfigMixin.__iter__s9==7==? KD%+  s=?cT|jjd|jS)N )r+rto_json_stringrUs r__repr__z QuantizationConfigMixin.__repr__s(..))*!D,?,?,A+BCCruse_diffc|dur|j}n|j}tj|dddzS)a Serializes this instance to a JSON string. Args: use_diff (`bool`, *optional*, defaults to `True`): If set to `True`, only the difference between the config instance and the default `PretrainedConfig()` is serialized to JSON string. Returns: `str`: String containing all the attributes that make up this configuration instance in JSON format. Tr rCrF) to_diff_dictrHrIrJ)r)r]r7s rr[z&QuantizationConfigMixin.to_json_strings= t ++-K,,.Kzz+a4@4GGrc g}|jD]0\}}t||st||||j|2|jDcic] \}}||vs ||}}}|Scc}}w)a Updates attributes of this class instance with attributes from `kwargs` if they match existing attributes, returning all the unused kwargs. Args: kwargs (`Dict[str, Any]`): Dictionary of attributes to tentatively update this class. Returns: `Dict[str, Any]`: Dictionary containing all the key-value pairs that were not used to update the instance. )r1r2r3r4)r)r9r;r<r= unused_kwargss rupdatezQuantizationConfigMixin.updates}  ,,. &JCtS!c5)  % & 7=lln] US\H\e] ]^s  A0&A0N)F)T)rrr__doc__r__annotations___exclude_attributes_at_init classmethodr>r strosPathLikerNrrrHrXr\boolr[rbrrrr.r.=s|%$"$@&5bkk1A+B&",c3h, DHtHsH$rr.ceZdZdZgdZ ddZedZejde fdZedZ e jde fd Z d Z d Z d Z d eeeffdZdZd eeeffdZy)BitsAndBytesConfiga This is a wrapper class about all possible attributes and features that you can play with a model that has been loaded using `bitsandbytes`. This replaces `load_in_8bit` or `load_in_4bit` therefore both options are mutually exclusive. Currently only supports `LLM.int8()`, `FP4`, and `NF4` quantization. If more methods are added to `bitsandbytes`, then more arguments will be added to this class. Args: load_in_8bit (`bool`, *optional*, defaults to `False`): This flag is used to enable 8-bit quantization with LLM.int8(). load_in_4bit (`bool`, *optional*, defaults to `False`): This flag is used to enable 4-bit quantization by replacing the Linear layers with FP4/NF4 layers from `bitsandbytes`. llm_int8_threshold (`float`, *optional*, defaults to 6.0): This corresponds to the outlier threshold for outlier detection as described in `LLM.int8() : 8-bit Matrix Multiplication for Transformers at Scale` paper: https://huggingface.co/papers/2208.07339 Any hidden states value that is above this threshold will be considered an outlier and the operation on those values will be done in fp16. Values are usually normally distributed, that is, most values are in the range [-3.5, 3.5], but there are some exceptional systematic outliers that are very differently distributed for large models. These outliers are often in the interval [-60, -6] or [6, 60]. Int8 quantization works well for values of magnitude ~5, but beyond that, there is a significant performance penalty. A good default threshold is 6, but a lower threshold might be needed for more unstable models (small models, fine-tuning). llm_int8_skip_modules (`List[str]`, *optional*): An explicit list of the modules that we do not want to convert in 8-bit. This is useful for models such as Jukebox that has several heads in different places and not necessarily at the last position. For example for `CausalLM` models, the last `lm_head` is typically kept in its original `dtype`. llm_int8_enable_fp32_cpu_offload (`bool`, *optional*, defaults to `False`): This flag is used for advanced use cases and users that are aware of this feature. If you want to split your model in different parts and run some parts in int8 on GPU and some parts in fp32 on CPU, you can use this flag. This is useful for offloading large models such as `google/flan-t5-xxl`. Note that the int8 operations will not be run on CPU. llm_int8_has_fp16_weight (`bool`, *optional*, defaults to `False`): This flag runs LLM.int8() with 16-bit main weights. This is useful for fine-tuning as the weights do not have to be converted back and forth for the backward pass. bnb_4bit_compute_dtype (`torch.dtype` or str, *optional*, defaults to `torch.float32`): This sets the computational type which might be different than the input type. For example, inputs might be fp32, but computation can be set to bf16 for speedups. bnb_4bit_quant_type (`str`, *optional*, defaults to `"fp4"`): This sets the quantization data type in the bnb.nn.Linear4Bit layers. Options are FP4 and NF4 data types which are specified by `fp4` or `nf4`. bnb_4bit_use_double_quant (`bool`, *optional*, defaults to `False`): This flag is used for nested quantization where the quantization constants from the first quantization are quantized again. bnb_4bit_quant_storage (`torch.dtype` or str, *optional*, defaults to `torch.uint8`): This sets the storage type to pack the quanitzed 4-bit prarams. kwargs (`Dict[str, Any]`, *optional*): Additional parameters from which to initialize the configuration object. ) _load_in_4bit _load_in_8bitr/Nc  Ntj_|r |r td|_|_|_|_|_|_ |_ | _ |tj_nSt|t rt#t|_n-t|tj$r|_n td| tj&_nbt| t r%| dvr tdt#t| _n-t| tj$r| _n td| rQt+fd| Ds=t,j/dt1| j3dj4d j7y) NVload_in_4bit and load_in_8bit are both True, but only one can be used at the same timez8bnb_4bit_compute_dtype must be a string or a torch.dtype)float16float32int8uint8float64bfloat16zv`bnb_4bit_quant_storage` must be a valid string (one of 'float16', 'float32', 'int8', 'uint8', 'float64', 'bfloat16') z8bnb_4bit_quant_storage must be a string or a torch.dtypec3:K|]}|jvywr$)re).0kr)s r z.BitsAndBytesConfig.__init__..sTa4#C#CCTszUnused kwargs: z. These kwargs are not used in .)rrr/ ValueErrorrnrmllm_int8_thresholdllm_int8_skip_modules llm_int8_enable_fp32_cpu_offloadllm_int8_has_fp16_weightbnb_4bit_quant_typebnb_4bit_use_double_quanttorchrrbnb_4bit_compute_dtyper%rggetattrdtypertbnb_4bit_quant_storageallloggerwarninglistkeysr+ post_init) r) load_in_8bit load_in_4bitr}r~rrrrrrr9s ` r__init__zBitsAndBytesConfig.__init__sm/== Luv v))"4%:"0P-(@%#6 )B& ! )*/--D ' . 4*1%9O*PD ' . <*@D 'WX X ! )*/++D ' . 4%-kk M+2%9O*PD ' . <*@D 'WX X #TVTT NN_T&++--@,AA`aeaoao`ppqr s rc|jSr$)rmrUs rrzBitsAndBytesConfig.load_in_4bit!!!rr=czt|ts td|jr |r t d||_y)Nload_in_4bit must be a booleanrp)r%rj TypeErrorrr|rmr)r=s rrzBitsAndBytesConfig.load_in_4bit"7%&<= =   uv v"rc|jSr$)rnrUs rrzBitsAndBytesConfig.load_in_8bit+rrczt|ts td|jr |r t d||_y)Nload_in_8bit must be a booleanrp)r%rjrrr|rnrs rrzBitsAndBytesConfig.load_in_8bit/rrct|jts tdt|jts tdt|j t s td|j%t|jts tdt|jts tdt|jts td|j/t|jtjs tdt|jts td t|j ts td |jrTt#j$t&j(j#d t#j$d k\s t+d yy)z~ Safety checker that arguments are correct - also replaces some NoneType arguments with their default values. rrz"llm_int8_threshold must be a floatNz/llm_int8_skip_modules must be a list of stringsz2llm_int8_enable_fp32_cpu_offload must be a booleanz*llm_int8_has_fp16_weight must be a booleanz*bnb_4bit_compute_dtype must be torch.dtypez$bnb_4bit_quant_type must be a stringz+bnb_4bit_use_double_quant must be a booleanrz0.39.0z[4 bit quantization requires bitsandbytes>=0.39.0 - please upgrade your bitsandbytes version)r%rrjrrr}floatr~rrrrrrrrgrr parse importlibmetadatar|rUs rrzBitsAndBytesConfig.post_init8st$++T2<= =$++T2<= =$1159@A A  % % 1*TE_E_ae:fMN N$??FPQ Q$77>HI I  & & 2:dFaFachcncn;oHI I$22C8BC C$88$?IJ J   W]]93E3E3M3Mn3]%^bibobo c & m & rc6|jxs |jS)zP Returns `True` if the model is quantizable, `False` otherwise. )rrrUs ris_quantizablez!BitsAndBytesConfig.is_quantizable]s  5D$5$55rc|jry|jr|jdk(ry|jr|jdk(ryy)z This method returns the quantization method used for the model. If the model is not quantizable, it returns `None`. llm_int8fp4nf4N)rrrrUs rquantization_methodz&BitsAndBytesConfig.quantization_methodcsE      4#;#;u#D   4#;#;u#DrrOc tj|j}t|dj dd|d<t|dj dd|d<|j |d<|j |d<|S)rQrr{rrr)rRrSrTrgsplitrr)r)outputs rrHzBitsAndBytesConfig.to_dictqs t}}-+.v6N/O+P+V+VWZ+[\]+^'(+.v6N/O+P+V+VWZ+[\]+^'(!%!2!2~!%!2!2~ rc|j}|jjdtj|dddS)NrZr TrCrF)rHr+rrIrJr)r7s rr\zBitsAndBytesConfig.__repr__~s;lln ..))*!DJJ{1X\,]+^^`aarc|j}tj}i}|jD]\}}|||k7s|||<|S)a' Removes all attributes from config which correspond to the default config attributes for better readability and serializes to a Python dictionary. Returns: `Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance, )rHrlr1)r)r7default_config_dictserializable_config_dictr<r=s rr_zBitsAndBytesConfig.to_diff_dictshlln 12::<#% &++- 6JC+C0005(- 6('r) FFg@NFFNrFN)rrrrcrerpropertyrsetterrjrrrrrrgrrHr\r_rrrrlrls1f#U").!&#!"'#5n""#$##""#$###J6  c3h b(d38n(rrlc$eZdZdZddedfdZy)GGUFQuantizationConfigaIThis is a config class for GGUF Quantization techniques. Args: compute_dtype: (`torch.dtype`, defaults to `torch.float32`): This sets the computational type which might be different than the input type. For example, inputs might be fp32, but computation can be set to bf16 for speedups. N compute_dtypez torch.dtypectj|_||_d|_d|_|jt j|_yy)NT)rrr/r pre_quantizedmodules_to_not_convertrrr)r)rs rrzGGUFQuantizationConfig.__init__sG.33*!'+#    %!&D  &rr$)rrrrcr rrrrrrs /h}&= /rrcdeZdZdZd dedeeeddfdZedZ e de fdZ d Z d Zy) TorchAoConfigaJ This is a config class for torchao quantization/sparsity techniques. Args: quant_type (`str`): The type of quantization we want to use, currently supporting: - **Integer quantization:** - Full function names: `int4_weight_only`, `int8_dynamic_activation_int4_weight`, `int8_weight_only`, `int8_dynamic_activation_int8_weight` - Shorthands: `int4wo`, `int4dq`, `int8wo`, `int8dq` - **Floating point 8-bit quantization:** - Full function names: `float8_weight_only`, `float8_dynamic_activation_float8_weight`, `float8_static_activation_float8_weight` - Shorthands: `float8wo`, `float8wo_e5m2`, `float8wo_e4m3`, `float8dq`, `float8dq_e4m3`, `float8_e4m3_tensor`, `float8_e4m3_row`, - **Floating point X-bit quantization:** - Full function names: `fpx_weight_only` - Shorthands: `fpX_eAwB`, where `X` is the number of bits (between `1` to `7`), `A` is the number of exponent bits and `B` is the number of mantissa bits. The constraint of `X == A + B + 1` must be satisfied for a given shorthand notation. - **Unsigned Integer quantization:** - Full function names: `uintx_weight_only` - Shorthands: `uint1wo`, `uint2wo`, `uint3wo`, `uint4wo`, `uint5wo`, `uint6wo`, `uint7wo` modules_to_not_convert (`List[str]`, *optional*, default to `None`): The list of modules to not quantize, useful for quantizing models that explicitly require to have some modules left in their original precision. kwargs (`Dict[str, Any]`, *optional*): The keyword arguments for the chosen type of quantization, for example, int4_weight_only quantization supports two keyword arguments `group_size` and `inner_k_tiles` currently. More API examples and documentation of arguments can be found in https://github.com/pytorch/ao/tree/main/torchao/quantization#other-available-quantization-techniques Example: ```python from diffusers import FluxTransformer2DModel, TorchAoConfig quantization_config = TorchAoConfig("int8wo") transformer = FluxTransformer2DModel.from_pretrained( "black-forest-labs/Flux.1-Dev", subfolder="transformer", quantization_config=quantization_config, torch_dtype=torch.bfloat16, ) ``` N quant_typerrOc tj|_||_||_d|vr |d|_n||_|j }|j|jvr||jjdxs|jjd}|r)|jstd|jdtd|jd||j}tj|}|jjDchc]N}|jtj j"tj j$fvr |j&P} }t)|j j| z } t+| dkDrtd|d | d | d ycc}w) Nquant_type_kwargsrfpzRequested quantization type: z is not supported on GPUs with CUDA capability <= 8.9. You can check the CUDA capability of your GPU using `torch.cuda.get_device_capability()`.z is not supported or is an incorrect `quant_type` name. If you think the provided quantization type should be supported, please open an issue at https://github.com/huggingface/diffusers/issues.rzThe quantization method "z4" does not support the following keyword arguments: z2. The following keywords arguments are supported: r{)rrr/rrr!_get_torchao_quant_type_to_methodr startswith&_is_xpu_or_cuda_capability_atleast_8_9r|inspect signature parametersvalueskind Parameter KEYWORD_ONLYPOSITIONAL_OR_KEYWORDr&rlen) r)rrr9TORCHAO_QUANT_TYPE_METHODSis_floating_quant_typemethodrparam all_kwargsunsupported_kwargss rrzTorchAoConfig.__init__s.66$&<# & (%+,?%@D "%+D "%)%K%K%M" ??"<"A"A"C C%)__%?%?%H%lDOOLfLfgkLl "%d.Y.Y.[ 3DOO3DElm //@AKL  ,DOO<%%f- #--446 zzg//<O>O>e>eff JJ  "$"8"8"="="?*"LM ! "Q &+J<7k%&&XYcXddeg  ' s"AGc xtr!ddlmm}m}mm}m}m}m }m }ddl m m dtjffd }dt ffd } ||||d} ||||d } t#|tj$ |t#|tj$ t#|tj& t#tj&tj& d |tj&d |i| d| d| d| d| d} |t#|tj(t#|tj*t#|tj,t#|tj.t#|tj0t#|tj2t#|tj4d} i}|j7| |j7| |j7| |j9r|j7| |St;d)z` Returns supported torchao quantization types with all commonly used notations. r) 'float8_dynamic_activation_float8_weight&float8_static_activation_float8_weightfloat8_weight_onlyfpx_weight_onlyint4_weight_only#int8_dynamic_activation_int4_weight#int8_dynamic_activation_int8_weightint8_weight_onlyuintx_weight_only)PerRow PerTensorrc |tjk(rdnd}i}fD].}|urdnd}t||||f|d|d|<0|S)Ne5m2e4m3tensorrow)activation_dtype weight_dtype granularity float8dq__)r float8_e5m2r)rr&typesgranularity_clsgranularity_namerrrs rgenerate_float8dq_typeszPTorchAoConfig._get_torchao_quant_type_to_method..generate_float8dq_types!s}!&%*;*;!;v(16':O3Bi3OxUZ$CJ?).%*%4%68I$J DEIdV1-=,>?@ rbitsc i}td|D]$}||z dz }t|||d|d|d|<&|dz }|dzdz}||z }t|||d|<|S)Nr)ebitsmbitsr_emr )ranger)rrrr non_sign_bits default_ebits default_mbitsrs rgenerate_fpx_quantization_typeszXTorchAoConfig._get_torchao_quant_type_to_method..generate_fpx_quantization_types1s"1d^mE 5L1,E:A/Y^fk:lEBtfBugQug67m!%q !.!2q 8 - = %,_MYf%g4&k" r)int4worint4dqr)int8worint8dqr)r)rr)float8wor float8wo_e5m2 float8wo_e4m3float8dqr float8dq_e4m3r)r)ruint1wouint2wouint3wouint4wouint5wouint6wouint7wozYTorchAoConfig requires torchao to be installed, please install with `pip install torchao`)rtorchao.quantizationrrrrrrrrrtorchao.quantization.observerrrrrintrr float8_e4m3fnuint1uint2uint3uint4uint5uint6uint7rbrr|)r6rrrrrrrrrINT4_QUANTIZATION_TYPESINT8_QUANTIZATION_TYPESFLOATX_QUANTIZATION_TYPESUINTX_QUANTIZATION_DTYPESQUANTIZATION_TYPESrrrrs @@@@rrz/TorchAoConfig._get_torchao_quant_type_to_method s !    H u{{  c  +$4=7Z ' #+$4=7Z ' #$$6UEVEVW&8!();%J[J[!\!();%J]J]!^C;b");%*%8%8!&!4!4")#)2*%*=*=>3#)69:`7#)<2!4=#)>2!4?#)@2!4A#)B2!4C#)D2!4E#) %L&7"#4EKKH"#4EKKH"#4EKKH"#4EKKH"#4EKKH"#4EKKH"#4EKKH ) %"$   % %&= >  % %&= >  % %&? @99;"))*CD% %k rctjjr0tjj\}}|dk(r|dk\S|dk\Stjjryt d)N TzPTorchAO requires a CUDA compatible GPU or Intel XPU and installation of PyTorch.)rcuda is_availableget_device_capabilityxpu RuntimeError)majorminors rrz4TorchAoConfig._is_xpu_or_cuda_capability_atleast_8_9sb :: " " $ ::;;=LE5zz!A:  YY # # %qr rrc`|j}||jdi|jS)Nr)rrr)r)rs rget_apply_tensor_subclassz'TorchAoConfig.get_apply_tensor_subclasss1%)%K%K%M":)$//:TT=S=STTrc|j}|jjdtj|ddt dS)a_ Example of how this looks for `TorchAoConfig("uint4wo", group_size=32)`: ``` TorchAoConfig { "modules_to_not_convert": null, "quant_method": "torchao", "quant_type": "uint4wo", "quant_type_kwargs": { "group_size": 32 } } ``` rZr T)rDrEr6rF)rHr+rrIrJr"rs rr\zTorchAoConfig.__repr__sBlln ~~&&'qKUY_q)r(ssu v rr$)rrrrcrgr rrrfr staticmethodrjrr&r\rrrrrsm.`&3&c@S&im&PDDL sD s sU rrc8eZdZdZ ddedeeefdZdZy) QuantoConfiga This is a wrapper class about all possible attributes and features that you can play with a model that has been loaded using `quanto`. Args: weights_dtype (`str`, *optional*, defaults to `"int8"`): The target dtype for the weights after quantization. Supported values are ("float8","int8","int4","int2") modules_to_not_convert (`list`, *optional*, default to `None`): The list of modules to not quantize, useful for quantizing models that explicitly require to have some modules left in their original precision (e.g. Whisper encoder, Llava encoder, Mixtral gate layers). N weights_dtyperc jtj|_||_||_|j yr$)rrr/r+rr)r)r+rr9s rrzQuantoConfig.__init__s, /55*&<# rc^gd}|j|vrtd|d|jy)z; Safety checker that arguments are correct )float8rsint4int2zOnly support weights in z but found N)r+r|)r)accepted_weightss rrzQuantoConfig.post_initsA>   %5 578H7IUYUgUgThij j 6r)rsN) rrrrcrgr rrrrrrr*r*s5 $6:  !)c 3 krr*)(rcrRimportlib.metadatarrrIrh dataclassesrenumr functoolsrtypingrrrr r packagingr utilsr rrr get_loggerrrrgr%torchao.quantization.quant_primitivesr JSONEncoderr"r.rlrrr*rrrr<s$   !33EE   H %dA(T--( oo od e(0e( e(P /4/ /, B +B  B J k*k kr