U Kcm@sddlZddlZddlZddlZddlmZddlmmmZ ddl m Z ddl m Z mZmZmZmZmZmZmZddlmZmZddlmZmZddlmZmZmZmZm Z m!Z!ddl"m#Z#ej$ej%j$ej&ej%j&iej%j$ejj$ej%j&ejj&id Z'd d Z(d d Z)d7ddZ*d8ddZ+ddZ,ddZ-d9ddZ.d:ddZ/ddZ0ddZ1d;d d!Z2d"d#Z3d$d%Z4dd,d-Z9d?d/d0Z:d@d1d2Z;d3d4Z.get_activation_post_processcSst|do|jdk S)Nr&r<r&)mrrrneeds_observationsz(add_observer_..needs_observationcs>|r:t|ts:|d|j|t|t|jddS)zn Adds an activation post process module and register a pre or post hook that calls the module r5rDN)rrZ add_moduler&rFr)rMrJrIrKrNrrinsert_activation_post_processsz5add_observer_..insert_activation_post_process)N)N)rget_unique_devices_lenr=formatnextiterr*rnnZDropoutnnqZFloatFunctionalZ QFunctionalr&r5rrr r from_floatsetattrr add_observer_r rZ Sequential) r#qconfig_propagation_listZnon_leaf_module_listrIcustom_module_class_mappingdevicesrQr0r1rJZobserved_childrrPrr[sP           r[cCs$dd|Ddd|DBS)NcSsh|] }|jqSrrI.0prrr sz&get_unique_devices_..cSsh|] }|jqSrr_r`rrrrcs) parametersbuffersr"rrrrRsrRcCsDt|r t|dr |jr t|S|D]\}}t||j|<q(|S)a{Wrap the leaf child module in QuantWrapper if it has a valid qconfig Note that this function will modify the children of module inplace and it can return a new module which wraps the input module as well. Args: module: input module with qconfig attributes for all the leaf modules that we want to quantize Return: Either the inplace modified module with submodules wrapped in `QuantWrapper` based on qconfig or a new `QuantWrapper` module which wraps the input module, the latter case only happens when the input module is a leaf module and we want to quantize it. r&)r r<r&rr*add_quant_dequant_modules)r#r0r1rrrrfs rfcCstjd|dkrt}|di}|s4t|}|}|dkrFt}t|ddt dd| Dsrt dt ||||d|S) aPrepares a copy of the model for quantization calibration or quantization-aware training. Quantization configuration should be assigned preemptively to individual submodules in `.qconfig` attribute. The model will be attached with observer or fake quant modules, and qconfig will be propagated. Args: `model`: input model to be modified in-place `inplace`: carry out model transformations in-place, the original module is mutated `allow_list`: list of quantizable modules `observer_non_leaf_module_list`: list of non-leaf modules we want to add observer `prepare_custom_config_dict`: customization configuration dictionary for prepare function .. code-block:: python # Example of prepare_custom_config_dict: prepare_custom_config_dict = { # user will manually define the corresponding observed # module class which has a from_float class method that converts # float custom module to observed custom module "float_to_observed_custom_module_class": { CustomModule: ObservedCustomModule } } z!quantization_api.quantize.prepareNrr-css|]}t|do|jVqdS)r&NrL)rarMrrr %szprepare..zNone of the submodule got qconfig applied. Make sure you passed correct configuration through `qconfig_dict` or by assigning the `.qconfig` attribute directly on submodules)r])r_C_log_api_usage_oncerr(copydeepcopyrr3anymoduleswarningswarnr[)modelinplaceZ allow_listobserver_non_leaf_module_listr/r]r\rrrprepares&     rucsDtdrtjrtddfdd }|dd|dddS)Nr5Fcs^|r jnj}|rtnt}t}|D]\}}||kr*||q*|D]}||qJdSr)r?rCr:r9setitemsaddpop)rDZhook_mapZ observer_hookZhandle_ids_to_removeZ handle_idhook_fnr"rr remove_hooks7s  z5_remove_activation_post_process..remove_hooksTrO)F)r<r$r5delattr)r#r{rr"r_remove_activation_post_process/s   r}cCs0|D] }t|qt|dr$|`t|dS)zClean up the qconfig left in the module so that new qconfig can be propagated. Args: module: module to be cleaned up r&N)children_remove_qconfigr<r&r})r#r1rrrrEs    rcCs\tjd|dkrt}|s(t|}|t|dd||f|t||dd|S)aQuantize the input float model with post training static quantization. First it will prepare the model for calibration, then it calls `run_fn` which will run the calibration step, after that we will convert the model to a quantized model. Args: model: input float model run_fn: a calibration function for calibrating the prepared model run_args: positional arguments for `run_fn` inplace: carry out model transformations in-place, the original module is mutated mapping: correspondence between original module types and quantized counterparts Return: Quantized model. z"quantization_api.quantize.quantizeNTrs) rrjrkrrlrmevalruconvert)rrrun_fnrun_argsmappingrsrrrquantizeTs   rc Csrtjd|dkr|tjkrHtjttjttjttj ttj ttj ti}nt|tj kr|tjt tjt tjt tj t tj t tj t i}n@|tjkrtjttjti}n$|tjkrtjti}ntd|npt|tr.|tjkrt}n@|tj krt }n0|tjkrt}n |tjkr t}ntdt|tt|t|}|dkr>t}|sNt |}|!t"||t#||dd|S)avConverts a float model to dynamic (i.e. weights-only) quantized model. Replaces specified modules with dynamic weight-only quantized versions and output the quantized model. For simplest usage provide `dtype` argument that can be float16 or qint8. Weight-only quantization by default is performed for layers with large weights size - i.e. Linear and RNN variants. Fine grained control is possible with `qconfig` and `mapping` that act similarly to `quantize()`. If `qconfig` is provided, the `dtype` argument is ignored. Args: model: input model qconfig_spec: Either: - A dictionary that maps from name or type of submodule to quantization configuration, qconfig applies to all submodules of a given module unless qconfig for the submodules are specified (when the submodule already has qconfig attribute). Entries in the dictionary need to be QConfig instances. - A set of types and/or submodule names to apply dynamic quantization to, in which case the `dtype` argument is used to specify the bit-width inplace: carry out model transformations in-place, the original module is mutated mapping: maps type of a submodule to a type of corresponding dynamically quantized version with which the submodule needs to be replaced z*quantization_api.quantize.quantize_dynamicNzTDon't know how to quantize with default settings for {}. Provide full qconfig pleasez.Unknown dtype specified for quantize_dynamic: Tr)$rrjrkqint8rWZLinearrLSTMZGRUZLSTMCellZRNNCellZGRUCellZfloat16rZquint8Z EmbeddingBagrZ EmbeddingZquint4x2r ValueErrorrTrrv RuntimeErrorstrdictzip itertoolsrepeatrrlrmrr3r)rrZ qconfig_specZdtyperrsZdefault_qconfigrrrquantize_dynamicps|             rcCsltjd|jstd|dkr(t}|s6t|}t|ddt ||dddt |t | dd|S) a Prepares a copy of the model for quantization calibration or quantization-aware training and converts it to quantized version. Quantization configuration should be assigned preemptively to individual submodules in `.qconfig` attribute. Args: model: input model to be modified in-place mapping: dictionary that maps float modules to quantized modules to be replaced. inplace: carry out model transformations in-place, the original module is mutated z%quantization_api.quantize.prepare_qatz1prepare_qat only works on models in training modeNrhTF)rrsremove_qconfig)rtrs) rrjrkZtrainingr=rrlrmr3rrurvvalues)rrrrsrrr prepare_qats   rcCsLtjd|st|}|t|dd||f|t|dd|S)agDo quantization aware training and output a quantized model Args: model: input model run_fn: a function for evaluating the prepared model, can be a function that simply runs the prepared model or a training loop run_args: positional arguments for `run_fn` Return: Quantized model. z&quantization_api.quantize.quantize_qatTr)rrjrkrlrmZtrainrr)rrrrrsrrr quantize_qats    rTcCs<tjd|st|}t||d||d|r8t||S)agConverts submodules in input module to a different module according to `mapping` by calling `from_float` method on the target module class. And remove qconfig at the end if remove_qconfig is set to True. Args: `module`: prepared and calibrated module `mapping`: a dictionary that maps from source module type to target module type, can be overwritten to allow swapping user defined Modules `inplace`: carry out model transformations in-place, the original module is mutated `convert_custom_config_dict`: custom configuration dictionary for convert function .. code-block:: python # Example of convert_custom_config_dict: convert_custom_config_dict = { # user will manually define the corresponding quantized # module class which has a from_observed class method that converts # observed custom module to quantized custom module "observed_to_quantized_custom_module_class": { ObservedCustomModule: QuantizedCustomModule } } z!quantization_api.quantize.convertT)rs is_referenceconvert_custom_config_dict)rrjrkrlrm_convertr)r#rrsrrrrrrrs  rc Cs|dkr|rtnt}|dkr&t}|di}|s@t|}i}|D]>\}}t|tszt ||krzt ||d||t |||||<qL| D]\} } | |j | <q|S)aConverts submodules in input module to a different module according to `mapping` by calling `from_float` method on the target module class Args: module: input module mapping: a dictionary that maps from source module type to target module type, can be overwritten to allow swapping user defined Modules inplace: carry out model transformations in-place, the original module is mutated is_reference: a flag to enable quantized reference module NrT)rrrr(rlrmr*rrrr swap_modulerwrg) r#rrsrrr]Zreassignr0modkeyvaluerrrrs,      rc CsJ|}t|drF|jdk rFd}t||krD|t||}d}nnt||kr|t|}t|dr|jr|jdk szt|j}||jt|}|||}n ||}d}|rF|j D]}| |q|j D]} | t k r|| qt|} t| dkstd| t| dkr2tt| nd} | rF|| |S) a Swaps the module if it has a quantized counterpart and it has an `observer` attached. Args: mod: input module mapping: a dictionary that maps from nn module to nnq module Return: The corresponding quantized module of `mod` r&NFT _IS_REFERENCEr zQswap_module only works with cpu or single-device CUDA modules, but got devices {}r)r<r&rZ from_observedrr=Zweightr rYr?rr>rCr9rBrRrSrTrUrVrH) rrr]new_modZswappedZqmodZweight_post_processZweight_qparamsZ pre_hook_fnrzr^rIrrrrFs@          rcCsZdd}t|dr$|j|||d<|D](\}}|rD|||n|}t|||q,dS)a,Traverse the modules and save all observers into dict. This is mainly used for quantization accuracy debug Args: mod: the top module we want to save all observers prefix: the prefix for the current module target_dict: the dictionary used to save all the observers cSs|dkr |S|dS)Nr%r'r)r.rrr get_prefixsz%get_observer_dict..get_prefixr5N)r<r5r*get_observer_dict)r target_dictr.rr0r1r2rrrrxs  r)Nr%N)NN)F)NNNN)FNNN)NF)NF)F)NFTFN)NFFN)r%)>rlrrprZtorch.nnrWZtorch.ao.nn.quantizedr Z quantizedrXZtorch.nn.intrinsicrZ+torch.ao.quantization.quantization_mappingsrrrrrrr r utilsr r Ztorch.ao.quantization.stubsrrZtorch.ao.quantization.qconfigrrrrrrZtorch.nn.utils.parametrizerrZ quantizableZMultiheadAttentionrrr$r,r3r9r:rFr[rRrfrur}rrrrrrrrrrrrrrsz  (     (  V 8  U   ( (2