PyTorch TEQ Weight-only 3x API Implementation (#1598)

Signed-off-by: Kaihui-intel <kaihui.tang@intel.com>
Signed-off-by: Tang, Kaihui <kaihui.tang@intel.com>

Author: Kaihui-intel

neural_compressor/torch/algorithms/weight_only/__init__.py

@@ -15,6 +15,7 @@
 from .rtn import rtn_quantize
 from .gptq import gptq_quantize
 from .awq import awq_quantize
+from .teq import teq_quantize
 from .hqq import hqq_quantize
 from .modules import WeightOnlyLinear
 from .utility import *

neural_compressor/torch/algorithms/weight_only/teq.py

@@ -0,0 +1,343 @@
+#
+# -*- coding: utf-8 -*-
+#
+# Copyright (c) 2024 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import torch
+import transformers
+
+from neural_compressor.torch.utils import logger
+
+from .modules import MulLinear, TEQLinearFakeQuant
+from .utility import get_module, quant_tensor, set_module
+
+
+class TEQuantizer:
+    """Weight-only quantization, Trainable Equivalent Transformation (TEQ): linear wrapper to apply scale to input."""
+
+    def __init__(self, model, weight_config={}, absorb_to_layer={}, folding=True, example_inputs=None):
+        """
+        :param model: the model for quantization
+        :param weight_config (dict, optional): contains all info required by RTN. Defaults to {}.
+        :param example_inputs: inputs for trace
+        """
+        self.model = model
+        self.weight_config = weight_config
+        self.folding = folding
+        self.example_inputs = example_inputs
+        self.device, self.dtype = self._get_device()
+        self.model.eval()
+        self.trained_alphas = {}
+        self.absorb_to_layer = absorb_to_layer
+
+    def _get_device(self):
+        """Get the model device
+        :return:Model device."""
+        for _, p in self.model.named_parameters():
+            return p.data.device, p.data.dtype
+
+    def add_tuning_scale(self, sqrt_w_init=False):
+        """The main entry of smooth quant
+        to the paper for more details
+        :param sqrt_w_init: use sqrt weight to init."""
+
+        # freeze model.
+        for n, p in self.model.named_parameters():
+            p.requires_grad = False
+
+        for layer_norm in self.absorb_to_layer:
+            layer_0_name = self.absorb_to_layer[layer_norm][0]
+
+            module = get_module(self.model, layer_0_name)
+
+            if sqrt_w_init:  # pragma: no cover
+                weights = []
+                for layer_name in self.absorb_to_layer[layer_norm]:
+                    module = get_module(self.model, layer_name)
+                    weights.append(module.weight)
+
+                weights = torch.cat(weights, dim=0)
+                max_value = torch.sqrt(torch.max(torch.abs(weights), dim=0).values)
+                max_value[max_value == 0] = 1.0
+                max_value = 1.0 / max_value
+
+                alpha = torch.nn.Parameter(max_value)
+                alpha = alpha.to(self.device)
+            else:
+                alpha = torch.nn.Parameter(torch.ones(module.weight.shape[1], device=self.device))
+
+            self.trained_alphas[layer_norm] = alpha
+            for layer_name in self.absorb_to_layer[layer_norm]:
+                if self.weight_config.get(layer_name) is None:  # pragma: no cover
+                    logger.info(f"layer {layer_name} not in weight config, skip.")
+                    continue
+                num_bits = self.weight_config[layer_name]["bits"]
+                group_size = self.weight_config[layer_name]["group_size"]
+                scheme = self.weight_config[layer_name]["scheme"]
+
+                module = get_module(self.model, layer_name)
+                wrapper_module = TEQLinearFakeQuant(
+                    orig_layer=module, alpha=alpha, num_bits=num_bits, group_size=group_size, scheme=scheme
+                )
+                set_module(self.model, layer_name, wrapper_module)
+
+        for n, m in self.model.named_modules():
+            if isinstance(m, torch.nn.Linear) and "orig_layer" not in n:
+                if self.weight_config.get(n) is None:  # pragma: no cover
+                    logger.info(f"out of absorbed layer {n} not in weight config, skip.")
+                    continue
+                num_bits = self.weight_config[layer_name]["bits"]
+                group_size = self.weight_config[layer_name]["group_size"]
+                scheme = self.weight_config[layer_name]["scheme"]
+
+                alpha = torch.nn.Parameter(torch.ones(m.weight.shape[1], device=self.device))
+                alpha.requires_grad_(False)
+                wrapper_module = TEQLinearFakeQuant(
+                    orig_layer=m, alpha=alpha, num_bits=num_bits, group_size=group_size, scheme=scheme
+                )
+                set_module(self.model, n, wrapper_module)
+
+    @torch.no_grad()
+    def _absorb_scales(self, layer, scale, layer_name=""):
+        """Absorb the scale to the layer at output channel
+        :param layer: The module
+        :param scale: The scale to be absorbed
+        :param layer_name: The layer name."""
+        # for insert mul
+        if not self.folding:  # pragma: no cover
+            if isinstance(layer, MulLinear):
+                set_module(self.model, layer_name, layer.linear)  ##recover
+            else:
+                new_module = MulLinear(layer, scale)
+                set_module(self.model, layer_name, new_module)
+            self.weight_config[layer_name + ".linear"] = self.weight_config[layer_name]
+            return
+
+        if (
+            isinstance(layer, torch.nn.BatchNorm2d)
+            or isinstance(layer, torch.nn.GroupNorm)
+            or isinstance(layer, torch.nn.InstanceNorm2d)
+        ):
+            if layer.affine:  # pragma: no cover
+                layer.weight *= scale
+                layer.bias *= scale
+            else:  # pragma: no cover
+                layer.affine = True
+                weight = torch.ones(layer.num_features, device=self.device, dtype=self.dtype) * scale
+                layer.weight = torch.nn.Parameter(weight, requires_grad=False)
+                bias = torch.zeros(layer.num_features, device=self.device, dtype=self.dtype)
+                layer.bias = torch.nn.Parameter(bias, requires_grad=False)
+        elif isinstance(layer, torch.nn.LayerNorm):
+            if layer.elementwise_affine:
+                layer.weight *= scale
+                layer.bias *= scale
+            else:  # pragma: no cover
+                layer.elementwise_affine = True
+                weight = torch.ones(layer.num_features, device=self.device, dtype=self.dtype) * scale
+                layer.weight = torch.nn.Parameter(torch.ones(weight, requires_grad=False))
+                bias = torch.zeros(layer.num_features, device=self.device, dtype=self.dtype)
+                layer.bias = torch.nn.Parameter(bias, requires_grad=False)
+
+        elif isinstance(layer, torch.nn.Conv2d):  # pragma: no cover
+            ## the order could not be changed
+            if hasattr(layer, "bias") and (layer.bias is not None):
+                layer.bias *= scale
+            scale = scale.view(scale.shape[0], 1, 1, 1)
+            layer.weight *= scale
+
+        elif isinstance(layer, torch.nn.Linear):  # pragma: no cover
+            if hasattr(layer, "bias") and (layer.bias is not None):
+                layer.bias *= scale
+            scale = scale.view(scale.shape[0], 1)
+            layer.weight *= scale
+
+        elif layer.__class__.__name__ == "LlamaRMSNorm" or layer.__class__.__name__ == "T5LayerNorm":  ##quite tricky
+            layer.weight *= scale
+
+        else:  # pragma: no cover
+            logger.info(
+                f"found unsupported layer {type(layer)}, try to multiply scale to "
+                f"weight and bias directly, this may introduce accuracy issue, please have a check "
+            )
+            if hasattr(layer, "weight") and layer.weight is not None:
+                layer.weight *= scale
+            if hasattr(layer, "bias") and layer.bias is not None:
+                layer.bias *= scale
+
+    @torch.no_grad()
+    def _scale_layer_weight(self, layer, scale):  ##input channel
+        """Scale the layer weights at input channel, depthwise conv output channel
+        :param layer_name: The layer name
+        :param scale: The scale to be multiplied
+        :return:"""
+        if layer.__class__.__name__ == "MulLinear":
+            layer = layer.linear
+
+        if layer.__class__.__name__ == "TEQLinearFakeQuant":
+            layer = layer.orig_layer
+
+        scale = scale.view(1, scale.shape[0])
+        layer.weight = torch.nn.Parameter(layer.weight * scale)
+        return scale
+
+    @torch.no_grad()
+    def transform(self):
+        """Apply alpha/scale."""
+        for ln_name, layer_names in self.absorb_to_layer.items():
+            module = get_module(self.model, ln_name)
+            scale = self.trained_alphas[ln_name]
+            scale = torch.clip(scale, 1e-5)
+            input_scale = 1.0 / scale
+            if hasattr(module, "orig_layer"):
+                module = module.orig_layer
+
+            self._absorb_scales(module, input_scale, layer_name=ln_name)
+            weight_scale = scale
+            for layer_name in layer_names:
+                layer_module = get_module(self.model, layer_name)
+                self._scale_layer_weight(layer_module, weight_scale)
+
+        # for Folding = True
+        for n, m in self.model.named_modules():
+            if isinstance(m, TEQLinearFakeQuant):
+                set_module(self.model, n, m.orig_layer)
+
+    def train(
+        self,
+        dataloader,
+        train_steps=1000,
+        lr=1e-3,
+        warmup_ratio=0.05,
+        gradient_accumulation_steps=1,
+        logging_steps=10,
+        betas=[0.9, 0.9],
+        weight_decay=0,
+        lr_scheduler_type="linear",
+    ):
+        """Train function."""
+        trained_alphas_list = []
+        for item in self.trained_alphas.items():
+            trained_alphas_list.append(item[1])
+        optimizer = torch.optim.Adam(trained_alphas_list, lr=lr, weight_decay=weight_decay, betas=betas)
+
+        lr_scheduler = transformers.get_scheduler(  # pylint: disable=E1111
+            name=lr_scheduler_type,
+            optimizer=optimizer,
+            num_warmup_steps=int(train_steps * warmup_ratio) // gradient_accumulation_steps,
+            num_training_steps=train_steps // gradient_accumulation_steps,
+        )
+
+        logger.info("start training")
+        self.model.train()
+        global_steps = 0
+
+        while global_steps <= train_steps:
+            for inputs in dataloader:
+                if isinstance(inputs, torch.Tensor):
+                    input_id = inputs
+                elif isinstance(inputs, dict):
+                    input_id = inputs["input_ids"]
+                else:
+                    input_id = inputs[0]
+
+                input_id = input_id.to(self.device)
+                output = self.model(input_id, labels=input_id)
+                loss = output[0] / gradient_accumulation_steps
+                loss.backward()
+                global_steps += 1
+
+                if global_steps % logging_steps == 0:
+                    logger.info("steps: {}, loss: {}".format(global_steps, loss.detach().cpu().item()))
+
+                if global_steps % gradient_accumulation_steps == 0:
+                    optimizer.step()
+                    optimizer.zero_grad()
+                    lr_scheduler.step()
+
+                if global_steps >= train_steps:  # pragma: no cover
+                    break
+
+        logger.info("finish training")
+        self.model.eval()
+        return None
+
+    @torch.no_grad()
+    def quantize(self):
+        """quantization."""
+
+        for n, m in self.model.named_modules():
+            if self.weight_config.get(n) is None:  # pragma: no cover
+                logger.info(f"quantize layer {n} not in weight config, skip.")
+                continue
+            num_bits = self.weight_config[n]["bits"]
+            group_size = self.weight_config[n]["group_size"]
+            scheme = self.weight_config[n]["scheme"]
+            if isinstance(m, torch.nn.Linear):  # pragma: no cover
+                quant_tensor(m.weight.data, num_bits=num_bits, group_size=group_size, scheme=scheme)
+
+    def save(self, save_scale_file="", save_state_dict_file=""):
+        """
+        save alpha/scale or model weight
+        :param save_scale_file: save alpha/scale with torch.save
+        :param save_state_dict_file: save model state_dict
+        """
+        if save_scale_file:  # pragma: no cover
+            torch.save(self.trained_alphas, save_scale_file)
+
+        if save_state_dict_file:  # pragma: no cover
+            torch.save(self.model.state_dict(), save_state_dict_file)
+
+
+def teq_quantize(
+    model, weight_config={}, absorb_to_layer={}, folding=True, dataloader=None, calib_func=None, example_inputs=None
+):
+    """Run TEQ weight-only quantization."""
+    assert isinstance(model, torch.nn.Module), "only support torch module"
+    logger.info("TEQ quantizing start.")
+    if example_inputs is None:
+        if dataloader is None:  # pragma: no cover
+            assert False, "Please provide dataloader or example_inputs for TEQ algorithm."
+        try:
+            for idx, (input, label) in enumerate(dataloader):
+                example_inputs = input
+                break
+        except:  # pragma: no cover
+            for idx, input in enumerate(dataloader):
+                example_inputs = input
+                break
+
+    teq_quantizer = TEQuantizer(model, weight_config, absorb_to_layer, folding, example_inputs)
+
+    # 1. wrapper tuning scale to model
+    teq_quantizer.add_tuning_scale()
+
+    # 2. tuning
+    # custom train function, there calls calib_func
+    if calib_func:  # pragma: no cover
+        calib_func(teq_quantizer.model)
+    else:
+        if dataloader is None:  # pragma: no cover
+            assert False, "Please provide dataloader to train."
+        teq_quantizer.train(dataloader)
+
+    # 3. apply scale to model
+    teq_quantizer.transform()
+
+    # 4. get quantized model
+    teq_quantizer.quantize()
+
+    logger.info("TEQ quantizing done.")
+    return teq_quantizer.model

neural_compressor/torch/algorithms/weight_only/utility.py

@@ -149,6 +149,7 @@ def qdq_weight_asym(weight, bits=4, quantile=1.0, return_int=False, **kwargs):
     zp.unsqueeze_(dim=-1)
     weight.div_(scale)
     weight.round_()
+    weight.add_(zp)
     weight.clamp_(0, maxq)
     keep_scale = kwargs.get("double_quant", False)
     if return_int or keep_scale:

neural_compressor/torch/quantization/__init__.py

@@ -25,6 +25,8 @@
     get_default_static_config,
     SmoothQuantConfig,
     get_default_sq_config,
+    TEQConfig,
+    get_default_teq_config,
     HQQConfig,
     get_default_hqq_config,
 )