first commit

config.json

@@ -0,0 +1,28 @@
+{
+  "_name_or_path": "THUDM/chatglm-6b",
+  "architectures": [
+    "ChatGLMForConditionalGeneration"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_chatglm.ChatGLMConfig",
+    "AutoModel": "modeling_chatglm.ChatGLMForConditionalGeneration",
+    "AutoModelForSeq2SeqLM": "modeling_chatglm.ChatGLMForConditionalGeneration"
+  },
+  "bos_token_id": 150004,
+  "eos_token_id": 150005,
+  "hidden_size": 4096,
+  "inner_hidden_size": 16384,
+  "layernorm_epsilon": 1e-05,
+  "max_sequence_length": 2048,
+  "model_type": "chatglm",
+  "num_attention_heads": 32,
+  "num_layers": 28,
+  "pad_token_id": 20003,
+  "position_encoding_2d": true,
+  "quantization_bit": 4,
+  "quantization_embeddings": false,
+  "torch_dtype": "float32",
+  "transformers_version": "4.27.1",
+  "use_cache": true,
+  "vocab_size": 150528
+}

configuration_chatglm.py

@@ -0,0 +1,96 @@
+""" ChatGLM model configuration """
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class ChatGLMConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`~ChatGLMModel`].
+    It is used to instantiate an ChatGLM model according to the specified arguments, defining the model
+    architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of
+    the ChatGLM-6B [THUDM/ChatGLM-6B](https://huggingface.co/THUDM/chatglm-6b) architecture.
+
+    Configuration objects inherit from  [`PretrainedConfig`] and can be used
+    to control the model outputs. Read the documentation from  [`PretrainedConfig`]
+    for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 150528):
+            Vocabulary size of the ChatGLM-6B model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`~ChatGLMModel`] or
+            [`~TFChatGLMModel`].
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 28):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        inner_hidden_size (`int`, *optional*, defaults to 16384):
+            Dimension of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        max_sequence_length (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with.
+            Typically set this to something large just in case (e.g., 512 or 1024 or 2048).
+        layernorm_epsilon (`float`, *optional*, defaults to 1e-5):
+            The epsilon used by the layer normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether the model should return the last key/values attentions (not used by all models).
+        Example:
+
+    ```python
+    >>> from configuration_chatglm import ChatGLMConfig
+    >>> from modeling_chatglm import ChatGLMModel
+
+    >>> # Initializing a ChatGLM-6B THUDM/ChatGLM-6B style configuration
+    >>> configuration = ChatGLMConfig()
+
+    >>> # Initializing a model from the THUDM/ChatGLM-6B style configuration
+    >>> model = ChatGLMModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```
+"""
+    model_type = "chatglm"
+
+    def __init__(
+            self,
+            vocab_size=150528,
+            hidden_size=4096,
+            num_layers=28,
+            num_attention_heads=32,
+            layernorm_epsilon=1e-5,
+            use_cache=False,
+            bos_token_id=150004,
+            eos_token_id=150005,
+            pad_token_id=0,
+            max_sequence_length=2048,
+            inner_hidden_size=16384,
+            position_encoding_2d=True,
+            quantization_bit=0,
+            quantization_embeddings=False,
+            **kwargs
+    ):
+        self.num_layers = num_layers
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_attention_heads = num_attention_heads
+        self.max_sequence_length = max_sequence_length
+        self.layernorm_epsilon = layernorm_epsilon
+        self.inner_hidden_size = inner_hidden_size
+        self.use_cache = use_cache
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.pad_token_id = pad_token_id
+        self.position_encoding_2d = position_encoding_2d
+        self.quantization_bit=quantization_bit
+        self.quantization_embeddings=quantization_embeddings
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            **kwargs
+        )

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 150004,
+  "eos_token_id": 150005,
+  "pad_token_id": 20003,
+  "transformers_version": "4.27.1"
+}

modeling_chatglm.py

@@ -0,0 +1,1318 @@
+""" PyTorch ChatGLM model. """
+
+import math
+import copy
+import os
+import warnings
+import re
+
+import torch
+import torch.utils.checkpoint
+import torch.nn.functional as F
+from torch import nn
+from torch.nn import CrossEntropyLoss, LayerNorm
+from torch.nn.utils import skip_init
+from typing import Optional, Tuple, Union, List, Callable
+
+from transformers.utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    BaseModelOutputWithPastAndCrossAttentions,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from transformers.generation.logits_process import LogitsProcessor
+from transformers.generation.utils import LogitsProcessorList, StoppingCriteriaList, GenerationConfig
+
+from configuration_chatglm import ChatGLMConfig
+
+
+# flags required to enable jit fusion kernels
+torch._C._jit_set_profiling_mode(False)
+torch._C._jit_set_profiling_executor(False)
+torch._C._jit_override_can_fuse_on_cpu(True)
+torch._C._jit_override_can_fuse_on_gpu(True)
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "THUDM/ChatGLM-6B"
+_CONFIG_FOR_DOC = "ChatGLM6BConfig"
+
+CHATGLM_6B_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "THUDM/chatglm-6b",
+    # See all ChatGLM-6B models at https://huggingface.co/models?filter=chatglm
+]
+
+
+class InvalidScoreLogitsProcessor(LogitsProcessor):
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        if torch.isnan(scores).any() or torch.isinf(scores).any():
+            scores.zero_()
+            scores[..., 20005] = 5e4
+        return scores
+
+
+def load_tf_weights_in_chatglm_6b(model, config, tf_checkpoint_path):
+    """Load tf checkpoints in a pytorch model."""
+    try:
+        import re
+
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error(
+            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions."
+        )
+        raise
+    tf_path = os.path.abspath(tf_checkpoint_path)
+    logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        logger.info(f"Loading TF weight {name} with shape {shape}")
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array)
+
+    for name, array in zip(names, arrays):
+        name = name.split("/")
+        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
+        # which are not required for using pretrained model
+        if any(
+                n in ["adam_v", "adam_m", "AdamWeightDecayOptimizer", "AdamWeightDecayOptimizer_1", "global_step"]
+                for n in name
+        ):
+            logger.info(f"Skipping {'/'.join(name)}")
+            continue
+        pointer = model
+        for m_name in name:
+            if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
+                scope_names = re.split(r"_(\d+)", m_name)
+            else:
+                scope_names = [m_name]
+            if scope_names[0] == "kernel" or scope_names[0] == "gamma":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
+                pointer = getattr(pointer, "bias")
+            elif scope_names[0] == "output_weights":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "squad":
+                pointer = getattr(pointer, "classifier")
+            else:
+                try:
+                    pointer = getattr(pointer, scope_names[0])
+                except AttributeError:
+                    logger.info(f"Skipping {'/'.join(name)}")
+                    continue
+            if len(scope_names) >= 2:
+                num = int(scope_names[1])
+                pointer = pointer[num]
+        if m_name[-11:] == "_embeddings":
+            pointer = getattr(pointer, "weight")
+        elif m_name == "kernel":
+            array = np.transpose(array)
+        try:
+            assert (
+                    pointer.shape == array.shape
+            ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        logger.info(f"Initialize PyTorch weight {name}")
+        pointer.data = torch.from_numpy(array)
+    return model
+
+
+@torch.jit.script
+def gelu_impl(x):
+    """OpenAI's gelu implementation."""
+    return 0.5 * x * (1.0 + torch.tanh(0.7978845608028654 * x *
+                                       (1.0 + 0.044715 * x * x)))
+
+
+def gelu(x):
+    return gelu_impl(x)
+
+
+class RotaryEmbedding(torch.nn.Module):
+    def __init__(self, dim, base=10000, precision=torch.half, learnable=False):
+        super().__init__()
+        inv_freq = 1. / (base ** (torch.arange(0, dim, 2).float() / dim))
+        inv_freq = inv_freq.half()
+        self.learnable = learnable
+        if learnable:
+            self.inv_freq = torch.nn.Parameter(inv_freq)
+            self.max_seq_len_cached = None
+        else:
+            self.register_buffer('inv_freq', inv_freq)
+            self.max_seq_len_cached = None
+            self.cos_cached = None
+            self.sin_cached = None
+        self.precision = precision
+
+    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys,
+                              error_msgs):
+        pass
+
+    def forward(self, x, seq_dim=1, seq_len=None):
+        if seq_len is None:
+            seq_len = x.shape[seq_dim]
+        if self.max_seq_len_cached is None or (seq_len > self.max_seq_len_cached):
+            self.max_seq_len_cached = None if self.learnable else seq_len
+            t = torch.arange(seq_len, device=x.device, dtype=self.inv_freq.dtype)
+            freqs = torch.einsum('i,j->ij', t, self.inv_freq)
+            # Different from paper, but it uses a different permutation in order to obtain the same calculation
+            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
+            if self.precision == torch.bfloat16:
+                emb = emb.float()
+
+            # [sx, 1 (b * np), hn]
+            cos_cached = emb.cos()[:, None, :]
+            sin_cached = emb.sin()[:, None, :]
+            if self.precision == torch.bfloat16:
+                cos_cached = cos_cached.bfloat16()
+                sin_cached = sin_cached.bfloat16()
+            if self.learnable:
+                return cos_cached, sin_cached
+            self.cos_cached, self.sin_cached = cos_cached, sin_cached
+        return self.cos_cached[:seq_len, ...], self.sin_cached[:seq_len, ...]
+
+    def _apply(self, fn):
+        if self.cos_cached is not None:
+            self.cos_cached = fn(self.cos_cached)
+        if self.sin_cached is not None:
+            self.sin_cached = fn(self.sin_cached)
+        return super()._apply(fn)
+
+def rotate_half(x):
+    x1, x2 = x[..., :x.shape[-1] // 2], x[..., x.shape[-1] // 2:]
+    return torch.cat((-x2, x1), dim=x1.ndim - 1)  # dim=-1 triggers a bug in earlier torch versions
+
+
+@torch.jit.script
+def apply_rotary_pos_emb_index(q, k, cos, sin, position_id):
+    # position_id: [sq, b], q, k: [sq, b, np, hn], cos: [sq, 1, hn] -> [sq, b, 1, hn]
+    cos, sin = F.embedding(position_id, cos.squeeze(1)).unsqueeze(2), \
+        F.embedding(position_id, sin.squeeze(1)).unsqueeze(2)
+    q, k = (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin)
+    return q, k
+
+
+def attention_fn(
+        self,
+        query_layer,
+        key_layer,
+        value_layer,
+        attention_mask,
+        hidden_size_per_partition,
+        layer_id,
+        layer_past=None,
+        scaling_attention_score=True,
+        use_cache=False,
+):
+    if layer_past is not None:
+        past_key, past_value = layer_past
+        key_layer = torch.cat((past_key, key_layer), dim=0)
+        value_layer = torch.cat((past_value, value_layer), dim=0)
+
+    # seqlen, batch, num_attention_heads, hidden_size_per_attention_head
+    seq_len, b, nh, hidden_size = key_layer.shape
+
+    if use_cache:
+        present = (key_layer, value_layer)
+    else:
+        present = None
+
+    query_key_layer_scaling_coeff = float(layer_id + 1)
+    if scaling_attention_score:
+        query_layer = query_layer / (math.sqrt(hidden_size) * query_key_layer_scaling_coeff)
+
+    # ===================================
+    # Raw attention scores. [b, np, s, s]
+    # ===================================
+
+    # [b, np, sq, sk]
+    output_size = (query_layer.size(1), query_layer.size(2), query_layer.size(0), key_layer.size(0))
+
+    # [sq, b, np, hn] -> [sq, b * np, hn]
+    query_layer = query_layer.view(output_size[2], output_size[0] * output_size[1], -1)
+    # [sk, b, np, hn] -> [sk, b * np, hn]
+    key_layer = key_layer.view(output_size[3], output_size[0] * output_size[1], -1)
+
+    matmul_result = torch.empty(
+        output_size[0] * output_size[1],
+        output_size[2],
+        output_size[3],
+        dtype=query_layer.dtype,
+        device=query_layer.device,
+    )
+
+    matmul_result = torch.baddbmm(
+        matmul_result,
+        query_layer.transpose(0, 1),  # [b * np, sq, hn]
+        key_layer.transpose(0, 1).transpose(1, 2),  # [b * np, hn, sk]
+        beta=0.0,
+        alpha=1.0,
+    )
+
+    # change view to [b, np, sq, sk]
+    attention_scores = matmul_result.view(*output_size)
+
+    if self.scale_mask_softmax:
+        self.scale_mask_softmax.scale = query_key_layer_scaling_coeff
+        attention_probs = self.scale_mask_softmax(attention_scores, attention_mask.contiguous())
+    else:
+        if not (attention_mask == 0).all():
+            # if auto-regressive, skip
+            attention_scores.masked_fill_(attention_mask, -10000.0)
+        dtype = attention_scores.type()
+        attention_scores = attention_scores.float()
+        attention_scores = attention_scores * query_key_layer_scaling_coeff
+
+        attention_probs = F.softmax(attention_scores, dim=-1)
+
+        attention_probs = attention_probs.type(dtype)
+
+    # =========================
+    # Context layer. [sq, b, hp]
+    # =========================
+
+    # value_layer -> context layer.
+    # [sk, b, np, hn] --> [b, np, sq, hn]
+
+    # context layer shape: [b, np, sq, hn]
+    output_size = (value_layer.size(1), value_layer.size(2), query_layer.size(0), value_layer.size(3))
+
+    # change view [sk, b * np, hn]
+    value_layer = value_layer.view(value_layer.size(0), output_size[0] * output_size[1], -1)
+
+    # change view [b * np, sq, sk]
+    attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1)
+
+    # matmul: [b * np, sq, hn]
+    context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1))
+
+    # change view [b, np, sq, hn]
+    context_layer = context_layer.view(*output_size)
+
+    # [b, np, sq, hn] --> [sq, b, np, hn]
+    context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
+
+    # [sq, b, np, hn] --> [sq, b, hp]
+    new_context_layer_shape = context_layer.size()[:-2] + (hidden_size_per_partition,)
+    context_layer = context_layer.view(*new_context_layer_shape)
+
+    outputs = (context_layer, present, attention_probs)
+
+    return outputs
+
+
+class SelfAttention(torch.nn.Module):
+    def __init__(self, hidden_size, num_attention_heads,
+                 layer_id, hidden_size_per_attention_head=None, bias=True,
+                 params_dtype=torch.float, position_encoding_2d=True):
+        super(SelfAttention, self).__init__()
+
+        self.layer_id = layer_id
+        self.hidden_size = hidden_size
+        self.hidden_size_per_partition = hidden_size
+        self.num_attention_heads = num_attention_heads
+        self.num_attention_heads_per_partition = num_attention_heads
+        self.position_encoding_2d = position_encoding_2d
+        self.rotary_emb = RotaryEmbedding(
+            self.hidden_size // (self.num_attention_heads * 2)
+            if position_encoding_2d
+            else self.hidden_size // self.num_attention_heads,
+            base=10000,
+            precision=torch.half,
+            learnable=False,
+        )
+
+        self.scale_mask_softmax = None
+
+        if hidden_size_per_attention_head is None:
+            self.hidden_size_per_attention_head = hidden_size // num_attention_heads
+        else:
+            self.hidden_size_per_attention_head = hidden_size_per_attention_head
+
+        self.inner_hidden_size = num_attention_heads * self.hidden_size_per_attention_head
+
+        # Strided linear layer.
+        self.query_key_value = skip_init(
+            torch.nn.Linear,
+            hidden_size,
+            3 * self.inner_hidden_size,
+            bias=bias,
+            dtype=params_dtype,
+        )
+
+        self.dense = skip_init(
+            torch.nn.Linear,
+            self.inner_hidden_size,
+            hidden_size,
+            bias=bias,
+            dtype=params_dtype,
+        )
+
+    @staticmethod
+    def attention_mask_func(attention_scores, attention_mask):
+        attention_scores.masked_fill_(attention_mask, -10000.0)
+        return attention_scores
+
+    def split_tensor_along_last_dim(self, tensor, num_partitions,
+                                    contiguous_split_chunks=False):
+        """Split a tensor along its last dimension.
+        Arguments:
+            tensor: input tensor.
+            num_partitions: number of partitions to split the tensor
+            contiguous_split_chunks: If True, make each chunk contiguous
+                                    in memory.
+        """
+        # Get the size and dimension.
+        last_dim = tensor.dim() - 1
+        last_dim_size = tensor.size()[last_dim] // num_partitions
+        # Split.
+        tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
+        # Note: torch.split does not create contiguous tensors by default.
+        if contiguous_split_chunks:
+            return tuple(chunk.contiguous() for chunk in tensor_list)
+
+        return tensor_list
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            position_ids,
+            attention_mask: torch.Tensor,
+            layer_id,
+            layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+            use_cache: bool = False,
+            output_attentions: bool = False,
+    ):
+        """
+        hidden_states: [seq_len, batch, hidden_size]
+        attention_mask: [(1, 1), seq_len, seq_len]
+        """
+
+        # [seq_len, batch, 3 * hidden_size]
+        mixed_raw_layer = self.query_key_value(hidden_states)
+
+        # [seq_len, batch, 3 * hidden_size] --> [seq_len, batch, num_attention_heads, 3 * hidden_size_per_attention_head]
+        new_tensor_shape = mixed_raw_layer.size()[:-1] + (
+            self.num_attention_heads_per_partition,
+            3 * self.hidden_size_per_attention_head,
+        )
+        mixed_raw_layer = mixed_raw_layer.view(*new_tensor_shape)
+
+        # [seq_len, batch, num_attention_heads, hidden_size_per_attention_head]
+        (query_layer, key_layer, value_layer) = self.split_tensor_along_last_dim(mixed_raw_layer, 3)
+
+        if self.position_encoding_2d:
+            q1, q2 = query_layer.chunk(2, dim=(query_layer.ndim - 1))
+            k1, k2 = key_layer.chunk(2, dim=(key_layer.ndim - 1))
+            cos, sin = self.rotary_emb(q1, seq_len=position_ids.max() + 1)
+            position_ids, block_position_ids = position_ids[:, 0, :].transpose(0, 1).contiguous(), \
+                position_ids[:, 1, :].transpose(0, 1).contiguous()
+            q1, k1 = apply_rotary_pos_emb_index(q1, k1, cos, sin, position_ids)
+            q2, k2 = apply_rotary_pos_emb_index(q2, k2, cos, sin, block_position_ids)
+            query_layer = torch.concat([q1, q2], dim=(q1.ndim - 1))
+            key_layer = torch.concat([k1, k2], dim=(k1.ndim - 1))
+        else:
+            position_ids = position_ids.transpose(0, 1)
+            cos, sin = self.rotary_emb(value_layer, seq_len=position_ids.max() + 1)
+            # [seq_len, batch, num_attention_heads, hidden_size_per_attention_head]
+            query_layer, key_layer = apply_rotary_pos_emb_index(query_layer, key_layer, cos, sin, position_ids)
+
+        # [seq_len, batch, hidden_size]
+        context_layer, present, attention_probs = attention_fn(
+            self=self,
+            query_layer=query_layer,
+            key_layer=key_layer,
+            value_layer=value_layer,
+            attention_mask=attention_mask,
+            hidden_size_per_partition=self.hidden_size_per_partition,
+            layer_id=layer_id,
+            layer_past=layer_past,
+            use_cache=use_cache
+        )
+
+        output = self.dense(context_layer)
+
+        outputs = (output, present)
+
+        if output_attentions:
+            outputs += (attention_probs,)
+
+        return outputs  # output, present, attention_probs
+
+
+class GEGLU(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.activation_fn = F.gelu
+
+    def forward(self, x):
+        # dim=-1 breaks in jit for pt<1.10
+        x1, x2 = x.chunk(2, dim=(x.ndim - 1))
+        return x1 * self.activation_fn(x2)
+
+
+class GLU(torch.nn.Module):
+    def __init__(self, hidden_size, inner_hidden_size=None,
+                 layer_id=None, bias=True, activation_func=gelu, params_dtype=torch.float):
+        super(GLU, self).__init__()
+        self.layer_id = layer_id
+        self.activation_func = activation_func
+
+        # Project to 4h.
+        self.hidden_size = hidden_size
+        if inner_hidden_size is None:
+            inner_hidden_size = 4 * hidden_size
+        self.inner_hidden_size = inner_hidden_size
+        self.dense_h_to_4h = skip_init(
+            torch.nn.Linear,
+            self.hidden_size,
+            self.inner_hidden_size,
+            bias=bias,
+            dtype=params_dtype,
+        )
+        # Project back to h.
+        self.dense_4h_to_h = skip_init(
+            torch.nn.Linear,
+            self.inner_hidden_size,
+            self.hidden_size,
+            bias=bias,
+            dtype=params_dtype,
+        )
+
+    def forward(self, hidden_states):
+        """
+        hidden_states: [seq_len, batch, hidden_size]
+        """
+
+        # [seq_len, batch, inner_hidden_size]
+        intermediate_parallel = self.dense_h_to_4h(hidden_states)
+
+        intermediate_parallel = self.activation_func(intermediate_parallel)
+
+        output = self.dense_4h_to_h(intermediate_parallel)
+
+        return output
+
+
+class GLMBlock(torch.nn.Module):
+    def __init__(
+            self,
+            hidden_size,
+            num_attention_heads,
+            layernorm_epsilon,
+            layer_id,
+            inner_hidden_size=None,
+            hidden_size_per_attention_head=None,
+            layernorm=LayerNorm,
+            use_bias=True,
+            params_dtype=torch.float,
+            num_layers=28,
+            position_encoding_2d=True
+    ):
+        super(GLMBlock, self).__init__()
+        # Set output layer initialization if not provided.
+
+        self.layer_id = layer_id
+
+        # Layernorm on the input data.
+        self.input_layernorm = layernorm(hidden_size, eps=layernorm_epsilon)
+
+        self.position_encoding_2d = position_encoding_2d
+
+        # Self attention.
+        self.attention = SelfAttention(
+            hidden_size,
+            num_attention_heads,
+            layer_id,
+            hidden_size_per_attention_head=hidden_size_per_attention_head,
+            bias=use_bias,
+            params_dtype=params_dtype,
+            position_encoding_2d=self.position_encoding_2d
+        )
+
+        # Layernorm on the input data.
+        self.post_attention_layernorm = layernorm(hidden_size, eps=layernorm_epsilon)
+
+        self.num_layers = num_layers
+
+        # GLU
+        self.mlp = GLU(
+            hidden_size,
+            inner_hidden_size=inner_hidden_size,
+            bias=use_bias,
+            layer_id=layer_id,
+            params_dtype=params_dtype,
+        )
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            position_ids,
+            attention_mask: torch.Tensor,
+            layer_id,
+            layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+            use_cache: bool = False,
+            output_attentions: bool = False,
+    ):
+        """
+        hidden_states: [seq_len, batch, hidden_size]
+        attention_mask: [(1, 1), seq_len, seq_len]
+        """
+
+        # Layer norm at the begining of the transformer layer.
+        # [seq_len, batch, hidden_size]
+        attention_input = self.input_layernorm(hidden_states)
+
+        # Self attention.
+        attention_outputs = self.attention(
+            attention_input,
+            position_ids,
+            attention_mask=attention_mask,
+            layer_id=layer_id,
+            layer_past=layer_past,
+            use_cache=use_cache,
+            output_attentions=output_attentions
+        )
+
+        attention_output = attention_outputs[0]
+
+        outputs = attention_outputs[1:]
+
+        # Residual connection.
+        alpha = (2 * self.num_layers) ** 0.5
+        hidden_states = attention_input * alpha + attention_output
+
+        mlp_input = self.post_attention_layernorm(hidden_states)
+
+        # MLP.
+        mlp_output = self.mlp(mlp_input)
+
+        # Second residual connection.
+        output = mlp_input * alpha + mlp_output
+
+        if use_cache:
+            outputs = (output,) + outputs
+        else:
+            outputs = (output,) + outputs[1:]
+
+        return outputs  # hidden_states, present, attentions
+
+
+class ChatGLMPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and
+    a simple interface for downloading and loading pretrained models.
+    """
+
+    is_parallelizable = False
+    supports_gradient_checkpointing = False
+    config_class = ChatGLMConfig
+    base_model_prefix = "transformer"
+    _no_split_modules = ["GLM6BBlock"]
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module: nn.Module):
+        """Initialize the weights."""
+        return
+
+
+CHATGLM_6B_START_DOCSTRING = r"""
+    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general
+    usage and behavior.
+
+    Parameters:
+        config ([`~ChatGLM6BConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+CHATGLM_6B_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`ChatGLM6BTokenizer`].
+            See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0, 1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings.
+            Selected in the range `[0, config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert *input_ids* indices into associated vectors
+            than the model's internal embedding lookup matrix.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare ChatGLM-6B Model transformer outputting raw hidden-states without any specific head on top.",
+    CHATGLM_6B_START_DOCSTRING,
+)
+class ChatGLMModel(ChatGLMPreTrainedModel):
+    """
+
+    The model can behave as an encoder (with only self-attention) as well
+    as a decoder, in which case a layer of cross-attention is added between
+    the self-attention layers, following the architecture described in [Attention is
+    all you need](https://arxiv.org/abs/1706.03762) by Ashish Vaswani,
+    Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
+
+    To behave as an decoder the model needs to be initialized with the
+    `is_decoder` argument of the configuration set to `True`.
+    To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder`
+    argument and `add_cross_attention` set to `True`; an
+    `encoder_hidden_states` is then expected as an input to the forward pass.
+    """
+
+    def __init__(self, config: ChatGLMConfig):
+        super().__init__(config)
+
+        # recording parameters
+        self.max_sequence_length = config.max_sequence_length
+        self.hidden_size = config.hidden_size
+        self.params_dtype = torch.half
+        self.num_attention_heads = config.num_attention_heads
+        self.vocab_size = config.vocab_size
+        self.num_layers = config.num_layers
+        self.layernorm_epsilon = config.layernorm_epsilon
+        self.inner_hidden_size = config.inner_hidden_size
+        self.hidden_size_per_attention_head = self.hidden_size // self.num_attention_heads
+        self.position_encoding_2d = config.position_encoding_2d
+
+        self.word_embeddings = skip_init(
+            torch.nn.Embedding,
+            num_embeddings=self.vocab_size, embedding_dim=self.hidden_size,
+            dtype=self.params_dtype
+        )
+
+        def get_layer(layer_id):
+            return GLMBlock(
+                self.hidden_size,
+                self.num_attention_heads,
+                self.layernorm_epsilon,
+                layer_id,
+                inner_hidden_size=self.inner_hidden_size,
+                hidden_size_per_attention_head=self.hidden_size_per_attention_head,
+                layernorm=LayerNorm,
+                use_bias=True,
+                params_dtype=self.params_dtype,
+                position_encoding_2d=self.position_encoding_2d,
+            )
+
+        self.layers = torch.nn.ModuleList(
+            [get_layer(layer_id) for layer_id in range(self.num_layers)]
+        )
+
+        # Final layer norm before output.
+        self.final_layernorm = LayerNorm(self.hidden_size, eps=self.layernorm_epsilon)
+
+    def get_input_embeddings(self):
+        return self.word_embeddings
+
+    def set_input_embeddings(self, new_embeddings: torch.Tensor):
+        self.word_embeddings = new_embeddings
+
+    def get_masks(self, seq, device):
+        context_length = seq.index(self.config.bos_token_id) + 1
+
+        attention_mask = torch.ones((1, len(seq), len(seq)), device=device)
+        attention_mask.tril_()
+        attention_mask[..., :context_length - 1] = 1
+        attention_mask.unsqueeze_(1)
+        attention_mask = (attention_mask < 0.5).bool()
+
+        return attention_mask
+
+    def get_position_ids(self, seq, mask_position, device, gmask=False):
+        context_length = seq.index(self.config.bos_token_id) + 1
+        if self.position_encoding_2d:
+            seq_length = seq.index(self.config.bos_token_id)
+            position_ids = torch.arange(context_length, dtype=torch.long, device=device)
+            if not gmask:
+                position_ids[seq_length:] = mask_position
+            block_position_ids = torch.cat((
+                torch.zeros(seq_length, dtype=torch.long, device=device),
+                torch.arange(context_length - seq_length, dtype=torch.long, device=device) + 1
+            ))
+            position_ids = torch.stack((position_ids, block_position_ids), dim=0)
+        else:
+            position_ids = torch.arange(context_length, dtype=torch.long, device=device)
+            if not gmask:
+                position_ids[context_length - 1:] = mask_position
+
+        position_ids = position_ids.unsqueeze(0)
+
+        return position_ids
+
+    @add_start_docstrings_to_model_forward(CHATGLM_6B_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPastAndCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+            self,
+            input_ids: Optional[torch.LongTensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+            inputs_embeds: Optional[torch.LongTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor, ...], BaseModelOutputWithPast]:
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape[:2]
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape[:2]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if past_key_values is None:
+            past_key_values = tuple([None] * len(self.layers))
+            seq = input_ids[0].tolist()
+
+            if attention_mask is None:
+                attention_mask = self.get_masks(
+                    seq=seq,
+                    device=input_ids.device
+                )
+
+            if position_ids is None:
+                MASK, gMASK = 150000, 150001
+                mask_token = MASK if MASK in input_ids else gMASK
+                use_gmask = False if MASK in input_ids else gMASK
+
+                mask_position = seq.index(mask_token)
+                position_ids = self.get_position_ids(
+                    seq=seq,
+                    mask_position=mask_position,
+                    device=input_ids.device,
+                    gmask=use_gmask
+                )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+
+        # [seq_len, batch, hidden_size]
+        hidden_states = inputs_embeds.transpose(0, 1)
+
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+
+        seq_length_with_past = seq_length
+        past_key_values_length = 0
+        if past_key_values[0] is not None:
+            past_key_values_length = past_key_values[0][0].shape[0]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+        if attention_mask is None:
+            attention_mask = torch.zeros(1, 1, device=input_ids.device).bool()
+
+        else:
+            attention_mask = attention_mask.to(input_ids.device)
+
+        for i, layer in enumerate(self.layers):
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_ret = layer(
+                hidden_states,
+                position_ids=position_ids,
+                attention_mask=attention_mask,
+                layer_id=torch.tensor(i),
+                layer_past=past_key_values[i],
+                use_cache=use_cache,
+                output_attentions=output_attentions
+            )
+
+            hidden_states = layer_ret[0]
+
+            if use_cache:
+                presents = presents + (layer_ret[1],)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_ret[2 if use_cache else 1],)
+
+        # Final layer norm.
+        hidden_states = self.final_layernorm(hidden_states)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
+    def __init__(self, config: ChatGLMConfig):
+        super().__init__(config)
+
+        # self.hidden_size = config.hidden_size
+        # self.params_dtype = torch.half
+        # self.vocab_size = config.vocab_size
+        self.max_sequence_length = config.max_sequence_length
+
+        self.position_encoding_2d = config.position_encoding_2d
+
+        self.transformer = ChatGLMModel(config)
+
+        self.lm_head = skip_init(
+            nn.Linear,
+            config.hidden_size,
+            config.vocab_size,
+            bias=False,
+            dtype=torch.half
+        )
+
+        self.config = config
+
+        self.quantized = False
+
+        if self.config.quantization_bit:
+            self.quantize(self.config.quantization_bit, self.config.quantization_embeddings, use_quantization_cache=True, empty_init=True)
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def get_masks_and_position_ids(self, seq, mask_position, context_length, device, gmask=False):
+        attention_mask = torch.ones((1, context_length, context_length), device=device)
+        attention_mask.tril_()
+        attention_mask[..., :context_length - 1] = 1
+        attention_mask.unsqueeze_(1)
+        attention_mask = (attention_mask < 0.5).bool()
+
+        if self.position_encoding_2d:
+            seq_length = seq.index(self.config.bos_token_id)
+            position_ids = torch.arange(context_length, dtype=torch.long, device=device)
+            if not gmask:
+                position_ids[seq_length:] = mask_position
+            block_position_ids = torch.cat((
+                torch.zeros(seq_length, dtype=torch.long, device=device),
+                torch.arange(context_length - seq_length, dtype=torch.long, device=device) + 1
+            ))
+            position_ids = torch.stack((position_ids, block_position_ids), dim=0)
+        else:
+            position_ids = torch.arange(context_length, dtype=torch.long, device=device)
+            if not gmask:
+                position_ids[context_length - 1:] = mask_position
+
+        position_ids = position_ids.unsqueeze(0)
+
+        return attention_mask, position_ids
+
+    def prepare_inputs_for_generation(
+            self,
+            input_ids: torch.LongTensor,
+            past: Optional[torch.Tensor] = None,
+            past_key_values: Optional[torch.Tensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            **kwargs
+    ) -> dict:
+
+        MASK, gMASK = 150000, 150001
+        mask_token = MASK if MASK in input_ids else gMASK
+        use_gmask = False if MASK in input_ids else gMASK
+        seq = input_ids[0].tolist()
+        mask_position = seq.index(mask_token)
+
+        if mask_token not in seq:
+            raise ValueError("You have to add either [MASK] or [gMASK] in your input")
+
+        # only last token for input_ids if past is not None
+        if past is not None or past_key_values is not None:
+            context_length = seq.index(self.config.bos_token_id)
+            last_token = input_ids[:, -1].unsqueeze(-1)
+            if self.position_encoding_2d:
+                position_ids = torch.tensor([[[mask_position], [len(seq) - context_length]]], dtype=torch.long,
+                                            device=input_ids.device)
+            else:
+                position_ids = torch.tensor([[mask_position]], dtype=torch.long, device=input_ids.device)
+
+            if past is None:
+                past = past_key_values
+            return {
+                "input_ids": last_token,
+                "past_key_values": past,
+                "position_ids": position_ids,
+            }
+        else:
+            attention_mask, position_ids = self.get_masks_and_position_ids(
+                seq=seq,
+                mask_position=mask_position,
+                context_length=len(seq),
+                device=input_ids.device,
+                gmask=use_gmask
+            )
+
+            return {
+                "input_ids": input_ids,
+                "past_key_values": past,
+                "position_ids": position_ids,
+                "attention_mask": attention_mask
+            }
+
+    def forward(
+            self,
+            input_ids: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.Tensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_values: Optional[Tuple[torch.FloatTensor]] = None,
+            inputs_embeds: Optional[torch.Tensor] = None,
+            labels: Optional[torch.Tensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ):
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = transformer_outputs[0]
+
+        lm_logits = self.lm_head(hidden_states).permute(1, 0, 2).contiguous()
+
+        loss = None
+        if labels is not None:
+            lm_logits = lm_logits.to(torch.float32)
+
+            # Shift so that tokens < n predict n
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+            lm_logits = lm_logits.to(hidden_states.dtype)
+            loss = loss.to(hidden_states.dtype)
+
+        if not return_dict:
+            output = (lm_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(
+            past: Tuple[Tuple[torch.Tensor, torch.Tensor], ...], beam_idx: torch.LongTensor
+    ) -> Tuple[Tuple[torch.Tensor, torch.Tensor], ...]:
+        """
+        This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
+        [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
+        beam_idx at every generation step.
+
+        Output shares the same memory storage as `past`.
+        """
+        return tuple(
+            (
+                layer_past[0].index_select(1, beam_idx.to(layer_past[0].device)),
+                layer_past[1].index_select(1, beam_idx.to(layer_past[1].device)),
+            )
+            for layer_past in past
+        )
+
+    def process_response(self, response):
+        response = response.strip()
+        response = response.replace("[[训练时间]]", "2023年")
+        punkts = [
+            [",", "，"],
+            ["!", "！"],
+            [":", "："],
+            [";", "；"],
+            ["\?", "？"],
+        ]
+        for item in punkts:
+            response = re.sub(r"([\u4e00-\u9fff])%s" % item[0], r"\1%s" % item[1], response)
+            response = re.sub(r"%s([\u4e00-\u9fff])" % item[0], r"%s\1" % item[1], response)
+        return response
+
+    @torch.no_grad()
+    def chat(self, tokenizer, query: str, history: List[Tuple[str, str]] = None, max_length: int = 2048, num_beams=1,
+             do_sample=True, top_p=0.7, temperature=0.95, logits_processor=None, **kwargs):
+        if history is None:
+            history = []
+        if logits_processor is None:
+            logits_processor = LogitsProcessorList()
+        logits_processor.append(InvalidScoreLogitsProcessor())
+        gen_kwargs = {"max_length": max_length, "num_beams": num_beams, "do_sample": do_sample, "top_p": top_p,
+                      "temperature": temperature, "logits_processor": logits_processor, **kwargs}
+        if not history:
+            prompt = query
+        else:
+            prompt = ""
+            for i, (old_query, response) in enumerate(history):
+                prompt += "[Round {}]\n问：{}\n答：{}\n".format(i, old_query, response)
+            prompt += "[Round {}]\n问：{}\n答：".format(len(history), query)
+        input_ids = tokenizer([prompt], return_tensors="pt", padding=True)
+        input_ids = input_ids.to(self.device)
+        outputs = self.generate(**input_ids, **gen_kwargs)
+        outputs = outputs.tolist()[0][len(input_ids["input_ids"][0]):]
+        response = tokenizer.decode(outputs)
+        response = self.process_response(response)
+        history = history + [(query, response)]
+        return response, history
+
+    @torch.no_grad()
+    def stream_chat(self, tokenizer, query: str, history: List[Tuple[str, str]] = None, max_length: int = 2048,
+                    do_sample=True, top_p=0.7, temperature=0.95, logits_processor=None, **kwargs):
+        if history is None:
+            history = []
+        if logits_processor is None:
+            logits_processor = LogitsProcessorList()
+        logits_processor.append(InvalidScoreLogitsProcessor())
+        gen_kwargs = {"max_length": max_length, "do_sample": do_sample, "top_p": top_p,
+                      "temperature": temperature, "logits_processor": logits_processor, **kwargs}
+        if not history:
+            prompt = query
+        else:
+            prompt = ""
+            for i, (old_query, response) in enumerate(history):
+                prompt += "[Round {}]\n问：{}\n答：{}\n".format(i, old_query, response)
+            prompt += "[Round {}]\n问：{}\n答：".format(len(history), query)
+        input_ids = tokenizer([prompt], return_tensors="pt", padding=True)
+        input_ids = input_ids.to(self.device)
+        for outputs in self.stream_generate(**input_ids, **gen_kwargs):
+            outputs = outputs.tolist()[0][len(input_ids["input_ids"][0]):]
+            response = tokenizer.decode(outputs)
+            response = self.process_response(response)
+            new_history = history + [(query, response)]
+            yield response, new_history
+
+    @torch.no_grad()
+    def stream_generate(
+            self,
+            input_ids,
+            generation_config: Optional[GenerationConfig] = None,
+            logits_processor: Optional[LogitsProcessorList] = None,
+            stopping_criteria: Optional[StoppingCriteriaList] = None,
+            prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
+            **kwargs,
+    ):
+        batch_size, input_ids_seq_length = input_ids.shape[0], input_ids.shape[-1]
+
+        if generation_config is None:
+            generation_config = self.generation_config
+        generation_config = copy.deepcopy(generation_config)
+        model_kwargs = generation_config.update(**kwargs)
+        bos_token_id, eos_token_id = generation_config.bos_token_id, generation_config.eos_token_id
+
+        if isinstance(eos_token_id, int):
+            eos_token_id = [eos_token_id]
+
+        has_default_max_length = kwargs.get("max_length") is None and generation_config.max_length is not None
+        if has_default_max_length and generation_config.max_new_tokens is None:
+            warnings.warn(
+                f"Using `max_length`'s default ({generation_config.max_length}) to control the generation length. "
+                "This behaviour is deprecated and will be removed from the config in v5 of Transformers -- we"
+                " recommend using `max_new_tokens` to control the maximum length of the generation.",
+                UserWarning,
+            )
+        elif generation_config.max_new_tokens is not None:
+            generation_config.max_length = generation_config.max_new_tokens + input_ids_seq_length
+            if not has_default_max_length:
+                logger.warn(
+                    f"Both `max_new_tokens` (={generation_config.max_new_tokens}) and `max_length`(="
+                    f"{generation_config.max_length}) seem to have been set. `max_new_tokens` will take precedence. "
+                    "Please refer to the documentation for more information. "
+                    "(https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)",
+                    UserWarning,
+                )
+
+        if input_ids_seq_length >= generation_config.max_length:
+            input_ids_string = "decoder_input_ids" if self.config.is_encoder_decoder else "input_ids"
+            logger.warning(
+                f"Input length of {input_ids_string} is {input_ids_seq_length}, but `max_length` is set to"
+                f" {generation_config.max_length}. This can lead to unexpected behavior. You should consider"
+                " increasing `max_new_tokens`."
+            )
+
+        # 2. Set generation parameters if not already defined
+        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
+        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
+
+        logits_processor = self._get_logits_processor(
+            generation_config=generation_config,
+            input_ids_seq_length=input_ids_seq_length,
+            encoder_input_ids=input_ids,
+            prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
+            logits_processor=logits_processor,
+        )
+
+        stopping_criteria = self._get_stopping_criteria(
+            generation_config=generation_config, stopping_criteria=stopping_criteria
+        )
+        logits_warper = self._get_logits_warper(generation_config)
+
+        unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1)
+        scores = None
+        while True:
+            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)
+            # forward pass to get next token
+            outputs = self(
+                **model_inputs,
+                return_dict=True,
+                output_attentions=False,
+                output_hidden_states=False,
+            )
+
+            next_token_logits = outputs.logits[:, -1, :]
+
+            # pre-process distribution
+            next_token_scores = logits_processor(input_ids, next_token_logits)
+            next_token_scores = logits_warper(input_ids, next_token_scores)
+
+            # sample
+            probs = nn.functional.softmax(next_token_scores, dim=-1)
+            if generation_config.do_sample:
+                next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)
+            else:
+                next_tokens = torch.argmax(probs, dim=-1)
+
+            # update generated ids, model inputs, and length for next step
+            input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
+            model_kwargs = self._update_model_kwargs_for_generation(
+                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
+            )
+            unfinished_sequences = unfinished_sequences.mul((sum(next_tokens != i for i in eos_token_id)).long())
+
+            # stop when each sentence is finished, or if we exceed the maximum length
+            if unfinished_sequences.max() == 0 or stopping_criteria(input_ids, scores):
+                break
+            yield input_ids
+
+    def quantize(self, bits: int, quantize_embeddings=False, use_quantization_cache=False, empty_init=False, **kwargs):
+        if bits == 0:
+            return
+
+        from quantization import quantize, QuantizedEmbedding, QuantizedLinear, load_cpu_kernel
+
+        if self.quantized:
+            if self.device == torch.device("cpu"):
+                logger.info("Already quantized, reloading cpu kernel.")
+                load_cpu_kernel(**kwargs)
+            else:
+                logger.info("Already quantized.")
+            return self
+
+        self.quantized = True
+
+        self.config.quantization_bit = bits
+        self.config.quantization_embeddings = quantize_embeddings
+
+        self.transformer = quantize(self.transformer, bits, use_quantization_cache=use_quantization_cache, empty_init=empty_init, **kwargs)
+
+        if quantize_embeddings:
+            logger.info("Applying quantization to embeddings")
+            self.transformer.word_embeddings = QuantizedEmbedding(
+                weight_bit_width=bits,
+                weight_tensor=self.transformer.word_embeddings.weight.to(self.device),
+                num_embeddings=self.transformer.word_embeddings.num_embeddings,
+                embedding_dim=self.transformer.word_embeddings.embedding_dim,
+                dtype=torch.half,
+                empty_init=True,
+                device=self.transformer.word_embeddings.weight.device,
+            )
+            self.lm_head =  QuantizedLinear(
+                weight_bit_width=bits,
+                weight_tensor=self.lm_head.weight.to(self.device),
+                bias_tensor=None,
+                in_features=self.lm_head.in_features,
+                out_features=self.lm_head.out_features,
+                bias=False,
+                quantized_weight=self.transformer.word_embeddings.weight,
+                quantized_weight_scale=self.transformer.word_embeddings.weight_scale,
+                dtype=torch.half,
+                empty_init=True,
+                device=self.lm_head.weight.device,
+            )
+
+        return self

quantization.py

@@ -0,0 +1,476 @@
+from torch.nn import Linear, Embedding
+from torch.nn.parameter import Parameter
+import torch.nn.functional as F
+
+import os
+import bz2
+import torch
+import base64
+import ctypes
+
+from typing import List
+from functools import partial
+
+try:
+    from cpm_kernels.kernels.base import LazyKernelCModule, KernelFunction, round_up
+
+    class Kernel:
+        def __init__(self, code: bytes, function_names: List[str]):
+            self.code = code
+            self._function_names = function_names
+            self._cmodule = LazyKernelCModule(self.code)
+
+            for name in self._function_names:
+                setattr(self, name, KernelFunction(self._cmodule, name))
+
+    quantization_code = "$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"
+
+    kernels = Kernel(
+        bz2.decompress(base64.b64decode(quantization_code)),
+        [
+            "int4WeightCompression",
+            "int4WeightExtractionFloat",
+            "int4WeightExtractionHalf",
+            "int8WeightExtractionFloat",
+            "int8WeightExtractionHalf",
+        ],
+    )
+except Exception as exception:
+    kernels = None
+    print("Failed to load cpm_kernels:", exception)
+
+
+class W8A16Linear(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, inp: torch.Tensor, quant_w: torch.Tensor, scale_w: torch.Tensor, weight_bit_width):
+        ctx.inp_shape = inp.size()
+        ctx.weight_shape = quant_w.size()
+        ctx.weight_bit_width = weight_bit_width
+        out_features = quant_w.size(0)
+        inp = inp.contiguous().view(-1, inp.size(-1))
+        weight = extract_weight_to_half(quant_w, scale_w, weight_bit_width)
+        output = inp.mm(weight.t())
+        ctx.save_for_backward(inp, quant_w, scale_w)
+        return output.view(*(ctx.inp_shape[:-1] + (out_features,)))
+
+    @staticmethod
+    def backward(ctx, grad_output: torch.Tensor):
+        inp, quant_w, scale_w = ctx.saved_tensors
+        weight = extract_weight_to_half(quant_w, scale_w, ctx.weight_bit_width)
+        grad_output = grad_output.contiguous().view(-1, weight.size(0))
+        grad_input = grad_output.mm(weight)
+        grad_weight = grad_output.t().mm(inp)
+        return grad_input.view(ctx.inp_shape), grad_weight.view(ctx.weight_shape), None
+
+
+class W8A16LinearCPU(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, inp: torch.Tensor, quant_w: torch.Tensor, scale_w: torch.Tensor, weight_bit_width, quantization_cache=None):
+        ctx.inp_shape = inp.size()
+        ctx.weight_shape = quant_w.size()
+        ctx.weight_bit_width = weight_bit_width
+        out_features = quant_w.size(0)
+        inp = inp.contiguous().view(-1, inp.size(-1))
+        weight = extract_weight_to_float(quant_w, scale_w, weight_bit_width, quantization_cache=quantization_cache)
+        output = inp.mm(weight.t())
+        ctx.save_for_backward(inp, quant_w, scale_w)
+        return output.view(*(ctx.inp_shape[:-1] + (out_features,)))
+
+    @staticmethod
+    def backward(ctx, grad_output: torch.Tensor):
+        inp, quant_w, scale_w = ctx.saved_tensors
+        weight = extract_weight_to_float(quant_w, scale_w, ctx.weight_bit_width)
+        grad_output = grad_output.contiguous().view(-1, weight.size(0))
+        grad_input = grad_output.mm(weight)
+        grad_weight = grad_output.t().mm(inp)
+        return grad_input.view(ctx.inp_shape), grad_weight.view(ctx.weight_shape), None
+
+
+default_cpu_kernel_code_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "quantization_kernels.c")
+default_cpu_kernel_code = "QlpoOTFBWSZTWXLbSoQAAgzbgERwQXxmTwAAr/ff3kABt0Q2oRVT0hpo9RtEAAAAyBEiSQ9EGjQGQAAAwANGhowjJoNGmgMEUplMTNSMJ5TQaDJpsoMyRMj8P4mZzFSVVwqSXG8GG7MlVwiToYEQwVD7noBxMhNfkeZYtYFtbgOBUSIGtIQjhNHCEnPJsadhb3yBmRIOD3TeAtNLSaU5GgvKUBWSNuuOIHmVt0YhW6rsmDMDUjeUJGJ64R1Jm5lrh0Aa0tKjhFwPdWcGogxLDSXPWQUWTM8Sd3Qz1HMYNxx3HMeiNqNo4jeRDEfZ3gUSHIcU/heomq0vEzL1Msz5KKGxH8FrNOYw3KaxdqaEmNHYMxJFgQbR0DyRknL2L4kwUSxKRdhjRpEtUqilVfggFL1klaMS3PPRDfNqbBOPWO7m4JTVGhS9QTBDDJaEbLbrUQNB+IpJSKQbG5SZZ5gkwJEhJ3aYKJipZ/i7kinChIOW2lQg"
+default_cpu_parallel_kernel_code_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "quantization_kernels_parallel.c")
+default_cpu_parallel_kernel_code = "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"
+
+cpu_kernels = None
+
+
+class CPUKernel:
+    def __init__(self, kernel_file="", source_code=default_cpu_kernel_code_path, compile_parallel_kernel=None, parallel_num=None):
+        self.load =False
+        self.int8WeightExtractionFloat = None
+        self.int4WeightExtractionFloat = None
+        self.int4WeightCompression = None
+        self.SetNumThreads = None
+
+        try:
+            if not os.path.exists(default_cpu_kernel_code_path):
+                with open(default_cpu_kernel_code_path, "w", encoding="utf-8") as file:
+                    code = default_cpu_kernel_code
+                    cpu_quantization_code = bz2.decompress(base64.b64decode(code)).decode()
+                    file.write(cpu_quantization_code)
+
+            if not os.path.exists(default_cpu_parallel_kernel_code_path):
+                with open(default_cpu_parallel_kernel_code_path, "w", encoding="utf-8") as file:
+                    code = default_cpu_parallel_kernel_code
+                    cpu_quantization_code = bz2.decompress(base64.b64decode(code)).decode()
+                    file.write(cpu_quantization_code)
+
+        except Exception as ex:
+            print("Error when generating default cpu kernel code(can be ignored when using custom kernels).")
+
+        if compile_parallel_kernel is None:
+            compile_parallel_kernel = bool(int(os.cpu_count()) >= 4)
+
+        if compile_parallel_kernel and source_code == default_cpu_kernel_code_path:
+            source_code = default_cpu_parallel_kernel_code_path
+
+        if (not kernel_file) or (not os.path.exists(kernel_file)):
+            print("No compiled kernel found.")
+            try:
+                if os.path.exists(source_code):
+                    print("Compiling kernels :", source_code)
+                    kernel_file = source_code[:-2] + ".so"
+                    if compile_parallel_kernel:
+                        compile_command = "gcc -O3 -fPIC -pthread -fopenmp -std=c99 {} -shared -o {}".format(source_code, kernel_file)
+                        print("Compiling", compile_command)
+                        exit_state = os.system(compile_command)
+                        if exit_state:
+                            print("Compile failed, using default cpu kernel code.")
+                            compile_parallel_kernel = False
+                            source_code = default_cpu_kernel_code_path
+                            kernel_file = source_code[:-2] + ".so"
+                            compile_command = "gcc -O3 -fPIC -std=c99 {} -shared -o {}".format(source_code, kernel_file)
+                            print("Compiling", compile_command)
+                    else:
+                        compile_command = "gcc -O3 -fPIC -std=c99 {} -shared -o {}".format(source_code, kernel_file)
+                        print("Compiling", compile_command)
+                        exit_state = os.system(compile_command)
+
+                    print("Kernels compiled :", kernel_file)
+                else:
+                    print("Kernel source code not found.")
+                    return
+            except:
+                print("Failed to build kernel.")
+                return
+        if kernel_file:
+            kernels = ctypes.cdll.LoadLibrary(kernel_file)
+            self.int8WeightExtractionFloat = kernels.extract_int8_weight_to_float
+            self.int4WeightExtractionFloat = kernels.extract_int4_weight_to_float
+            self.int4WeightCompression = kernels.compress_int4_weight
+            if compile_parallel_kernel:
+                try:
+                    self.SetNumThreads = kernels.set_num_threads
+                except:
+                    print("No set_num_threads() found in kernel.")
+                    self.SetNumThreads = lambda x: x
+            self.load = True
+            print("Load kernel :", kernel_file)
+        else:
+            print("Failed to load kernel.")
+        
+        if compile_parallel_kernel:
+            if parallel_num is None:
+                parallel_num = max(os.cpu_count() // 2, 1)
+            print("Setting CPU quantization kernel threads to", parallel_num)
+            if parallel_num < 4:
+                print("Parallel kernel is not recommended when parallel num < 4.")
+            self.SetNumThreads(parallel_num)
+        
+        self.parallel_num = parallel_num
+
+
+def compress_int4_weight(weight: torch.Tensor):  # (n, m)
+    """compress weight on cpu or cuda to int4"""
+    if weight.device == torch.device("cpu"):
+        assert isinstance(cpu_kernels, CPUKernel)
+        n, m = weight.size(0), weight.size(1)
+        assert m % 2 == 0
+        m = m // 2
+        out = torch.empty(n, m, dtype=torch.int8, device="cpu")
+        cpu_kernels.int4WeightCompression(
+            ctypes.c_void_p(weight.data_ptr()),
+            ctypes.c_void_p(out.data_ptr()),
+            ctypes.c_int32(n),
+            ctypes.c_int32(m)
+        )
+        return out
+    else:
+        with torch.cuda.device(weight.device):
+            n, m = weight.size(0), weight.size(1)
+            assert m % 2 == 0
+            m = m // 2
+            out = torch.empty(n, m, dtype=torch.int8, device="cuda")
+            stream = torch.cuda.current_stream()
+
+            gridDim = (n, 1, 1)
+            blockDim = (min(round_up(m, 32), 1024), 1, 1)
+
+            kernels.int4WeightCompression(
+                gridDim,
+                blockDim,
+                0,
+                stream,
+                [ctypes.c_void_p(weight.data_ptr()), ctypes.c_void_p(out.data_ptr()), ctypes.c_int32(n), ctypes.c_int32(m)],
+            )
+            return out
+
+
+def extract_weight_to_half(weight: torch.Tensor, scale_list: torch.Tensor, source_bit_width: int):
+    if source_bit_width == 8:
+        func = kernels.int8WeightExtractionHalf
+    elif source_bit_width == 4:
+        func = kernels.int4WeightExtractionHalf
+    else:
+        assert False, "Unsupported bit-width"
+
+    with torch.cuda.device(weight.device):
+        n, m = weight.size(0), weight.size(1)
+        out = torch.empty(n, m * (8 // source_bit_width), dtype=torch.half, device="cuda")
+        stream = torch.cuda.current_stream()
+
+        gridDim = (n, 1, 1)
+        blockDim = (min(round_up(m, 32), 1024), 1, 1)
+
+        func(
+            gridDim,
+            blockDim,
+            0,
+            stream,
+            [
+                ctypes.c_void_p(weight.data_ptr()),
+                ctypes.c_void_p(scale_list.data_ptr()),
+                ctypes.c_void_p(out.data_ptr()),
+                ctypes.c_int32(n),
+                ctypes.c_int32(m),
+            ],
+        )
+        return out
+
+
+def extract_weight_to_float(weight: torch.Tensor, scale_list: torch.Tensor, source_bit_width: int, quantization_cache=None):
+    """extract weight on cpu to float32"""
+    if source_bit_width == 8:
+        func = cpu_kernels.int8WeightExtractionFloat
+    elif source_bit_width == 4:
+        func = cpu_kernels.int4WeightExtractionFloat
+    else:
+        assert False, "Unsupported bit-width"
+
+    n, m = weight.size(0), weight.size(1)
+
+    if quantization_cache is not None:
+        out = quantization_cache
+        func(
+            ctypes.c_void_p(weight.data_ptr()),
+            ctypes.c_void_p(scale_list.data_ptr()),
+            ctypes.c_void_p(out.data_ptr()),
+            ctypes.c_int32(n),
+            ctypes.c_int32(m)
+        )
+        return out.tensor
+    else:
+        out = torch.empty(n, m * (8 // source_bit_width), dtype=torch.float, device="cpu")
+        func(
+            ctypes.c_void_p(weight.data_ptr()),
+            ctypes.c_void_p(scale_list.data_ptr()),
+            ctypes.c_void_p(out.data_ptr()),
+            ctypes.c_int32(n),
+            ctypes.c_int32(m)
+        )
+        return out
+
+
+class CacheTensor():
+    def __init__(self, *args, **kwargs):
+        self.tensor = torch.empty(*args, **kwargs)
+    
+    def to(self, *args, **kwargs):
+        self.tensor = self.tensor.to(*args, **kwargs)
+        
+    def data_ptr(self):
+        return self.tensor.data_ptr()
+
+
+class QuantizedLinear(Linear):
+    def __init__(self, weight_bit_width: int, weight_tensor=None, bias_tensor=None, quantized_weight=None, quantized_weight_scale=None, quantization_cache=None, empty_init=False, *args, **kwargs):
+        super(QuantizedLinear, self).__init__(*args, **kwargs)
+        self.weight_bit_width = weight_bit_width
+        self.quantization_cache = quantization_cache
+
+        if (quantized_weight is not None) and (quantized_weight_scale is not None):
+            del self.weight
+            self.weight = Parameter(quantized_weight.to(kwargs["device"]), requires_grad=False)
+            self.weight_scale = Parameter(quantized_weight_scale.to(kwargs["device"]), requires_grad=False)
+        else:
+            shape = self.weight.shape
+            del self.weight
+
+            if weight_tensor is None or empty_init:
+                self.weight = torch.empty(
+                    shape[0], shape[1] * weight_bit_width // 8, dtype=torch.int8, device=kwargs["device"]
+                )
+                self.weight_scale = torch.empty(shape[0], dtype=kwargs["dtype"], device=kwargs["device"])
+            else:
+                self.weight_scale = (weight_tensor.abs().max(dim=-1).values / ((2 ** (weight_bit_width - 1)) - 1)).to(kwargs["dtype"])
+                self.weight = torch.round(weight_tensor / self.weight_scale[:, None]).to(torch.int8)
+                if weight_bit_width == 4:
+                    self.weight = compress_int4_weight(self.weight)
+
+            self.weight = Parameter(self.weight.to(kwargs["device"]), requires_grad=False)
+            self.weight_scale = Parameter(self.weight_scale.to(kwargs["device"]), requires_grad=False)
+
+        if bias_tensor is not None:
+            self.bias = Parameter(bias_tensor.to(kwargs["device"]), requires_grad=False)
+        else:
+            self.bias = None
+
+    def reset_parameters(self):
+        """To accelerate initialization"""
+        pass
+
+    def forward(self, input):
+        if self.weight.device == torch.device("cpu"):
+            output = W8A16LinearCPU.apply(input, self.weight, self.weight_scale, self.weight_bit_width, self.quantization_cache)
+        else:
+            output = W8A16Linear.apply(input, self.weight, self.weight_scale, self.weight_bit_width)
+        if self.bias is not None:
+            output = output + self.bias
+        return output
+
+    def _apply(self, fn):
+        self_obj = super()._apply(fn)
+        if self.quantization_cache is not None:
+            self.quantization_cache.to(self_obj.weight.device)
+            self.quantization_cache.to(self_obj.weight_scale.dtype)
+        return self_obj
+
+
+class QuantizedEmbedding(Embedding):  # TODO: backward, check empty_init
+    def __init__(self, weight_bit_width: int, weight_tensor=None, quantized_weight=None, quantized_weight_scale=None, empty_init=False, *args, **kwargs):
+        super(QuantizedEmbedding, self).__init__(*args, **kwargs)
+        self.weight_bit_width = weight_bit_width
+
+        if (quantized_weight is not None) and (quantized_weight_scale is not None):
+            del self.weight
+            self.weight = Parameter(quantized_weight.to(kwargs["device"]), requires_grad=False)
+            self.weight_scale = Parameter(quantized_weight_scale.to(kwargs["device"]), requires_grad=False)
+        else:
+            shape = self.weight.shape
+            del self.weight
+
+            if weight_tensor is None or empty_init:
+                self.weight = torch.empty(
+                    shape[0], shape[1] * weight_bit_width // 8, dtype=torch.int8, device=kwargs["device"]
+                )
+                self.weight_scale = torch.empty(shape[0], dtype=kwargs["dtype"], device=kwargs["device"])
+            else:
+                self.weight_scale = (weight_tensor.abs().max(dim=-1).values / ((2 ** (weight_bit_width - 1)) - 1)).half()
+                self.weight = torch.round(weight_tensor / self.weight_scale[:, None]).to(torch.int8)
+                if weight_bit_width == 4:
+                    self.weight = compress_int4_weight(self.weight)
+
+            self.weight = Parameter(self.weight.to(kwargs["device"]), requires_grad=False)
+            self.weight_scale = Parameter(self.weight_scale.to(kwargs["device"]), requires_grad=False)
+
+    def forward(self, input):
+        if self.weight.device == torch.device("cpu"):
+            original_weight = extract_weight_to_float(weight=self.weight, scale_list=self.weight_scale, source_bit_width=self.weight_bit_width)
+        else:
+            original_weight = extract_weight_to_half(weight=self.weight, scale_list=self.weight_scale, source_bit_width=self.weight_bit_width)
+        output = F.embedding(
+            input, original_weight, self.padding_idx, self.max_norm,
+            self.norm_type, self.scale_grad_by_freq, self.sparse
+        )
+        return output
+
+
+def load_cpu_kernel(**kwargs):
+    global cpu_kernels
+    cpu_kernels = CPUKernel(**kwargs)
+    assert cpu_kernels.load
+
+
+def quantize(model, weight_bit_width, use_quantization_cache=False, empty_init=False, **kwargs):
+    """Replace fp16 linear with quantized linear"""
+    
+    query_key_value_quantization_cache = None
+    dense_quantization_cache = None
+    dense_h_to_4h_quantization_cache = None
+    dense_4h_to_h_quantization_cache = None
+
+    try:
+        load_cpu_kernel(**kwargs)
+    except:
+        print("Cannot load cpu kernel, don't use quantized model on cpu.")
+        if kernels is None:  # CUDA kernels failed
+            print("Cannot load cuda kernel, quantization failed.")
+            return model
+
+    current_device = model.device
+
+    if model.device == torch.device("cpu"):
+        dtype=torch.float32
+    else:
+        dtype = torch.half
+
+    QuantizedLinearWithPara = partial(
+        QuantizedLinear,
+        weight_bit_width=weight_bit_width,
+        bias=True,
+        dtype=dtype,
+        empty_init=empty_init
+    )
+
+    if use_quantization_cache:
+        print("Using quantization cache")
+        layer = model.layers[0]
+        weight = layer.attention.query_key_value.weight
+        n, m = weight.size(0), weight.size(1)
+        query_key_value_quantization_cache = CacheTensor(n, m, dtype=dtype, device=current_device, requires_grad=False)
+        weight = layer.attention.dense.weight
+        n, m = weight.size(0), weight.size(1)
+        dense_quantization_cache = CacheTensor(n, m, dtype=dtype, device=current_device, requires_grad=False)
+        weight = layer.mlp.dense_h_to_4h.weight
+        n, m = weight.size(0), weight.size(1)
+        dense_h_to_4h_quantization_cache = CacheTensor(n, m, dtype=dtype, device=current_device, requires_grad=False)
+        weight = layer.mlp.dense_4h_to_h.weight
+        n, m = weight.size(0), weight.size(1)
+        dense_4h_to_h_quantization_cache = CacheTensor(n, m, dtype=dtype, device=current_device, requires_grad=False)
+
+    print("Applying quantization to glm layers")
+
+    for layer in model.layers:
+        layer.attention.query_key_value = QuantizedLinearWithPara(
+            weight_tensor=layer.attention.query_key_value.weight.to(current_device),
+            bias_tensor=layer.attention.query_key_value.bias,
+            in_features=layer.attention.query_key_value.in_features,
+            out_features=layer.attention.query_key_value.out_features,
+            device=layer.attention.query_key_value.weight.device,
+            quantization_cache=query_key_value_quantization_cache
+        )
+        layer.attention.dense = QuantizedLinearWithPara(
+            weight_tensor=layer.attention.dense.weight.to(current_device),
+            bias_tensor=layer.attention.dense.bias,
+            in_features=layer.attention.dense.in_features,
+            out_features=layer.attention.dense.out_features,
+            device=layer.attention.dense.weight.device,
+            quantization_cache=dense_quantization_cache
+        )
+        layer.mlp.dense_h_to_4h = QuantizedLinearWithPara(
+            weight_tensor=layer.mlp.dense_h_to_4h.weight.to(current_device),
+            bias_tensor=layer.mlp.dense_h_to_4h.bias,
+            in_features=layer.mlp.dense_h_to_4h.in_features,
+            out_features=layer.mlp.dense_h_to_4h.out_features,
+            device=layer.mlp.dense_h_to_4h.weight.device,
+            quantization_cache=dense_h_to_4h_quantization_cache
+        )
+        layer.mlp.dense_4h_to_h = QuantizedLinearWithPara(
+            weight_tensor=layer.mlp.dense_4h_to_h.weight.to(current_device),
+            bias_tensor=layer.mlp.dense_4h_to_h.bias,
+            in_features=layer.mlp.dense_4h_to_h.in_features,
+            out_features=layer.mlp.dense_4h_to_h.out_features,
+            device=layer.mlp.dense_4h_to_h.weight.device,
+            quantization_cache=dense_4h_to_h_quantization_cache
+        )
+    return model

quantization_kernels.c

@@ -0,0 +1,34 @@
+void compress_int4_weight(void *weight, void *out, int n, int m)
+{
+    for(int i=0;i<n*m;i++)
+    {
+        (*(unsigned char*)(out)) = ((*(unsigned char*)(weight)) << 4);
+        weight += sizeof(char);
+        (*(unsigned char*)(out)) |= ((*(unsigned char*)(weight)) & 15);
+        weight += sizeof(char);
+        out += sizeof(char);
+    }
+}
+
+void extract_int8_weight_to_float(void *weight, void *scale_list, void *out, int n, int m)
+{
+	for(int i=0;i<n;i++)
+        for(int j=0;j<m;j++)
+            (*(float*)(out + sizeof(float) * (i * m + j))) = (*(float*)(scale_list + sizeof(float) * i)) * (*(char*)(weight + sizeof(char) * (i * m + j)));
+}
+
+void extract_int4_weight_to_float(void *weight, void *scale_list, void *out, int n, int m)
+{
+	for(int i=0;i<n;i++)
+    {
+        for(int j=0;j<m;j++)
+        {
+            (*(float*)(out)) = (*(float*)(scale_list)) * ((*(char*)(weight)) >> 4);
+            out += sizeof(float);
+            (*(float*)(out)) = (*(float*)(scale_list)) * (((char)((*(unsigned char*)(weight)) << 4))>> 4);
+            out += sizeof(float);
+            weight += sizeof(char);
+        }
+        scale_list += sizeof(float);
+    }
+}

quantization_kernels_parallel.c

@@ -0,0 +1,47 @@
+#include <omp.h>
+
+void set_num_threads(int n_threads)
+{
+    omp_set_num_threads(n_threads);
+}
+
+int get_num_threads()
+{
+    return omp_get_num_threads();
+}
+
+void compress_int4_weight(void *weight, void *out, int n, int m)
+{
+    #pragma omp parallel for
+	for(int i=0;i<n;i++)
+    {
+        for(int j=0;j<m;j++)
+        {
+            (*(unsigned char*)(out + sizeof(unsigned char) * (i * m + j))) |= ((*(unsigned char*)(weight + sizeof(unsigned char) * (i * (m << 1) + (j << 1)))) << 4);
+            (*(unsigned char*)(out + sizeof(unsigned char) * (i * m + j))) |= (((*(unsigned char*)(weight + sizeof(unsigned char) * (i * (m << 1) + ((j << 1) | 1)))) & 15));
+        }
+    }
+}
+
+void extract_int8_weight_to_float(void *weight, void *scale_list, void *out, int n, int m)
+{
+    #pragma omp parallel for
+	for(int i=0;i<n;i++)
+    {
+        for(int j=0;j<m;j++)
+            (*(float*)(out + sizeof(float) * (i * m + j))) = (*(float*)(scale_list + sizeof(float) * i)) * (*(char*)(weight + sizeof(char) * (i * m + j)));
+    }
+}
+
+void extract_int4_weight_to_float(void *weight, void *scale_list, void *out, int n, int m)
+{
+    #pragma omp parallel for
+	for(int i=0;i<n;i++)
+    {
+        for(int j=0;j<m;j++)
+        {
+            (*(float*)(out + sizeof(float) * (i * (m << 1) + (j << 1)))) = (*(float*)(scale_list + sizeof(float) * i)) * ((*(char*)(weight + sizeof(char) * (i * m + j))) >> 4);
+            (*(float*)(out + sizeof(float) * (i * (m << 1) + ((j << 1) | 1)))) = (*(float*)(scale_list + sizeof(float) * i)) * (((char)((*(unsigned char*)(weight + sizeof(char) * (i * m + j))) << 4))>> 4);
+        }
+    }
+}