# Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. from dataclasses import dataclass from concurrent import futures from fnmatch import fnmatch from functools import partial import io import math from multiprocessing import cpu_count import typing as tp import zlib import torch class BaseQuantizer: @dataclass class _QuantizedParam: name: str param: torch.nn.Parameter module: torch.nn.Module # If a Parameter is used multiple times, `other` can be used # to share state between the different Quantizers other: tp.Optional[tp.Any] def __init__(self, model: torch.nn.Module, min_size: float = 0.01, float16: bool = False, exclude: tp.Optional[tp.List[str]] = [], detect_bound: bool = True): self.model = model self.min_size = min_size self.float16 = float16 self.exclude = exclude self.detect_bound = detect_bound self._quantized = False self._pre_handle = self.model.register_forward_pre_hook(self._forward_pre_hook) self._post_handle = self.model.register_forward_hook(self._forward_hook) self._quantized_state = None self._qparams = [] self._float16 = [] self._others = [] self._rnns = [] self._saved = [] self._find_params() def _find_params(self): min_params = self.min_size * 2**20 // 4 previous = {} for module_name, module in self.model.named_modules(): if isinstance(module, torch.nn.RNNBase): self._rnns.append(module) for name, param in list(module.named_parameters(recurse=False)): full_name = f"{module_name}.{name}" matched = False for pattern in self.exclude: if fnmatch(full_name, pattern) or fnmatch(name, pattern): matched = True break if param.numel() <= min_params or matched: if id(param) in previous: continue if self.detect_bound: previous[id(param)] = None if self.float16: self._float16.append(param) else: self._others.append(param) else: qparam = self._register_param(name, param, module, previous.get(id(param))) if self.detect_bound: previous[id(param)] = qparam self._qparams.append(qparam) def _register_param(self, name, param, module, other): return self.__class__._QuantizedParam(name, param, module, other) def _forward_pre_hook(self, module, input): if self.model.training: self._quantized_state = None if self._quantized: self.unquantize() if self._pre_forward_train(): self._fix_rnns() else: self.quantize() def _forward_hook(self, module, input, output): if self.model.training: if self._post_forward_train(): self._fix_rnns(flatten=False) # Hacky, next forward will flatten def quantize(self, save=True): """ Immediately apply quantization to the model parameters. If `save` is True, save a copy of the unquantized parameters, that can be restored with `unquantize()`. """ if self._quantized: return if save: self._saved = [qp.param.data.to('cpu', copy=True) for qp in self._qparams if qp.other is None] self.restore_quantized_state(self.get_quantized_state()) self._quantized = True self._fix_rnns() def unquantize(self): """ Revert a previous call to `quantize()`. """ if not self._quantized: raise RuntimeError("Can only be called on a quantized model.") if not self._saved: raise RuntimeError("Nothing to restore.") for qparam in self._qparams: if qparam.other is None: qparam.param.data[:] = self._saved.pop(0) assert len(self._saved) == 0 self._quantized = False self._fix_rnns() def _pre_forward_train(self) -> bool: """ Called once before each forward for continuous quantization. Should return True if parameters were changed. """ return False def _post_forward_train(self) -> bool: """ Called once after each forward (to restore state for instance). Should return True if parameters were changed. """ return False def _fix_rnns(self, flatten=True): """ To be called after quantization happened to fix RNNs. """ for rnn in self._rnns: rnn._flat_weights = [ (lambda wn: getattr(rnn, wn) if hasattr(rnn, wn) else None)(wn) for wn in rnn._flat_weights_names] if flatten: rnn.flatten_parameters() def get_quantized_state(self): """ Returns sufficient quantized information to rebuild the model state. ..Note:: To achieve maximum compression, you should compress this with gzip or other, as quantized weights are not optimally coded! """ if self._quantized_state is None: self._quantized_state = self._get_quantized_state() return self._quantized_state def _get_quantized_state(self): """ Actual implementation for `get_quantized_state`. """ float16_params = [] for p in self._float16: q = p.data.half() float16_params.append(q) return { "quantized": [self._quantize_param(qparam) for qparam in self._qparams if qparam.other is None], "float16": float16_params, "others": [p.data.clone() for p in self._others], } def _quantize_param(self, qparam: _QuantizedParam) -> tp.Any: """ To be overriden. """ raise NotImplementedError() def _unquantize_param(self, qparam: _QuantizedParam, quantized: tp.Any) -> torch.Tensor: """ To be overriden. """ raise NotImplementedError() def restore_quantized_state(self, state) -> None: """ Restore the state of the model from the quantized state. """ for p, q in zip(self._float16, state["float16"]): p.data[:] = q.to(p) for p, q in zip(self._others, state["others"]): p.data[:] = q remaining = list(state["quantized"]) for qparam in self._qparams: if qparam.other is not None: # Only unquantize first appearance of nn.Parameter. continue quantized = remaining.pop(0) qparam.param.data[:] = self._unquantize_param(qparam, quantized) self._fix_rnns() def detach(self) -> None: """ Detach from the model, removes hooks and anything else. """ self._pre_handle.remove() self._post_handle.remove() def model_size(self) -> torch.Tensor: """ Returns an estimate of the quantized model size. """ total = torch.tensor(0.) for p in self._float16: total += 16 * p.numel() for p in self._others: total += 32 * p.numel() return total / 2**20 / 8 # bits to MegaBytes def true_model_size(self) -> float: """ Return the true quantized model size, in MB, without extra compression. """ return self.model_size().item() def compressed_model_size(self, compress_level=-1, num_workers=8) -> float: """ Return the compressed quantized model size, in MB. Args: compress_level (int): compression level used with zlib, see `zlib.compress` for details. num_workers (int): will split the final big byte representation in that many chunks processed in parallels. """ out = io.BytesIO() torch.save(self.get_quantized_state(), out) ms = _parallel_compress_len(out.getvalue(), compress_level, num_workers) return ms / 2 ** 20 def _compress_len(data, compress_level): return len(zlib.compress(data, level=compress_level)) def _parallel_compress_len(data, compress_level, num_workers): num_workers = min(cpu_count(), num_workers) chunk_size = int(math.ceil(len(data) / num_workers)) chunks = [data[offset:offset + chunk_size] for offset in range(0, len(data), chunk_size)] with futures.ProcessPoolExecutor(num_workers) as pool: return sum(pool.map(partial(_compress_len, compress_level=compress_level), chunks))