oMjSrSSKJr SSKrSSKrSSKrSSKrSSKrSSK J r SSK J r SSK JrJr SSKJr SS KJrJrJrJr \"5(aSSKr\R0"\5r"S S \\ 5r\ "S S 55r\ "SS\55r\ "SS\55r\ "SS\55r \ "SS\55r!\ "SS\55r"\ "SS\55r#\ "SS\55r$"SS\5r%g)z Adapted from https://github.com/huggingface/transformers/blob/52cb4034ada381fe1ffe8d428a1076e5411a8026/src/transformers/utils/quantization_config.py ) annotationsN) dataclass)Enum)AnyCallable)version) deprecateis_torch_availableis_torchao_versionloggingc4\rSrSrSrSrSrSrSrSr Sr S r S r g ) QuantizationMethod- bitsandbytesgguf nunchaku_litetorchaoquantomodeloptz auto-roundsdnqN) __name__ __module__ __qualname____firstlineno__BITS_AND_BYTESGGUF NUNCHAKU_LITETORCHAOQUANTOMODELOPT AUTOROUNDSDNQ__static_attributes__rڶC:\Users\johnn\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.13_qbz5n2kfra8p0\LocalCache\local-packages\Python313\site-packages\diffusers/quantizers/quantization_config.pyrr-s*#N D#MG FHI Dr&rcp\rSrSr%SrS\S'/r\SSj5rSSjr SSjr Sr S r SSS jjr S rS rg )QuantizationConfigMixin8z% Mixin class for quantization config r quant_methodc U"S0UD6n/nUR5H4upg[XF5(dM[XFU5 URU5 M6 UHnUR US5 M U(aXC4$U$)an Instantiates a [`QuantizationConfigMixin`] from a Python dictionary of parameters. Args: config_dict (`dict[str, Any]`): Dictionary that will be used to instantiate the configuration object. return_unused_kwargs (`bool`, *optional*, defaults to `False`): Whether or not to return a list of unused keyword arguments. Used for `from_pretrained` method in `PreTrainedModel`. kwargs (`dict[str, Any]`): Additional parameters from which to initialize the configuration object. Returns: [`QuantizationConfigMixin`]: The configuration object instantiated from those parameters. Nr)itemshasattrsetattrappendpop)cls config_dictreturn_unused_kwargskwargsconfig to_removekeyvalues r' from_dict!QuantizationConfigMixin.from_dictAsw$#{#  ,,.JCv##U+  %)C JJsD ! > !Mr&c[USSS9nUR5n[R"USSS9S-nUR U5 SSS5 g!,(df  g=f) 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)selfjson_file_pathwriterr3 json_strings r' to_json_file$QuantizationConfigMixin.to_json_filebsL.# 8F,,.K**[dKdRK LL % 9 8 8s ;A AcB[R"UR5$) Serializes this instance to a Python dictionary. Returns: `dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance. )copydeepcopy__dict__rHs r'rDQuantizationConfigMixin.to_dictss }}T]]++r&c## [R"UR5R5H upX4v M g7f)zTallows `dict(obj)` for situations where obj may be a dict or QuantizationConfigMixinN)rPrQrRr-)rHattrr9s r'__iter__ QuantizationConfigMixin.__iter__zs0==7==?KD+ @s>AcTURRSUR53$)N ) __class__rto_json_stringrSs r'__repr__ QuantizationConfigMixin.__repr__s(..))*!D,?,?,A+BCCr&cUSLaUR5nOUR5n[R"USSS9S-$)ar 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 r?rB) to_diff_dictrDrErF)rHuse_diffr3s r'r\&QuantizationConfigMixin.to_json_strings= t ++-K,,.Kzz+a4@4GGr&c /nUR5H4up4[X5(dM[XU5 URU5 M6 UR5VVs0sHup4X2;dM X4_M nnnU$s snnf)ao Updates attributes of this class instance with attributes from `kwargs` if they match existing attributes, returning all the unused kwargs. Args: kwargs (`dict[str, Any]`): Dictionary of attributes to tentatively update this class. Returns: `dict[str, Any]`: Dictionary containing all the key-value pairs that were not used to update the instance. )r-r.r/r0)rHr5r7r8r9 unused_kwargss r'updateQuantizationConfigMixin.updatesq  ,,.JCt!!5)  %) 7=lln]n H\n ]^s  A8-A8rNF)rIzstr | os.PathLikereturnzdict[str, Any])T)raboolristr)rrrr__doc____annotations___exclude_attributes_at_init classmethodr:rLrDrWr]r\rer%rr&r'r)r)8sH%$"$@&", DH$r&r)c\rSrSrSr/SQrSSjr\S5r\RSSj5r\S5r \ RSS j5r S r S r S r SS jrSrSSjrSrg)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  ^[RTlU(aU(a [S5eUTlUTlUTlUTlUTlUTl UTl U Tl Uc[RTlO][U[ 5(a[#[U5TlO2[U[R$5(aUTlO [S5eU c[R&TlOn[U [ 5(a'U S;a [S5e[#[U 5TlO2[U [R$5(aU TlO [S5eU (aW[+U4SjU 55(d=[,R/S[1U R355STR4S 35 TR75 g) NVload_in_4bit and load_in_8bit are both True, but only one can be used at the same timez8bnb_4bit_compute_dtype must be a string or a torch.dtype)float16float32int8uint8float64bfloat16zv`bnb_4bit_quant_storage` must be a valid string (one of 'float16', 'float32', 'int8', 'uint8', 'float64', 'bfloat16') z8bnb_4bit_quant_storage must be a string or a torch.dtypec3@># UHoTR;v M g7fN)rn).0krHs r' .BitsAndBytesConfig.__init__..sTV4#C#CCVszUnused kwargs: z. These kwargs are not used in .)rrr+ ValueErrorrtrsllm_int8_thresholdllm_int8_skip_modules llm_int8_enable_fp32_cpu_offloadllm_int8_has_fp16_weightbnb_4bit_quant_typebnb_4bit_use_double_quanttorchrxbnb_4bit_compute_dtype isinstancerkgetattrdtyperzbnb_4bit_quant_storageallloggerwarninglistkeysr[ post_init) rH load_in_8bit load_in_4bitrrrrrrrrr5s ` r'__init__BitsAndBytesConfig.__init__s/== Luv v))"4%:"0P-(@%#6 )B& ! )*/--D ' . 4 4*1%9O*PD ' . < <*@D 'WX X ! )*/++D ' . 4 4%.!M+2%9O*PD ' . < <*@D 'WX X #TVTTT NN_T&++--@,AA`aeaoao`ppqr s r&cUR$r~)rsrSs r'rBitsAndBytesConfig.load_in_4bit !!!r&c[U[5(d [S5eUR(aU(a [ S5eXlg)Nload_in_4bit must be a booleanrv)rrj TypeErrorrrrsrHr9s r'rr$8%&&<= =   uv v"r&cUR$r~)rtrSs r'rBitsAndBytesConfig.load_in_8bit-rr&c[U[5(d [S5eUR(aU(a [ S5eXlg)Nload_in_8bit must be a booleanrv)rrjrrrrtrs r'rr1rr&c[UR[5(d [S5e[UR[5(d [S5e[UR [ 5(d [S5eURb*[UR[5(d [S5e[UR[5(d [S5e[UR[5(d [S5eURb4[UR[R5(d [S5e[UR[5(d [S 5e[UR [5(d [S 5eUR(aW["R$"[&R(R#S 55["R$"S 5:d [+S 5egg)zn 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$bnb_4bit_quant_type must be a stringz+bnb_4bit_use_double_quant must be a booleanrz0.39.0z[4 bit quantization requires bitsandbytes>=0.39.0 - please upgrade your bitsandbytes version)rrrjrrrfloatrrrrrrrrrkrrparse importlibmetadatarrSs r'rBitsAndBytesConfig.post_init:s$++T22<= =$++T22<= =$11599@A A  % % 1*TE_E_ae:f:fMN N$??FFPQ Q$77>>HI I  & & 2:dFaFachcncn;o;oHI I$22C88BC C$88$??IJ J   W]]93E3E3M3Mn3]%^bibobo c & m & r&c@UR=(d UR$)z@ Returns `True` if the model is quantizable, `False` otherwise. )rrrSs r'is_quantizable!BitsAndBytesConfig.is_quantizable_s  5D$5$55r&cUR(agUR(aURS:XagUR(aURS:Xagg)zu This method returns the quantization method used for the model. If the model is not quantizable, it returns `None`. llm_int8fp4nf4N)rrrrSs r'quantization_method&BitsAndBytesConfig.quantization_methodesE      4#;#;u#D   4#;#;u#Dr&c[R"UR5n[US5R S5SUS'[US5R S5SUS'UR US'UR US'U$)rOrrrrr)rPrQrRrksplitrr)rHoutputs r'rDBitsAndBytesConfig.to_dictss t}}-+.v6N/O+P+V+VWZ+[\]+^'(+.v6N/O+P+V+VWZ+[\]+^'(!%!2!2~!%!2!2~ r&cUR5nURRS[R"USSS9S3$NrZr Tr?rBrDr[rrErFrHr3s r'r]BitsAndBytesConfig.__repr__;lln ..))*!DJJ{1X\,]+^^`aar&cUR5n[5R5n0nUR5HupEXRU:wdMXSU'M U$)z 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, )rDrqr-)rHr3default_config_dictserializable_config_dictr8r9s r'r`BitsAndBytesConfig.to_diff_dicts[lln 12::<#% &++-JCC0005-.('r&) rsrtrrrrrrrrr+) FFg@NFFNrFN)r9rjrh)rrrrrlrnrpropertyrsetterrrrrrDr]r`r%rr&r'rqrqs1f#U").!&#!"'#<|""##""###J6  b(r&rqc&\rSrSrSrSSSjjrSrg)GGUFQuantizationConfigia5This 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. Nc[RUlXlSUlSUlURc[ RUlgg)NT)rrr+ compute_dtype pre_quantizedmodules_to_not_convertrrx)rHrs r'rGGUFQuantizationConfig.__init__sE.33*!'+#    %!&D  &r&)rrrr+r~)rz'torch.dtype' | None)rrrrrlrr%rr&r'rrs / /r&rcT^\rSrSrSrSSSjjrSrS U4SjjrSrU=r $) NunchakuLiteQuantizationConfigiaConfiguration for loading Nunchaku Lite checkpoints. Nunchaku Lite support in Diffusers loads prequantized checkpoints. To create a compatible checkpoint, use [`diffuse-compressor`](https://github.com/rootonchair/diffuse-compressor) to choose or adapt a target configuration for the model architecture, quantize and export the transformer, and package it as a Diffusers pipeline with the compact `quantization_config` stored in `config.json`. The exported state dict must match the target Diffusers model architecture exactly. Checkpoints quantized with fused QKV projections won't load into a model config that expects separate Q, K, and V projection modules. Example compact `config.json` config: ```json { "_class_name": "ErnieImageTransformer2DModel", "quantization_config": { "quant_method": "nunchaku_lite", "compute_dtype": "bfloat16", "svdq_w4a4": { "precision": "nvfp4", "group_size": 16, "rank": 32, "targets": ["layers.0.self_attention.to_q"] }, "awq_w4a16": { "precision": "int4", "group_size": 64, "targets": ["final_linear"] } } } ``` Args: compute_dtype (`torch.dtype`, defaults to `torch.bfloat16`): Runtime dtype used by the floating-point buffers in the quantized modules. svdq_w4a4 (`dict`, *optional*): Explicit SVDQ W4A4 target configuration with `precision`, `group_size`, `rank`, and `targets`. awq_w4a16 (`dict`, *optional*): Explicit AWQ W4A16 target configuration with `precision`, `group_size`, and `targets`. c [RUlUc[Rn[ U[ 5(a5[[U5(d[SU<S35e[[U5n[ U[R5(d [S5eXl SUl X l X0lUR5 g)Nz$Unsupported Nunchaku compute dtype: rz9Nunchaku compute_dtype must be a string or a torch.dtype.T)rrr+rr|rrkr.rrrrr svdq_w4a4 awq_w4a16r)rHrrrr5s r'r'NunchakuLiteQuantizationConfig.__init__s/<<  !NNM mS ) )5-00 #G GXXY!Z[[#E=9M-55XY Y*!"" r&c URcURc [S5eSUR4SUR44GHupUcM [U[5(d[SU<S35eS[ 4S[ 4S[44HRup4X2;a[SU<S U<S 35e[X#U5(aM3[SU<S U<S URS 35e USnUSnUSnUS ;a[SU<SU<S 35eUS::a[SU<S35eU(d[SU<S35e[SU55(d[SU<S35eUS:XaSU;a[SU<S35e[US[ 5(d[SU<S35eUSS:a[SU<S35eUS:XaSOSnXh:wa[SU<SUSUS 35eGMUS:wdGM[S 5e g)!Nz]Nunchaku compact quantization config must include `svdq_w4a4.targets` or `awq_w4a16.targets`.rrz Nunchaku compact config section z must be a JSON object. precision group_sizetargetsz is missing required field rz field z must be )int4nvfp4zUnsupported Nunchaku precision z for rz must have positive group_size.z" must contain at least one target.c3B# UHn[U[5v M g7fr~)rrk)rtargets r'r;NunchakuLiteQuantizationConfig.post_init..sEW6z&#..Wsz targets must be strings.rankz" is missing required field 'rank'.z field 'rank' must be int.z must have non-negative rank.r@z$Nunchaku SVDQ config with precision=z requires group_size=z, got rz.Nunchaku AWQ target requires precision='int4'.) rrrrdictrkintrrr) rHoprawr8 expected_typerrrexpected_group_sizes r'r(NunchakuLiteQuantizationConfig.post_initsm >> !dnn&<o %dnn5 T^^7TUGB{c4(( #CB6I`!abb(3S'9L#;NQZ\`Pa&b">$'GvMhilhoop%qrr!#(M::$:2&wiXeXnXnWoopq 'cK(I\*J)nG 11 #B9-uUWTZZ[!\]]Q #CB6Ih!ijj #CB6Ik!lmmEWEEE #CB6Ib!cdd[ $$'GvMo%pqq!#f+s33$'GvMg%hiiv;?$'GvMj%kll,5,@bb#4$>ymL&&9%:& AO5 f$ !QRRQVr&cj>[TU]5n[US5RS5SUS'U$)Nrrr)superrDrkrrHrr[s r'rD&NunchakuLiteQuantizationConfig.to_dict%s8""%f_&=">"D"DS"I!"L r&)rrrr+r)NNN)rz'torch.dtype' | str | Nonerdict[str, Any] | Nonerrrh) rrrrrlrrrDr% __classcell__r[s@r'rrsF(X59+/+/ 1)) ..S`r&rcl^\rSrSrSrS S SjjrSrU4Sjr\S Sj5r Sr Sr S r U=r $) TorchAoConfigi+aThis is a config class for torchao quantization/sparsity techniques. Args: quant_type (`AOBaseConfig`): An `AOBaseConfig` subclass instance specifying the quantization type. See the [torchao documentation](https://docs.pytorch.org/ao/main/api_ref_quantization.html#inference-apis-for-quantize) for available config classes (e.g. `Int4WeightOnlyConfig`, `Int8WeightOnlyConfig`, `Float8WeightOnlyConfig`, `Float8DynamicActivationFloat8WeightConfig`, etc.). modules_to_not_convert (`list[str]`, *optional*, default to `None`): The list of modules to not quantize, useful for quantizing models that explicitly require to have some modules left in their original precision. Example: ```python from diffusers import FluxTransformer2DModel, TorchAoConfig from torchao.quantization import Int8WeightOnlyConfig quantization_config = TorchAoConfig(Int8WeightOnlyConfig()) transformer = FluxTransformer2DModel.from_pretrained( "black-forest-labs/Flux.1-Dev", subfolder="transformer", quantization_config=quantization_config, torch_dtype=torch.bfloat16, ) ``` c f[RUlXlX lUR 5 gr~)rr r+ quant_typerr)rHrrr5s r'rTorchAoConfig.__init__Hs' /66$&<# r&c[SS5(a [S5eSSKJn [ UR U5(d+[ S[UR 5R35eg)N<0.15.0zWTorchAoConfig requires torchao >= 0.15.0. Please upgrade with `pip install -U torchao`.r) AOBaseConfigz1quant_type must be an AOBaseConfig instance, got ) r rtorchao.quantization.quant_apirrrrtyper)rHrs r'rTorchAoConfig.post_initTs[ c8 , ,vw w?$//<88OPTUYUdUdPePnPnOopq q9r&c^>[TU]5nSSKJn SU"UR50US'U$)z&Convert configuration to a dictionary.r)config_to_dictdefaultr)rrDtorchao.core.configrr)rHdrr[s r'rDTorchAoConfig.to_dict]s2 GO  7 %nT__&EF,r&c [SS5(d [S5eUR5nURS5n[ U5S:XaSU;dS5eUSnSS KJn U"U5nU"S SU0UD6$) z'Create configuration from a dictionary.z>=rzCTorchAoConfig requires torchao >= 0.15.0 for construction from dictrrrz8Expected only one key 'default' in quant_type dictionaryr)config_from_dictr)r NotImplementedErrorrPr1lenrr)r2r3r4r5rrs r'r:TorchAoConfig.from_dictks"$11%&kl l!&&(  __\2 :!# Z(? F ? *  9%j1 8j8K88r&cUR$)zBCreate the appropriate quantization method based on configuration.)rrSs r'get_apply_tensor_subclass'TorchAoConfig.get_apply_tensor_subclasss r&cUR5nURRS[R"USSS9S3$rrrs r'r]TorchAoConfig.__repr__rr&)rr+rr~)rz'AOBaseConfig'rlist[str] | NoneriNonerg)rrrrrlrrrDror:rr]r%rrs@r'rr+s^<48 " !1   r 99*bbr&rc6\rSrSrSrSSSjjrSrSrg) QuantoConfigia] 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). Nc Sn[SSU5 [RUlXlX lUR 5 g)NzB`QuantoConfig` is deprecated and will be removed in version 1.0.0.r z1.0.0)r rr!r+ weights_dtyperr)rHrrr5deprecation_messages r'rQuantoConfig.__init__s< c.'+>?.55*&<# r&cb/SQnURU;a[SUSUR35eg)z+ Safety checker that arguments are correct )float8ryrint2zOnly support weights in z but found N)rr)rHaccepted_weightss r'rQuantoConfig.post_initsA>   %5 578H7IUYUgUgThij j 6r&)rr+r)ryN)rrkrr )rrrrrlrrr%rr&r'r r s, $37  !1 kr&r c\rSrSrSrSSSSSS.rSSS.rSSS jjrSSS jjrSS jr SS jr Sr g )NVIDIAModelOptConfigia9This is a config class to use nvidia modelopt for quantization. Args: quant_type (`str`): The type of quantization we want to use, following is how to use: **weightquant_activationquant ==> FP8_FP8** In the above example we have use FP8 for both weight and activation quantization. Following are the all the options: - FP8 - INT8 - INT4 - NF4 - NVFP4 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 weight_only (`bool`, *optional*, default to `False`): If set to `True`, the quantization will be applied only to the weights of the model. channel_quantize (`int`, *optional*, default to `None`): The channel quantization axis, useful for quantizing models across different axes. block_quantize (`int`, *optional*, default to `None`): The block size, useful to further quantize each channel/axes into blocks. scale_channel_quantize (`int`, *optional*, default to `None`): The scale channel quantization axis, useful for quantizing calculated scale across different axes. scale_block_quantize (`int`, *optional*, default to `None`): The scale block size, useful for quantizing each scale channel/axes into blocks. algorithm (`str`, *optional*, default to `"max"`): The algorithm to use for quantization, currently only supports `"max"`. forward_loop (`Callable`, *optional*, default to `None`): The forward loop function to use for calibration during quantization. modelopt_config (`dict`, *optional*, default to `None`): The modelopt config, useful for passing custom configs to modelopt. disable_conv_quantization (`bool`, *optional*, default to `False`): If set to `True`, the quantization will be disabled for convolutional layers. kwargs (`dict[str, Any]`, *optional*): Additional parameters which are to be used for calibration. )r)r r)FP8INT8INT4NF4NVFP4)rr Nc  [RUlURU5 X lX0lX@lXPlSU0UlXl X`l Xpl U (dUR5OU Ul Xlg)Nmethod)rr"r+_normalize_quant_typer weight_onlychannel_quantizeblock_quantize calib_cfg forward_loopscale_channel_quantizescale_block_quantizeget_config_from_quant_typemodelopt_configdisable_conv_quantization) rHrrr$r%r&r)r* algorithmr(r,r-r5s r'rNVIDIAModelOptConfig.__init__sv/77 "":.&<#& 0, i )&<#$8!HWt>>@]l)B&r&chSSKJn [U5S:XaSUS3n[R"U5 gg)Nr)_PATCHED_CLASSESzNot z weights in modelopt format. This might cause unreliable behavior.Please make sure to run the following code before loading/saving model weights: from modelopt.torch.opt import enable_huggingface_checkpointing enable_huggingface_checkpointing() )&modelopt.torch.opt.plugins.huggingfacer1rwarningswarn)rH operationr1 warning_msgs r'check_model_patching)NVIDIAModelOptConfig.check_model_patchings<K  A %yk";;  MM+ & &r&cURS5nUSn[U5S:aUSOSn[U5S:a[RSUS35 SnSnOaU[R ;a[RS US 35 SnUb/U[R ;a[RS US 35 SnX4bSU-OS -Ulg)av Validates and normalizes the quantization type string. Splits the quant_type into weight and activation components, verifies them against supported types, and replaces unsupported values with safe defaults. Args: quant_type (str): The input quantization type string (e.g., 'FP8_INT8'). Returns: str: A valid quantization type string (e.g., 'FP8_INT8' or 'FP8'). _rrNr zQuantization type z. is not supported. Picking FP8_INT8 as defaultrzWeight Quantization type z) is not supported. Picking FP8 as defaultzActivation Quantization type z* is not supported. Picking INT8 as default)rrrrrquanttype_to_numbitsr)rHrpartsw_typeact_types r'r#*NVIDIAModelOptConfig._normalize_quant_type s  %q"5zA~584 u:> NN/ |;ij kFH1FFF!:6(Bklm#8L8a8a(a!>xjHrst 6JC(NPRSr&cSSKJs Jn SS00SS00000S.URREUR S.nUSnUR (a#UHnS U;dM X4(aMSX4S'M URRS 5nUSn[U5S :aUS RS S 5OSnUHcnXARR;dMSX4;dM(US:XaUb[RUX4S'MK[RUX4S'Me URbLURb?URUR0US S'URURSS0USS'O9URb,URUS S'URUSS'SUSS'URbUR bU[R";aAUS SR%[R"UURUR 0S.5 U(aUU[R";aAUSSR%[R"UURUR 0S.5 U$)z, Get the config from the quantization type. rN fake_quantFenable)*weight_quantizer*input_quantizerz*output_quantizerz*q_bmm_quantizerz*k_bmm_quantizerz*v_bmm_quantizerz*softmax_quantizer) quant_cfgr.rFrDr:rAr;rEnum_bits block_sizesrdynamicaxis) scale_bitsscale_block_sizes)modelopt.torch.quantizationr quantizationr6_default_disabled_quantizer_cfgr'r$rrrreplacerr<r&r%r)r*quanttype_to_scalingbitsre)rHmtq BASE_CONFIGrFrr=r>r?s r'r+/NVIDIAModelOptConfig.get_config_from_quant_type*s 21'3E%:$&&.%6$&$&$&&( **<<    ,   &a/ -2IL*%%c*q03E Q58##C,DA BBBxW`WcGc**+3G3\3\]e3f Z0+?+T+TU[+\ Z(     *t/D/D/P=A=R=RTXTgTg 'T:Ar&rc^\rSrSrSr/SQrS S SjjrSrS U4Sjjr\ S S U4Sjjj5r Sr U=r $)AutoRoundConfiginuConfiguration class for AutoRound quantization. AutoRound is a weight-only quantization algorithm that uses sign gradient descent to jointly optimize weight rounding and min-max values. This config targets the W4A16 (4-bit weights, 16-bit activations) setting. Reference: https://github.com/intel/auto-round Args: bits (`int`, *optional*, defaults to `4`): The number of bits to quantize weights to. For W4A16 this should be 4. group_size (`int`, *optional*, defaults to `128`): The group size for weight quantization. Weights in each group share the same scale and zero-point. Common choices: 32, 64, 128, -1 (per-channel). sym (`bool`, *optional*, defaults to `True`): Whether to use symmetric quantization (zero-point fixed at 0) or asymmetric quantization (zero-point is learned). backend (`str`, *optional*, defaults to `"auto"`): The backend kernel to use for quantized inference. Available backends: - `"auto"`: Automatically select the best available backend for the current device. - `"torch"`: Pure PyTorch kernel — works on CPU and CUDA. - `"tritonv2"`: Triton-based kernel — requires CUDA. - `"exllamav2"`: Exllamav2 kernel via GPTQModel — requires CUDA and `gptqmodel>=5.8.0`. Offers good CUDA inference performance. - `"marlin"`: Marlin kernel via GPTQModel — requires CUDA and `gptqmodel>=5.8.0`. Offers the best CUDA inference performance. kwargs (`dict[str, Any]`, *optional*): Additional keyword arguments forwarded to AutoRound (e.g. `iters`, `seqlen`, `batch_size`, `lr`, `minmax_lr` for calibration when quantizing from scratch). )autortritonv2 exllamav2marlinc [RUlURU5 XlX lX0lX@lUR5Hupg[XU5 M gr~) rr#r+_validate_backendbitsrsymbackendr-r/)rHrarrbrcr5rvs r'rAutoRoundConfig.__init__sP/88 w' $ LLNDA DQ #r&cXXR;a[SUSUR35eg)NzInvalid backend 'z'. Valid options are: )VALID_BACKENDSr)rHrcs r'r`!AutoRoundConfig._validate_backends5 -- -0 9OPTPcPcOdef f .r&cH>[TU]5nUSRUS'U$)zSerialize the config to a JSON-compatible dict. Output: A dict containing all config fields. The `quant_method` is stored as its string value so it can be round-tripped through JSON. r+)rrDr9rs r'rDAutoRoundConfig.to_dicts+ "!'!7!=!=~ r&c >UR5VVs0sHupEUS:wdM XE_M nnn[TU]"U4SU0UD6$s snnf)zInstantiate an AutoRoundConfig from a dictionary. Input: config_dict with keys like bits, group_size, sym, etc. Output: An AutoRoundConfig instance (and optionally unused kwargs). r+r4)r-rr:)r2r3r4r5rrdr[s r'r:AutoRoundConfig.from_dictsS)4(9(9(;S(;qN?Rtqt(; Sw bCWb[abbTs AA)rcrarr+rb)rTr[) rarrrrbrjrcrkrir )rirrg)r3rr4rj) rrrrrlrgrr`rDror:r%rrs@r'rZrZnsz<JN          "g c c cr&rZc\rSrSrSrSrSrg) SDNQConfigiaConfiguration class for SDNQ (SD.Next Quantization). The `sdnq` library ships its own diffusers-compatible config and quantizer; this class is a thin factory that defers to them so that `quant_method="sdnq"` checkpoints load natively with diffusers. All arguments are forwarded to `sdnq.SDNQConfig`. Requires the `sdnq` library: `pip install sdnq`. Reference: https://github.com/Disty0/sdnq Args: weights_dtype (`str`, *optional*, defaults to `"int8"`): The target dtype for the weights after quantization, e.g. `"int8"`, `"uint4"`, `"float8_e4m3fn"`. See `sdnq.common.accepted_weight_dtypes` for all supported values. group_size (`int`, *optional*, defaults to `0`): How many elements of a tensor share the same quantization group. `0` auto-selects based on `weights_dtype`, `-1` disables grouping and uses row-wise quantization. use_svd (`bool`, *optional*, defaults to `False`): Whether to apply the SVDQuant algorithm on top of SDNQ quantization. use_quantized_matmul (`bool`, *optional*, defaults to `False`): Whether to use quantized INT8 / FP8 / FP16 matmul on the forward pass instead of BF16 / FP16. modules_to_not_convert (`list`, *optional*, defaults to `None`): The list of modules to skip during quantization. kwargs (`dict[str, Any]`, *optional*): Additional keyword arguments forwarded to `sdnq.SDNQConfig` (e.g. `quantized_matmul_dtype`, `svd_rank`, `use_hadamard`, `quant_conv`, `quant_embedding`, `modules_dtype_dict`). c8SSKJn U"5 SSKJn U"U0UD6$)Nr)_check_sdnq_requirementr)ro)sdnq.sdnq_quantizerrqrro)r2argsr5rq SDNQLibConfigs r'__new__SDNQConfig.__new__s@!4d-f--r&rN)rrrrrlrur%rr&r'roros 4.r&ro)&rl __future__rrPimportlib.metadatarrEosr3 dataclassesrenumrtypingrr packagingrutilsr r r r r get_loggerrrrkrr)rqrrrr rrZrorr&r'rsh$ #  ! NN   H %d oo od l(0l( l(^ /4/ /, u%<u up [b+[b [b| !k*!k !kH {2{ {| Jc-Jc JcZ!.(!.r&