first

models/12B-V2/tokenizer_config.json

+{
+  "name_or_path": "ChinaTelecom/telechat-12b",
+  "tokenizer_class": "TelechatTokenizer",
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_telechat.TelechatTokenizer",
+      null
+    ]
+  },
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "use_fast": false,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "<_end>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": true
+  },
+  "model_max_length": 100000000,
+  "sp_model_kwargs": {},
+  "pad_token": {
+    "__type": "AddedToken",
+    "content": "<_pad>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": true
+  },
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<_end>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": true
+  }
+}

models/12B-V2_8bit/config.json

+{
+  "apply_residual_connection_post_layernorm": false,
+  "architectures": [
+    "TelechatForCausalLM"
+  ],
+  "auto_map": {
+	"AutoConfig": "configuration_telechat.TelechatConfig",
+	"AutoModelForCausalLM": "modeling_telechat.TelechatForCausalLM"
+  },
+  "attention_dropout": 0.0,
+  "attention_softmax_in_fp32": true,
+  "bias_dropout_fusion": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_dropout": 0.0,
+  "hidden_size": 5120,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "masked_softmax_fusion": true,
+  "model_type": "telechat",
+  "n_head": 32,
+  "n_inner": null,
+  "n_layer": 38,
+  "offset_alibi": 100,
+  "pad_token_id": 3,
+  "pretraining_tp": 2,
+  "seq_length": 8192,
+  "skip_bias_add": true,
+  "skip_bias_add_qkv": false,
+  "slow_but_exact": false,
+  "transformers_version": "4.24.0",
+  "unk_token_id": 0,
+  "use_cache": true,
+  "vocab_size": 120000,
+  "ffn_hidden_size": 12288,
+  "flash_attn":true,
+  "tie_word_embeddings":false,
+  "training_seqlen":8192,
+  "logn":false,
+  "semi_causal":false,
+  "embed_layernorm":false
+}
+

models/12B-V2_8bit/configuration_telechat.py

+# coding=utf-8
+# Copyright 2022 the Big Science Workshop and HuggingFace Inc. team.  All rights reserved.
+#
+# 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.
+
+""" Telechat configuration"""
+
+from packaging import version
+from collections import OrderedDict
+from transformers.utils import is_torch_available, logging
+from transformers.configuration_utils import PretrainedConfig
+from typing import TYPE_CHECKING, Any, List, Mapping, Optional
+
+logger = logging.get_logger(__name__)
+
+class TelechatConfig(PretrainedConfig):
+    """
+    Args:
+        vocab_size (`int`, *optional*, defaults to 160256): Vocabulary size of the Telechat model.
+        hidden_size (`int`, *optional*, defaults to 4096): Dimensionality of the embeddings and hidden states.
+        ffn_hidden_size (`int`, *optional*, defaults to 12288): Dimensionality of the feed-forward hidden states.
+        n_layer (`int`, *optional*, defaults to 30): Number of hidden layers in the Transformer
+        n_head (`int`, *optional*, defaults to 32): Number of attention heads for each attention layer.
+        layer_norm_epsilon (`float`, *optional*, defaults to 1e-5): The epsilon to use in the layer normalization layers.
+        initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        apply_residual_connection_post_layernorm (`bool`, *optional*, defaults to `False`): If enabled, use the layer norm of the hidden states as the residual in the transformer blocks
+        hidden_dropout (`float`, *optional*, defaults to 0.0): Dropout rate of the dropout function on the bias dropout.
+        attention_dropout (`float`, *optional*, defaults to 0.0): Dropout rate applied to the attention probs
+        use_cache (`bool`, *optional*, defaults to `True`): Whether or not the model should return the last key/values attentions.
+        training_seqlen (`int`, *optional*, defaults to 8192): Sequence length during last finetuning.
+        logn (`bool`, *optional*, defaults to `True`): Whether or not to use logN during extrapolation.
+        embed_layernorm (`bool`, *optional*, defaults to `True`): Whether or not to use embedding layernorm.
+
+    """
+
+    model_type = "telechat"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "num_hidden_layers": "n_layer",
+        "num_attention_heads": "n_head",
+    }
+
+    def __init__(
+        self,
+        vocab_size=160256,
+        hidden_size=4096,
+        n_layer=30,
+        n_head=32,
+        layer_norm_epsilon=1e-5,
+        initializer_range=0.02,
+        use_cache=True,
+        bos_token_id=1,
+        eos_token_id=2,
+        apply_residual_connection_post_layernorm=False,
+        hidden_dropout=0.0,
+        attention_dropout=0.0,
+        ffn_hidden_size=12288,
+        training_seqlen = 8192,
+        logn = True,
+        embed_layernorm = False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        n_embed = kwargs.pop("n_embed", None)
+        self.hidden_size = hidden_size if n_embed is None else n_embed
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.use_cache = use_cache
+        self.apply_residual_connection_post_layernorm = apply_residual_connection_post_layernorm
+        self.hidden_dropout = hidden_dropout
+        self.attention_dropout = attention_dropout
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.logn = logn
+        self.ffn_hidden_size = ffn_hidden_size
+        self.training_seqlen = training_seqlen
+        self.embed_layernorm = embed_layernorm
+
+
+        super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+

models/12B-V2_8bit/generation_config.json

+{
+  "max_length": 8192,
+  "do_sample": false,
+  "use_cache": true,
+  "temperature": 0.3,
+  "top_k": 5,
+  "top_p": 0.85,
+  "repetition_penalty": 1.003,
+  "pad_token_id": 3,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "user_token_id": 20,
+  "bot_token_id": 21
+}

models/12B-V2_8bit/generation_utils.py

+from typing import Optional
+from collections import deque
+from queue import Queue
+import copy
+
+
+class History:
+
+    def __init__(self, tokenizer, history):
+        '''
+        init from a list of dict
+        '''
+        # use deque to meet some special situation
+        self.input_history = deque()
+        self.tokenizer = tokenizer
+        if history:
+            self._transfer_from_list(history)
+
+    def _transfer_from_list(self, history):
+        for message in history:
+            content = message.get("content")
+            # the token result may not be equal to the result model gen
+            message.update(self.tokenizer(content))
+            self.input_history.append(message)
+
+    def append(self, message):
+        content = message.get("content")
+        if "input_ids" not in message or "attention_mask" not in message:
+            message.update(self.tokenizer(content))
+        self.input_history.append(message)
+
+    def append_left(self, message):
+        content = message.get("content")
+        if "input_ids" not in message or "attention_mask" not in message:
+            message.update(self.tokenizer(content))
+        self.input_history.appendleft(message)
+
+    def pop(self):
+        x = self.input_history.pop()
+        return x
+
+    def pop_left(self):
+        x = self.pop_left()
+        return x
+
+    def update(self, message):
+        self.input_history.pop()
+        self.append(message)
+
+    def __len__(self):
+        return self.input_history.__len__()
+
+    def __str__(self):
+        return self.input_history.__str__()
+
+    def __copy__(self):
+        new_instance = type(self)(self.tokenizer, [])
+        new_instance.input_history = copy.copy(self.input_history)
+        return new_instance
+
+    def __deepcopy__(self, memodict={}):
+        new_instance = type(self)(self.tokenizer, [])
+        new_instance.input_history = copy.deepcopy(self.input_history)
+        return new_instance
+
+
+class TelechatIterTextStreamer:
+    """
+    With reference to the TextIterStreamers in transformers, we have rewritten this class
+    """
+
+    def __init__(
+            self, tokenizer, history: History = None, skip_prompt: bool = False, timeout: Optional[float] = None,
+            **decode_kwargs
+    ):
+
+        self.tokenizer = tokenizer
+        self.history = history
+        self.skip_prompt = skip_prompt
+        self.timeout = timeout
+        self.decode_kwargs = decode_kwargs
+
+        self.text_queue = Queue()
+        self.cache_time = 0
+        self.text_until = ""
+        self.token_until = []
+        self.stop_signal = None
+        self.next_tokens_are_prompt = True
+
+        self.history.append({"role": "bot", "content": self.text_until})
+
+    def put(self, value):
+        """
+        put printable text into queue
+        """
+        if len(value.shape) > 1 and value.shape[0] > 1:
+            raise ValueError("TextStreamer only supports batch size 1")
+        elif len(value.shape) > 1:
+            value = value[0]
+
+        if self.skip_prompt and self.next_tokens_are_prompt:
+            self.next_tokens_are_prompt = False
+            return
+
+        if value[-1] == self.tokenizer.eos_token_id:
+            return
+
+        # there may be some smart way to decode.
+        self.token_until.extend(value.tolist())
+        text = self.tokenizer.decode(self.token_until, **self.decode_kwargs)
+
+
+        if self._is_printable(text) or self.cache_time >= 6:
+            output_text = text[len(self.text_until):]
+            self.text_until = text
+
+        else:
+            self.cache_time+=1
+            return
+
+        self.on_finalized_text(output_text)
+
+    def end(self):
+        """Flushes any remaining cache and prints a newline to stdout."""
+        # Flush the cache, if it exists
+        text = self.tokenizer.decode(self.token_until, **self.decode_kwargs)
+        output_text = text[len(self.text_until):]
+        self.text_until = text
+        self.on_finalized_text(output_text, stream_end=True)
+        self.clear_cache()
+
+    def clear_cache(self):
+        self.cache_time = 0
+        self.token_until = []
+        self.text_until = ""
+        self.history = None
+        self.next_tokens_are_prompt = True
+
+    def on_finalized_text(self, text: str, stream_end: bool = False):
+        """Put the text tuple in the queue."""
+        self.history.update({"role": "bot", "content": self.text_until, "input_ids": self.token_until,
+                             "attention_mask": [1] * len(self.token_until)})
+        self.text_queue.put((text, self.history), timeout=self.timeout)
+        if stream_end:
+            self.text_queue.put((self.stop_signal, self.history), timeout=self.timeout)
+
+    @staticmethod
+    def _is_printable(cp):
+        """Checks whether tokens can be decoded or not"""
+        if "�" in cp:
+            return False
+        return True
+
+    def __iter__(self):
+        return self
+
+    def __next__(self):
+        value_now, history_until = self.text_queue.get(timeout=self.timeout)
+        if value_now == self.stop_signal:
+            raise StopIteration()
+        else:
+            return value_now, history_until

models/12B-V2_8bit/modeling_telechat.py

+# coding=utf-8
+# Copyright 2022 HuggingFace Inc. team and BigScience workshop.
+#
+# 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.
+
+# Copyright (c) 2023, NVIDIA CORPORATION.  All rights reserved.
+
+# Copyright (c) 2021 EleutherAI
+# This file is based on code by the authors denoted below and has been modified from its original version.
+#
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# 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.
+
+
+"""PyTorch TELECHAT model."""
+
+import warnings
+from typing import Optional, Tuple, Union, List, Dict
+from threading import Thread
+
+import torch
+import math
+import copy
+from torch import nn
+import torch.utils.checkpoint
+from torch.nn import functional as F
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, LayerNorm, MSELoss
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from transformers import GenerationConfig
+
+from .configuration_telechat import TelechatConfig
+from .generation_utils import History, TelechatIterTextStreamer
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "telechat"
+_CONFIG_FOR_DOC = "TelechatConfig"
+
+TELECHAT_PRETRAINED_MODEL_ARCHIVE_LIST = []
+
+try:
+    from einops import rearrange
+except ImportError:
+    rearrange = None
+
+use_flash_attn = True
+try:
+    from flash_attn.flash_attn_interface import flash_attn_unpadded_func
+except ImportError:
+    try:
+        from flash_attn.flash_attn_interface import flash_attn_varlen_func as flash_attn_unpadded_func
+    except ImportError:
+        flash_attn_unpadded_func = None
+
+
+class RotaryEmbedding(torch.nn.Module):
+    # Extracted from: https://github.com/EleutherAI/gpt-neox
+    def __init__(self, dim, config, base=10000,precision=torch.half):
+        super().__init__()
+        self.config = config
+        self.dim = dim
+        self.base = base
+        self.max_seq_len_cached = None
+        self.cos_cached = None
+        self.sin_cached = None
+        self.precision = precision
+
+    def get_mscale(self, scale=1):
+        if scale <= 1:
+            return 1.0
+        return 0.1 * math.log(scale) + 1.0
+
+    def get_ntk_alpha(self, true_seq_len):
+        context_value = math.log(true_seq_len / 8192, 2) + 1
+        # ntk_alpha = 2 ** context_value - 1
+        ntk_alpha = 2 ** math.ceil(context_value) - 1
+        ntk_alpha = max(ntk_alpha, 1)
+        return ntk_alpha
+
+    def forward(self, x, seq_dim=0, seq_len=None):
+        if seq_len is None:
+            seq_len = x.shape[seq_dim]
+        seq_len = max(seq_len, self.config.training_seqlen)
+        ntk_alpha = self.get_ntk_alpha(seq_len)
+        mscale = float(self.get_mscale(seq_len / self.config.training_seqlen))
+        base = self.base * ntk_alpha ** (self.dim / (self.dim - 2))
+        inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2, device=x.device).float( )/ self.dim ))
+        max_seq_len_cached = seq_len
+        t = torch.arange(max_seq_len_cached, device=x.device, dtype=inv_freq.dtype)
+        freqs = torch.einsum('i,j->ij', t, 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 = mscale *emb.cos()[:, None, :].half()
+        sin_cached = mscale *emb.sin()[:, None, :].half()
+        if self.precision == torch.bfloat16:
+            cos_cached = cos_cached.bfloat16()
+            sin_cached = sin_cached.bfloat16()
+        return cos_cached[:seq_len, ...], sin_cached[:seq_len, ...]
+
+
+# rotary pos emb helpers:
+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
+
+
+def apply_rotary_pos_emb_torch(q, k, cos, sin, offset: int = 0):  # jitting fails with bf16
+    cos, sin = cos[offset:q.shape[0] + offset, ...], sin[offset:q.shape[0] + offset, ...]
+    return (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin)
+
+
+class MixedFusedRMSNorm(nn.Module):
+    # Extracted from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py
+    def __init__(self, hidden_size, eps=1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+class FlashSelfAttention(torch.nn.Module):
+    # Extracted from https://github.com/microsoft/Megatron-DeepSpeed/blob/main/megatron/model/transformer.py
+    """Implement the scaled dot product attention with softmax.
+    Arguments
+    ---------
+        softmax_scale: The temperature to use for the softmax attention.
+                      (default: 1/sqrt(d_keys) where d_keys is computed at
+                      runtime)
+        attention_dropout: The dropout rate to apply to the attention
+                           (default: 0.0)
+    """
+
+    def __init__(self, causal=False, softmax_scale=None, attention_dropout=0.0,
+                 device=None, dtype=None):
+        super().__init__()
+        assert flash_attn_unpadded_func is not None, ('Please install FlashAttention first, '
+                                                      'e.g., with pip install flash-attn')
+        assert rearrange is not None, 'Please install einops first, e.g., with pip install einops'
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.dropout_p = attention_dropout
+
+    def forward(self, q, k, v):
+        """Implements the multihead softmax attention.
+        Arguments
+        ---------
+            q, k, v: The tensor containing the query, key, and value. (B, S, H, D)
+        """
+        assert all((i.dtype in [torch.float16, torch.bfloat16] for i in (q, k, v)))
+        assert all((i.is_cuda for i in (q, k, v)))
+
+        batch_size, seqlen_q = q.shape[0], q.shape[1]
+        seqlen_k = k.shape[1]
+
+        q, k, v = [rearrange(x, 'b s ... -> (b s) ...') for x in [q, k, v]]
+        cu_seqlens_q = torch.arange(0, (batch_size + 1) * seqlen_q, step=seqlen_q, dtype=torch.int32,
+                                    device=q.device)
+        if self.training:
+            # during training q,k,v always have same seqlen
+            assert seqlen_k == seqlen_q
+
+            is_causal = self.causal
+            cu_seqlens_k = cu_seqlens_q
+            dropout_p = self.dropout_p
+        else:
+            # turn off FA causal mask after first inference autoregressive iteration
+            # only on first autoregressive step q,k,v have same seqlen
+            is_causal = seqlen_q == seqlen_k
+            cu_seqlens_k = torch.arange(0, (batch_size + 1) * seqlen_k, step=seqlen_k, dtype=torch.int32,
+                                        device=q.device)
+            dropout_p = 0
+
+        output = flash_attn_unpadded_func(
+            q, k, v, cu_seqlens_q, cu_seqlens_k, seqlen_q, seqlen_k,
+            dropout_p=dropout_p,
+            softmax_scale=self.softmax_scale, causal=is_causal
+        )
+
+        output = rearrange(output, '(b s) ... -> b s ...', b=batch_size)
+        return output
+
+
+def _make_causal_mask(
+        input_ids_shape: torch.Size, device: torch.device, past_key_values_length: int
+) -> torch.BoolTensor:
+    """
+    Make causal mask used for self-attention.
+    """
+    batch_size, target_length = input_ids_shape
+    mask = torch.empty((target_length, target_length + past_key_values_length), dtype=torch.bool, device=device)
+    # ONNX doesn't support `torch.Tensor.triu` properly, thus we use this workaround
+    seq_ids = torch.arange(target_length, device=device)
+    mask[:, past_key_values_length:] = seq_ids[:, None] < seq_ids[None, :]
+
+    if past_key_values_length > 0:
+        mask[:, :past_key_values_length] = False
+
+    expanded_mask = mask[None, None, :, :].expand(batch_size, 1, target_length, target_length + past_key_values_length)
+    return expanded_mask
+
+
+def _expand_mask(mask: torch.Tensor, tgt_length: int) -> torch.BoolTensor:
+    """
+    Expands attention_mask from `[batch_size, src_length]` to `[batch_size, 1, tgt_length, src_length]`.
+    """
+    batch_size, src_length = mask.shape
+    tgt_length = tgt_length if tgt_length is not None else src_length
+
+    expanded_mask = ~(mask[:, None, None, :].to(torch.bool))
+    return expanded_mask.expand(batch_size, 1, tgt_length, src_length)
+
+
+def dropout_add(x: torch.Tensor, residual: torch.Tensor, prob: float, training: bool) -> torch.Tensor:
+    """
+    Dropout add function
+
+    Args:
+        x (`torch.tensor`, *required*):
+            input tensor
+        residual (`torch.tensor`, *required*):
+            residual tensor
+        prob (`float`, *required*):
+            dropout probability
+        training (`bool`, *required*):
+            training mode
+    """
+    out = F.dropout(x, p=prob, training=training)
+    out = residual + out
+    return out
+
+
+def telechat_gelu_forward(x: torch.Tensor) -> torch.Tensor:
+    """
+    Custom bias GELU function. Adapted from Megatron-DeepSpeed code. Here we use a simple implementation (inference) to
+    make the model jitable.
+
+    Args:
+        x (`torch.tensor`, *required*):
+            input hidden states
+    """
+    return x * 0.5 * (1.0 + torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x)))
+
+
+def telechat_gelu_back(g: torch.Tensor, x: torch.Tensor) -> torch.Tensor:
+    """
+    gradient of tanh approximation of gelu gradient of actual gelu is: 0.5 * (1. + torch.erf(x * 0.70710678)) +
+    0.3989423 * x * torch.exp(-0.5 * x * x)
+
+    Args:
+        g (`torch.tensor`, *required*):
+            gradient output tensor
+        x (`torch.tensor`, *required*):
+            input tensor
+    """
+    x = x[0]  # x is a tuple of 1 element, needs to unpack it first
+    tanh_out = torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x))
+    # sqrt(2/pi) * 3 * 0.044715 -> 0.1070322243
+    ff = 0.5 * x * ((1 - tanh_out * tanh_out) * (0.79788456 + 0.1070322243 * x * x)) + 0.5 * (1 + tanh_out)
+    return ff * g
+
+
+class GeLUFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, input: torch.Tensor) -> torch.Tensor:
+        ctx.save_for_backward(input)
+        return telechat_gelu_forward(input)
+
+    @staticmethod
+    def backward(ctx, grad_output: torch.Tensor) -> torch.Tensor:
+        input = ctx.saved_tensors
+        tmp = telechat_gelu_back(grad_output, input)
+        return tmp
+
+
+class TelechatGelu(nn.Module):
+    """
+    TelechatBiasGelu wrapper function that make use of the simple function on inference mode to make the model
+    torchscriptable and use the autograd function in training mode to get the accurate results of the gradients Partly
+    copied from Megatron-DeepSpeed code and adapted for our needs
+
+    See here why autograd functions are not torchscriptable: https://github.com/pytorch/pytorch/issues/22329
+    """
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.training:
+            return GeLUFunction.apply(x)
+        else:
+            return telechat_gelu_forward(x)
+
+
+class TelechatAttention(nn.Module):
+    def __init__(self, config: TelechatConfig, layer_idx):
+        super().__init__()
+        self.kv_cache = None
+        self.layer_idx = layer_idx
+
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.n_head
+        self.head_dim = self.hidden_size // self.num_heads
+        self.split_size = self.hidden_size
+        self.hidden_dropout = config.hidden_dropout
+        self.config = config
+
+        if self.head_dim * self.num_heads != self.hidden_size:
+            raise ValueError(
+                f"`hidden_size` must be divisible by num_heads (got `hidden_size`: {self.hidden_size} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+
+        # Layer-wise attention scaling
+        self.inv_norm_factor = 1.0 / math.sqrt(self.head_dim)
+        self.beta = 1.0
+
+        self.num_key_value_heads = self.num_heads
+        kv_projection_size = self.head_dim * self.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.query = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+        self.key_value = nn.Linear(self.hidden_size, kv_projection_size * 2, bias=False)
+        self.dense = nn.Linear(self.hidden_size, self.hidden_size)
+        self.attention_dropout = nn.Dropout(config.attention_dropout)
+        self.rotary_emb = RotaryEmbedding(self.head_dim, config=config)
+
+        self.core_attention_flash = FlashSelfAttention(
+            causal=True, attention_dropout=config.attention_dropout
+        )
+
+        self.last_key_layer = None
+
+    def repeat_kv(self, hidden_states, n_rep):
+        slen, batch, num_key_value_heads_per_partition, head_dim = hidden_states.shape
+        if n_rep == 1:
+            return hidden_states
+        hidden_states = hidden_states[:, :, :, None, :].expand(slen, batch, num_key_value_heads_per_partition, n_rep,
+                                                               head_dim)
+        return hidden_states.reshape(slen, batch, num_key_value_heads_per_partition * n_rep, head_dim)
+
+    def split_tensor_along_last_dim(self,
+                                    tensor: torch.Tensor,
+                                    num_partitions: int,
+                                    contiguous_split_chunks: bool = False,
+                                    ):
+
+        # 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 _merge_heads(self, x: torch.Tensor) -> torch.Tensor:
+        batch_size_and_num_heads, seq_length, _ = x.shape
+        batch_size = batch_size_and_num_heads // self.num_heads
+        x = x.view(batch_size, self.num_heads, seq_length, self.head_dim)
+        x = x.permute(0, 2, 1, 3)
+        return x.reshape(batch_size, seq_length, self.num_heads * self.head_dim)
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            residual: torch.Tensor,
+            attention_mask: torch.Tensor,
+            layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+            use_cache: bool = False,
+            output_attentions: bool = False,
+    ):
+        hidden_states = hidden_states.transpose(1, 0)
+        query_layer = self.query(hidden_states)
+        new_tensor_shape = query_layer.size()[:-1] + \
+                           (self.num_heads,
+                            self.head_dim)
+        query_layer = query_layer.view(*new_tensor_shape)
+
+        mixed_kv_layer = self.key_value(hidden_states)
+        new_tensor_shape = mixed_kv_layer.size()[:-1] + \
+                           (self.num_key_value_heads,
+                            2 * self.head_dim)
+        mixed_kv_layer = mixed_kv_layer.view(*new_tensor_shape)
+        (key_layer, value_layer) = self.split_tensor_along_last_dim(mixed_kv_layer, 2)
+
+        output_size = (query_layer.size(1),
+                       query_layer.size(2),
+                       query_layer.size(0),
+                       key_layer.size(0))
+
+        query_layer = query_layer.view(output_size[2], output_size[0] * output_size[1], -1)
+        key_layer = key_layer.view(output_size[3], output_size[0] * output_size[1], -1)
+
+        apply_rotary_fn = apply_rotary_pos_emb_torch
+
+        seq_len = key_layer.shape[0]
+        offset = 0
+
+        if use_cache and layer_past != None:
+            past_key, past_value = layer_past
+            offset = past_key.shape[0]
+            seq_len += offset
+
+        cos, sin = self.rotary_emb(value_layer)
+
+        query_layer, key_layer = apply_rotary_fn(query_layer, key_layer, cos, sin, offset=offset)
+        if use_cache:
+            if layer_past != None:
+                past_key, past_value = layer_past
+                key_layer = torch.cat((past_key, key_layer[-1, ...].unsqueeze(0)), dim=0)
+                value_layer = torch.cat((past_value, value_layer[-1, ...].unsqueeze(0)), dim=0)
+            layer_past = key_layer, value_layer
+        s, bz, head, dim = value_layer.shape
+        s_key = key_layer.shape[0]
+        s_query = query_layer.shape[0]
+        query_layer = query_layer.reshape((s_query, bz, head, dim))
+        key_layer = key_layer.reshape((s_key, bz, head, dim))
+
+        if self.config.flash_attn:
+            q, k, v = [rearrange(x, 's b ... -> b s ...').contiguous() for x in
+                       (query_layer, key_layer, value_layer)]
+            context_layer = self.core_attention_flash(q, k, v)
+            context_layer = rearrange(context_layer, 'b s h d -> b s (h d)').contiguous()
+        else:
+            ##[sq, b, np, hn] -> [sq, b * np, hn]
+            query_layer = query_layer.reshape(s_query, bz * self.num_heads, dim)
+            # [sk, b, np, hn] -> [sk, b * np, hn]
+            key_layer = key_layer.reshape(s_key, bz * self.num_heads, dim)
+            matmul_result = self.inv_norm_factor * torch.einsum('bik,bkj->bij', query_layer.transpose(0, 1),
+                                                                key_layer.transpose(0, 1).transpose(1, 2))
+
+            attention_scores = matmul_result.view(bz, self.num_heads, s_query, s_key)
+
+            input_dtype = attention_scores.dtype
+            if input_dtype == torch.float16 or input_dtype == torch.bfloat16:
+                attention_scores = attention_scores.to(torch.float)
+            attn_weights = torch.masked_fill(attention_scores, attention_mask, torch.finfo(attention_scores.dtype).min)
+            attention_probs = F.softmax(attn_weights, dim=-1).to(input_dtype)  ##dtype = torch.float32
+            attention_probs = self.attention_dropout(attention_probs)
+            attention_probs_reshaped = attention_probs.view(bz * self.num_heads, s_query, s_key)
+
+            value_layer = value_layer.reshape(s_key, bz * self.num_heads, dim)
+            context_layer = torch.bmm(attention_probs_reshaped, value_layer.transpose(0, 1))
+            context_layer = self._merge_heads(context_layer)
+
+        output_tensor = self.dense(context_layer)
+
+        output_tensor = dropout_add(output_tensor, residual, self.hidden_dropout, self.training)
+        present = None
+        outputs = (output_tensor, present)
+        if output_attentions:
+            outputs += (attention_probs,)
+
+        return output_tensor, layer_past
+
+
+class TelechatMLP(nn.Module):
+    def __init__(self, config: TelechatConfig):
+        super().__init__()
+        hidden_size = config.hidden_size
+        self.gate_proj = nn.Linear(hidden_size, config.ffn_hidden_size, bias=False)
+        self.up_proj = nn.Linear(hidden_size, config.ffn_hidden_size, bias=False)
+        self.down_proj = nn.Linear(config.ffn_hidden_size, hidden_size, bias=True)
+        self.hidden_dropout = config.hidden_dropout
+
+    def forward(self, hidden_states: torch.Tensor, residual: torch.Tensor) -> torch.Tensor:
+        intermediate_output = self.down_proj(F.silu(self.gate_proj(hidden_states)) * self.up_proj(hidden_states))
+        output = dropout_add(intermediate_output, residual, self.hidden_dropout, self.training)
+        return output
+
+
+class TelechatBlock(nn.Module):
+    def __init__(self, config: TelechatConfig, layer_idx):
+        super().__init__()
+        hidden_size = config.hidden_size
+
+        self.input_layernorm = MixedFusedRMSNorm(hidden_size, eps=config.layer_norm_epsilon)
+        self.num_heads = config.n_head
+        self.layer_idx = layer_idx
+        self.self_attention = TelechatAttention(config, layer_idx)
+        self.post_attention_layernorm = MixedFusedRMSNorm(hidden_size, eps=config.layer_norm_epsilon)
+
+        self.mlp = TelechatMLP(config)
+
+        self.apply_residual_connection_post_layernorm = config.apply_residual_connection_post_layernorm
+        self.hidden_dropout = config.hidden_dropout
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            attention_mask: torch.Tensor,
+            layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+            use_cache: bool = False,
+            output_attentions: bool = False,
+    ):
+        layernorm_output = self.input_layernorm(hidden_states)
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = hidden_states
+
+        attn_outputs = self.self_attention(
+            layernorm_output,
+            residual,
+            layer_past=layer_past,
+            attention_mask=attention_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+        )
+
+        attention_output = attn_outputs[0]
+        outputs = attn_outputs[1:]
+        layernorm_output = self.post_attention_layernorm(attention_output)
+
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = attention_output
+        output = self.mlp(layernorm_output, residual)
+
+        if use_cache:
+            outputs = (output,) + outputs
+        else:
+            outputs = (output,) + outputs[1:]
+
+        return outputs
+
+
+class TelechatPreTrainedModel(PreTrainedModel):
+    config_class = TelechatConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["TelechatBlock"]
+    _skip_keys_device_placement = "past_key_values"
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module: nn.Module):
+        """Initialize the weights."""
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+        elif isinstance(module, LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    def _set_gradient_checkpointing(self, module: nn.Module, value: bool = False):
+        if isinstance(module, TelechatModel):
+            module.gradient_checkpointing = value
+
+
+class TelechatModel(TelechatPreTrainedModel):
+    def __init__(self, config: TelechatConfig):
+        super().__init__(config)
+
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.n_head
+        self.config = config
+        self.word_embeddings = nn.Embedding(config.vocab_size, self.embed_dim)
+        if self.config.embed_layernorm:
+            self.word_embeddings_layernorm = MixedFusedRMSNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+
+        self.h = nn.ModuleList([TelechatBlock(config, _) for _ in range(config.num_hidden_layers)])
+        self.ln_f = MixedFusedRMSNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+        self.gradient_checkpointing = False
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.word_embeddings
+
+    def _prepare_attn_mask(
+            self, attention_mask: torch.Tensor, input_shape: Tuple[int, int], past_key_values_length: int
+    ) -> torch.BoolTensor:
+        combined_attention_mask = None
+        device = attention_mask.device
+        _, src_length = input_shape
+
+        if src_length > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape, device=device, past_key_values_length=past_key_values_length
+            )
+        expanded_attn_mask = _expand_mask(attention_mask, tgt_length=src_length)
+        combined_attention_mask = (
+            expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask | combined_attention_mask
+        )
+
+        return combined_attention_mask
+
+    def set_input_embeddings(self, new_embeddings: torch.Tensor):
+        self.word_embeddings = new_embeddings
+
+    def forward(
+            self,
+            input_ids: Optional[torch.LongTensor] = None,
+            past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+            attention_mask: Optional[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,
+            **deprecated_arguments,
+    ) -> Union[Tuple[torch.Tensor, ...], BaseModelOutputWithPastAndCrossAttentions]:
+
+        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:
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+
+        if past_key_values is None:
+            past_key_values = tuple([None] * len(self.h))
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        hidden_states = inputs_embeds
+
+        if self.config.embed_layernorm:
+            hidden_states = self.word_embeddings_layernorm(inputs_embeds)
+
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                use_cache = False
+
+        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[2]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+        if attention_mask is None:
+            attention_mask = torch.ones((batch_size, seq_length_with_past), device=hidden_states.device)
+        else:
+            attention_mask = attention_mask.to(hidden_states.device)
+        causal_mask = self._prepare_attn_mask(
+            attention_mask,
+            input_shape=(batch_size, seq_length),
+            past_key_values_length=past_key_values_length,
+        )
+
+        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, use_cache=use_cache, output_attentions=output_attentions)
+
+                    return custom_forward
+
+                outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    causal_mask,
+                    layer_past,
+                )
+            else:
+                outputs = block(
+                    hidden_states,
+                    layer_past=layer_past,
+                    attention_mask=causal_mask,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = outputs[0]
+            if use_cache is True:
+                presents = presents + (outputs[1],)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
+        hidden_states = self.ln_f(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 BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+class TelechatForCausalLM(TelechatPreTrainedModel):
+    # _tied_weights_keys = ["lm_head.weight"]
+    _keys_to_ignore_on_load_missing = [r"lm_head.weight"]
+
+    def __init__(self, config: TelechatConfig):
+        super().__init__(config)
+        self.transformer = TelechatModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings: torch.Tensor):
+        self.lm_head = new_embeddings
+
+    def prepare_inputs_for_generation(
+            self,
+            input_ids: torch.LongTensor,
+            past_key_values: Optional[torch.Tensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            inputs_embeds: Optional[torch.Tensor] = None,
+            **kwargs,
+    ) -> dict:
+        if past_key_values:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    def forward(
+            self,
+            input_ids: Optional[torch.LongTensor] = None,
+            past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+            attention_mask: Optional[torch.Tensor] = 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,
+            **deprecated_arguments,
+    ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            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)
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(lm_logits.device)
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            batch_size, seq_length, vocab_size = shift_logits.shape
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(
+                shift_logits.view(batch_size * seq_length, vocab_size), shift_labels.view(batch_size * seq_length)
+            )
+
+        if not return_dict:
+            output = (lm_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    def chat(self, tokenizer, question: str = '', history: Union[List[Dict], History] = None, stream: bool = False,
+             generation_config: Optional[GenerationConfig] = None, **kwargs):
+        """
+        Args:
+            tokenizer:  the tokenizer of  telechat
+            question: question which the model reply in this turn
+            history: history which will format the input for telechat
+            stream: if return the full text at last or yield the text in token
+            generation_config:  configuration for generation
+            **kwargs: args which will update the generation config or pass to model forward
+        """
+        generation_config = generation_config or self.generation_config
+        if not generation_config:
+            logger.error("generation_config is None")
+            raise ValueError("generation_config must not be None")
+        if not question:
+            logger.error("question is empty")
+            raise ValueError("question must not be empty")
+        if history is None:
+            history = []
+
+        # we update and check generate_config here for building inputs.
+
+        generation_config = copy.deepcopy(generation_config)
+        user_id = generation_config.user_token_id
+        bot_id = generation_config.bot_token_id
+        model_kwargs = generation_config.update(**kwargs)
+        generation_config.validate()
+
+        # transfer to History
+        if not isinstance(history, History):
+            history = History(tokenizer, history)
+
+        inputs = self.build_inputs_for_chat(tokenizer, question, history, generation_config, user_id, bot_id)
+        history.append({"role": "user", "content": question})
+        if stream:
+            streamer = TelechatIterTextStreamer(tokenizer, history,skip_prompt=True)
+            Thread(target=self.generate, kwargs=dict(
+                inputs=inputs.to(self.device), streamer=streamer,
+                generation_config=generation_config, **model_kwargs
+            )).start()
+            return streamer
+        else:
+            outputs = self.generate(inputs.to(self.device), generation_config=generation_config, **model_kwargs)
+            response = tokenizer.decode(outputs[0][len(inputs[0]):-1])
+            history.append({"role": "bot", "content": response})
+            return response, history
+
+    def build_inputs_for_chat(self, tokenizer, question, history, generation_config, usr_id, bot_id):
+        """
+        check history and  build inputs here
+        """
+        # first tokenize question
+        q_token = tokenizer(question)
+        qa_history = copy.deepcopy(history)
+
+        # get the max length we should build our inputs in
+        model_max_length = self.config.seq_length
+        build_max_length = max(0, model_max_length - generation_config.max_new_tokens) \
+            if generation_config.max_new_tokens else max(0, generation_config.max_length)
+        if build_max_length < 3:
+            logger.warning("the model can not meet the  requirements of input length,Please check config")
+            raise ValueError("")
+
+        # trunc left
+        input_tokens = [usr_id] + q_token["input_ids"][-build_max_length + 1:] + [bot_id]
+        length = len(input_tokens)
+
+        while len(qa_history) != 0:
+            message = qa_history.pop()
+            if message["role"] == "user":
+                tokens = [usr_id] + message["input_ids"]
+            elif message["role"] == "bot":
+                tokens = [bot_id] + message["input_ids"] + [generation_config.eos_token_id]
+            else:
+                tokens = []
+            if len(tokens) + length >= build_max_length:
+                break
+            else:
+                input_tokens = tokens + input_tokens
+
+        return torch.tensor([input_tokens], dtype=torch.int64)

models/12B-V2_8bit/quantize_config.json

+{
+  "bits": 8,
+  "group_size": 128,
+  "damp_percent": 0.01,
+  "desc_act": false,
+  "sym": true,
+  "true_sequential": true,
+  "model_name_or_path": null,
+  "model_file_base_name": null
+}
\ No newline at end of file

models/12B-V2_8bit/specail_tokens_map.json

+{
+  "bos_token": {
+    "content": "<_start>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<_end>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<_pad>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<_unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}
\ No newline at end of file

models/12B-V2_8bit/tokenization_telechat.py

+import os
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple
+import sentencepiece as spm
+from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model"}
+
+# TODO: when we get download url from huggingface, refresh the map
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {},
+    "tokenizer_file": {},
+}
+
+
+class TelechatTokenizer(PreTrainedTokenizer):
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        unk_token="<unk>",
+        bos_token="<_start>",
+        eos_token="<_end>",
+        pad_token="<_pad>",
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
+        add_bos_token=True,
+        add_eos_token=False,
+        clean_up_tokenization_spaces=False,
+        **kwargs,
+    ):
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+        bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
+        unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
+        pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(vocab_file)
+        super().__init__(
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            add_bos_token=add_bos_token,
+            add_eos_token=add_eos_token,
+            sp_model_kwargs=self.sp_model_kwargs,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+        self.vocab_file = vocab_file
+        self.add_bos_token = add_bos_token
+        self.add_eos_token = add_eos_token
+
+
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["sp_model"] = None
+        return state
+
+    def __setstate__(self, d):
+        self.__dict__ = d
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(self.vocab_file)
+
+    @property
+    def vocab_size(self):
+        """Returns vocab size"""
+        return self.sp_model.get_piece_size()
+
+    def get_vocab(self):
+        """Returns vocab as a dict"""
+        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def _tokenize(self, text):
+        """Returns a tokenized string."""
+        return self.sp_model.encode(text, out_type=str)
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.sp_model.piece_to_id(token)
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        token = self.sp_model.IdToPiece(index)
+        return token
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        current_sub_tokens = []
+        out_string = ""
+        prev_is_special = False
+        for i, token in enumerate(tokens):
+            # make sure that special tokens are not decoded using sentencepiece model
+            if token in self.all_special_tokens:
+                if not prev_is_special and i != 0:
+                    out_string += " "
+                out_string += self.sp_model.decode(current_sub_tokens) + token
+                prev_is_special = True
+                current_sub_tokens = []
+            else:
+                current_sub_tokens.append(token)
+                prev_is_special = False
+        out_string += self.sp_model.decode(current_sub_tokens)
+        return out_string
+
+    def save_vocabulary(self, save_directory, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        """
+        Save the vocabulary and special tokens file to a directory.
+
+        Args:
+            save_directory (`str`):
+                The directory in which to save the vocabulary.
+
+        Returns:
+            `Tuple(str)`: Paths to the files saved.
+        """
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+        elif not os.path.isfile(self.vocab_file):
+            with open(out_vocab_file, "wb") as fi:
+                content_spiece_model = self.sp_model.serialized_model_proto()
+                fi.write(content_spiece_model)
+
+        return (out_vocab_file,)
+
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+        eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+        output = bos_token_id + token_ids_0 + eos_token_id
+
+        if token_ids_1 is not None:
+            output = output + bos_token_id + token_ids_1 + eos_token_id
+
+        return output
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        bos_token_id = [1] if self.add_bos_token else []
+        eos_token_id = [1] if self.add_eos_token else []
+
+        if token_ids_1 is None:
+            return bos_token_id + ([0] * len(token_ids_0)) + eos_token_id
+        return (
+            bos_token_id
+            + ([0] * len(token_ids_0))
+            + eos_token_id
+            + bos_token_id
+            + ([0] * len(token_ids_1))
+            + eos_token_id
+        )
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Creates a mask from the two sequences passed to be used in a sequence-pair classification task. An ALBERT
+        sequence pair mask has the following format:
+
+        ```
+        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+        | first sequence    | second sequence |
+        ```
+
+        if token_ids_1 is None, only returns the first portion of the mask (0s).
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of ids.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
+        """
+        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+        eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+        output = [0] * len(bos_token_id + token_ids_0 + eos_token_id)
+
+        if token_ids_1 is not None:
+            output += [1] * len(bos_token_id + token_ids_1 + eos_token_id)
+
+        return output

models/12B-V2_8bit/tokenizer_config.json

+{
+  "name_or_path": "ChinaTelecom/telechat-12b",
+  "tokenizer_class": "TelechatTokenizer",
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_telechat.TelechatTokenizer",
+      null
+    ]
+  },
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "use_fast": false,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "<_end>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": true
+  },
+  "model_max_length": 100000000,
+  "sp_model_kwargs": {},
+  "pad_token": {
+    "__type": "AddedToken",
+    "content": "<_pad>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": true
+  },
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<_end>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": true
+  }
+}

models/12B/config.json

+{
+  "apply_residual_connection_post_layernorm": false,
+  "architectures": [
+    "TelechatForCausalLM"
+  ],
+  "auto_map": {
+	"AutoConfig": "configuration_telechat.TelechatConfig",
+	"AutoModelForCausalLM": "modeling_telechat.TelechatForCausalLM"
+  },
+  "attention_dropout": 0.0,
+  "attention_softmax_in_fp32": true,
+  "bias_dropout_fusion": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_dropout": 0.0,
+  "hidden_size": 5120,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "masked_softmax_fusion": true,
+  "model_type": "telechat",
+  "n_head": 32,
+  "n_inner": null,
+  "n_layer": 38,
+  "offset_alibi": 100,
+  "pad_token_id": 3,
+  "pretraining_tp": 2,
+  "seq_length": 8192,
+  "skip_bias_add": true,
+  "skip_bias_add_qkv": false,
+  "slow_but_exact": false,
+  "transformers_version": "4.24.0",
+  "unk_token_id": 0,
+  "use_cache": true,
+  "vocab_size": 120000,
+  "ffn_hidden_size": 12288,
+  "flash_attn":true,
+  "tie_word_embeddings":false,
+  "training_seqlen":8192,
+  "logn":false,
+  "semi_causal":false,
+  "embed_layernorm":false
+}
+

models/12B/configuration_telechat.py

+# coding=utf-8
+# Copyright 2022 the Big Science Workshop and HuggingFace Inc. team.  All rights reserved.
+#
+# 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.
+
+""" Telechat configuration"""
+
+from packaging import version
+from collections import OrderedDict
+from transformers.utils import is_torch_available, logging
+from transformers.configuration_utils import PretrainedConfig
+from typing import TYPE_CHECKING, Any, List, Mapping, Optional
+
+logger = logging.get_logger(__name__)
+
+class TelechatConfig(PretrainedConfig):
+    """
+    Args:
+        vocab_size (`int`, *optional*, defaults to 160256): Vocabulary size of the Telechat model.
+        hidden_size (`int`, *optional*, defaults to 4096): Dimensionality of the embeddings and hidden states.
+        ffn_hidden_size (`int`, *optional*, defaults to 12288): Dimensionality of the feed-forward hidden states.
+        n_layer (`int`, *optional*, defaults to 30): Number of hidden layers in the Transformer
+        n_head (`int`, *optional*, defaults to 32): Number of attention heads for each attention layer.
+        layer_norm_epsilon (`float`, *optional*, defaults to 1e-5): The epsilon to use in the layer normalization layers.
+        initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        apply_residual_connection_post_layernorm (`bool`, *optional*, defaults to `False`): If enabled, use the layer norm of the hidden states as the residual in the transformer blocks
+        hidden_dropout (`float`, *optional*, defaults to 0.0): Dropout rate of the dropout function on the bias dropout.
+        attention_dropout (`float`, *optional*, defaults to 0.0): Dropout rate applied to the attention probs
+        use_cache (`bool`, *optional*, defaults to `True`): Whether or not the model should return the last key/values attentions.
+        training_seqlen (`int`, *optional*, defaults to 8192): Sequence length during last finetuning.
+        logn (`bool`, *optional*, defaults to `True`): Whether or not to use logN during extrapolation.
+        embed_layernorm (`bool`, *optional*, defaults to `True`): Whether or not to use embedding layernorm.
+
+    """
+
+    model_type = "telechat"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "num_hidden_layers": "n_layer",
+        "num_attention_heads": "n_head",
+    }
+
+    def __init__(
+        self,
+        vocab_size=160256,
+        hidden_size=4096,
+        n_layer=30,
+        n_head=32,
+        layer_norm_epsilon=1e-5,
+        initializer_range=0.02,
+        use_cache=True,
+        bos_token_id=1,
+        eos_token_id=2,
+        apply_residual_connection_post_layernorm=False,
+        hidden_dropout=0.0,
+        attention_dropout=0.0,
+        ffn_hidden_size=12288,
+        training_seqlen = 8192,
+        logn = True,
+        embed_layernorm = False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        n_embed = kwargs.pop("n_embed", None)
+        self.hidden_size = hidden_size if n_embed is None else n_embed
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.use_cache = use_cache
+        self.apply_residual_connection_post_layernorm = apply_residual_connection_post_layernorm
+        self.hidden_dropout = hidden_dropout
+        self.attention_dropout = attention_dropout
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.logn = logn
+        self.ffn_hidden_size = ffn_hidden_size
+        self.training_seqlen = training_seqlen
+        self.embed_layernorm = embed_layernorm
+
+
+        super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+

models/12B/generation_config.json

+{
+  "max_length": 8192,
+  "do_sample": false,
+  "use_cache": true,
+  "temperature": 0.3,
+  "top_k": 5,
+  "top_p": 0.85,
+  "repetition_penalty": 1.01,
+  "pad_token_id": 3,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "user_token_id": 20,
+  "bot_token_id": 21
+}

models/12B/generation_utils.py

+from typing import Optional
+from collections import deque
+from queue import Queue
+import copy
+
+
+class History:
+
+    def __init__(self, tokenizer, history):
+        '''
+        init from a list of dict
+        '''
+        # use deque to meet some special situation
+        self.input_history = deque()
+        self.tokenizer = tokenizer
+        if history:
+            self._transfer_from_list(history)
+
+    def _transfer_from_list(self, history):
+        for message in history:
+            content = message.get("content")
+            # the token result may not be equal to the result model gen
+            message.update(self.tokenizer(content))
+            self.input_history.append(message)
+
+    def append(self, message):
+        content = message.get("content")
+        if "input_ids" not in message or "attention_mask" not in message:
+            message.update(self.tokenizer(content))
+        self.input_history.append(message)
+
+    def append_left(self, message):
+        content = message.get("content")
+        if "input_ids" not in message or "attention_mask" not in message:
+            message.update(self.tokenizer(content))
+        self.input_history.appendleft(message)
+
+    def pop(self):
+        x = self.input_history.pop()
+        return x
+
+    def pop_left(self):
+        x = self.pop_left()
+        return x
+
+    def update(self, message):
+        self.input_history.pop()
+        self.append(message)
+
+    def __len__(self):
+        return self.input_history.__len__()
+
+    def __str__(self):
+        return self.input_history.__str__()
+
+    def __copy__(self):
+        new_instance = type(self)(self.tokenizer, [])
+        new_instance.input_history = copy.copy(self.input_history)
+        return new_instance
+
+    def __deepcopy__(self, memodict={}):
+        new_instance = type(self)(self.tokenizer, [])
+        new_instance.input_history = copy.deepcopy(self.input_history)
+        return new_instance
+
+
+class TelechatIterTextStreamer:
+    """
+    With reference to the TextIterStreamers in transformers, we have rewritten this class
+    """
+
+    def __init__(
+            self, tokenizer, history: History = None, skip_prompt: bool = False, timeout: Optional[float] = None,
+            **decode_kwargs
+    ):
+
+        self.tokenizer = tokenizer
+        self.history = history
+        self.skip_prompt = skip_prompt
+        self.timeout = timeout
+        self.decode_kwargs = decode_kwargs
+
+        self.text_queue = Queue()
+        self.cache_time = 0
+        self.text_until = ""
+        self.token_until = []
+        self.stop_signal = None
+        self.next_tokens_are_prompt = True
+
+        self.history.append({"role": "bot", "content": self.text_until})
+
+    def put(self, value):
+        """
+        put printable text into queue
+        """
+        if len(value.shape) > 1 and value.shape[0] > 1:
+            raise ValueError("TextStreamer only supports batch size 1")
+        elif len(value.shape) > 1:
+            value = value[0]
+
+        if self.skip_prompt and self.next_tokens_are_prompt:
+            self.next_tokens_are_prompt = False
+            return
+
+        if value[-1] == self.tokenizer.eos_token_id:
+            return
+
+        # there may be some smart way to decode.
+        self.token_until.extend(value.tolist())
+        text = self.tokenizer.decode(self.token_until, **self.decode_kwargs)
+
+
+        if self._is_printable(text) or self.cache_time >= 6:
+            output_text = text[len(self.text_until):]
+            self.text_until = text
+
+        else:
+            self.cache_time+=1
+            return
+
+        self.on_finalized_text(output_text)
+
+    def end(self):
+        """Flushes any remaining cache and prints a newline to stdout."""
+        # Flush the cache, if it exists
+        text = self.tokenizer.decode(self.token_until, **self.decode_kwargs)
+        output_text = text[len(self.text_until):]
+        self.text_until = text
+        self.on_finalized_text(output_text, stream_end=True)
+        self.clear_cache()
+
+    def clear_cache(self):
+        self.cache_time = 0
+        self.token_until = []
+        self.text_until = ""
+        self.history = None
+        self.next_tokens_are_prompt = True
+
+    def on_finalized_text(self, text: str, stream_end: bool = False):
+        """Put the text tuple in the queue."""
+        self.history.update({"role": "bot", "content": self.text_until, "input_ids": self.token_until,
+                             "attention_mask": [1] * len(self.token_until)})
+        self.text_queue.put((text, self.history), timeout=self.timeout)
+        if stream_end:
+            self.text_queue.put((self.stop_signal, self.history), timeout=self.timeout)
+
+    @staticmethod
+    def _is_printable(cp):
+        """Checks whether tokens can be decoded or not"""
+        if "�" in cp:
+            return False
+        return True
+
+    def __iter__(self):
+        return self
+
+    def __next__(self):
+        value_now, history_until = self.text_queue.get(timeout=self.timeout)
+        if value_now == self.stop_signal:
+            raise StopIteration()
+        else:
+            return value_now, history_until

models/12B/modeling_telechat.py

+# coding=utf-8
+# Copyright 2022 HuggingFace Inc. team and BigScience workshop.
+#
+# 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.
+
+# Copyright (c) 2023, NVIDIA CORPORATION.  All rights reserved.
+
+# Copyright (c) 2021 EleutherAI
+# This file is based on code by the authors denoted below and has been modified from its original version.
+#
+# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+#
+# 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.
+
+
+"""PyTorch TELECHAT model."""
+
+import warnings
+from typing import Optional, Tuple, Union, List, Dict
+from threading import Thread
+
+import torch
+import math
+import copy
+from torch import nn
+import torch.utils.checkpoint
+from torch.nn import functional as F
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, LayerNorm, MSELoss
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from transformers import GenerationConfig
+
+from .configuration_telechat import TelechatConfig
+from .generation_utils import History, TelechatIterTextStreamer
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "telechat"
+_CONFIG_FOR_DOC = "TelechatConfig"
+
+TELECHAT_PRETRAINED_MODEL_ARCHIVE_LIST = []
+
+try:
+    from einops import rearrange
+except ImportError:
+    rearrange = None
+
+use_flash_attn = True
+try:
+    from flash_attn.flash_attn_interface import flash_attn_unpadded_func
+except ImportError:
+    try:
+        from flash_attn.flash_attn_interface import flash_attn_varlen_func as flash_attn_unpadded_func
+    except ImportError:
+        flash_attn_unpadded_func = None
+
+
+class RotaryEmbedding(torch.nn.Module):
+    # Extracted from: https://github.com/EleutherAI/gpt-neox
+    def __init__(self, dim, config, base=10000,precision=torch.half):
+        super().__init__()
+        self.config = config
+        self.dim = dim
+        self.base = base
+        self.max_seq_len_cached = None
+        self.cos_cached = None
+        self.sin_cached = None
+        self.precision = precision
+
+    def get_mscale(self, scale=1):
+        if scale <= 1:
+            return 1.0
+        return 0.1 * math.log(scale) + 1.0
+
+    def get_ntk_alpha(self, true_seq_len):
+        context_value = math.log(true_seq_len / 8192, 2) + 1
+        # ntk_alpha = 2 ** context_value - 1
+        ntk_alpha = 2 ** math.ceil(context_value) - 1
+        ntk_alpha = max(ntk_alpha, 1)
+        return ntk_alpha
+
+    def forward(self, x, seq_dim=0, seq_len=None):
+        if seq_len is None:
+            seq_len = x.shape[seq_dim]
+        seq_len = max(seq_len, self.config.training_seqlen)
+        ntk_alpha = self.get_ntk_alpha(seq_len)
+        mscale = float(self.get_mscale(seq_len / self.config.training_seqlen))
+        base = self.base * ntk_alpha ** (self.dim / (self.dim - 2))
+        inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2, device=x.device).float( )/ self.dim ))
+        max_seq_len_cached = seq_len
+        t = torch.arange(max_seq_len_cached, device=x.device, dtype=inv_freq.dtype)
+        freqs = torch.einsum('i,j->ij', t, 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 = mscale *emb.cos()[:, None, :].half()
+        sin_cached = mscale *emb.sin()[:, None, :].half()
+        if self.precision == torch.bfloat16:
+            cos_cached = cos_cached.bfloat16()
+            sin_cached = sin_cached.bfloat16()
+        return cos_cached[:seq_len, ...], sin_cached[:seq_len, ...]
+
+
+# rotary pos emb helpers:
+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
+
+
+def apply_rotary_pos_emb_torch(q, k, cos, sin, offset: int = 0):  # jitting fails with bf16
+    cos, sin = cos[offset:q.shape[0] + offset, ...], sin[offset:q.shape[0] + offset, ...]
+    return (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin)
+
+
+class MixedFusedRMSNorm(nn.Module):
+    # Extracted from https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py
+    def __init__(self, hidden_size, eps=1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+class FlashSelfAttention(torch.nn.Module):
+    # Extracted from https://github.com/microsoft/Megatron-DeepSpeed/blob/main/megatron/model/transformer.py
+    """Implement the scaled dot product attention with softmax.
+    Arguments
+    ---------
+        softmax_scale: The temperature to use for the softmax attention.
+                      (default: 1/sqrt(d_keys) where d_keys is computed at
+                      runtime)
+        attention_dropout: The dropout rate to apply to the attention
+                           (default: 0.0)
+    """
+
+    def __init__(self, causal=False, softmax_scale=None, attention_dropout=0.0,
+                 device=None, dtype=None):
+        super().__init__()
+        assert flash_attn_unpadded_func is not None, ('Please install FlashAttention first, '
+                                                      'e.g., with pip install flash-attn')
+        assert rearrange is not None, 'Please install einops first, e.g., with pip install einops'
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+        self.dropout_p = attention_dropout
+
+    def forward(self, q, k, v):
+        """Implements the multihead softmax attention.
+        Arguments
+        ---------
+            q, k, v: The tensor containing the query, key, and value. (B, S, H, D)
+        """
+        assert all((i.dtype in [torch.float16, torch.bfloat16] for i in (q, k, v)))
+        assert all((i.is_cuda for i in (q, k, v)))
+
+        batch_size, seqlen_q = q.shape[0], q.shape[1]
+        seqlen_k = k.shape[1]
+
+        q, k, v = [rearrange(x, 'b s ... -> (b s) ...') for x in [q, k, v]]
+        cu_seqlens_q = torch.arange(0, (batch_size + 1) * seqlen_q, step=seqlen_q, dtype=torch.int32,
+                                    device=q.device)
+        if self.training:
+            # during training q,k,v always have same seqlen
+            assert seqlen_k == seqlen_q
+
+            is_causal = self.causal
+            cu_seqlens_k = cu_seqlens_q
+            dropout_p = self.dropout_p
+        else:
+            # turn off FA causal mask after first inference autoregressive iteration
+            # only on first autoregressive step q,k,v have same seqlen
+            is_causal = seqlen_q == seqlen_k
+            cu_seqlens_k = torch.arange(0, (batch_size + 1) * seqlen_k, step=seqlen_k, dtype=torch.int32,
+                                        device=q.device)
+            dropout_p = 0
+
+        output = flash_attn_unpadded_func(
+            q, k, v, cu_seqlens_q, cu_seqlens_k, seqlen_q, seqlen_k,
+            dropout_p=dropout_p,
+            softmax_scale=self.softmax_scale, causal=is_causal
+        )
+
+        output = rearrange(output, '(b s) ... -> b s ...', b=batch_size)
+        return output
+
+
+def _make_causal_mask(
+        input_ids_shape: torch.Size, device: torch.device, past_key_values_length: int
+) -> torch.BoolTensor:
+    """
+    Make causal mask used for self-attention.
+    """
+    batch_size, target_length = input_ids_shape
+    mask = torch.empty((target_length, target_length + past_key_values_length), dtype=torch.bool, device=device)
+    # ONNX doesn't support `torch.Tensor.triu` properly, thus we use this workaround
+    seq_ids = torch.arange(target_length, device=device)
+    mask[:, past_key_values_length:] = seq_ids[:, None] < seq_ids[None, :]
+
+    if past_key_values_length > 0:
+        mask[:, :past_key_values_length] = False
+
+    expanded_mask = mask[None, None, :, :].expand(batch_size, 1, target_length, target_length + past_key_values_length)
+    return expanded_mask
+
+
+def _expand_mask(mask: torch.Tensor, tgt_length: int) -> torch.BoolTensor:
+    """
+    Expands attention_mask from `[batch_size, src_length]` to `[batch_size, 1, tgt_length, src_length]`.
+    """
+    batch_size, src_length = mask.shape
+    tgt_length = tgt_length if tgt_length is not None else src_length
+
+    expanded_mask = ~(mask[:, None, None, :].to(torch.bool))
+    return expanded_mask.expand(batch_size, 1, tgt_length, src_length)
+
+
+def dropout_add(x: torch.Tensor, residual: torch.Tensor, prob: float, training: bool) -> torch.Tensor:
+    """
+    Dropout add function
+
+    Args:
+        x (`torch.tensor`, *required*):
+            input tensor
+        residual (`torch.tensor`, *required*):
+            residual tensor
+        prob (`float`, *required*):
+            dropout probability
+        training (`bool`, *required*):
+            training mode
+    """
+    out = F.dropout(x, p=prob, training=training)
+    out = residual + out
+    return out
+
+
+def telechat_gelu_forward(x: torch.Tensor) -> torch.Tensor:
+    """
+    Custom bias GELU function. Adapted from Megatron-DeepSpeed code. Here we use a simple implementation (inference) to
+    make the model jitable.
+
+    Args:
+        x (`torch.tensor`, *required*):
+            input hidden states
+    """
+    return x * 0.5 * (1.0 + torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x)))
+
+
+def telechat_gelu_back(g: torch.Tensor, x: torch.Tensor) -> torch.Tensor:
+    """
+    gradient of tanh approximation of gelu gradient of actual gelu is: 0.5 * (1. + torch.erf(x * 0.70710678)) +
+    0.3989423 * x * torch.exp(-0.5 * x * x)
+
+    Args:
+        g (`torch.tensor`, *required*):
+            gradient output tensor
+        x (`torch.tensor`, *required*):
+            input tensor
+    """
+    x = x[0]  # x is a tuple of 1 element, needs to unpack it first
+    tanh_out = torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x))
+    # sqrt(2/pi) * 3 * 0.044715 -> 0.1070322243
+    ff = 0.5 * x * ((1 - tanh_out * tanh_out) * (0.79788456 + 0.1070322243 * x * x)) + 0.5 * (1 + tanh_out)
+    return ff * g
+
+
+class GeLUFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, input: torch.Tensor) -> torch.Tensor:
+        ctx.save_for_backward(input)
+        return telechat_gelu_forward(input)
+
+    @staticmethod
+    def backward(ctx, grad_output: torch.Tensor) -> torch.Tensor:
+        input = ctx.saved_tensors
+        tmp = telechat_gelu_back(grad_output, input)
+        return tmp
+
+
+class TelechatGelu(nn.Module):
+    """
+    TelechatBiasGelu wrapper function that make use of the simple function on inference mode to make the model
+    torchscriptable and use the autograd function in training mode to get the accurate results of the gradients Partly
+    copied from Megatron-DeepSpeed code and adapted for our needs
+
+    See here why autograd functions are not torchscriptable: https://github.com/pytorch/pytorch/issues/22329
+    """
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.training:
+            return GeLUFunction.apply(x)
+        else:
+            return telechat_gelu_forward(x)
+
+
+class TelechatAttention(nn.Module):
+    def __init__(self, config: TelechatConfig, layer_idx):
+        super().__init__()
+        self.kv_cache = None
+        self.layer_idx = layer_idx
+
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.n_head
+        self.head_dim = self.hidden_size // self.num_heads
+        self.split_size = self.hidden_size
+        self.hidden_dropout = config.hidden_dropout
+        self.config = config
+
+        if self.head_dim * self.num_heads != self.hidden_size:
+            raise ValueError(
+                f"`hidden_size` must be divisible by num_heads (got `hidden_size`: {self.hidden_size} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+
+        # Layer-wise attention scaling
+        self.inv_norm_factor = 1.0 / math.sqrt(self.head_dim)
+        self.beta = 1.0
+
+        self.num_key_value_heads = self.num_heads
+        kv_projection_size = self.head_dim * self.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.query = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+        self.key_value = nn.Linear(self.hidden_size, kv_projection_size * 2, bias=False)
+        self.dense = nn.Linear(self.hidden_size, self.hidden_size)
+        self.attention_dropout = nn.Dropout(config.attention_dropout)
+        self.rotary_emb = RotaryEmbedding(self.head_dim, config=config)
+
+        self.core_attention_flash = FlashSelfAttention(
+            causal=True, attention_dropout=config.attention_dropout
+        )
+
+        self.last_key_layer = None
+
+    def repeat_kv(self, hidden_states, n_rep):
+        slen, batch, num_key_value_heads_per_partition, head_dim = hidden_states.shape
+        if n_rep == 1:
+            return hidden_states
+        hidden_states = hidden_states[:, :, :, None, :].expand(slen, batch, num_key_value_heads_per_partition, n_rep,
+                                                               head_dim)
+        return hidden_states.reshape(slen, batch, num_key_value_heads_per_partition * n_rep, head_dim)
+
+    def split_tensor_along_last_dim(self,
+                                    tensor: torch.Tensor,
+                                    num_partitions: int,
+                                    contiguous_split_chunks: bool = False,
+                                    ):
+
+        # 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 _merge_heads(self, x: torch.Tensor) -> torch.Tensor:
+        batch_size_and_num_heads, seq_length, _ = x.shape
+        batch_size = batch_size_and_num_heads // self.num_heads
+        x = x.view(batch_size, self.num_heads, seq_length, self.head_dim)
+        x = x.permute(0, 2, 1, 3)
+        return x.reshape(batch_size, seq_length, self.num_heads * self.head_dim)
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            residual: torch.Tensor,
+            attention_mask: torch.Tensor,
+            layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+            use_cache: bool = False,
+            output_attentions: bool = False,
+    ):
+        hidden_states = hidden_states.transpose(1, 0)
+        query_layer = self.query(hidden_states)
+        new_tensor_shape = query_layer.size()[:-1] + \
+                           (self.num_heads,
+                            self.head_dim)
+        query_layer = query_layer.view(*new_tensor_shape)
+
+        mixed_kv_layer = self.key_value(hidden_states)
+        new_tensor_shape = mixed_kv_layer.size()[:-1] + \
+                           (self.num_key_value_heads,
+                            2 * self.head_dim)
+        mixed_kv_layer = mixed_kv_layer.view(*new_tensor_shape)
+        (key_layer, value_layer) = self.split_tensor_along_last_dim(mixed_kv_layer, 2)
+
+        output_size = (query_layer.size(1),
+                       query_layer.size(2),
+                       query_layer.size(0),
+                       key_layer.size(0))
+
+        query_layer = query_layer.view(output_size[2], output_size[0] * output_size[1], -1)
+        key_layer = key_layer.view(output_size[3], output_size[0] * output_size[1], -1)
+
+        apply_rotary_fn = apply_rotary_pos_emb_torch
+
+        seq_len = key_layer.shape[0]
+        offset = 0
+
+        if use_cache and layer_past != None:
+            past_key, past_value = layer_past
+            offset = past_key.shape[0]
+            seq_len += offset
+
+        cos, sin = self.rotary_emb(value_layer)
+
+        query_layer, key_layer = apply_rotary_fn(query_layer, key_layer, cos, sin, offset=offset)
+        if use_cache:
+            if layer_past != None:
+                past_key, past_value = layer_past
+                key_layer = torch.cat((past_key, key_layer[-1, ...].unsqueeze(0)), dim=0)
+                value_layer = torch.cat((past_value, value_layer[-1, ...].unsqueeze(0)), dim=0)
+            layer_past = key_layer, value_layer
+        s, bz, head, dim = value_layer.shape
+        s_key = key_layer.shape[0]
+        s_query = query_layer.shape[0]
+        query_layer = query_layer.reshape((s_query, bz, head, dim))
+        key_layer = key_layer.reshape((s_key, bz, head, dim))
+
+        if self.config.flash_attn:
+            q, k, v = [rearrange(x, 's b ... -> b s ...').contiguous() for x in
+                       (query_layer, key_layer, value_layer)]
+            context_layer = self.core_attention_flash(q, k, v)
+            context_layer = rearrange(context_layer, 'b s h d -> b s (h d)').contiguous()
+        else:
+            ##[sq, b, np, hn] -> [sq, b * np, hn]
+            query_layer = query_layer.reshape(s_query, bz * self.num_heads, dim)
+            # [sk, b, np, hn] -> [sk, b * np, hn]
+            key_layer = key_layer.reshape(s_key, bz * self.num_heads, dim)
+            matmul_result = self.inv_norm_factor * torch.einsum('bik,bkj->bij', query_layer.transpose(0, 1),
+                                                                key_layer.transpose(0, 1).transpose(1, 2))
+
+            attention_scores = matmul_result.view(bz, self.num_heads, s_query, s_key)
+
+            input_dtype = attention_scores.dtype
+            if input_dtype == torch.float16 or input_dtype == torch.bfloat16:
+                attention_scores = attention_scores.to(torch.float)
+            attn_weights = torch.masked_fill(attention_scores, attention_mask, torch.finfo(attention_scores.dtype).min)
+            attention_probs = F.softmax(attn_weights, dim=-1).to(input_dtype)  ##dtype = torch.float32
+            attention_probs = self.attention_dropout(attention_probs)
+            attention_probs_reshaped = attention_probs.view(bz * self.num_heads, s_query, s_key)
+
+            value_layer = value_layer.reshape(s_key, bz * self.num_heads, dim)
+            context_layer = torch.bmm(attention_probs_reshaped, value_layer.transpose(0, 1))
+            context_layer = self._merge_heads(context_layer)
+
+        output_tensor = self.dense(context_layer)
+
+        output_tensor = dropout_add(output_tensor, residual, self.hidden_dropout, self.training)
+        present = None
+        outputs = (output_tensor, present)
+        if output_attentions:
+            outputs += (attention_probs,)
+
+        return output_tensor, layer_past
+
+
+class TelechatMLP(nn.Module):
+    def __init__(self, config: TelechatConfig):
+        super().__init__()
+        hidden_size = config.hidden_size
+        self.gate_proj = nn.Linear(hidden_size, config.ffn_hidden_size, bias=False)
+        self.up_proj = nn.Linear(hidden_size, config.ffn_hidden_size, bias=False)
+        self.down_proj = nn.Linear(config.ffn_hidden_size, hidden_size, bias=True)
+        self.hidden_dropout = config.hidden_dropout
+
+    def forward(self, hidden_states: torch.Tensor, residual: torch.Tensor) -> torch.Tensor:
+        intermediate_output = self.down_proj(F.silu(self.gate_proj(hidden_states)) * self.up_proj(hidden_states))
+        output = dropout_add(intermediate_output, residual, self.hidden_dropout, self.training)
+        return output
+
+
+class TelechatBlock(nn.Module):
+    def __init__(self, config: TelechatConfig, layer_idx):
+        super().__init__()
+        hidden_size = config.hidden_size
+
+        self.input_layernorm = MixedFusedRMSNorm(hidden_size, eps=config.layer_norm_epsilon)
+        self.num_heads = config.n_head
+        self.layer_idx = layer_idx
+        self.self_attention = TelechatAttention(config, layer_idx)
+        self.post_attention_layernorm = MixedFusedRMSNorm(hidden_size, eps=config.layer_norm_epsilon)
+
+        self.mlp = TelechatMLP(config)
+
+        self.apply_residual_connection_post_layernorm = config.apply_residual_connection_post_layernorm
+        self.hidden_dropout = config.hidden_dropout
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            attention_mask: torch.Tensor,
+            layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+            use_cache: bool = False,
+            output_attentions: bool = False,
+    ):
+        layernorm_output = self.input_layernorm(hidden_states)
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = hidden_states
+
+        attn_outputs = self.self_attention(
+            layernorm_output,
+            residual,
+            layer_past=layer_past,
+            attention_mask=attention_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+        )
+
+        attention_output = attn_outputs[0]
+        outputs = attn_outputs[1:]
+        layernorm_output = self.post_attention_layernorm(attention_output)
+
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = attention_output
+        output = self.mlp(layernorm_output, residual)
+
+        if use_cache:
+            outputs = (output,) + outputs
+        else:
+            outputs = (output,) + outputs[1:]
+
+        return outputs
+
+
+class TelechatPreTrainedModel(PreTrainedModel):
+    config_class = TelechatConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["TelechatBlock"]
+    _skip_keys_device_placement = "past_key_values"
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module: nn.Module):
+        """Initialize the weights."""
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+        elif isinstance(module, LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    def _set_gradient_checkpointing(self, module: nn.Module, value: bool = False):
+        if isinstance(module, TelechatModel):
+            module.gradient_checkpointing = value
+
+
+class TelechatModel(TelechatPreTrainedModel):
+    def __init__(self, config: TelechatConfig):
+        super().__init__(config)
+
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.n_head
+        self.config = config
+        self.word_embeddings = nn.Embedding(config.vocab_size, self.embed_dim)
+        if self.config.embed_layernorm:
+            self.word_embeddings_layernorm = MixedFusedRMSNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+
+        self.h = nn.ModuleList([TelechatBlock(config, _) for _ in range(config.num_hidden_layers)])
+        self.ln_f = MixedFusedRMSNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+        self.gradient_checkpointing = False
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.word_embeddings
+
+    def _prepare_attn_mask(
+            self, attention_mask: torch.Tensor, input_shape: Tuple[int, int], past_key_values_length: int
+    ) -> torch.BoolTensor:
+        combined_attention_mask = None
+        device = attention_mask.device
+        _, src_length = input_shape
+
+        if src_length > 1:
+            combined_attention_mask = _make_causal_mask(
+                input_shape, device=device, past_key_values_length=past_key_values_length
+            )
+        expanded_attn_mask = _expand_mask(attention_mask, tgt_length=src_length)
+        combined_attention_mask = (
+            expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask | combined_attention_mask
+        )
+
+        return combined_attention_mask
+
+    def set_input_embeddings(self, new_embeddings: torch.Tensor):
+        self.word_embeddings = new_embeddings
+
+    def forward(
+            self,
+            input_ids: Optional[torch.LongTensor] = None,
+            past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+            attention_mask: Optional[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,
+            **deprecated_arguments,
+    ) -> Union[Tuple[torch.Tensor, ...], BaseModelOutputWithPastAndCrossAttentions]:
+
+        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:
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+
+        if past_key_values is None:
+            past_key_values = tuple([None] * len(self.h))
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        hidden_states = inputs_embeds
+
+        if self.config.embed_layernorm:
+            hidden_states = self.word_embeddings_layernorm(inputs_embeds)
+
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                use_cache = False
+
+        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[2]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+        if attention_mask is None:
+            attention_mask = torch.ones((batch_size, seq_length_with_past), device=hidden_states.device)
+        else:
+            attention_mask = attention_mask.to(hidden_states.device)
+        causal_mask = self._prepare_attn_mask(
+            attention_mask,
+            input_shape=(batch_size, seq_length),
+            past_key_values_length=past_key_values_length,
+        )
+
+        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, use_cache=use_cache, output_attentions=output_attentions)
+
+                    return custom_forward
+
+                outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    causal_mask,
+                    layer_past,
+                )
+            else:
+                outputs = block(
+                    hidden_states,
+                    layer_past=layer_past,
+                    attention_mask=causal_mask,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = outputs[0]
+            if use_cache is True:
+                presents = presents + (outputs[1],)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
+        hidden_states = self.ln_f(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 BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+class TelechatForCausalLM(TelechatPreTrainedModel):
+    # _tied_weights_keys = ["lm_head.weight"]
+    _keys_to_ignore_on_load_missing = [r"lm_head.weight"]
+
+    def __init__(self, config: TelechatConfig):
+        super().__init__(config)
+        self.transformer = TelechatModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.post_init()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings: torch.Tensor):
+        self.lm_head = new_embeddings
+
+    def prepare_inputs_for_generation(
+            self,
+            input_ids: torch.LongTensor,
+            past_key_values: Optional[torch.Tensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            inputs_embeds: Optional[torch.Tensor] = None,
+            **kwargs,
+    ) -> dict:
+        if past_key_values:
+            input_ids = input_ids[:, -1].unsqueeze(-1)
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    def forward(
+            self,
+            input_ids: Optional[torch.LongTensor] = None,
+            past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+            attention_mask: Optional[torch.Tensor] = 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,
+            **deprecated_arguments,
+    ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            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)
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(lm_logits.device)
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            batch_size, seq_length, vocab_size = shift_logits.shape
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(
+                shift_logits.view(batch_size * seq_length, vocab_size), shift_labels.view(batch_size * seq_length)
+            )
+
+        if not return_dict:
+            output = (lm_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    def chat(self, tokenizer, question: str = '', history: Union[List[Dict], History] = None, stream: bool = False,
+             generation_config: Optional[GenerationConfig] = None, **kwargs):
+        """
+        Args:
+            tokenizer:  the tokenizer of  telechat
+            question: question which the model reply in this turn
+            history: history which will format the input for telechat
+            stream: if return the full text at last or yield the text in token
+            generation_config:  configuration for generation
+            **kwargs: args which will update the generation config or pass to model forward
+        """
+        generation_config = generation_config or self.generation_config
+        if not generation_config:
+            logger.error("generation_config is None")
+            raise ValueError("generation_config must not be None")
+        if not question:
+            logger.error("question is empty")
+            raise ValueError("question must not be empty")
+        if history is None:
+            history = []
+
+        # we update and check generate_config here for building inputs.
+
+        generation_config = copy.deepcopy(generation_config)
+        user_id = generation_config.user_token_id
+        bot_id = generation_config.bot_token_id
+        model_kwargs = generation_config.update(**kwargs)
+        generation_config.validate()
+
+        # transfer to History
+        if not isinstance(history, History):
+            history = History(tokenizer, history)
+
+        inputs = self.build_inputs_for_chat(tokenizer, question, history, generation_config, user_id, bot_id)
+        history.append({"role": "user", "content": question})
+        if stream:
+            streamer = TelechatIterTextStreamer(tokenizer, history,skip_prompt=True)
+            Thread(target=self.generate, kwargs=dict(
+                inputs=inputs.to(self.device), streamer=streamer,
+                generation_config=generation_config, **model_kwargs
+            )).start()
+            return streamer
+        else:
+            outputs = self.generate(inputs.to(self.device), generation_config=generation_config, **model_kwargs)
+            response = tokenizer.decode(outputs[0][len(inputs[0]):-1])
+            history.append({"role": "bot", "content": response})
+            return response, history
+
+    def build_inputs_for_chat(self, tokenizer, question, history, generation_config, usr_id, bot_id):
+        """
+        check history and  build inputs here
+        """
+        # first tokenize question
+        q_token = tokenizer(question)
+        qa_history = copy.deepcopy(history)
+
+        # get the max length we should build our inputs in
+        model_max_length = self.config.seq_length
+        build_max_length = max(0, model_max_length - generation_config.max_new_tokens) \
+            if generation_config.max_new_tokens else max(0, generation_config.max_length)
+        if build_max_length < 3:
+            logger.warning("the model can not meet the  requirements of input length,Please check config")
+            raise ValueError("")
+
+        # trunc left
+        input_tokens = [usr_id] + q_token["input_ids"][-build_max_length + 1:] + [bot_id]
+        length = len(input_tokens)
+
+        while len(qa_history) != 0:
+            message = qa_history.pop()
+            if message["role"] == "user":
+                tokens = [usr_id] + message["input_ids"]
+            elif message["role"] == "bot":
+                tokens = [bot_id] + message["input_ids"] + [generation_config.eos_token_id]
+            else:
+                tokens = []
+            if len(tokens) + length >= build_max_length:
+                break
+            else:
+                input_tokens = tokens + input_tokens
+
+        return torch.tensor([input_tokens], dtype=torch.int64)

models/12B/pytorch_model.bin.index.json

+{
+  "metadata": {
+    "total_size": 24772864000
+  },
+  "weight_map": {
+    "lm_head.weight": "pytorch_model_00041-of-00041.bin",
+    "transformer.h.0.input_layernorm.weight": "pytorch_model_00002-of-00041.bin",
+    "transformer.h.0.mlp.down_proj.bias": "pytorch_model_00002-of-00041.bin",
+    "transformer.h.0.mlp.down_proj.weight": "pytorch_model_00002-of-00041.bin",
+    "transformer.h.0.mlp.gate_proj.weight": "pytorch_model_00002-of-00041.bin",
+    "transformer.h.0.mlp.up_proj.weight": "pytorch_model_00002-of-00041.bin",
+    "transformer.h.0.post_attention_layernorm.weight": "pytorch_model_00002-of-00041.bin",
+    "transformer.h.0.self_attention.dense.bias": "pytorch_model_00002-of-00041.bin",
+    "transformer.h.0.self_attention.dense.weight": "pytorch_model_00002-of-00041.bin",
+    "transformer.h.0.self_attention.key_value.weight": "pytorch_model_00002-of-00041.bin",
+    "transformer.h.0.self_attention.query.weight": "pytorch_model_00002-of-00041.bin",
+    "transformer.h.1.input_layernorm.weight": "pytorch_model_00003-of-00041.bin",
+    "transformer.h.1.mlp.down_proj.bias": "pytorch_model_00003-of-00041.bin",
+    "transformer.h.1.mlp.down_proj.weight": "pytorch_model_00003-of-00041.bin",
+    "transformer.h.1.mlp.gate_proj.weight": "pytorch_model_00003-of-00041.bin",
+    "transformer.h.1.mlp.up_proj.weight": "pytorch_model_00003-of-00041.bin",
+    "transformer.h.1.post_attention_layernorm.weight": "pytorch_model_00003-of-00041.bin",
+    "transformer.h.1.self_attention.dense.bias": "pytorch_model_00003-of-00041.bin",
+    "transformer.h.1.self_attention.dense.weight": "pytorch_model_00003-of-00041.bin",
+    "transformer.h.1.self_attention.key_value.weight": "pytorch_model_00003-of-00041.bin",
+    "transformer.h.1.self_attention.query.weight": "pytorch_model_00003-of-00041.bin",
+    "transformer.h.10.input_layernorm.weight": "pytorch_model_00012-of-00041.bin",
+    "transformer.h.10.mlp.down_proj.bias": "pytorch_model_00012-of-00041.bin",
+    "transformer.h.10.mlp.down_proj.weight": "pytorch_model_00012-of-00041.bin",
+    "transformer.h.10.mlp.gate_proj.weight": "pytorch_model_00012-of-00041.bin",
+    "transformer.h.10.mlp.up_proj.weight": "pytorch_model_00012-of-00041.bin",
+    "transformer.h.10.post_attention_layernorm.weight": "pytorch_model_00012-of-00041.bin",
+    "transformer.h.10.self_attention.dense.bias": "pytorch_model_00012-of-00041.bin",
+    "transformer.h.10.self_attention.dense.weight": "pytorch_model_00012-of-00041.bin",
+    "transformer.h.10.self_attention.key_value.weight": "pytorch_model_00012-of-00041.bin",
+    "transformer.h.10.self_attention.query.weight": "pytorch_model_00012-of-00041.bin",
+    "transformer.h.11.input_layernorm.weight": "pytorch_model_00013-of-00041.bin",
+    "transformer.h.11.mlp.down_proj.bias": "pytorch_model_00013-of-00041.bin",
+    "transformer.h.11.mlp.down_proj.weight": "pytorch_model_00013-of-00041.bin",
+    "transformer.h.11.mlp.gate_proj.weight": "pytorch_model_00013-of-00041.bin",
+    "transformer.h.11.mlp.up_proj.weight": "pytorch_model_00013-of-00041.bin",
+    "transformer.h.11.post_attention_layernorm.weight": "pytorch_model_00013-of-00041.bin",
+    "transformer.h.11.self_attention.dense.bias": "pytorch_model_00013-of-00041.bin",
+    "transformer.h.11.self_attention.dense.weight": "pytorch_model_00013-of-00041.bin",
+    "transformer.h.11.self_attention.key_value.weight": "pytorch_model_00013-of-00041.bin",
+    "transformer.h.11.self_attention.query.weight": "pytorch_model_00013-of-00041.bin",
+    "transformer.h.12.input_layernorm.weight": "pytorch_model_00014-of-00041.bin",
+    "transformer.h.12.mlp.down_proj.bias": "pytorch_model_00014-of-00041.bin",
+    "transformer.h.12.mlp.down_proj.weight": "pytorch_model_00014-of-00041.bin",
+    "transformer.h.12.mlp.gate_proj.weight": "pytorch_model_00014-of-00041.bin",
+    "transformer.h.12.mlp.up_proj.weight": "pytorch_model_00014-of-00041.bin",
+    "transformer.h.12.post_attention_layernorm.weight": "pytorch_model_00014-of-00041.bin",
+    "transformer.h.12.self_attention.dense.bias": "pytorch_model_00014-of-00041.bin",
+    "transformer.h.12.self_attention.dense.weight": "pytorch_model_00014-of-00041.bin",
+    "transformer.h.12.self_attention.key_value.weight": "pytorch_model_00014-of-00041.bin",
+    "transformer.h.12.self_attention.query.weight": "pytorch_model_00014-of-00041.bin",
+    "transformer.h.13.input_layernorm.weight": "pytorch_model_00015-of-00041.bin",
+    "transformer.h.13.mlp.down_proj.bias": "pytorch_model_00015-of-00041.bin",
+    "transformer.h.13.mlp.down_proj.weight": "pytorch_model_00015-of-00041.bin",
+    "transformer.h.13.mlp.gate_proj.weight": "pytorch_model_00015-of-00041.bin",
+    "transformer.h.13.mlp.up_proj.weight": "pytorch_model_00015-of-00041.bin",
+    "transformer.h.13.post_attention_layernorm.weight": "pytorch_model_00015-of-00041.bin",
+    "transformer.h.13.self_attention.dense.bias": "pytorch_model_00015-of-00041.bin",
+    "transformer.h.13.self_attention.dense.weight": "pytorch_model_00015-of-00041.bin",
+    "transformer.h.13.self_attention.key_value.weight": "pytorch_model_00015-of-00041.bin",
+    "transformer.h.13.self_attention.query.weight": "pytorch_model_00015-of-00041.bin",
+    "transformer.h.14.input_layernorm.weight": "pytorch_model_00016-of-00041.bin",
+    "transformer.h.14.mlp.down_proj.bias": "pytorch_model_00016-of-00041.bin",
+    "transformer.h.14.mlp.down_proj.weight": "pytorch_model_00016-of-00041.bin",
+    "transformer.h.14.mlp.gate_proj.weight": "pytorch_model_00016-of-00041.bin",
+    "transformer.h.14.mlp.up_proj.weight": "pytorch_model_00016-of-00041.bin",
+    "transformer.h.14.post_attention_layernorm.weight": "pytorch_model_00016-of-00041.bin",
+    "transformer.h.14.self_attention.dense.bias": "pytorch_model_00016-of-00041.bin",
+    "transformer.h.14.self_attention.dense.weight": "pytorch_model_00016-of-00041.bin",
+    "transformer.h.14.self_attention.key_value.weight": "pytorch_model_00016-of-00041.bin",
+    "transformer.h.14.self_attention.query.weight": "pytorch_model_00016-of-00041.bin",
+    "transformer.h.15.input_layernorm.weight": "pytorch_model_00017-of-00041.bin",
+    "transformer.h.15.mlp.down_proj.bias": "pytorch_model_00017-of-00041.bin",
+    "transformer.h.15.mlp.down_proj.weight": "pytorch_model_00017-of-00041.bin",
+    "transformer.h.15.mlp.gate_proj.weight": "pytorch_model_00017-of-00041.bin",
+    "transformer.h.15.mlp.up_proj.weight": "pytorch_model_00017-of-00041.bin",
+    "transformer.h.15.post_attention_layernorm.weight": "pytorch_model_00017-of-00041.bin",
+    "transformer.h.15.self_attention.dense.bias": "pytorch_model_00017-of-00041.bin",
+    "transformer.h.15.self_attention.dense.weight": "pytorch_model_00017-of-00041.bin",
+    "transformer.h.15.self_attention.key_value.weight": "pytorch_model_00017-of-00041.bin",
+    "transformer.h.15.self_attention.query.weight": "pytorch_model_00017-of-00041.bin",
+    "transformer.h.16.input_layernorm.weight": "pytorch_model_00018-of-00041.bin",
+    "transformer.h.16.mlp.down_proj.bias": "pytorch_model_00018-of-00041.bin",
+    "transformer.h.16.mlp.down_proj.weight": "pytorch_model_00018-of-00041.bin",
+    "transformer.h.16.mlp.gate_proj.weight": "pytorch_model_00018-of-00041.bin",
+    "transformer.h.16.mlp.up_proj.weight": "pytorch_model_00018-of-00041.bin",
+    "transformer.h.16.post_attention_layernorm.weight": "pytorch_model_00018-of-00041.bin",
+    "transformer.h.16.self_attention.dense.bias": "pytorch_model_00018-of-00041.bin",
+    "transformer.h.16.self_attention.dense.weight": "pytorch_model_00018-of-00041.bin",
+    "transformer.h.16.self_attention.key_value.weight": "pytorch_model_00018-of-00041.bin",
+    "transformer.h.16.self_attention.query.weight": "pytorch_model_00018-of-00041.bin",
+    "transformer.h.17.input_layernorm.weight": "pytorch_model_00019-of-00041.bin",
+    "transformer.h.17.mlp.down_proj.bias": "pytorch_model_00019-of-00041.bin",
+    "transformer.h.17.mlp.down_proj.weight": "pytorch_model_00019-of-00041.bin",
+    "transformer.h.17.mlp.gate_proj.weight": "pytorch_model_00019-of-00041.bin",
+    "transformer.h.17.mlp.up_proj.weight": "pytorch_model_00019-of-00041.bin",
+    "transformer.h.17.post_attention_layernorm.weight": "pytorch_model_00019-of-00041.bin",
+    "transformer.h.17.self_attention.dense.bias": "pytorch_model_00019-of-00041.bin",
+    "transformer.h.17.self_attention.dense.weight": "pytorch_model_00019-of-00041.bin",
+    "transformer.h.17.self_attention.key_value.weight": "pytorch_model_00019-of-00041.bin",
+    "transformer.h.17.self_attention.query.weight": "pytorch_model_00019-of-00041.bin",
+    "transformer.h.18.input_layernorm.weight": "pytorch_model_00020-of-00041.bin",
+    "transformer.h.18.mlp.down_proj.bias": "pytorch_model_00020-of-00041.bin",
+    "transformer.h.18.mlp.down_proj.weight": "pytorch_model_00020-of-00041.bin",
+    "transformer.h.18.mlp.gate_proj.weight": "pytorch_model_00020-of-00041.bin",
+    "transformer.h.18.mlp.up_proj.weight": "pytorch_model_00020-of-00041.bin",
+    "transformer.h.18.post_attention_layernorm.weight": "pytorch_model_00020-of-00041.bin",
+    "transformer.h.18.self_attention.dense.bias": "pytorch_model_00020-of-00041.bin",
+    "transformer.h.18.self_attention.dense.weight": "pytorch_model_00020-of-00041.bin",
+    "transformer.h.18.self_attention.key_value.weight": "pytorch_model_00020-of-00041.bin",
+    "transformer.h.18.self_attention.query.weight": "pytorch_model_00020-of-00041.bin",
+    "transformer.h.19.input_layernorm.weight": "pytorch_model_00021-of-00041.bin",
+    "transformer.h.19.mlp.down_proj.bias": "pytorch_model_00021-of-00041.bin",
+    "transformer.h.19.mlp.down_proj.weight": "pytorch_model_00021-of-00041.bin",
+    "transformer.h.19.mlp.gate_proj.weight": "pytorch_model_00021-of-00041.bin",
+    "transformer.h.19.mlp.up_proj.weight": "pytorch_model_00021-of-00041.bin",
+    "transformer.h.19.post_attention_layernorm.weight": "pytorch_model_00021-of-00041.bin",
+    "transformer.h.19.self_attention.dense.bias": "pytorch_model_00021-of-00041.bin",
+    "transformer.h.19.self_attention.dense.weight": "pytorch_model_00021-of-00041.bin",
+    "transformer.h.19.self_attention.key_value.weight": "pytorch_model_00021-of-00041.bin",
+    "transformer.h.19.self_attention.query.weight": "pytorch_model_00021-of-00041.bin",
+    "transformer.h.2.input_layernorm.weight": "pytorch_model_00004-of-00041.bin",
+    "transformer.h.2.mlp.down_proj.bias": "pytorch_model_00004-of-00041.bin",
+    "transformer.h.2.mlp.down_proj.weight": "pytorch_model_00004-of-00041.bin",
+    "transformer.h.2.mlp.gate_proj.weight": "pytorch_model_00004-of-00041.bin",
+    "transformer.h.2.mlp.up_proj.weight": "pytorch_model_00004-of-00041.bin",
+    "transformer.h.2.post_attention_layernorm.weight": "pytorch_model_00004-of-00041.bin",
+    "transformer.h.2.self_attention.dense.bias": "pytorch_model_00004-of-00041.bin",
+    "transformer.h.2.self_attention.dense.weight": "pytorch_model_00004-of-00041.bin",
+    "transformer.h.2.self_attention.key_value.weight": "pytorch_model_00004-of-00041.bin",
+    "transformer.h.2.self_attention.query.weight": "pytorch_model_00004-of-00041.bin",
+    "transformer.h.20.input_layernorm.weight": "pytorch_model_00022-of-00041.bin",
+    "transformer.h.20.mlp.down_proj.bias": "pytorch_model_00022-of-00041.bin",
+    "transformer.h.20.mlp.down_proj.weight": "pytorch_model_00022-of-00041.bin",
+    "transformer.h.20.mlp.gate_proj.weight": "pytorch_model_00022-of-00041.bin",
+    "transformer.h.20.mlp.up_proj.weight": "pytorch_model_00022-of-00041.bin",
+    "transformer.h.20.post_attention_layernorm.weight": "pytorch_model_00022-of-00041.bin",
+    "transformer.h.20.self_attention.dense.bias": "pytorch_model_00022-of-00041.bin",
+    "transformer.h.20.self_attention.dense.weight": "pytorch_model_00022-of-00041.bin",
+    "transformer.h.20.self_attention.key_value.weight": "pytorch_model_00022-of-00041.bin",
+    "transformer.h.20.self_attention.query.weight": "pytorch_model_00022-of-00041.bin",
+    "transformer.h.21.input_layernorm.weight": "pytorch_model_00023-of-00041.bin",
+    "transformer.h.21.mlp.down_proj.bias": "pytorch_model_00023-of-00041.bin",
+    "transformer.h.21.mlp.down_proj.weight": "pytorch_model_00023-of-00041.bin",
+    "transformer.h.21.mlp.gate_proj.weight": "pytorch_model_00023-of-00041.bin",
+    "transformer.h.21.mlp.up_proj.weight": "pytorch_model_00023-of-00041.bin",
+    "transformer.h.21.post_attention_layernorm.weight": "pytorch_model_00023-of-00041.bin",
+    "transformer.h.21.self_attention.dense.bias": "pytorch_model_00023-of-00041.bin",
+    "transformer.h.21.self_attention.dense.weight": "pytorch_model_00023-of-00041.bin",
+    "transformer.h.21.self_attention.key_value.weight": "pytorch_model_00023-of-00041.bin",
+    "transformer.h.21.self_attention.query.weight": "pytorch_model_00023-of-00041.bin",
+    "transformer.h.22.input_layernorm.weight": "pytorch_model_00024-of-00041.bin",
+    "transformer.h.22.mlp.down_proj.bias": "pytorch_model_00024-of-00041.bin",
+    "transformer.h.22.mlp.down_proj.weight": "pytorch_model_00024-of-00041.bin",
+    "transformer.h.22.mlp.gate_proj.weight": "pytorch_model_00024-of-00041.bin",
+    "transformer.h.22.mlp.up_proj.weight": "pytorch_model_00024-of-00041.bin",
+    "transformer.h.22.post_attention_layernorm.weight": "pytorch_model_00024-of-00041.bin",
+    "transformer.h.22.self_attention.dense.bias": "pytorch_model_00024-of-00041.bin",
+    "transformer.h.22.self_attention.dense.weight": "pytorch_model_00024-of-00041.bin",
+    "transformer.h.22.self_attention.key_value.weight": "pytorch_model_00024-of-00041.bin",
+    "transformer.h.22.self_attention.query.weight": "pytorch_model_00024-of-00041.bin",
+    "transformer.h.23.input_layernorm.weight": "pytorch_model_00025-of-00041.bin",
+    "transformer.h.23.mlp.down_proj.bias": "pytorch_model_00025-of-00041.bin",
+    "transformer.h.23.mlp.down_proj.weight": "pytorch_model_00025-of-00041.bin",
+    "transformer.h.23.mlp.gate_proj.weight": "pytorch_model_00025-of-00041.bin",
+    "transformer.h.23.mlp.up_proj.weight": "pytorch_model_00025-of-00041.bin",
+    "transformer.h.23.post_attention_layernorm.weight": "pytorch_model_00025-of-00041.bin",
+    "transformer.h.23.self_attention.dense.bias": "pytorch_model_00025-of-00041.bin",
+    "transformer.h.23.self_attention.dense.weight": "pytorch_model_00025-of-00041.bin",
+    "transformer.h.23.self_attention.key_value.weight": "pytorch_model_00025-of-00041.bin",
+    "transformer.h.23.self_attention.query.weight": "pytorch_model_00025-of-00041.bin",
+    "transformer.h.24.input_layernorm.weight": "pytorch_model_00026-of-00041.bin",
+    "transformer.h.24.mlp.down_proj.bias": "pytorch_model_00026-of-00041.bin",
+    "transformer.h.24.mlp.down_proj.weight": "pytorch_model_00026-of-00041.bin",
+    "transformer.h.24.mlp.gate_proj.weight": "pytorch_model_00026-of-00041.bin",
+    "transformer.h.24.mlp.up_proj.weight": "pytorch_model_00026-of-00041.bin",
+    "transformer.h.24.post_attention_layernorm.weight": "pytorch_model_00026-of-00041.bin",
+    "transformer.h.24.self_attention.dense.bias": "pytorch_model_00026-of-00041.bin",
+    "transformer.h.24.self_attention.dense.weight": "pytorch_model_00026-of-00041.bin",
+    "transformer.h.24.self_attention.key_value.weight": "pytorch_model_00026-of-00041.bin",
+    "transformer.h.24.self_attention.query.weight": "pytorch_model_00026-of-00041.bin",
+    "transformer.h.25.input_layernorm.weight": "pytorch_model_00027-of-00041.bin",
+    "transformer.h.25.mlp.down_proj.bias": "pytorch_model_00027-of-00041.bin",
+    "transformer.h.25.mlp.down_proj.weight": "pytorch_model_00027-of-00041.bin",
+    "transformer.h.25.mlp.gate_proj.weight": "pytorch_model_00027-of-00041.bin",
+    "transformer.h.25.mlp.up_proj.weight": "pytorch_model_00027-of-00041.bin",
+    "transformer.h.25.post_attention_layernorm.weight": "pytorch_model_00027-of-00041.bin",
+    "transformer.h.25.self_attention.dense.bias": "pytorch_model_00027-of-00041.bin",
+    "transformer.h.25.self_attention.dense.weight": "pytorch_model_00027-of-00041.bin",
+    "transformer.h.25.self_attention.key_value.weight": "pytorch_model_00027-of-00041.bin",
+    "transformer.h.25.self_attention.query.weight": "pytorch_model_00027-of-00041.bin",
+    "transformer.h.26.input_layernorm.weight": "pytorch_model_00028-of-00041.bin",
+    "transformer.h.26.mlp.down_proj.bias": "pytorch_model_00028-of-00041.bin",
+    "transformer.h.26.mlp.down_proj.weight": "pytorch_model_00028-of-00041.bin",
+    "transformer.h.26.mlp.gate_proj.weight": "pytorch_model_00028-of-00041.bin",
+    "transformer.h.26.mlp.up_proj.weight": "pytorch_model_00028-of-00041.bin",
+    "transformer.h.26.post_attention_layernorm.weight": "pytorch_model_00028-of-00041.bin",
+    "transformer.h.26.self_attention.dense.bias": "pytorch_model_00028-of-00041.bin",
+    "transformer.h.26.self_attention.dense.weight": "pytorch_model_00028-of-00041.bin",
+    "transformer.h.26.self_attention.key_value.weight": "pytorch_model_00028-of-00041.bin",
+    "transformer.h.26.self_attention.query.weight": "pytorch_model_00028-of-00041.bin",
+    "transformer.h.27.input_layernorm.weight": "pytorch_model_00029-of-00041.bin",
+    "transformer.h.27.mlp.down_proj.bias": "pytorch_model_00029-of-00041.bin",
+    "transformer.h.27.mlp.down_proj.weight": "pytorch_model_00029-of-00041.bin",
+    "transformer.h.27.mlp.gate_proj.weight": "pytorch_model_00029-of-00041.bin",
+    "transformer.h.27.mlp.up_proj.weight": "pytorch_model_00029-of-00041.bin",
+    "transformer.h.27.post_attention_layernorm.weight": "pytorch_model_00029-of-00041.bin",
+    "transformer.h.27.self_attention.dense.bias": "pytorch_model_00029-of-00041.bin",
+    "transformer.h.27.self_attention.dense.weight": "pytorch_model_00029-of-00041.bin",
+    "transformer.h.27.self_attention.key_value.weight": "pytorch_model_00029-of-00041.bin",
+    "transformer.h.27.self_attention.query.weight": "pytorch_model_00029-of-00041.bin",
+    "transformer.h.28.input_layernorm.weight": "pytorch_model_00030-of-00041.bin",
+    "transformer.h.28.mlp.down_proj.bias": "pytorch_model_00030-of-00041.bin",
+    "transformer.h.28.mlp.down_proj.weight": "pytorch_model_00030-of-00041.bin",
+    "transformer.h.28.mlp.gate_proj.weight": "pytorch_model_00030-of-00041.bin",
+    "transformer.h.28.mlp.up_proj.weight": "pytorch_model_00030-of-00041.bin",
+    "transformer.h.28.post_attention_layernorm.weight": "pytorch_model_00030-of-00041.bin",
+    "transformer.h.28.self_attention.dense.bias": "pytorch_model_00030-of-00041.bin",
+    "transformer.h.28.self_attention.dense.weight": "pytorch_model_00030-of-00041.bin",
+    "transformer.h.28.self_attention.key_value.weight": "pytorch_model_00030-of-00041.bin",
+    "transformer.h.28.self_attention.query.weight": "pytorch_model_00030-of-00041.bin",
+    "transformer.h.29.input_layernorm.weight": "pytorch_model_00031-of-00041.bin",
+    "transformer.h.29.mlp.down_proj.bias": "pytorch_model_00031-of-00041.bin",
+    "transformer.h.29.mlp.down_proj.weight": "pytorch_model_00031-of-00041.bin",
+    "transformer.h.29.mlp.gate_proj.weight": "pytorch_model_00031-of-00041.bin",
+    "transformer.h.29.mlp.up_proj.weight": "pytorch_model_00031-of-00041.bin",
+    "transformer.h.29.post_attention_layernorm.weight": "pytorch_model_00031-of-00041.bin",
+    "transformer.h.29.self_attention.dense.bias": "pytorch_model_00031-of-00041.bin",
+    "transformer.h.29.self_attention.dense.weight": "pytorch_model_00031-of-00041.bin",
+    "transformer.h.29.self_attention.key_value.weight": "pytorch_model_00031-of-00041.bin",
+    "transformer.h.29.self_attention.query.weight": "pytorch_model_00031-of-00041.bin",
+    "transformer.h.3.input_layernorm.weight": "pytorch_model_00005-of-00041.bin",
+    "transformer.h.3.mlp.down_proj.bias": "pytorch_model_00005-of-00041.bin",
+    "transformer.h.3.mlp.down_proj.weight": "pytorch_model_00005-of-00041.bin",
+    "transformer.h.3.mlp.gate_proj.weight": "pytorch_model_00005-of-00041.bin",
+    "transformer.h.3.mlp.up_proj.weight": "pytorch_model_00005-of-00041.bin",
+    "transformer.h.3.post_attention_layernorm.weight": "pytorch_model_00005-of-00041.bin",
+    "transformer.h.3.self_attention.dense.bias": "pytorch_model_00005-of-00041.bin",
+    "transformer.h.3.self_attention.dense.weight": "pytorch_model_00005-of-00041.bin",
+    "transformer.h.3.self_attention.key_value.weight": "pytorch_model_00005-of-00041.bin",
+    "transformer.h.3.self_attention.query.weight": "pytorch_model_00005-of-00041.bin",
+    "transformer.h.30.input_layernorm.weight": "pytorch_model_00032-of-00041.bin",
+    "transformer.h.30.mlp.down_proj.bias": "pytorch_model_00032-of-00041.bin",
+    "transformer.h.30.mlp.down_proj.weight": "pytorch_model_00032-of-00041.bin",
+    "transformer.h.30.mlp.gate_proj.weight": "pytorch_model_00032-of-00041.bin",
+    "transformer.h.30.mlp.up_proj.weight": "pytorch_model_00032-of-00041.bin",
+    "transformer.h.30.post_attention_layernorm.weight": "pytorch_model_00032-of-00041.bin",
+    "transformer.h.30.self_attention.dense.bias": "pytorch_model_00032-of-00041.bin",
+    "transformer.h.30.self_attention.dense.weight": "pytorch_model_00032-of-00041.bin",
+    "transformer.h.30.self_attention.key_value.weight": "pytorch_model_00032-of-00041.bin",
+    "transformer.h.30.self_attention.query.weight": "pytorch_model_00032-of-00041.bin",
+    "transformer.h.31.input_layernorm.weight": "pytorch_model_00033-of-00041.bin",
+    "transformer.h.31.mlp.down_proj.bias": "pytorch_model_00033-of-00041.bin",
+    "transformer.h.31.mlp.down_proj.weight": "pytorch_model_00033-of-00041.bin",
+    "transformer.h.31.mlp.gate_proj.weight": "pytorch_model_00033-of-00041.bin",
+    "transformer.h.31.mlp.up_proj.weight": "pytorch_model_00033-of-00041.bin",
+    "transformer.h.31.post_attention_layernorm.weight": "pytorch_model_00033-of-00041.bin",
+    "transformer.h.31.self_attention.dense.bias": "pytorch_model_00033-of-00041.bin",
+    "transformer.h.31.self_attention.dense.weight": "pytorch_model_00033-of-00041.bin",
+    "transformer.h.31.self_attention.key_value.weight": "pytorch_model_00033-of-00041.bin",
+    "transformer.h.31.self_attention.query.weight": "pytorch_model_00033-of-00041.bin",
+    "transformer.h.32.input_layernorm.weight": "pytorch_model_00034-of-00041.bin",
+    "transformer.h.32.mlp.down_proj.bias": "pytorch_model_00034-of-00041.bin",
+    "transformer.h.32.mlp.down_proj.weight": "pytorch_model_00034-of-00041.bin",
+    "transformer.h.32.mlp.gate_proj.weight": "pytorch_model_00034-of-00041.bin",
+    "transformer.h.32.mlp.up_proj.weight": "pytorch_model_00034-of-00041.bin",
+    "transformer.h.32.post_attention_layernorm.weight": "pytorch_model_00034-of-00041.bin",
+    "transformer.h.32.self_attention.dense.bias": "pytorch_model_00034-of-00041.bin",
+    "transformer.h.32.self_attention.dense.weight": "pytorch_model_00034-of-00041.bin",
+    "transformer.h.32.self_attention.key_value.weight": "pytorch_model_00034-of-00041.bin",
+    "transformer.h.32.self_attention.query.weight": "pytorch_model_00034-of-00041.bin",
+    "transformer.h.33.input_layernorm.weight": "pytorch_model_00035-of-00041.bin",
+    "transformer.h.33.mlp.down_proj.bias": "pytorch_model_00035-of-00041.bin",
+    "transformer.h.33.mlp.down_proj.weight": "pytorch_model_00035-of-00041.bin",
+    "transformer.h.33.mlp.gate_proj.weight": "pytorch_model_00035-of-00041.bin",
+    "transformer.h.33.mlp.up_proj.weight": "pytorch_model_00035-of-00041.bin",
+    "transformer.h.33.post_attention_layernorm.weight": "pytorch_model_00035-of-00041.bin",
+    "transformer.h.33.self_attention.dense.bias": "pytorch_model_00035-of-00041.bin",
+    "transformer.h.33.self_attention.dense.weight": "pytorch_model_00035-of-00041.bin",
+    "transformer.h.33.self_attention.key_value.weight": "pytorch_model_00035-of-00041.bin",
+    "transformer.h.33.self_attention.query.weight": "pytorch_model_00035-of-00041.bin",
+    "transformer.h.34.input_layernorm.weight": "pytorch_model_00036-of-00041.bin",
+    "transformer.h.34.mlp.down_proj.bias": "pytorch_model_00036-of-00041.bin",
+    "transformer.h.34.mlp.down_proj.weight": "pytorch_model_00036-of-00041.bin",
+    "transformer.h.34.mlp.gate_proj.weight": "pytorch_model_00036-of-00041.bin",
+    "transformer.h.34.mlp.up_proj.weight": "pytorch_model_00036-of-00041.bin",
+    "transformer.h.34.post_attention_layernorm.weight": "pytorch_model_00036-of-00041.bin",
+    "transformer.h.34.self_attention.dense.bias": "pytorch_model_00036-of-00041.bin",
+    "transformer.h.34.self_attention.dense.weight": "pytorch_model_00036-of-00041.bin",
+    "transformer.h.34.self_attention.key_value.weight": "pytorch_model_00036-of-00041.bin",
+    "transformer.h.34.self_attention.query.weight": "pytorch_model_00036-of-00041.bin",
+    "transformer.h.35.input_layernorm.weight": "pytorch_model_00037-of-00041.bin",
+    "transformer.h.35.mlp.down_proj.bias": "pytorch_model_00037-of-00041.bin",
+    "transformer.h.35.mlp.down_proj.weight": "pytorch_model_00037-of-00041.bin",
+    "transformer.h.35.mlp.gate_proj.weight": "pytorch_model_00037-of-00041.bin",
+    "transformer.h.35.mlp.up_proj.weight": "pytorch_model_00037-of-00041.bin",
+    "transformer.h.35.post_attention_layernorm.weight": "pytorch_model_00037-of-00041.bin",
+    "transformer.h.35.self_attention.dense.bias": "pytorch_model_00037-of-00041.bin",
+    "transformer.h.35.self_attention.dense.weight": "pytorch_model_00037-of-00041.bin",
+    "transformer.h.35.self_attention.key_value.weight": "pytorch_model_00037-of-00041.bin",
+    "transformer.h.35.self_attention.query.weight": "pytorch_model_00037-of-00041.bin",
+    "transformer.h.36.input_layernorm.weight": "pytorch_model_00038-of-00041.bin",
+    "transformer.h.36.mlp.down_proj.bias": "pytorch_model_00038-of-00041.bin",
+    "transformer.h.36.mlp.down_proj.weight": "pytorch_model_00038-of-00041.bin",
+    "transformer.h.36.mlp.gate_proj.weight": "pytorch_model_00038-of-00041.bin",
+    "transformer.h.36.mlp.up_proj.weight": "pytorch_model_00038-of-00041.bin",
+    "transformer.h.36.post_attention_layernorm.weight": "pytorch_model_00038-of-00041.bin",
+    "transformer.h.36.self_attention.dense.bias": "pytorch_model_00038-of-00041.bin",
+    "transformer.h.36.self_attention.dense.weight": "pytorch_model_00038-of-00041.bin",
+    "transformer.h.36.self_attention.key_value.weight": "pytorch_model_00038-of-00041.bin",
+    "transformer.h.36.self_attention.query.weight": "pytorch_model_00038-of-00041.bin",
+    "transformer.h.37.input_layernorm.weight": "pytorch_model_00039-of-00041.bin",
+    "transformer.h.37.mlp.down_proj.bias": "pytorch_model_00039-of-00041.bin",
+    "transformer.h.37.mlp.down_proj.weight": "pytorch_model_00039-of-00041.bin",
+    "transformer.h.37.mlp.gate_proj.weight": "pytorch_model_00039-of-00041.bin",
+    "transformer.h.37.mlp.up_proj.weight": "pytorch_model_00039-of-00041.bin",
+    "transformer.h.37.post_attention_layernorm.weight": "pytorch_model_00039-of-00041.bin",
+    "transformer.h.37.self_attention.dense.bias": "pytorch_model_00039-of-00041.bin",
+    "transformer.h.37.self_attention.dense.weight": "pytorch_model_00039-of-00041.bin",
+    "transformer.h.37.self_attention.key_value.weight": "pytorch_model_00039-of-00041.bin",
+    "transformer.h.37.self_attention.query.weight": "pytorch_model_00039-of-00041.bin",
+    "transformer.h.4.input_layernorm.weight": "pytorch_model_00006-of-00041.bin",
+    "transformer.h.4.mlp.down_proj.bias": "pytorch_model_00006-of-00041.bin",
+    "transformer.h.4.mlp.down_proj.weight": "pytorch_model_00006-of-00041.bin",
+    "transformer.h.4.mlp.gate_proj.weight": "pytorch_model_00006-of-00041.bin",
+    "transformer.h.4.mlp.up_proj.weight": "pytorch_model_00006-of-00041.bin",
+    "transformer.h.4.post_attention_layernorm.weight": "pytorch_model_00006-of-00041.bin",
+    "transformer.h.4.self_attention.dense.bias": "pytorch_model_00006-of-00041.bin",
+    "transformer.h.4.self_attention.dense.weight": "pytorch_model_00006-of-00041.bin",
+    "transformer.h.4.self_attention.key_value.weight": "pytorch_model_00006-of-00041.bin",
+    "transformer.h.4.self_attention.query.weight": "pytorch_model_00006-of-00041.bin",
+    "transformer.h.5.input_layernorm.weight": "pytorch_model_00007-of-00041.bin",
+    "transformer.h.5.mlp.down_proj.bias": "pytorch_model_00007-of-00041.bin",
+    "transformer.h.5.mlp.down_proj.weight": "pytorch_model_00007-of-00041.bin",
+    "transformer.h.5.mlp.gate_proj.weight": "pytorch_model_00007-of-00041.bin",
+    "transformer.h.5.mlp.up_proj.weight": "pytorch_model_00007-of-00041.bin",
+    "transformer.h.5.post_attention_layernorm.weight": "pytorch_model_00007-of-00041.bin",
+    "transformer.h.5.self_attention.dense.bias": "pytorch_model_00007-of-00041.bin",
+    "transformer.h.5.self_attention.dense.weight": "pytorch_model_00007-of-00041.bin",
+    "transformer.h.5.self_attention.key_value.weight": "pytorch_model_00007-of-00041.bin",
+    "transformer.h.5.self_attention.query.weight": "pytorch_model_00007-of-00041.bin",
+    "transformer.h.6.input_layernorm.weight": "pytorch_model_00008-of-00041.bin",
+    "transformer.h.6.mlp.down_proj.bias": "pytorch_model_00008-of-00041.bin",
+    "transformer.h.6.mlp.down_proj.weight": "pytorch_model_00008-of-00041.bin",
+    "transformer.h.6.mlp.gate_proj.weight": "pytorch_model_00008-of-00041.bin",
+    "transformer.h.6.mlp.up_proj.weight": "pytorch_model_00008-of-00041.bin",
+    "transformer.h.6.post_attention_layernorm.weight": "pytorch_model_00008-of-00041.bin",
+    "transformer.h.6.self_attention.dense.bias": "pytorch_model_00008-of-00041.bin",
+    "transformer.h.6.self_attention.dense.weight": "pytorch_model_00008-of-00041.bin",
+    "transformer.h.6.self_attention.key_value.weight": "pytorch_model_00008-of-00041.bin",
+    "transformer.h.6.self_attention.query.weight": "pytorch_model_00008-of-00041.bin",
+    "transformer.h.7.input_layernorm.weight": "pytorch_model_00009-of-00041.bin",
+    "transformer.h.7.mlp.down_proj.bias": "pytorch_model_00009-of-00041.bin",
+    "transformer.h.7.mlp.down_proj.weight": "pytorch_model_00009-of-00041.bin",
+    "transformer.h.7.mlp.gate_proj.weight": "pytorch_model_00009-of-00041.bin",
+    "transformer.h.7.mlp.up_proj.weight": "pytorch_model_00009-of-00041.bin",
+    "transformer.h.7.post_attention_layernorm.weight": "pytorch_model_00009-of-00041.bin",
+    "transformer.h.7.self_attention.dense.bias": "pytorch_model_00009-of-00041.bin",
+    "transformer.h.7.self_attention.dense.weight": "pytorch_model_00009-of-00041.bin",
+    "transformer.h.7.self_attention.key_value.weight": "pytorch_model_00009-of-00041.bin",
+    "transformer.h.7.self_attention.query.weight": "pytorch_model_00009-of-00041.bin",
+    "transformer.h.8.input_layernorm.weight": "pytorch_model_00010-of-00041.bin",
+    "transformer.h.8.mlp.down_proj.bias": "pytorch_model_00010-of-00041.bin",
+    "transformer.h.8.mlp.down_proj.weight": "pytorch_model_00010-of-00041.bin",
+    "transformer.h.8.mlp.gate_proj.weight": "pytorch_model_00010-of-00041.bin",
+    "transformer.h.8.mlp.up_proj.weight": "pytorch_model_00010-of-00041.bin",
+    "transformer.h.8.post_attention_layernorm.weight": "pytorch_model_00010-of-00041.bin",
+    "transformer.h.8.self_attention.dense.bias": "pytorch_model_00010-of-00041.bin",
+    "transformer.h.8.self_attention.dense.weight": "pytorch_model_00010-of-00041.bin",
+    "transformer.h.8.self_attention.key_value.weight": "pytorch_model_00010-of-00041.bin",
+    "transformer.h.8.self_attention.query.weight": "pytorch_model_00010-of-00041.bin",
+    "transformer.h.9.input_layernorm.weight": "pytorch_model_00011-of-00041.bin",
+    "transformer.h.9.mlp.down_proj.bias": "pytorch_model_00011-of-00041.bin",
+    "transformer.h.9.mlp.down_proj.weight": "pytorch_model_00011-of-00041.bin",
+    "transformer.h.9.mlp.gate_proj.weight": "pytorch_model_00011-of-00041.bin",
+    "transformer.h.9.mlp.up_proj.weight": "pytorch_model_00011-of-00041.bin",
+    "transformer.h.9.post_attention_layernorm.weight": "pytorch_model_00011-of-00041.bin",
+    "transformer.h.9.self_attention.dense.bias": "pytorch_model_00011-of-00041.bin",
+    "transformer.h.9.self_attention.dense.weight": "pytorch_model_00011-of-00041.bin",
+    "transformer.h.9.self_attention.key_value.weight": "pytorch_model_00011-of-00041.bin",
+    "transformer.h.9.self_attention.query.weight": "pytorch_model_00011-of-00041.bin",
+    "transformer.ln_f.weight": "pytorch_model_00040-of-00041.bin",
+    "transformer.word_embeddings.weight": "pytorch_model_00001-of-00041.bin"
+  }
+}