update model

checkpoint-2000/config.json

@@ -0,0 +1,31 @@
+{
+  "_name_or_path": "THUDM/chatglm-6b",
+  "architectures": [
+    "ChatGLMForConditionalGeneration"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_chatglm.ChatGLMConfig",
+    "AutoModel": "modeling_chatglm.ChatGLMForConditionalGeneration",
+    "AutoModelForSeq2SeqLM": "modeling_chatglm.ChatGLMForConditionalGeneration"
+  },
+  "bos_token_id": 130004,
+  "eos_token_id": 130005,
+  "gmask_token_id": 130001,
+  "hidden_size": 4096,
+  "inner_hidden_size": 16384,
+  "layernorm_epsilon": 1e-05,
+  "mask_token_id": 130000,
+  "max_sequence_length": 2048,
+  "model_type": "chatglm",
+  "num_attention_heads": 32,
+  "num_layers": 28,
+  "pad_token_id": 3,
+  "position_encoding_2d": true,
+  "pre_seq_len": null,
+  "prefix_projection": false,
+  "quantization_bit": 0,
+  "torch_dtype": "float16",
+  "transformers_version": "4.27.1",
+  "use_cache": true,
+  "vocab_size": 130528
+}

checkpoint-2000/configuration_chatglm.py

@@ -0,0 +1,103 @@
+""" ChatGLM model configuration """
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class ChatGLMConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`~ChatGLMModel`].
+    It is used to instantiate an ChatGLM model according to the specified arguments, defining the model
+    architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of
+    the ChatGLM-6B [THUDM/ChatGLM-6B](https://huggingface.co/THUDM/chatglm-6b) architecture.
+
+    Configuration objects inherit from  [`PretrainedConfig`] and can be used
+    to control the model outputs. Read the documentation from  [`PretrainedConfig`]
+    for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 150528):
+            Vocabulary size of the ChatGLM-6B model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`~ChatGLMModel`] or
+            [`~TFChatGLMModel`].
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 28):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        inner_hidden_size (`int`, *optional*, defaults to 16384):
+            Dimension of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        max_sequence_length (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with.
+            Typically set this to something large just in case (e.g., 512 or 1024 or 2048).
+        layernorm_epsilon (`float`, *optional*, defaults to 1e-5):
+            The epsilon used by the layer normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether the model should return the last key/values attentions (not used by all models).
+        Example:
+
+    ```python
+    >>> from configuration_chatglm import ChatGLMConfig
+    >>> from modeling_chatglm import ChatGLMModel
+
+    >>> # Initializing a ChatGLM-6B THUDM/ChatGLM-6B style configuration
+    >>> configuration = ChatGLMConfig()
+
+    >>> # Initializing a model from the THUDM/ChatGLM-6B style configuration
+    >>> model = ChatGLMModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```
+"""
+    model_type = "chatglm"
+
+    def __init__(
+            self,
+            vocab_size=150528,
+            hidden_size=4096,
+            num_layers=28,
+            num_attention_heads=32,
+            layernorm_epsilon=1e-5,
+            use_cache=False,
+            bos_token_id=150004,
+            eos_token_id=150005,
+            mask_token_id=150000,
+            gmask_token_id=150001,
+            pad_token_id=0,
+            max_sequence_length=2048,
+            inner_hidden_size=16384,
+            position_encoding_2d=True,
+            quantization_bit=0,
+            pre_seq_len=None,
+            prefix_projection=False,
+            **kwargs
+    ):
+        self.num_layers = num_layers
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_attention_heads = num_attention_heads
+        self.max_sequence_length = max_sequence_length
+        self.layernorm_epsilon = layernorm_epsilon
+        self.inner_hidden_size = inner_hidden_size
+        self.use_cache = use_cache
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.pad_token_id = pad_token_id
+        self.mask_token_id = mask_token_id
+        self.gmask_token_id = gmask_token_id
+        self.position_encoding_2d = position_encoding_2d
+        self.quantization_bit = quantization_bit
+        self.pre_seq_len = pre_seq_len
+        self.prefix_projection = prefix_projection
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            **kwargs
+        )

checkpoint-2000/generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 130004,
+  "eos_token_id": 130005,
+  "pad_token_id": 3,
+  "transformers_version": "4.27.1"
+}

checkpoint-2000/ice_text.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e974d9a69c242ce014c88c2b26089270f6198f3c0b700a887666cd3e816f17e
+size 2706249

checkpoint-2000/latest

@@ -0,0 +1 @@
+global_step2000

checkpoint-2000/modeling_chatglm.py

@@ -0,0 +1,1435 @@
+""" PyTorch ChatGLM model. """
+
+import math
+import copy
+import os
+import warnings
+import re
+import sys
+
+import torch
+import torch.utils.checkpoint
+import torch.nn.functional as F
+from torch import nn
+from torch.nn import CrossEntropyLoss, LayerNorm
+from torch.nn.utils import skip_init
+from typing import Optional, Tuple, Union, List, Callable, Dict, Any
+
+from transformers.utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    BaseModelOutputWithPastAndCrossAttentions,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from transformers.generation.logits_process import LogitsProcessor
+from transformers.generation.utils import LogitsProcessorList, StoppingCriteriaList, GenerationConfig, ModelOutput
+
+from .configuration_chatglm import ChatGLMConfig
+
+# flags required to enable jit fusion kernels
+
+if sys.platform != 'darwin':
+    torch._C._jit_set_profiling_mode(False)
+    torch._C._jit_set_profiling_executor(False)
+    torch._C._jit_override_can_fuse_on_cpu(True)
+    torch._C._jit_override_can_fuse_on_gpu(True)
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "THUDM/ChatGLM-6B"
+_CONFIG_FOR_DOC = "ChatGLM6BConfig"
+
+CHATGLM_6B_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "THUDM/chatglm-6b",
+    # See all ChatGLM-6B models at https://huggingface.co/models?filter=chatglm
+]
+
+
+class InvalidScoreLogitsProcessor(LogitsProcessor):
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        if torch.isnan(scores).any() or torch.isinf(scores).any():
+            scores.zero_()
+            scores[..., 5] = 5e4
+        return scores
+
+
+def load_tf_weights_in_chatglm_6b(model, config, tf_checkpoint_path):
+    """Load tf checkpoints in a pytorch model."""
+    try:
+        import re
+
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error(
+            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions."
+        )
+        raise
+    tf_path = os.path.abspath(tf_checkpoint_path)
+    logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        logger.info(f"Loading TF weight {name} with shape {shape}")
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array)
+
+    for name, array in zip(names, arrays):
+        name = name.split("/")
+        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
+        # which are not required for using pretrained model
+        if any(
+                n in ["adam_v", "adam_m", "AdamWeightDecayOptimizer", "AdamWeightDecayOptimizer_1", "global_step"]
+                for n in name
+        ):
+            logger.info(f"Skipping {'/'.join(name)}")
+            continue
+        pointer = model
+        for m_name in name:
+            if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
+                scope_names = re.split(r"_(\d+)", m_name)
+            else:
+                scope_names = [m_name]
+            if scope_names[0] == "kernel" or scope_names[0] == "gamma":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
+                pointer = getattr(pointer, "bias")
+            elif scope_names[0] == "output_weights":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "squad":
+                pointer = getattr(pointer, "classifier")
+            else:
+                try:
+                    pointer = getattr(pointer, scope_names[0])
+                except AttributeError:
+                    logger.info(f"Skipping {'/'.join(name)}")
+                    continue
+            if len(scope_names) >= 2:
+                num = int(scope_names[1])
+                pointer = pointer[num]
+        if m_name[-11:] == "_embeddings":
+            pointer = getattr(pointer, "weight")
+        elif m_name == "kernel":
+            array = np.transpose(array)
+        try:
+            assert (
+                    pointer.shape == array.shape
+            ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        logger.info(f"Initialize PyTorch weight {name}")
+        pointer.data = torch.from_numpy(array)
+    return model
+
+
+class PrefixEncoder(torch.nn.Module):
+    """
+    The torch.nn model to encode the prefix
+    Input shape: (batch-size, prefix-length)
+    Output shape: (batch-size, prefix-length, 2*layers*hidden)
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.prefix_projection = config.prefix_projection
+        if self.prefix_projection:
+            # Use a two-layer MLP to encode the prefix
+            self.embedding = torch.nn.Embedding(config.pre_seq_len, config.hidden_size)
+            self.trans = torch.nn.Sequential(
+                torch.nn.Linear(config.hidden_size, config.hidden_size),
+                torch.nn.Tanh(),
+                torch.nn.Linear(config.hidden_size, config.num_layers * config.hidden_size * 2)
+            )
+        else:
+            self.embedding = torch.nn.Embedding(config.pre_seq_len, config.num_layers * config.hidden_size * 2)
+
+    def forward(self, prefix: torch.Tensor):
+        if self.prefix_projection:
+            prefix_tokens = self.embedding(prefix)
+            past_key_values = self.trans(prefix_tokens)
+        else:
+            past_key_values = self.embedding(prefix)
+        return past_key_values
+
+
+@torch.jit.script
+def gelu_impl(x):
+    """OpenAI's gelu implementation."""
+    return 0.5 * x * (1.0 + torch.tanh(0.7978845608028654 * x *
+                                       (1.0 + 0.044715 * x * x)))
+
+
+def gelu(x):
+    return gelu_impl(x)
+
+
+class RotaryEmbedding(torch.nn.Module):
+    def __init__(self, dim, base=10000, precision=torch.half, learnable=False):
+        super().__init__()
+        inv_freq = 1. / (base ** (torch.arange(0, dim, 2).float() / dim))
+        inv_freq = inv_freq.half()
+        self.learnable = learnable
+        if learnable:
+            self.inv_freq = torch.nn.Parameter(inv_freq)
+            self.max_seq_len_cached = None
+        else:
+            self.register_buffer('inv_freq', inv_freq)
+            self.max_seq_len_cached = None
+            self.cos_cached = None
+            self.sin_cached = None
+        self.precision = precision
+
+    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys,
+                              error_msgs):
+        pass
+
+    def forward(self, x, seq_dim=1, seq_len=None):
+        if seq_len is None:
+            seq_len = x.shape[seq_dim]
+        if self.max_seq_len_cached is None or (seq_len > self.max_seq_len_cached):
+            self.max_seq_len_cached = None if self.learnable else seq_len
+            t = torch.arange(seq_len, device=x.device, dtype=self.inv_freq.dtype)
+            freqs = torch.einsum('i,j->ij', t, self.inv_freq)
+            # Different from paper, but it uses a different permutation in order to obtain the same calculation
+            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
+            if self.precision == torch.bfloat16:
+                emb = emb.float()
+
+            # [sx, 1 (b * np), hn]
+            cos_cached = emb.cos()[:, None, :]
+            sin_cached = emb.sin()[:, None, :]
+            if self.precision == torch.bfloat16:
+                cos_cached = cos_cached.bfloat16()
+                sin_cached = sin_cached.bfloat16()
+            if self.learnable:
+                return cos_cached, sin_cached
+            self.cos_cached, self.sin_cached = cos_cached, sin_cached
+        return self.cos_cached[:seq_len, ...], self.sin_cached[:seq_len, ...]
+
+    def _apply(self, fn):
+        if self.cos_cached is not None:
+            self.cos_cached = fn(self.cos_cached)
+        if self.sin_cached is not None:
+            self.sin_cached = fn(self.sin_cached)
+        return super()._apply(fn)
+
+
+def rotate_half(x):
+    x1, x2 = x[..., :x.shape[-1] // 2], x[..., x.shape[-1] // 2:]
+    return torch.cat((-x2, x1), dim=x1.ndim - 1)  # dim=-1 triggers a bug in earlier torch versions
+
+
+@torch.jit.script
+def apply_rotary_pos_emb_index(q, k, cos, sin, position_id):
+    # position_id: [sq, b], q, k: [sq, b, np, hn], cos: [sq, 1, hn] -> [sq, b, 1, hn]
+    cos, sin = F.embedding(position_id, cos.squeeze(1)).unsqueeze(2), \
+        F.embedding(position_id, sin.squeeze(1)).unsqueeze(2)
+    q, k = (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin)
+    return q, k
+
+
+def attention_fn(
+        self,
+        query_layer,
+        key_layer,
+        value_layer,
+        attention_mask,
+        hidden_size_per_partition,
+        layer_id,
+        layer_past=None,
+        scaling_attention_score=True,
+        use_cache=False,
+):
+    if layer_past is not None:
+        past_key, past_value = layer_past[0], layer_past[1]
+        key_layer = torch.cat((past_key, key_layer), dim=0)
+        value_layer = torch.cat((past_value, value_layer), dim=0)
+
+    # seqlen, batch, num_attention_heads, hidden_size_per_attention_head
+    seq_len, b, nh, hidden_size = key_layer.shape
+
+    if use_cache:
+        present = (key_layer, value_layer)
+    else:
+        present = None
+
+    query_key_layer_scaling_coeff = float(layer_id + 1)
+    if scaling_attention_score:
+        query_layer = query_layer / (math.sqrt(hidden_size) * query_key_layer_scaling_coeff)
+
+    # ===================================
+    # Raw attention scores. [b, np, s, s]
+    # ===================================
+
+    # [b, np, sq, sk]
+    output_size = (query_layer.size(1), query_layer.size(2), query_layer.size(0), key_layer.size(0))
+
+    # [sq, b, np, hn] -> [sq, b * np, hn]
+    query_layer = query_layer.view(output_size[2], output_size[0] * output_size[1], -1)
+    # [sk, b, np, hn] -> [sk, b * np, hn]
+    key_layer = key_layer.view(output_size[3], output_size[0] * output_size[1], -1)
+
+    matmul_result = torch.zeros(
+        1, 1, 1,
+        dtype=query_layer.dtype,
+        device=query_layer.device,
+    )
+
+    matmul_result = torch.baddbmm(
+        matmul_result,
+        query_layer.transpose(0, 1),  # [b * np, sq, hn]
+        key_layer.transpose(0, 1).transpose(1, 2),  # [b * np, hn, sk]
+        beta=0.0,
+        alpha=1.0,
+    )
+
+    # change view to [b, np, sq, sk]
+    attention_scores = matmul_result.view(*output_size)
+
+    if self.scale_mask_softmax:
+        self.scale_mask_softmax.scale = query_key_layer_scaling_coeff
+        attention_probs = self.scale_mask_softmax(attention_scores, attention_mask.contiguous())
+    else:
+        if not (attention_mask == 0).all():
+            # if auto-regressive, skip
+            attention_scores.masked_fill_(attention_mask, -10000.0)
+        dtype = attention_scores.dtype
+        attention_scores = attention_scores.float()
+        attention_scores = attention_scores * query_key_layer_scaling_coeff
+
+        attention_probs = F.softmax(attention_scores, dim=-1)
+
+        attention_probs = attention_probs.type(dtype)
+
+    # =========================
+    # Context layer. [sq, b, hp]
+    # =========================
+
+    # value_layer -> context layer.
+    # [sk, b, np, hn] --> [b, np, sq, hn]
+
+    # context layer shape: [b, np, sq, hn]
+    output_size = (value_layer.size(1), value_layer.size(2), query_layer.size(0), value_layer.size(3))
+
+    # change view [sk, b * np, hn]
+    value_layer = value_layer.view(value_layer.size(0), output_size[0] * output_size[1], -1)
+
+    # change view [b * np, sq, sk]
+    attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1)
+
+    # matmul: [b * np, sq, hn]
+    context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1))
+
+    # change view [b, np, sq, hn]
+    context_layer = context_layer.view(*output_size)
+
+    # [b, np, sq, hn] --> [sq, b, np, hn]
+    context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
+
+    # [sq, b, np, hn] --> [sq, b, hp]
+    new_context_layer_shape = context_layer.size()[:-2] + (hidden_size_per_partition,)
+    context_layer = context_layer.view(*new_context_layer_shape)
+
+    outputs = (context_layer, present, attention_probs)
+
+    return outputs
+
+
+def default_init(cls, *args, **kwargs):
+    return cls(*args, **kwargs)
+
+
+class SelfAttention(torch.nn.Module):
+    def __init__(self, hidden_size, num_attention_heads,
+                 layer_id, hidden_size_per_attention_head=None, bias=True,
+                 params_dtype=torch.float, position_encoding_2d=True, empty_init=True):
+        if empty_init:
+            init_method = skip_init
+        else:
+            init_method = default_init
+        super(SelfAttention, self).__init__()
+
+        self.layer_id = layer_id
+        self.hidden_size = hidden_size
+        self.hidden_size_per_partition = hidden_size
+        self.num_attention_heads = num_attention_heads
+        self.num_attention_heads_per_partition = num_attention_heads
+        self.position_encoding_2d = position_encoding_2d
+        self.rotary_emb = RotaryEmbedding(
+            self.hidden_size // (self.num_attention_heads * 2)
+            if position_encoding_2d
+            else self.hidden_size // self.num_attention_heads,
+            base=10000,
+            precision=torch.half,
+            learnable=False,
+        )
+
+        self.scale_mask_softmax = None
+
+        if hidden_size_per_attention_head is None:
+            self.hidden_size_per_attention_head = hidden_size // num_attention_heads
+        else:
+            self.hidden_size_per_attention_head = hidden_size_per_attention_head
+
+        self.inner_hidden_size = num_attention_heads * self.hidden_size_per_attention_head
+
+        # Strided linear layer.
+        self.query_key_value = init_method(
+            torch.nn.Linear,
+            hidden_size,
+            3 * self.inner_hidden_size,
+            bias=bias,
+            dtype=params_dtype,
+        )
+
+        self.dense = init_method(
+            torch.nn.Linear,
+            self.inner_hidden_size,
+            hidden_size,
+            bias=bias,
+            dtype=params_dtype,
+        )
+
+    @staticmethod
+    def attention_mask_func(attention_scores, attention_mask):
+        attention_scores.masked_fill_(attention_mask, -10000.0)
+        return attention_scores
+
+    def split_tensor_along_last_dim(self, tensor, num_partitions,
+                                    contiguous_split_chunks=False):
+        """Split a tensor along its last dimension.
+        Arguments:
+            tensor: input tensor.
+            num_partitions: number of partitions to split the tensor
+            contiguous_split_chunks: If True, make each chunk contiguous
+                                    in memory.
+        """
+        # Get the size and dimension.
+        last_dim = tensor.dim() - 1
+        last_dim_size = tensor.size()[last_dim] // num_partitions
+        # Split.
+        tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
+        # Note: torch.split does not create contiguous tensors by default.
+        if contiguous_split_chunks:
+            return tuple(chunk.contiguous() for chunk in tensor_list)
+
+        return tensor_list
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            position_ids,
+            attention_mask: torch.Tensor,
+            layer_id,
+            layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+            use_cache: bool = False,
+            output_attentions: bool = False,
+    ):
+        """
+        hidden_states: [seq_len, batch, hidden_size]
+        attention_mask: [(1, 1), seq_len, seq_len]
+        """
+
+        # [seq_len, batch, 3 * hidden_size]
+        mixed_raw_layer = self.query_key_value(hidden_states)
+
+        # [seq_len, batch, 3 * hidden_size] --> [seq_len, batch, num_attention_heads, 3 * hidden_size_per_attention_head]
+        new_tensor_shape = mixed_raw_layer.size()[:-1] + (
+            self.num_attention_heads_per_partition,
+            3 * self.hidden_size_per_attention_head,
+        )
+        mixed_raw_layer = mixed_raw_layer.view(*new_tensor_shape)
+
+        # [seq_len, batch, num_attention_heads, hidden_size_per_attention_head]
+        (query_layer, key_layer, value_layer) = self.split_tensor_along_last_dim(mixed_raw_layer, 3)
+
+        if self.position_encoding_2d:
+            q1, q2 = query_layer.chunk(2, dim=(query_layer.ndim - 1))
+            k1, k2 = key_layer.chunk(2, dim=(key_layer.ndim - 1))
+            cos, sin = self.rotary_emb(q1, seq_len=position_ids.max() + 1)
+            position_ids, block_position_ids = position_ids[:, 0, :].transpose(0, 1).contiguous(), \
+                position_ids[:, 1, :].transpose(0, 1).contiguous()
+            q1, k1 = apply_rotary_pos_emb_index(q1, k1, cos, sin, position_ids)
+            q2, k2 = apply_rotary_pos_emb_index(q2, k2, cos, sin, block_position_ids)
+            query_layer = torch.concat([q1, q2], dim=(q1.ndim - 1))
+            key_layer = torch.concat([k1, k2], dim=(k1.ndim - 1))
+        else:
+            position_ids = position_ids.transpose(0, 1)
+            cos, sin = self.rotary_emb(value_layer, seq_len=position_ids.max() + 1)
+            # [seq_len, batch, num_attention_heads, hidden_size_per_attention_head]
+            query_layer, key_layer = apply_rotary_pos_emb_index(query_layer, key_layer, cos, sin, position_ids)
+
+        # [seq_len, batch, hidden_size]
+        context_layer, present, attention_probs = attention_fn(
+            self=self,
+            query_layer=query_layer,
+            key_layer=key_layer,
+            value_layer=value_layer,
+            attention_mask=attention_mask,
+            hidden_size_per_partition=self.hidden_size_per_partition,
+            layer_id=layer_id,
+            layer_past=layer_past,
+            use_cache=use_cache
+        )
+
+        output = self.dense(context_layer)
+
+        outputs = (output, present)
+
+        if output_attentions:
+            outputs += (attention_probs,)
+
+        return outputs  # output, present, attention_probs
+
+
+class GEGLU(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.activation_fn = F.gelu
+
+    def forward(self, x):
+        # dim=-1 breaks in jit for pt<1.10
+        x1, x2 = x.chunk(2, dim=(x.ndim - 1))
+        return x1 * self.activation_fn(x2)
+
+
+class GLU(torch.nn.Module):
+    def __init__(self, hidden_size, inner_hidden_size=None,
+                 layer_id=None, bias=True, activation_func=gelu, params_dtype=torch.float, empty_init=True):
+        super(GLU, self).__init__()
+        if empty_init:
+            init_method = skip_init
+        else:
+            init_method = default_init
+        self.layer_id = layer_id
+        self.activation_func = activation_func
+
+        # Project to 4h.
+        self.hidden_size = hidden_size
+        if inner_hidden_size is None:
+            inner_hidden_size = 4 * hidden_size
+        self.inner_hidden_size = inner_hidden_size
+        self.dense_h_to_4h = init_method(
+            torch.nn.Linear,
+            self.hidden_size,
+            self.inner_hidden_size,
+            bias=bias,
+            dtype=params_dtype,
+        )
+        # Project back to h.
+        self.dense_4h_to_h = init_method(
+            torch.nn.Linear,
+            self.inner_hidden_size,
+            self.hidden_size,
+            bias=bias,
+            dtype=params_dtype,
+        )
+
+    def forward(self, hidden_states):
+        """
+        hidden_states: [seq_len, batch, hidden_size]
+        """
+
+        # [seq_len, batch, inner_hidden_size]
+        intermediate_parallel = self.dense_h_to_4h(hidden_states)
+
+        intermediate_parallel = self.activation_func(intermediate_parallel)
+
+        output = self.dense_4h_to_h(intermediate_parallel)
+
+        return output
+
+
+class GLMBlock(torch.nn.Module):
+    def __init__(
+            self,
+            hidden_size,
+            num_attention_heads,
+            layernorm_epsilon,
+            layer_id,
+            inner_hidden_size=None,
+            hidden_size_per_attention_head=None,
+            layernorm=LayerNorm,
+            use_bias=True,
+            params_dtype=torch.float,
+            num_layers=28,
+            position_encoding_2d=True,
+            empty_init=True
+    ):
+        super(GLMBlock, self).__init__()
+        # Set output layer initialization if not provided.
+
+        self.layer_id = layer_id
+
+        # Layernorm on the input data.
+        self.input_layernorm = layernorm(hidden_size, eps=layernorm_epsilon)
+
+        self.position_encoding_2d = position_encoding_2d
+
+        # Self attention.
+        self.attention = SelfAttention(
+            hidden_size,
+            num_attention_heads,
+            layer_id,
+            hidden_size_per_attention_head=hidden_size_per_attention_head,
+            bias=use_bias,
+            params_dtype=params_dtype,
+            position_encoding_2d=self.position_encoding_2d,
+            empty_init=empty_init
+        )
+
+        # Layernorm on the input data.
+        self.post_attention_layernorm = layernorm(hidden_size, eps=layernorm_epsilon)
+
+        self.num_layers = num_layers
+
+        # GLU
+        self.mlp = GLU(
+            hidden_size,
+            inner_hidden_size=inner_hidden_size,
+            bias=use_bias,
+            layer_id=layer_id,
+            params_dtype=params_dtype,
+            empty_init=empty_init
+        )
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            position_ids,
+            attention_mask: torch.Tensor,
+            layer_id,
+            layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+            use_cache: bool = False,
+            output_attentions: bool = False,
+    ):
+        """
+        hidden_states: [seq_len, batch, hidden_size]
+        attention_mask: [(1, 1), seq_len, seq_len]
+        """
+
+        # Layer norm at the begining of the transformer layer.
+        # [seq_len, batch, hidden_size]
+        attention_input = self.input_layernorm(hidden_states)
+
+        # Self attention.
+        attention_outputs = self.attention(
+            attention_input,
+            position_ids,
+            attention_mask=attention_mask,
+            layer_id=layer_id,
+            layer_past=layer_past,
+            use_cache=use_cache,
+            output_attentions=output_attentions
+        )
+
+        attention_output = attention_outputs[0]
+
+        outputs = attention_outputs[1:]
+
+        # Residual connection.
+        alpha = (2 * self.num_layers) ** 0.5
+        hidden_states = attention_input * alpha + attention_output
+
+        mlp_input = self.post_attention_layernorm(hidden_states)
+
+        # MLP.
+        mlp_output = self.mlp(mlp_input)
+
+        # Second residual connection.
+        output = mlp_input * alpha + mlp_output
+
+        if use_cache:
+            outputs = (output,) + outputs
+        else:
+            outputs = (output,) + outputs[1:]
+
+        return outputs  # hidden_states, present, attentions
+
+
+class ChatGLMPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and
+    a simple interface for downloading and loading pretrained models.
+    """
+
+    is_parallelizable = False
+    supports_gradient_checkpointing = True
+    config_class = ChatGLMConfig
+    base_model_prefix = "transformer"
+    _no_split_modules = ["GLMBlock"]
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module: nn.Module):
+        """Initialize the weights."""
+        return
+
+    def get_masks(self, input_ids, device):
+        batch_size, seq_length = input_ids.shape
+        context_lengths = [seq.tolist().index(self.config.bos_token_id) for seq in input_ids]
+        attention_mask = torch.ones((batch_size, seq_length, seq_length), device=device)
+        attention_mask.tril_()
+        for i, context_length in enumerate(context_lengths):
+            attention_mask[i, :, :context_length] = 1
+        attention_mask.unsqueeze_(1)
+        attention_mask = (attention_mask < 0.5).bool()
+
+        return attention_mask
+
+    def get_position_ids(self, input_ids, mask_positions, device, use_gmasks=None):
+        batch_size, seq_length = input_ids.shape
+        if use_gmasks is None:
+            use_gmasks = [False] * batch_size
+        context_lengths = [seq.tolist().index(self.config.bos_token_id) for seq in input_ids]
+        if self.position_encoding_2d:
+            position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1)
+            for i, context_length in enumerate(context_lengths):
+                position_ids[i, context_length:] = mask_positions[i]
+            block_position_ids = [torch.cat((
+                torch.zeros(context_length, dtype=torch.long, device=device),
+                torch.arange(seq_length - context_length, dtype=torch.long, device=device) + 1
+            )) for context_length in context_lengths]
+            block_position_ids = torch.stack(block_position_ids, dim=0)
+            position_ids = torch.stack((position_ids, block_position_ids), dim=1)
+        else:
+            position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1)
+            for i, context_length in enumerate(context_lengths):
+                if not use_gmasks[i]:
+                    position_ids[i, context_length:] = mask_positions[i]
+
+        return position_ids
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, ChatGLMModel):
+            module.gradient_checkpointing = value
+
+
+CHATGLM_6B_START_DOCSTRING = r"""
+    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general
+    usage and behavior.
+
+    Parameters:
+        config ([`~ChatGLM6BConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+CHATGLM_6B_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`ChatGLM6BTokenizer`].
+            See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0, 1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings.
+            Selected in the range `[0, config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert *input_ids* indices into associated vectors
+            than the model's internal embedding lookup matrix.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare ChatGLM-6B Model transformer outputting raw hidden-states without any specific head on top.",
+    CHATGLM_6B_START_DOCSTRING,
+)
+class ChatGLMModel(ChatGLMPreTrainedModel):
+    """
+
+    The model can behave as an encoder (with only self-attention) as well
+    as a decoder, in which case a layer of cross-attention is added between
+    the self-attention layers, following the architecture described in [Attention is
+    all you need](https://arxiv.org/abs/1706.03762) by Ashish Vaswani,
+    Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
+
+    To behave as an decoder the model needs to be initialized with the
+    `is_decoder` argument of the configuration set to `True`.
+    To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder`
+    argument and `add_cross_attention` set to `True`; an
+    `encoder_hidden_states` is then expected as an input to the forward pass.
+    """
+
+    def __init__(self, config: ChatGLMConfig, empty_init=True):
+        super().__init__(config)
+        if empty_init:
+            init_method = skip_init
+        else:
+            init_method = default_init
+        # recording parameters
+        self.max_sequence_length = config.max_sequence_length
+        self.hidden_size = config.hidden_size
+        self.params_dtype = torch.half
+        self.num_attention_heads = config.num_attention_heads
+        self.vocab_size = config.vocab_size
+        self.num_layers = config.num_layers
+        self.layernorm_epsilon = config.layernorm_epsilon
+        self.inner_hidden_size = config.inner_hidden_size
+        self.hidden_size_per_attention_head = self.hidden_size // self.num_attention_heads
+        self.position_encoding_2d = config.position_encoding_2d
+        self.pre_seq_len = config.pre_seq_len
+        self.prefix_projection = config.prefix_projection
+
+        self.word_embeddings = init_method(
+            torch.nn.Embedding,
+            num_embeddings=self.vocab_size, embedding_dim=self.hidden_size,
+            dtype=self.params_dtype
+        )
+        self.gradient_checkpointing = False
+
+        def get_layer(layer_id):
+            return GLMBlock(
+                self.hidden_size,
+                self.num_attention_heads,
+                self.layernorm_epsilon,
+                layer_id,
+                inner_hidden_size=self.inner_hidden_size,
+                hidden_size_per_attention_head=self.hidden_size_per_attention_head,
+                layernorm=LayerNorm,
+                use_bias=True,
+                params_dtype=self.params_dtype,
+                position_encoding_2d=self.position_encoding_2d,
+                empty_init=empty_init
+            )
+
+        self.layers = torch.nn.ModuleList(
+            [get_layer(layer_id) for layer_id in range(self.num_layers)]
+        )
+
+        # Final layer norm before output.
+        self.final_layernorm = LayerNorm(self.hidden_size, eps=self.layernorm_epsilon)
+
+        if self.pre_seq_len is not None:
+            for param in self.parameters():
+                param.requires_grad = False
+            self.prefix_tokens = torch.arange(self.pre_seq_len).long()
+            self.prefix_encoder = PrefixEncoder(config)
+            self.dropout = torch.nn.Dropout(0.1)
+
+            # total_params = sum(p.numel() for p in self.parameters())
+            # trainable_params = sum(p.numel() for p in self.parameters() if p.requires_grad)
+            # print("Using p-tuning v2: # trainable_params = {} / {}".format(trainable_params, total_params))
+
+    def get_input_embeddings(self):
+        return self.word_embeddings
+
+    def set_input_embeddings(self, new_embeddings: torch.Tensor):
+        self.word_embeddings = new_embeddings
+
+    def get_prompt(self, batch_size, device, dtype=torch.half):
+        prefix_tokens = self.prefix_tokens.unsqueeze(0).expand(batch_size, -1).to(device)
+        past_key_values = self.prefix_encoder(prefix_tokens).type(dtype)
+        past_key_values = past_key_values.view(
+            batch_size,
+            self.pre_seq_len,
+            self.num_layers * 2,
+            self.num_attention_heads,
+            self.hidden_size // self.num_attention_heads
+        )
+        # seq_len, b, nh, hidden_size
+        past_key_values = self.dropout(past_key_values)
+        past_key_values = past_key_values.permute([2, 1, 0, 3, 4]).split(2)
+        # past_key_values = [(v[0], v[1]) for v in past_key_values]
+        return past_key_values
+
+    @add_start_docstrings_to_model_forward(CHATGLM_6B_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPastAndCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+            self,
+            input_ids: Optional[torch.LongTensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+            inputs_embeds: Optional[torch.LongTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor, ...], BaseModelOutputWithPast]:
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape[:2]
+        elif inputs_embeds is not None:
+            batch_size, seq_length = inputs_embeds.shape[:2]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+
+        if past_key_values is None:
+            if self.pre_seq_len is not None:
+                past_key_values = self.get_prompt(batch_size=input_ids.shape[0], device=input_ids.device,
+                                                  dtype=inputs_embeds.dtype)
+            else:
+                past_key_values = tuple([None] * len(self.layers))
+
+            if attention_mask is None:
+                attention_mask = self.get_masks(
+                    input_ids,
+                    device=input_ids.device
+                )
+
+
+            if position_ids is None:
+                MASK, gMASK = self.config.mask_token_id, self.config.gmask_token_id
+                seqs = input_ids.tolist()
+
+                mask_positions, use_gmasks = [], []
+                for seq in seqs:
+                    mask_token = gMASK if gMASK in seq else MASK
+                    use_gmask = mask_token == gMASK
+                    mask_positions.append(seq.index(mask_token))
+                    use_gmasks.append(use_gmask)
+
+                position_ids = self.get_position_ids(
+                    input_ids,
+                    mask_positions=mask_positions,
+                    device=input_ids.device,
+                    use_gmasks=use_gmasks
+                )
+
+        if self.pre_seq_len is not None and attention_mask is not None:
+            prefix_attention_mask = torch.ones(batch_size, 1, input_ids.size(-1), self.pre_seq_len).to(
+                attention_mask.device)
+            prefix_attention_mask = (prefix_attention_mask < 0.5).bool()
+            attention_mask = torch.cat((prefix_attention_mask, attention_mask), dim=3)
+
+        # [seq_len, batch, hidden_size]
+        hidden_states = inputs_embeds.transpose(0, 1)
+
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+
+        if attention_mask is None:
+            attention_mask = torch.zeros(1, 1, device=input_ids.device).bool()
+        else:
+            attention_mask = attention_mask.to(hidden_states.device)
+
+        for i, layer in enumerate(self.layers):
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+            layer_past = past_key_values[i]
+
+            if self.gradient_checkpointing and self.training:
+                layer_ret = torch.utils.checkpoint.checkpoint(
+                    layer,
+                    hidden_states,
+                    position_ids,
+                    attention_mask,
+                    torch.tensor(i),
+                    layer_past,
+                    use_cache,
+                    output_attentions
+                )
+            else:
+                layer_ret = layer(
+                    hidden_states,
+                    position_ids=position_ids,
+                    attention_mask=attention_mask,
+                    layer_id=torch.tensor(i),
+                    layer_past=layer_past,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions
+                )
+
+            hidden_states = layer_ret[0]
+
+            if use_cache:
+                presents = presents + (layer_ret[1],)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_ret[2 if use_cache else 1],)
+
+        # Final layer norm.
+        hidden_states = self.final_layernorm(hidden_states)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
+    def __init__(self, config: ChatGLMConfig, empty_init=True):
+        super().__init__(config)
+        if empty_init:
+            init_method = skip_init
+        else:
+            init_method = default_init
+
+        # self.hidden_size = config.hidden_size
+        # self.params_dtype = torch.half
+        # self.vocab_size = config.vocab_size
+        self.max_sequence_length = config.max_sequence_length
+
+        self.position_encoding_2d = config.position_encoding_2d
+
+        self.transformer = ChatGLMModel(config, empty_init=empty_init)
+
+        self.lm_head = init_method(
+            nn.Linear,
+            config.hidden_size,
+            config.vocab_size,
+            bias=False,
+            dtype=torch.half
+        )
+
+        self.config = config
+
+        self.quantized = False
+
+        if self.config.quantization_bit:
+            self.quantize(self.config.quantization_bit, empty_init=True)
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def _update_model_kwargs_for_generation(
+        self,
+        outputs: ModelOutput,
+        model_kwargs: Dict[str, Any],
+        is_encoder_decoder: bool = False,
+        standardize_cache_format: bool = False,
+    ) -> Dict[str, Any]:
+        # update past_key_values
+        model_kwargs["past_key_values"] = self._extract_past_from_model_output(
+            outputs, standardize_cache_format=standardize_cache_format
+        )
+
+        # update attention mask
+        if "attention_mask" in model_kwargs:
+            attention_mask = model_kwargs["attention_mask"]
+            if attention_mask is not None and attention_mask.dtype == torch.bool:
+                attention_mask = torch.cat(
+                    [attention_mask, attention_mask.new_ones((*attention_mask.shape[:3], 1))], dim=3)
+                new_attention_mask = attention_mask[:, :, -1:].clone()
+                new_attention_mask[..., -1] = False
+                model_kwargs["attention_mask"] = torch.cat(
+                    [attention_mask, new_attention_mask], dim=2
+                )
+
+        # update position ids
+        if "position_ids" in model_kwargs:
+            position_ids = model_kwargs["position_ids"]
+            new_position_id = position_ids[..., -1:].clone()
+            new_position_id[:, 1, :] += 1
+            model_kwargs["position_ids"] = torch.cat(
+                [position_ids, new_position_id], dim=-1
+            )
+
+        return model_kwargs
+
+    def prepare_inputs_for_generation(
+            self,
+            input_ids: torch.LongTensor,
+            past: Optional[torch.Tensor] = None,
+            past_key_values: Optional[torch.Tensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.Tensor] = None,
+            **kwargs
+    ) -> dict:
+        batch_size, seq_length = input_ids.shape
+        MASK, gMASK = self.config.mask_token_id, self.config.gmask_token_id
+        seqs = input_ids.tolist()
+        mask_positions, use_gmasks = [], []
+        for seq in seqs:
+            mask_token = gMASK if gMASK in seq else MASK
+            use_gmask = mask_token == gMASK
+            mask_positions.append(seq.index(mask_token))
+            use_gmasks.append(use_gmask)
+
+        # only last token for input_ids if past is not None
+        if past is not None or past_key_values is not None:
+            last_token = input_ids[:, -1].unsqueeze(-1)
+            if attention_mask is not None and attention_mask.dtype == torch.bool:
+                attention_mask = attention_mask[:, :, -1:]
+            else:
+                attention_mask = None
+            if position_ids is not None:
+                position_ids = position_ids[..., -1:]
+            else:
+                context_lengths = [seq.index(self.config.bos_token_id) for seq in seqs]
+                if self.position_encoding_2d:
+                    position_ids = torch.tensor(
+                        [[mask_position, seq_length - context_length] for mask_position, context_length in
+                         zip(mask_positions, context_lengths)], dtype=torch.long, device=input_ids.device).unsqueeze(-1)
+                else:
+                    position_ids = torch.tensor([mask_position for mask_position in mask_positions], dtype=torch.long,
+                                                device=input_ids.device).unsqueeze(-1)
+
+            if past is None:
+                past = past_key_values
+            return {
+                "input_ids": last_token,
+                "past_key_values": past,
+                "position_ids": position_ids,
+                "attention_mask": attention_mask
+            }
+        else:
+            if attention_mask is not None and attention_mask.dtype != torch.bool:
+                logger.warning_once(f"The dtype of attention mask ({attention_mask.dtype}) is not bool")
+                attention_mask = None
+            if attention_mask is None:
+                attention_mask = self.get_masks(
+                    input_ids,
+                    device=input_ids.device
+                )
+            if position_ids is None:
+                position_ids = self.get_position_ids(
+                    input_ids,
+                    device=input_ids.device,
+                    mask_positions=mask_positions,
+                    use_gmasks=use_gmasks
+                )
+
+            return {
+                "input_ids": input_ids,
+                "past_key_values": past,
+                "position_ids": position_ids,
+                "attention_mask": attention_mask
+            }
+
+    def forward(
+            self,
+            input_ids: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.Tensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_values: Optional[Tuple[torch.FloatTensor]] = None,
+            inputs_embeds: Optional[torch.Tensor] = None,
+            labels: Optional[torch.Tensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ):
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = transformer_outputs[0]
+
+        lm_logits = self.lm_head(hidden_states).permute(1, 0, 2).contiguous()
+
+        loss = None
+        if labels is not None:
+            lm_logits = lm_logits.to(torch.float32)
+
+            # Shift so that tokens < n predict n
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss(ignore_index=-100)
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+            lm_logits = lm_logits.to(hidden_states.dtype)
+            loss = loss.to(hidden_states.dtype)
+
+        if not return_dict:
+            output = (lm_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(
+            past: Tuple[Tuple[torch.Tensor, torch.Tensor], ...], beam_idx: torch.LongTensor
+    ) -> Tuple[Tuple[torch.Tensor, torch.Tensor], ...]:
+        """
+        This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
+        [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
+        beam_idx at every generation step.
+
+        Output shares the same memory storage as `past`.
+        """
+        return tuple(
+            (
+                layer_past[0].index_select(1, beam_idx.to(layer_past[0].device)),
+                layer_past[1].index_select(1, beam_idx.to(layer_past[1].device)),
+            )
+            for layer_past in past
+        )
+
+    def process_response(self, response):
+        response = response.strip()
+        response = response.replace("[[训练时间]]", "2023年")
+        punkts = [
+            [",", "，"],
+            ["!", "！"],
+            [":", "："],
+            [";", "；"],
+            ["\?", "？"],
+        ]
+        for item in punkts:
+            response = re.sub(r"([\u4e00-\u9fff])%s" % item[0], r"\1%s" % item[1], response)
+            response = re.sub(r"%s([\u4e00-\u9fff])" % item[0], r"%s\1" % item[1], response)
+        return response
+
+    @torch.no_grad()
+    def chat(self, tokenizer, query: str, history: List[Tuple[str, str]] = None, max_length: int = 2048, num_beams=1,
+             do_sample=True, top_p=0.7, temperature=0.95, logits_processor=None, **kwargs):
+        if history is None:
+            history = []
+        if logits_processor is None:
+            logits_processor = LogitsProcessorList()
+        logits_processor.append(InvalidScoreLogitsProcessor())
+        gen_kwargs = {"max_length": max_length, "num_beams": num_beams, "do_sample": do_sample, "top_p": top_p,
+                      "temperature": temperature, "logits_processor": logits_processor, **kwargs}
+        if not history:
+            prompt = query
+        else:
+            prompt = ""
+            for i, (old_query, response) in enumerate(history):
+                prompt += "[Round {}]\n问：{}\n答：{}\n".format(i, old_query, response)
+            prompt += "[Round {}]\n问：{}\n答：".format(len(history), query)
+        inputs = tokenizer([prompt], return_tensors="pt")
+        inputs = inputs.to(self.device)
+        outputs = self.generate(**inputs, **gen_kwargs)
+        outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):]
+        response = tokenizer.decode(outputs)
+        response = self.process_response(response)
+        history = history + [(query, response)]
+        return response, history
+
+    @torch.no_grad()
+    def stream_chat(self, tokenizer, query: str, history: List[Tuple[str, str]] = None, max_length: int = 2048,
+                    do_sample=True, top_p=0.7, temperature=0.95, logits_processor=None, **kwargs):
+        if history is None:
+            history = []
+        if logits_processor is None:
+            logits_processor = LogitsProcessorList()
+        logits_processor.append(InvalidScoreLogitsProcessor())
+        gen_kwargs = {"max_length": max_length, "do_sample": do_sample, "top_p": top_p,
+                      "temperature": temperature, "logits_processor": logits_processor, **kwargs}
+        if not history:
+            prompt = query
+        else:
+            prompt = ""
+            for i, (old_query, response) in enumerate(history):
+                prompt += "[Round {}]\n问：{}\n答：{}\n".format(i, old_query, response)
+            prompt += "[Round {}]\n问：{}\n答：".format(len(history), query)
+        inputs = tokenizer([prompt], return_tensors="pt")
+        inputs = inputs.to(self.device)
+        for outputs in self.stream_generate(**inputs, **gen_kwargs):
+            outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):]
+            response = tokenizer.decode(outputs)
+            response = self.process_response(response)
+            new_history = history + [(query, response)]
+            yield response, new_history
+
+    @torch.no_grad()
+    def stream_generate(
+            self,
+            input_ids,
+            generation_config: Optional[GenerationConfig] = None,
+            logits_processor: Optional[LogitsProcessorList] = None,
+            stopping_criteria: Optional[StoppingCriteriaList] = None,
+            prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
+            **kwargs,
+    ):
+        batch_size, input_ids_seq_length = input_ids.shape[0], input_ids.shape[-1]
+
+        if generation_config is None:
+            generation_config = self.generation_config
+        generation_config = copy.deepcopy(generation_config)
+        model_kwargs = generation_config.update(**kwargs)
+        bos_token_id, eos_token_id = generation_config.bos_token_id, generation_config.eos_token_id
+
+        if isinstance(eos_token_id, int):
+            eos_token_id = [eos_token_id]
+
+        has_default_max_length = kwargs.get("max_length") is None and generation_config.max_length is not None
+        if has_default_max_length and generation_config.max_new_tokens is None:
+            warnings.warn(
+                f"Using `max_length`'s default ({generation_config.max_length}) to control the generation length. "
+                "This behaviour is deprecated and will be removed from the config in v5 of Transformers -- we"
+                " recommend using `max_new_tokens` to control the maximum length of the generation.",
+                UserWarning,
+            )
+        elif generation_config.max_new_tokens is not None:
+            generation_config.max_length = generation_config.max_new_tokens + input_ids_seq_length
+            if not has_default_max_length:
+                logger.warn(
+                    f"Both `max_new_tokens` (={generation_config.max_new_tokens}) and `max_length`(="
+                    f"{generation_config.max_length}) seem to have been set. `max_new_tokens` will take precedence. "
+                    "Please refer to the documentation for more information. "
+                    "(https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)",
+                    UserWarning,
+                )
+
+        if input_ids_seq_length >= generation_config.max_length:
+            input_ids_string = "decoder_input_ids" if self.config.is_encoder_decoder else "input_ids"
+            logger.warning(
+                f"Input length of {input_ids_string} is {input_ids_seq_length}, but `max_length` is set to"
+                f" {generation_config.max_length}. This can lead to unexpected behavior. You should consider"
+                " increasing `max_new_tokens`."
+            )
+
+        # 2. Set generation parameters if not already defined
+        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
+        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
+
+        logits_processor = self._get_logits_processor(
+            generation_config=generation_config,
+            input_ids_seq_length=input_ids_seq_length,
+            encoder_input_ids=input_ids,
+            prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
+            logits_processor=logits_processor,
+        )
+
+        stopping_criteria = self._get_stopping_criteria(
+            generation_config=generation_config, stopping_criteria=stopping_criteria
+        )
+        logits_warper = self._get_logits_warper(generation_config)
+
+        unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1)
+        scores = None
+        while True:
+            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)
+            # forward pass to get next token
+            outputs = self(
+                **model_inputs,
+                return_dict=True,
+                output_attentions=False,
+                output_hidden_states=False,
+            )
+
+            next_token_logits = outputs.logits[:, -1, :]
+
+            # pre-process distribution
+            next_token_scores = logits_processor(input_ids, next_token_logits)
+            next_token_scores = logits_warper(input_ids, next_token_scores)
+
+            # sample
+            probs = nn.functional.softmax(next_token_scores, dim=-1)
+            if generation_config.do_sample:
+                next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)
+            else:
+                next_tokens = torch.argmax(probs, dim=-1)
+
+            # update generated ids, model inputs, and length for next step
+            input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
+            model_kwargs = self._update_model_kwargs_for_generation(
+                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
+            )
+            unfinished_sequences = unfinished_sequences.mul((sum(next_tokens != i for i in eos_token_id)).long())
+
+            # stop when each sentence is finished, or if we exceed the maximum length
+            if unfinished_sequences.max() == 0 or stopping_criteria(input_ids, scores):
+                break
+            yield input_ids
+
+    def quantize(self, bits: int, empty_init=False, **kwargs):
+        if bits == 0:
+            return
+
+        from .quantization import quantize
+
+        if self.quantized:
+            logger.info("Already quantized.")
+            return self
+
+        self.quantized = True
+
+        self.config.quantization_bit = bits
+
+        self.transformer = quantize(self.transformer, bits, empty_init=empty_init, **kwargs)
+        return self

checkpoint-2000/pytorch_model-00001-of-00002.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e48d1194d23b4d83ab134a113cb070b6aa05b5d3588e7e5f858b677c5d43ede5
+size 12346621179

checkpoint-2000/pytorch_model-00002-of-00002.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42c3b591bed51fc1383a69a4c9d0218e14b02f913526955de448509464fec6b8
+size 12346585635

checkpoint-2000/pytorch_model.bin.index.json

@@ -0,0 +1,375 @@
+{
+  "metadata": {
+    "total_size": 13415859968
+  },
+  "weight_map": {
+    "lm_head.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.final_layernorm.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.final_layernorm.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.0.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.0.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.0.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.0.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.0.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.0.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.0.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.0.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.0.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.0.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.0.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.0.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.0.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.1.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.1.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.1.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.1.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.1.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.1.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.1.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.1.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.1.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.1.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.1.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.1.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.1.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.10.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.10.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.10.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.10.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.10.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.10.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.10.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.10.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.10.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.10.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.10.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.10.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.10.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.11.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.11.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.11.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.11.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.11.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.11.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.11.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.11.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.11.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.11.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.11.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.11.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.11.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.12.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.12.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.12.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.12.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.12.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.12.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.12.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.12.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.12.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.12.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.12.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.12.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.12.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.13.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.13.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.13.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.13.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.13.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.13.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.13.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.13.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.13.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.13.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.13.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.13.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.13.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.14.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.14.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.14.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.14.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.14.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.14.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.14.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.14.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.14.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.14.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.14.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.14.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.14.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.15.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.15.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.15.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.15.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.15.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.15.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.15.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.15.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.15.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.15.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.15.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.15.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.15.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.16.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.16.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.16.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.16.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.16.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.16.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.16.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.16.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.16.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.16.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.16.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.16.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.16.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.17.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.17.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.17.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.17.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.17.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.17.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.17.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.17.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.17.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.17.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.17.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.17.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.17.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.18.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.18.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.18.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.18.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.18.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.18.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.18.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.18.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.18.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.18.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.18.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.18.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.18.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.19.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.19.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.19.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.19.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.19.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.19.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.19.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.19.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.19.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.19.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.19.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.19.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.19.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.2.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.2.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.2.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.2.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.2.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.2.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.2.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.2.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.2.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.2.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.2.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.2.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.2.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.20.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.20.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.20.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.20.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.20.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.20.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.20.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.20.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.20.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.20.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.20.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.20.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.20.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.21.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.21.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.21.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.21.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.21.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.21.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.21.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.21.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.21.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.21.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.21.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.21.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.21.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.22.attention.dense.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.22.attention.dense.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.22.attention.query_key_value.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.22.attention.query_key_value.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.22.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.22.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.22.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.22.mlp.dense_4h_to_h.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.22.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.22.mlp.dense_h_to_4h.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.22.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.22.post_attention_layernorm.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.22.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.23.attention.dense.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.23.attention.dense.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.23.attention.query_key_value.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.23.attention.query_key_value.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.23.attention.rotary_emb.inv_freq": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.23.input_layernorm.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.23.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.23.mlp.dense_4h_to_h.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.23.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.23.mlp.dense_h_to_4h.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.23.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.23.post_attention_layernorm.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.23.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.24.attention.dense.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.24.attention.dense.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.24.attention.query_key_value.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.24.attention.query_key_value.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.24.attention.rotary_emb.inv_freq": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.24.input_layernorm.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.24.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.24.mlp.dense_4h_to_h.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.24.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.24.mlp.dense_h_to_4h.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.24.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.24.post_attention_layernorm.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.24.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.25.attention.dense.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.25.attention.dense.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.25.attention.query_key_value.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.25.attention.query_key_value.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.25.attention.rotary_emb.inv_freq": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.25.input_layernorm.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.25.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.25.mlp.dense_4h_to_h.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.25.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.25.mlp.dense_h_to_4h.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.25.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.25.post_attention_layernorm.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.25.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.26.attention.dense.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.26.attention.dense.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.26.attention.query_key_value.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.26.attention.query_key_value.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.26.attention.rotary_emb.inv_freq": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.26.input_layernorm.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.26.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.26.mlp.dense_4h_to_h.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.26.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.26.mlp.dense_h_to_4h.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.26.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.26.post_attention_layernorm.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.26.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.27.attention.dense.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.27.attention.dense.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.27.attention.query_key_value.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.27.attention.query_key_value.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.27.attention.rotary_emb.inv_freq": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.27.input_layernorm.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.27.input_layernorm.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.27.mlp.dense_4h_to_h.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.27.mlp.dense_4h_to_h.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.27.mlp.dense_h_to_4h.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.27.mlp.dense_h_to_4h.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.27.post_attention_layernorm.bias": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.27.post_attention_layernorm.weight": "pytorch_model-00002-of-00002.bin",
+    "transformer.layers.3.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.3.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.3.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.3.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.3.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.3.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.3.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.3.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.3.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.3.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.3.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.3.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.3.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.4.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.4.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.4.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.4.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.4.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.4.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.4.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.4.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.4.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.4.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.4.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.4.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.4.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.5.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.5.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.5.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.5.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.5.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.5.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.5.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.5.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.5.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.5.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.5.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.5.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.5.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.6.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.6.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.6.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.6.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.6.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.6.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.6.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.6.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.6.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.6.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.6.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.6.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.6.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.7.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.7.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.7.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.7.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.7.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.7.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.7.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.7.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.7.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.7.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.7.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.7.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.7.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.8.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.8.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.8.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.8.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.8.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.8.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.8.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.8.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.8.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.8.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.8.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.8.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.8.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.9.attention.dense.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.9.attention.dense.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.9.attention.query_key_value.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.9.attention.query_key_value.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.9.attention.rotary_emb.inv_freq": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.9.input_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.9.input_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.9.mlp.dense_4h_to_h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.9.mlp.dense_4h_to_h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.9.mlp.dense_h_to_4h.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.9.mlp.dense_h_to_4h.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.9.post_attention_layernorm.bias": "pytorch_model-00001-of-00002.bin",
+    "transformer.layers.9.post_attention_layernorm.weight": "pytorch_model-00001-of-00002.bin",
+    "transformer.word_embeddings.weight": "pytorch_model-00001-of-00002.bin"
+  }
+}

checkpoint-2000/quantization.py

@@ -0,0 +1,201 @@
+from torch.nn import Linear
+from torch.nn.parameter import Parameter
+
+import bz2
+import torch
+import base64
+import ctypes
+from transformers.utils import logging
+
+from typing import List
+from functools import partial
+
+logger = logging.get_logger(__name__)
+
+try:
+    from cpm_kernels.kernels.base import LazyKernelCModule, KernelFunction, round_up
+
+    class Kernel:
+        def __init__(self, code: bytes, function_names: List[str]):
+            self.code = code
+            self._function_names = function_names
+            self._cmodule = LazyKernelCModule(self.code)
+
+            for name in self._function_names:
+                setattr(self, name, KernelFunction(self._cmodule, name))
+
+    quantization_code = "$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"
+
+    kernels = Kernel(
+        bz2.decompress(base64.b64decode(quantization_code)),
+        [
+            "int4WeightCompression",
+            "int4WeightExtractionFloat",
+            "int4WeightExtractionHalf",
+            "int8WeightExtractionFloat",
+            "int8WeightExtractionHalf",
+        ],
+    )
+except Exception as exception:
+    kernels = None
+    logger.warning("Failed to load cpm_kernels:" + str(exception))
+
+
+class W8A16Linear(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, inp: torch.Tensor, quant_w: torch.Tensor, scale_w: torch.Tensor, weight_bit_width):
+        ctx.inp_shape = inp.size()
+        ctx.weight_bit_width = weight_bit_width
+        out_features = quant_w.size(0)
+        inp = inp.contiguous().view(-1, inp.size(-1))
+        weight = extract_weight_to_half(quant_w, scale_w, weight_bit_width)
+        ctx.weight_shape = weight.size()
+        output = inp.mm(weight.t())
+        ctx.save_for_backward(inp, quant_w, scale_w)
+        return output.view(*(ctx.inp_shape[:-1] + (out_features,)))
+
+    @staticmethod
+    def backward(ctx, grad_output: torch.Tensor):
+        inp, quant_w, scale_w = ctx.saved_tensors
+        weight = extract_weight_to_half(quant_w, scale_w, ctx.weight_bit_width)
+        grad_output = grad_output.contiguous().view(-1, weight.size(0))
+        grad_input = grad_output.mm(weight)
+        grad_weight = grad_output.t().mm(inp)
+        return grad_input.view(ctx.inp_shape), grad_weight.view(ctx.weight_shape), None, None
+
+
+def compress_int4_weight(weight: torch.Tensor):  # (n, m)
+    with torch.cuda.device(weight.device):
+        n, m = weight.size(0), weight.size(1)
+        assert m % 2 == 0
+        m = m // 2
+        out = torch.empty(n, m, dtype=torch.int8, device="cuda")
+        stream = torch.cuda.current_stream()
+
+        gridDim = (n, 1, 1)
+        blockDim = (min(round_up(m, 32), 1024), 1, 1)
+
+        kernels.int4WeightCompression(
+            gridDim,
+            blockDim,
+            0,
+            stream,
+            [ctypes.c_void_p(weight.data_ptr()), ctypes.c_void_p(out.data_ptr()), ctypes.c_int32(n), ctypes.c_int32(m)],
+        )
+        return out
+
+
+def extract_weight_to_half(weight: torch.Tensor, scale_list: torch.Tensor, source_bit_width: int):
+    if source_bit_width == 8:
+        func = kernels.int8WeightExtractionHalf
+    elif source_bit_width == 4:
+        func = kernels.int4WeightExtractionHalf
+    else:
+        assert False, "Unsupported bit-width"
+
+    with torch.cuda.device(weight.device):
+        n, m = weight.size(0), weight.size(1)
+        out = torch.empty(n, m * (8 // source_bit_width), dtype=torch.half, device="cuda")
+        stream = torch.cuda.current_stream()
+
+        gridDim = (n, 1, 1)
+        blockDim = (min(round_up(m, 32), 1024), 1, 1)
+
+        func(
+            gridDim,
+            blockDim,
+            0,
+            stream,
+            [
+                ctypes.c_void_p(weight.data_ptr()),
+                ctypes.c_void_p(scale_list.data_ptr()),
+                ctypes.c_void_p(out.data_ptr()),
+                ctypes.c_int32(n),
+                ctypes.c_int32(m),
+            ],
+        )
+        return out
+
+
+class QuantizedLinear(Linear):
+    def __init__(self, weight_bit_width: int, weight_tensor=None, bias_tensor=None, empty_init=False, *args, **kwargs):
+        super(QuantizedLinear, self).__init__(*args, **kwargs)
+        self.weight_bit_width = weight_bit_width
+
+        shape = self.weight.shape
+        del self.weight
+
+        if weight_tensor is None or empty_init:
+            self.weight = torch.empty(
+                shape[0], shape[1] * weight_bit_width // 8, dtype=torch.int8, device=kwargs["device"]
+            )
+            self.weight_scale = torch.empty(shape[0], dtype=kwargs["dtype"], device=kwargs["device"])
+        else:
+            self.weight_scale = (weight_tensor.abs().max(dim=-1).values / ((2 ** (weight_bit_width - 1)) - 1)).half()
+            self.weight = torch.round(weight_tensor / self.weight_scale[:, None]).to(torch.int8)
+            if weight_bit_width == 4:
+                self.weight = compress_int4_weight(self.weight)
+
+        self.weight = Parameter(self.weight.to(kwargs["device"]), requires_grad=False)
+        self.weight_scale = Parameter(self.weight_scale.to(kwargs["device"]), requires_grad=False)
+        if bias_tensor is not None:
+            self.bias = Parameter(bias_tensor.to(kwargs["device"]), requires_grad=False)
+        else:
+            self.bias = None
+
+    def forward(self, input):
+        output = W8A16Linear.apply(input, self.weight, self.weight_scale, self.weight_bit_width)
+        if self.bias is not None:
+            output = output + self.bias
+        return output
+
+
+def quantize(model, weight_bit_width, empty_init=False, **kwargs):
+    """Replace fp16 linear with quantized linear"""
+
+    for layer in model.layers:
+        layer.attention.query_key_value = QuantizedLinear(
+            weight_bit_width=weight_bit_width,
+            weight_tensor=layer.attention.query_key_value.weight.to(torch.cuda.current_device()),
+            bias_tensor=layer.attention.query_key_value.bias,
+            in_features=layer.attention.query_key_value.in_features,
+            out_features=layer.attention.query_key_value.out_features,
+            bias=True,
+            dtype=torch.half,
+            device=layer.attention.query_key_value.weight.device,
+            empty_init=empty_init
+        )
+        layer.attention.dense = QuantizedLinear(
+            weight_bit_width=weight_bit_width,
+            weight_tensor=layer.attention.dense.weight.to(torch.cuda.current_device()),
+            bias_tensor=layer.attention.dense.bias,
+            in_features=layer.attention.dense.in_features,
+            out_features=layer.attention.dense.out_features,
+            bias=True,
+            dtype=torch.half,
+            device=layer.attention.dense.weight.device,
+            empty_init=empty_init
+        )
+        layer.mlp.dense_h_to_4h = QuantizedLinear(
+            weight_bit_width=weight_bit_width,
+            weight_tensor=layer.mlp.dense_h_to_4h.weight.to(torch.cuda.current_device()),
+            bias_tensor=layer.mlp.dense_h_to_4h.bias,
+            in_features=layer.mlp.dense_h_to_4h.in_features,
+            out_features=layer.mlp.dense_h_to_4h.out_features,
+            bias=True,
+            dtype=torch.half,
+            device=layer.mlp.dense_h_to_4h.weight.device,
+            empty_init=empty_init
+        )
+        layer.mlp.dense_4h_to_h = QuantizedLinear(
+            weight_bit_width=weight_bit_width,
+            weight_tensor=layer.mlp.dense_4h_to_h.weight.to(torch.cuda.current_device()),
+            bias_tensor=layer.mlp.dense_4h_to_h.bias,
+            in_features=layer.mlp.dense_4h_to_h.in_features,
+            out_features=layer.mlp.dense_4h_to_h.out_features,
+            bias=True,
+            dtype=torch.half,
+            device=layer.mlp.dense_4h_to_h.weight.device,
+            empty_init=empty_init
+        )
+    return model

checkpoint-2000/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e75c96f06b249e57a701db73ce821398e69672027a86d3a44063830602a29ab4
+size 14583

checkpoint-2000/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6dae7f45b6bac644ac207a61f43cba6d4b919a4cac22022bbb02907914422f5d
+size 14583

checkpoint-2000/special_tokens_map.json

@@ -0,0 +1,7 @@
+{
+  "bos_token": "<sop>",
+  "eos_token": "<eop>",
+  "mask_token": "[MASK]",
+  "pad_token": "<pad>",
+  "unk_token": "<unk>"
+}

checkpoint-2000/tokenization_chatglm.py

@@ -0,0 +1,443 @@
+"""Tokenization classes for ChatGLM."""
+from typing import List, Optional, Union
+import os
+
+from transformers.tokenization_utils import PreTrainedTokenizer
+from transformers.utils import logging, PaddingStrategy
+from transformers.tokenization_utils_base import EncodedInput, BatchEncoding
+from typing import Dict
+import sentencepiece as spm
+import numpy as np
+
+logger = logging.get_logger(__name__)
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "THUDM/chatglm-6b": 2048,
+}
+
+
+class TextTokenizer:
+    def __init__(self, model_path):
+        self.sp = spm.SentencePieceProcessor()
+        self.sp.Load(model_path)
+        self.num_tokens = self.sp.vocab_size()
+
+    def encode(self, text):
+        return self.sp.EncodeAsIds(text)
+
+    def decode(self, ids: List[int]):
+        return self.sp.DecodeIds(ids)
+
+    def tokenize(self, text):
+        return self.sp.EncodeAsPieces(text)
+
+    def convert_tokens_to_string(self, tokens):
+        return self.sp.DecodePieces(tokens)
+
+    def convert_tokens_to_ids(self, tokens):
+        return [self.sp.PieceToId(token) for token in tokens]
+
+    def convert_token_to_id(self, token):
+        return self.sp.PieceToId(token)
+
+    def convert_id_to_token(self, idx):
+        return self.sp.IdToPiece(idx)
+
+    def __len__(self):
+        return self.num_tokens
+
+
+class SPTokenizer:
+    def __init__(
+            self,
+            vocab_file,
+            num_image_tokens=20000,
+            max_blank_length=80,
+            byte_fallback=True,
+    ):
+        assert vocab_file is not None
+        self.vocab_file = vocab_file
+        self.num_image_tokens = num_image_tokens
+        self.special_tokens = ["[MASK]", "[gMASK]", "[sMASK]", "<unused_0>", "<sop>", "<eop>", "<ENC>", "<dBLOCK>"]
+        self.max_blank_length = max_blank_length
+        self.byte_fallback = byte_fallback
+        self.text_tokenizer = TextTokenizer(vocab_file)
+
+    def _get_text_tokenizer(self):
+        return self.text_tokenizer
+
+    @staticmethod
+    def get_blank_token(length: int):
+        assert length >= 2
+        return f"<|blank_{length}|>"
+
+    @staticmethod
+    def get_tab_token():
+        return f"<|tab|>"
+
+    @property
+    def num_text_tokens(self):
+        return self.text_tokenizer.num_tokens
+
+    @property
+    def num_tokens(self):
+        return self.num_image_tokens + self.num_text_tokens
+
+    @staticmethod
+    def _encode_whitespaces(text: str, max_len: int = 80):
+        text = text.replace("\t", SPTokenizer.get_tab_token())
+        for i in range(max_len, 1, -1):
+            text = text.replace(" " * i, SPTokenizer.get_blank_token(i))
+        return text
+
+    def _preprocess(self, text: str, linebreak=True, whitespaces=True):
+        if linebreak:
+            text = text.replace("\n", "<n>")
+        if whitespaces:
+            text = self._encode_whitespaces(text, max_len=self.max_blank_length)
+        return text
+
+    def encode(
+            self, text: str, linebreak=True, whitespaces=True, add_dummy_prefix=True
+    ) -> List[int]:
+        """
+        @param text: Text to encode.
+        @param linebreak: Whether to encode newline (\n) in text.
+        @param whitespaces: Whether to encode multiple whitespaces or tab in text, useful for source code encoding.
+        @param special_tokens: Whether to encode special token ([MASK], [gMASK], etc.) in text.
+        @param add_dummy_prefix: Whether to add dummy blank space in the beginning.
+        """
+        text = self._preprocess(text, linebreak, whitespaces)
+        if not add_dummy_prefix:
+            text = "<n>" + text
+        tmp = self._get_text_tokenizer().encode(text)
+        tokens = [x + self.num_image_tokens for x in tmp]
+        return tokens if add_dummy_prefix else tokens[2:]
+
+    def postprocess(self, text):
+        text = text.replace("<n>", "\n")
+        text = text.replace(SPTokenizer.get_tab_token(), "\t")
+        for i in range(2, self.max_blank_length + 1):
+            text = text.replace(self.get_blank_token(i), " " * i)
+        return text
+
+    def decode(self, text_ids: List[int]) -> str:
+        ids = [int(_id) - self.num_image_tokens for _id in text_ids]
+        ids = [_id for _id in ids if _id >= 0]
+        text = self._get_text_tokenizer().decode(ids)
+        text = self.postprocess(text)
+        return text
+
+    def decode_tokens(self, tokens: List[str]) -> str:
+        text = self._get_text_tokenizer().convert_tokens_to_string(tokens)
+        text = self.postprocess(text)
+        return text
+
+    def tokenize(
+            self, text: str, linebreak=True, whitespaces=True, add_dummy_prefix=True
+    ) -> List[str]:
+        """
+        @param text: Text to encode.
+        @param linebreak: Whether to encode newline (\n) in text.
+        @param whitespaces: Whether to encode multiple whitespaces or tab in text, useful for source code encoding.
+        @param special_tokens: Whether to encode special token ([MASK], [gMASK], etc.) in text.
+        @param add_dummy_prefix: Whether to add dummy blank space in the beginning.
+        """
+        text = self._preprocess(text, linebreak, whitespaces)
+        if not add_dummy_prefix:
+            text = "<n>" + text
+        tokens = self._get_text_tokenizer().tokenize(text)
+        return tokens if add_dummy_prefix else tokens[2:]
+
+    def __getitem__(self, x: Union[int, str]):
+        if isinstance(x, int):
+            if x < self.num_image_tokens:
+                return "<image_{}>".format(x)
+            else:
+                return self.text_tokenizer.convert_id_to_token(x - self.num_image_tokens)
+        elif isinstance(x, str):
+            if x.startswith("<image_") and x.endswith(">") and x[7:-1].isdigit():
+                return int(x[7:-1])
+            else:
+                return self.text_tokenizer.convert_token_to_id(x) + self.num_image_tokens
+        else:
+            raise ValueError("The key should be str or int.")
+
+
+class ChatGLMTokenizer(PreTrainedTokenizer):
+    """
+    Construct a ChatGLM tokenizer. Based on byte-level Byte-Pair-Encoding.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+    """
+
+    vocab_files_names = {"vocab_file": "ice_text.model"}
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    model_input_names = ["input_ids", "attention_mask", "position_ids"]
+
+    def __init__(
+            self,
+            vocab_file,
+            do_lower_case=False,
+            remove_space=False,
+            bos_token='<sop>',
+            eos_token='<eop>',
+            end_token='</s>',
+            mask_token='[MASK]',
+            gmask_token='[gMASK]',
+            padding_side="left",
+            pad_token="<pad>",
+            unk_token="<unk>",
+            num_image_tokens=20000,
+            **kwargs
+    ) -> None:
+        super().__init__(
+            do_lower_case=do_lower_case,
+            remove_space=remove_space,
+            padding_side=padding_side,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            end_token=end_token,
+            mask_token=mask_token,
+            gmask_token=gmask_token,
+            pad_token=pad_token,
+            unk_token=unk_token,
+            num_image_tokens=num_image_tokens,
+            **kwargs
+        )
+
+        self.do_lower_case = do_lower_case
+        self.remove_space = remove_space
+        self.vocab_file = vocab_file
+
+        self.bos_token = bos_token
+        self.eos_token = eos_token
+        self.end_token = end_token
+        self.mask_token = mask_token
+        self.gmask_token = gmask_token
+
+        self.sp_tokenizer = SPTokenizer(vocab_file, num_image_tokens=num_image_tokens)
+
+        """ Initialisation """
+
+    @property
+    def gmask_token_id(self) -> Optional[int]:
+        if self.gmask_token is None:
+            return None
+        return self.convert_tokens_to_ids(self.gmask_token)
+
+    @property
+    def end_token_id(self) -> Optional[int]:
+        """
+        `Optional[int]`: Id of the end of context token in the vocabulary. Returns `None` if the token has not been
+        set.
+        """
+        if self.end_token is None:
+            return None
+        return self.convert_tokens_to_ids(self.end_token)
+
+    @property
+    def vocab_size(self):
+        """ Returns vocab size """
+        return self.sp_tokenizer.num_tokens
+
+    def get_vocab(self):
+        """ Returns vocab as a dict """
+        vocab = {self._convert_id_to_token(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def preprocess_text(self, inputs):
+        if self.remove_space:
+            outputs = " ".join(inputs.strip().split())
+        else:
+            outputs = inputs
+
+        if self.do_lower_case:
+            outputs = outputs.lower()
+
+        return outputs
+
+    def _tokenize(self, text, **kwargs):
+        """ Returns a tokenized string. """
+        text = self.preprocess_text(text)
+
+        seq = self.sp_tokenizer.tokenize(text)
+
+        return seq
+
+    def convert_tokens_to_string(self, tokens: List[str]) -> str:
+        return self.sp_tokenizer.decode_tokens(tokens)
+
+    def _decode(
+            self,
+            token_ids: Union[int, List[int]],
+            **kwargs
+    ) -> str:
+        if isinstance(token_ids, int):
+            token_ids = [token_ids]
+        if len(token_ids) == 0:
+            return ""
+        if self.pad_token_id in token_ids:  # remove pad
+            token_ids = list(filter((self.pad_token_id).__ne__, token_ids))
+        return super()._decode(token_ids, **kwargs)
+
+    def _convert_token_to_id(self, token):
+        """ Converts a token (str) in an id using the vocab. """
+        return self.sp_tokenizer[token]
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.sp_tokenizer[index]
+
+    def save_vocabulary(self, save_directory, filename_prefix=None):
+        """
+        Save the vocabulary and special tokens file to a directory.
+
+        Args:
+            save_directory (`str`):
+                The directory in which to save the vocabulary.
+            filename_prefix (`str`, *optional*):
+                An optional prefix to add to the named of the saved files.
+
+        Returns:
+            `Tuple(str)`: Paths to the files saved.
+        """
+        if os.path.isdir(save_directory):
+            vocab_file = os.path.join(
+                save_directory, self.vocab_files_names["vocab_file"]
+            )
+        else:
+            vocab_file = save_directory
+
+        with open(self.vocab_file, 'rb') as fin:
+            proto_str = fin.read()
+
+        with open(vocab_file, "wb") as writer:
+            writer.write(proto_str)
+
+        return (vocab_file,)
+
+    def build_inputs_with_special_tokens(
+            self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. A BERT sequence has the following format:
+
+        - single sequence: `[CLS] X [SEP]`
+        - pair of sequences: `[CLS] A [SEP] B [SEP]`
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+        gmask_id = self.sp_tokenizer[self.gmask_token]
+        eos_id = self.sp_tokenizer[self.eos_token]
+        token_ids_0 = token_ids_0 + [gmask_id, self.sp_tokenizer[self.bos_token]]
+        if token_ids_1 is not None:
+            token_ids_0 = token_ids_0 + token_ids_1 + [eos_id]
+        return token_ids_0
+
+    def _pad(
+            self,
+            encoded_inputs: Union[Dict[str, EncodedInput], BatchEncoding],
+            max_length: Optional[int] = None,
+            padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,
+            pad_to_multiple_of: Optional[int] = None,
+            return_attention_mask: Optional[bool] = None,
+    ) -> dict:
+        """
+        Pad encoded inputs (on left/right and up to predefined length or max length in the batch)
+
+        Args:
+            encoded_inputs:
+                Dictionary of tokenized inputs (`List[int]`) or batch of tokenized inputs (`List[List[int]]`).
+            max_length: maximum length of the returned list and optionally padding length (see below).
+                Will truncate by taking into account the special tokens.
+            padding_strategy: PaddingStrategy to use for padding.
+
+                - PaddingStrategy.LONGEST Pad to the longest sequence in the batch
+                - PaddingStrategy.MAX_LENGTH: Pad to the max length (default)
+                - PaddingStrategy.DO_NOT_PAD: Do not pad
+                The tokenizer padding sides are defined in self.padding_side:
+
+                    - 'left': pads on the left of the sequences
+                    - 'right': pads on the right of the sequences
+            pad_to_multiple_of: (optional) Integer if set will pad the sequence to a multiple of the provided value.
+                This is especially useful to enable the use of Tensor Core on NVIDIA hardware with compute capability
+                `>= 7.5` (Volta).
+            return_attention_mask:
+                (optional) Set to False to avoid returning attention mask (default: set to model specifics)
+        """
+        # Load from model defaults
+        bos_token_id = self.sp_tokenizer[self.bos_token]
+        mask_token_id = self.sp_tokenizer[self.mask_token]
+        gmask_token_id = self.sp_tokenizer[self.gmask_token]
+        assert self.padding_side == "left"
+
+        required_input = encoded_inputs[self.model_input_names[0]]
+        seq_length = len(required_input)
+
+        if padding_strategy == PaddingStrategy.LONGEST:
+            max_length = len(required_input)
+
+        if max_length is not None and pad_to_multiple_of is not None and (max_length % pad_to_multiple_of != 0):
+            max_length = ((max_length // pad_to_multiple_of) + 1) * pad_to_multiple_of
+
+        needs_to_be_padded = padding_strategy != PaddingStrategy.DO_NOT_PAD and len(required_input) != max_length
+
+        # Initialize attention mask if not present.
+        if max_length is not None:
+            if "attention_mask" not in encoded_inputs:
+                if bos_token_id in required_input:
+                    context_length = required_input.index(bos_token_id)
+                else:
+                    context_length = seq_length
+                attention_mask = np.ones((1, seq_length, seq_length))
+                attention_mask = np.tril(attention_mask)
+                attention_mask[:, :, :context_length] = 1
+                attention_mask = np.bool_(attention_mask < 0.5)
+                encoded_inputs["attention_mask"] = attention_mask
+
+            if "position_ids" not in encoded_inputs:
+                if bos_token_id in required_input:
+                    context_length = required_input.index(bos_token_id)
+                else:
+                    context_length = seq_length
+                position_ids = np.arange(seq_length, dtype=np.int64)
+                mask_token = mask_token_id if mask_token_id in required_input else gmask_token_id
+                if mask_token in required_input:
+                    mask_position = required_input.index(mask_token)
+                    position_ids[context_length:] = mask_position
+                block_position_ids = np.concatenate(
+                    [np.zeros(context_length, dtype=np.int64),
+                     np.arange(1, seq_length - context_length + 1, dtype=np.int64)])
+                encoded_inputs["position_ids"] = np.stack([position_ids, block_position_ids], axis=0)
+
+        if needs_to_be_padded:
+            difference = max_length - len(required_input)
+
+            if "attention_mask" in encoded_inputs:
+                encoded_inputs["attention_mask"] = np.pad(encoded_inputs["attention_mask"],
+                                                          pad_width=[(0, 0), (difference, 0), (difference, 0)],
+                                                          mode='constant', constant_values=True)
+            if "token_type_ids" in encoded_inputs:
+                encoded_inputs["token_type_ids"] = [self.pad_token_type_id] * difference + encoded_inputs[
+                    "token_type_ids"
+                ]
+            if "special_tokens_mask" in encoded_inputs:
+                encoded_inputs["special_tokens_mask"] = [1] * difference + encoded_inputs["special_tokens_mask"]
+            if "position_ids" in encoded_inputs:
+                encoded_inputs["position_ids"] = np.pad(encoded_inputs["position_ids"],
+                                                        pad_width=[(0, 0), (difference, 0)])
+            encoded_inputs[self.model_input_names[0]] = [self.pad_token_id] * difference + required_input
+
+        return encoded_inputs

checkpoint-2000/tokenizer_config.json

@@ -0,0 +1,22 @@
+{
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_chatglm.ChatGLMTokenizer",
+      null
+    ]
+  },
+  "bos_token": "<sop>",
+  "do_lower_case": false,
+  "end_token": "</s>",
+  "eos_token": "<eop>",
+  "gmask_token": "[gMASK]",
+  "mask_token": "[MASK]",
+  "model_max_length": 2048,
+  "num_image_tokens": 0,
+  "pad_token": "<pad>",
+  "padding_side": "left",
+  "remove_space": false,
+  "special_tokens_map_file": null,
+  "tokenizer_class": "ChatGLMTokenizer",
+  "unk_token": "<unk>"
+}

checkpoint-2000/trainer_state.json

@@ -0,0 +1,2416 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.5242463958060288,
+  "global_step": 2000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0,
+      "learning_rate": 9.999e-06,
+      "loss": 6.1641,
+      "step": 5
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 9.994000000000001e-06,
+      "loss": 5.275,
+      "step": 10
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 9.989e-06,
+      "loss": 4.8629,
+      "step": 15
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.984e-06,
+      "loss": 4.8023,
+      "step": 20
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.979e-06,
+      "loss": 4.7687,
+      "step": 25
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.974e-06,
+      "loss": 4.7188,
+      "step": 30
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.969e-06,
+      "loss": 4.6258,
+      "step": 35
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.964e-06,
+      "loss": 4.6254,
+      "step": 40
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.959e-06,
+      "loss": 4.5867,
+      "step": 45
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.954e-06,
+      "loss": 4.6207,
+      "step": 50
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.949e-06,
+      "loss": 4.6086,
+      "step": 55
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.944e-06,
+      "loss": 4.5559,
+      "step": 60
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.939000000000001e-06,
+      "loss": 4.5836,
+      "step": 65
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.934e-06,
+      "loss": 4.5121,
+      "step": 70
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.929000000000001e-06,
+      "loss": 4.5234,
+      "step": 75
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.924e-06,
+      "loss": 4.4992,
+      "step": 80
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.919000000000001e-06,
+      "loss": 4.4891,
+      "step": 85
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.914e-06,
+      "loss": 4.4688,
+      "step": 90
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 9.909000000000001e-06,
+      "loss": 4.4836,
+      "step": 95
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.904e-06,
+      "loss": 4.4363,
+      "step": 100
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.899000000000001e-06,
+      "loss": 4.4215,
+      "step": 105
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.894e-06,
+      "loss": 4.4469,
+      "step": 110
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.889000000000001e-06,
+      "loss": 4.3793,
+      "step": 115
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.884e-06,
+      "loss": 4.3934,
+      "step": 120
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.879000000000001e-06,
+      "loss": 4.3309,
+      "step": 125
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 9.874e-06,
+      "loss": 4.3875,
+      "step": 130
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.869000000000002e-06,
+      "loss": 4.4262,
+      "step": 135
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.864e-06,
+      "loss": 4.4285,
+      "step": 140
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.859e-06,
+      "loss": 4.3965,
+      "step": 145
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.854000000000001e-06,
+      "loss": 4.3359,
+      "step": 150
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.849e-06,
+      "loss": 4.4348,
+      "step": 155
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.844000000000001e-06,
+      "loss": 4.3152,
+      "step": 160
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.839e-06,
+      "loss": 4.3402,
+      "step": 165
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 9.834000000000001e-06,
+      "loss": 4.316,
+      "step": 170
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 9.829e-06,
+      "loss": 4.2969,
+      "step": 175
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 9.824000000000001e-06,
+      "loss": 4.2867,
+      "step": 180
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 9.819e-06,
+      "loss": 4.3902,
+      "step": 185
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 9.814000000000001e-06,
+      "loss": 4.2656,
+      "step": 190
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 9.809e-06,
+      "loss": 4.3797,
+      "step": 195
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 9.804000000000001e-06,
+      "loss": 4.2863,
+      "step": 200
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 9.799e-06,
+      "loss": 4.275,
+      "step": 205
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 9.794000000000001e-06,
+      "loss": 4.2891,
+      "step": 210
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 9.789e-06,
+      "loss": 4.3059,
+      "step": 215
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 9.784000000000002e-06,
+      "loss": 4.3832,
+      "step": 220
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 9.779e-06,
+      "loss": 4.3055,
+      "step": 225
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 9.774000000000002e-06,
+      "loss": 4.3008,
+      "step": 230
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 9.769e-06,
+      "loss": 4.25,
+      "step": 235
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 9.764000000000002e-06,
+      "loss": 4.2676,
+      "step": 240
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 9.759000000000001e-06,
+      "loss": 4.2422,
+      "step": 245
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 9.754000000000002e-06,
+      "loss": 4.3137,
+      "step": 250
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 9.749000000000001e-06,
+      "loss": 4.2555,
+      "step": 255
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 9.744000000000002e-06,
+      "loss": 4.2637,
+      "step": 260
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 9.739000000000001e-06,
+      "loss": 4.2996,
+      "step": 265
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 9.734000000000002e-06,
+      "loss": 4.3176,
+      "step": 270
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 9.729000000000001e-06,
+      "loss": 4.252,
+      "step": 275
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 9.724e-06,
+      "loss": 4.266,
+      "step": 280
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 9.719000000000001e-06,
+      "loss": 4.1748,
+      "step": 285
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 9.714e-06,
+      "loss": 4.3008,
+      "step": 290
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 9.709000000000001e-06,
+      "loss": 4.268,
+      "step": 295
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 9.704e-06,
+      "loss": 4.2523,
+      "step": 300
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 9.699e-06,
+      "loss": 4.3445,
+      "step": 305
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 9.694e-06,
+      "loss": 4.298,
+      "step": 310
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 9.689e-06,
+      "loss": 4.2809,
+      "step": 315
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 9.684e-06,
+      "loss": 4.2234,
+      "step": 320
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 9.679e-06,
+      "loss": 4.227,
+      "step": 325
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 9.674000000000001e-06,
+      "loss": 4.2605,
+      "step": 330
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 9.669e-06,
+      "loss": 4.2268,
+      "step": 335
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 9.664000000000001e-06,
+      "loss": 4.1734,
+      "step": 340
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 9.659e-06,
+      "loss": 4.2355,
+      "step": 345
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 9.654000000000001e-06,
+      "loss": 4.2123,
+      "step": 350
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 9.649e-06,
+      "loss": 4.1396,
+      "step": 355
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 9.644000000000001e-06,
+      "loss": 4.1869,
+      "step": 360
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 9.639e-06,
+      "loss": 4.2148,
+      "step": 365
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 9.634000000000001e-06,
+      "loss": 4.1201,
+      "step": 370
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 9.629e-06,
+      "loss": 4.1891,
+      "step": 375
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 9.624000000000001e-06,
+      "loss": 4.118,
+      "step": 380
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 9.619e-06,
+      "loss": 4.1359,
+      "step": 385
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 9.614000000000001e-06,
+      "loss": 4.1469,
+      "step": 390
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 9.609e-06,
+      "loss": 4.1941,
+      "step": 395
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 9.604000000000002e-06,
+      "loss": 4.1219,
+      "step": 400
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 9.599e-06,
+      "loss": 4.0951,
+      "step": 405
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 9.594000000000002e-06,
+      "loss": 4.1387,
+      "step": 410
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 9.589000000000001e-06,
+      "loss": 4.0973,
+      "step": 415
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 9.584000000000002e-06,
+      "loss": 4.1551,
+      "step": 420
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 9.579000000000001e-06,
+      "loss": 4.1883,
+      "step": 425
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 9.574000000000002e-06,
+      "loss": 4.2137,
+      "step": 430
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 9.569000000000001e-06,
+      "loss": 4.1748,
+      "step": 435
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 9.564e-06,
+      "loss": 4.1664,
+      "step": 440
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 9.559000000000001e-06,
+      "loss": 4.0812,
+      "step": 445
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 9.554e-06,
+      "loss": 4.2215,
+      "step": 450
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 9.549000000000001e-06,
+      "loss": 4.175,
+      "step": 455
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 9.544e-06,
+      "loss": 4.0766,
+      "step": 460
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 9.539e-06,
+      "loss": 4.0873,
+      "step": 465
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 9.534e-06,
+      "loss": 4.1316,
+      "step": 470
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 9.529e-06,
+      "loss": 4.108,
+      "step": 475
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 9.524e-06,
+      "loss": 4.1691,
+      "step": 480
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 9.519e-06,
+      "loss": 4.1154,
+      "step": 485
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 9.514e-06,
+      "loss": 4.1035,
+      "step": 490
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 9.509e-06,
+      "loss": 4.1293,
+      "step": 495
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 9.504e-06,
+      "loss": 4.1734,
+      "step": 500
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 9.499e-06,
+      "loss": 4.0504,
+      "step": 505
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 9.494000000000001e-06,
+      "loss": 4.048,
+      "step": 510
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 9.489e-06,
+      "loss": 4.1066,
+      "step": 515
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 9.484000000000001e-06,
+      "loss": 4.1354,
+      "step": 520
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 9.479e-06,
+      "loss": 4.1238,
+      "step": 525
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 9.474000000000001e-06,
+      "loss": 4.1232,
+      "step": 530
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 9.469e-06,
+      "loss": 4.1252,
+      "step": 535
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 9.464000000000001e-06,
+      "loss": 4.0975,
+      "step": 540
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 9.459e-06,
+      "loss": 4.1111,
+      "step": 545
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 9.454000000000001e-06,
+      "loss": 4.0047,
+      "step": 550
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 9.449e-06,
+      "loss": 4.0992,
+      "step": 555
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 9.444000000000001e-06,
+      "loss": 4.0734,
+      "step": 560
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 9.439e-06,
+      "loss": 4.0809,
+      "step": 565
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 9.434000000000001e-06,
+      "loss": 4.101,
+      "step": 570
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 9.429e-06,
+      "loss": 4.0662,
+      "step": 575
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 9.424000000000002e-06,
+      "loss": 4.1041,
+      "step": 580
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 9.419e-06,
+      "loss": 4.0564,
+      "step": 585
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 9.414000000000002e-06,
+      "loss": 4.0986,
+      "step": 590
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 9.409000000000001e-06,
+      "loss": 4.0309,
+      "step": 595
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 9.404e-06,
+      "loss": 4.0605,
+      "step": 600
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 9.399000000000001e-06,
+      "loss": 4.0857,
+      "step": 605
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 9.394e-06,
+      "loss": 4.1307,
+      "step": 610
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 9.389000000000001e-06,
+      "loss": 4.06,
+      "step": 615
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 9.384e-06,
+      "loss": 4.0039,
+      "step": 620
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 9.379000000000001e-06,
+      "loss": 4.0258,
+      "step": 625
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 9.374e-06,
+      "loss": 4.0738,
+      "step": 630
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 9.369e-06,
+      "loss": 4.0551,
+      "step": 635
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 9.364e-06,
+      "loss": 4.0518,
+      "step": 640
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 9.359e-06,
+      "loss": 4.0584,
+      "step": 645
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 9.354e-06,
+      "loss": 4.1109,
+      "step": 650
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 9.349e-06,
+      "loss": 3.9898,
+      "step": 655
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 9.344e-06,
+      "loss": 4.1406,
+      "step": 660
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 9.339e-06,
+      "loss": 4.0725,
+      "step": 665
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 9.334e-06,
+      "loss": 4.0207,
+      "step": 670
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 9.329e-06,
+      "loss": 4.0826,
+      "step": 675
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 9.324000000000001e-06,
+      "loss": 4.1059,
+      "step": 680
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 9.319e-06,
+      "loss": 3.9967,
+      "step": 685
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 9.314000000000001e-06,
+      "loss": 4.0328,
+      "step": 690
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 9.309e-06,
+      "loss": 3.9918,
+      "step": 695
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 9.304000000000001e-06,
+      "loss": 4.0434,
+      "step": 700
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 9.299e-06,
+      "loss": 3.9584,
+      "step": 705
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 9.294000000000001e-06,
+      "loss": 4.0551,
+      "step": 710
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 9.289e-06,
+      "loss": 3.9684,
+      "step": 715
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 9.284000000000001e-06,
+      "loss": 4.0221,
+      "step": 720
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 9.279e-06,
+      "loss": 3.985,
+      "step": 725
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 9.274000000000001e-06,
+      "loss": 4.0648,
+      "step": 730
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 9.269e-06,
+      "loss": 4.0109,
+      "step": 735
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 9.264000000000001e-06,
+      "loss": 3.9553,
+      "step": 740
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 9.259e-06,
+      "loss": 3.9904,
+      "step": 745
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 9.254000000000002e-06,
+      "loss": 3.9719,
+      "step": 750
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 9.249e-06,
+      "loss": 3.8973,
+      "step": 755
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 9.244e-06,
+      "loss": 3.9936,
+      "step": 760
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 9.239e-06,
+      "loss": 3.9498,
+      "step": 765
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 9.234e-06,
+      "loss": 3.9557,
+      "step": 770
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 9.229000000000001e-06,
+      "loss": 3.9266,
+      "step": 775
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 9.224e-06,
+      "loss": 3.9543,
+      "step": 780
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 9.219000000000001e-06,
+      "loss": 3.9732,
+      "step": 785
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 9.214e-06,
+      "loss": 3.9762,
+      "step": 790
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 9.209000000000001e-06,
+      "loss": 4.0695,
+      "step": 795
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 9.204e-06,
+      "loss": 3.9869,
+      "step": 800
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 9.199000000000001e-06,
+      "loss": 4.0061,
+      "step": 805
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 9.194e-06,
+      "loss": 4.0121,
+      "step": 810
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 9.189000000000001e-06,
+      "loss": 3.9105,
+      "step": 815
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 9.184e-06,
+      "loss": 3.8631,
+      "step": 820
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 9.179000000000001e-06,
+      "loss": 3.9498,
+      "step": 825
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 9.174e-06,
+      "loss": 3.9451,
+      "step": 830
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 9.169000000000001e-06,
+      "loss": 3.951,
+      "step": 835
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 9.164e-06,
+      "loss": 3.9297,
+      "step": 840
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 9.159000000000002e-06,
+      "loss": 3.9771,
+      "step": 845
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 9.154e-06,
+      "loss": 4.0842,
+      "step": 850
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 9.149000000000002e-06,
+      "loss": 3.8865,
+      "step": 855
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 9.144000000000001e-06,
+      "loss": 3.9312,
+      "step": 860
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 9.139000000000002e-06,
+      "loss": 3.8875,
+      "step": 865
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 9.134000000000001e-06,
+      "loss": 4.0389,
+      "step": 870
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 9.129000000000002e-06,
+      "loss": 3.9568,
+      "step": 875
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 9.124000000000001e-06,
+      "loss": 3.9541,
+      "step": 880
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 9.119000000000002e-06,
+      "loss": 3.9092,
+      "step": 885
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 9.114000000000001e-06,
+      "loss": 3.9404,
+      "step": 890
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 9.109e-06,
+      "loss": 3.9371,
+      "step": 895
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 9.104000000000001e-06,
+      "loss": 3.9477,
+      "step": 900
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 9.099e-06,
+      "loss": 3.9469,
+      "step": 905
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 9.094000000000001e-06,
+      "loss": 3.9191,
+      "step": 910
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 9.089e-06,
+      "loss": 3.9527,
+      "step": 915
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 9.084e-06,
+      "loss": 3.8934,
+      "step": 920
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 9.079e-06,
+      "loss": 3.9773,
+      "step": 925
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 9.074e-06,
+      "loss": 3.823,
+      "step": 930
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 9.069e-06,
+      "loss": 3.8857,
+      "step": 935
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 9.064e-06,
+      "loss": 3.9092,
+      "step": 940
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 9.059000000000001e-06,
+      "loss": 3.8338,
+      "step": 945
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 9.054e-06,
+      "loss": 3.9457,
+      "step": 950
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 9.049000000000001e-06,
+      "loss": 3.8869,
+      "step": 955
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 9.044e-06,
+      "loss": 3.8594,
+      "step": 960
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 9.039000000000001e-06,
+      "loss": 4.0318,
+      "step": 965
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 9.034e-06,
+      "loss": 3.8469,
+      "step": 970
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 9.029000000000001e-06,
+      "loss": 3.8367,
+      "step": 975
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 9.024e-06,
+      "loss": 3.8814,
+      "step": 980
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 9.019000000000001e-06,
+      "loss": 3.8818,
+      "step": 985
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 9.014e-06,
+      "loss": 3.908,
+      "step": 990
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 9.009000000000001e-06,
+      "loss": 3.9705,
+      "step": 995
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 9.004e-06,
+      "loss": 3.9086,
+      "step": 1000
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 8.999000000000001e-06,
+      "loss": 3.9795,
+      "step": 1005
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 8.994e-06,
+      "loss": 3.8629,
+      "step": 1010
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 8.989000000000002e-06,
+      "loss": 3.8287,
+      "step": 1015
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 8.984e-06,
+      "loss": 3.8717,
+      "step": 1020
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 8.979000000000002e-06,
+      "loss": 3.8865,
+      "step": 1025
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 8.974e-06,
+      "loss": 3.8344,
+      "step": 1030
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 8.969000000000002e-06,
+      "loss": 3.9541,
+      "step": 1035
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 8.964000000000001e-06,
+      "loss": 3.8318,
+      "step": 1040
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 8.959000000000002e-06,
+      "loss": 3.9328,
+      "step": 1045
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 8.954000000000001e-06,
+      "loss": 3.8621,
+      "step": 1050
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 8.949e-06,
+      "loss": 3.7871,
+      "step": 1055
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 8.944000000000001e-06,
+      "loss": 3.8988,
+      "step": 1060
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 8.939e-06,
+      "loss": 3.8232,
+      "step": 1065
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 8.934000000000001e-06,
+      "loss": 3.8816,
+      "step": 1070
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 8.929e-06,
+      "loss": 3.8775,
+      "step": 1075
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 8.924e-06,
+      "loss": 3.8115,
+      "step": 1080
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 8.919e-06,
+      "loss": 3.7941,
+      "step": 1085
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 8.914e-06,
+      "loss": 3.8678,
+      "step": 1090
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 8.909e-06,
+      "loss": 3.8215,
+      "step": 1095
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 8.904e-06,
+      "loss": 3.79,
+      "step": 1100
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 8.899e-06,
+      "loss": 3.8092,
+      "step": 1105
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 8.894e-06,
+      "loss": 3.79,
+      "step": 1110
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 8.889e-06,
+      "loss": 3.8162,
+      "step": 1115
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 8.884e-06,
+      "loss": 3.8568,
+      "step": 1120
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 8.879000000000001e-06,
+      "loss": 3.867,
+      "step": 1125
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 8.874e-06,
+      "loss": 3.7988,
+      "step": 1130
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 8.869000000000001e-06,
+      "loss": 3.8088,
+      "step": 1135
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 8.864e-06,
+      "loss": 3.7711,
+      "step": 1140
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 8.859000000000001e-06,
+      "loss": 3.7242,
+      "step": 1145
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 8.854e-06,
+      "loss": 3.8512,
+      "step": 1150
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 8.849000000000001e-06,
+      "loss": 3.8945,
+      "step": 1155
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 8.844e-06,
+      "loss": 3.8687,
+      "step": 1160
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 8.839000000000001e-06,
+      "loss": 3.7533,
+      "step": 1165
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 8.834e-06,
+      "loss": 3.8707,
+      "step": 1170
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 8.829000000000001e-06,
+      "loss": 3.8086,
+      "step": 1175
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 8.824e-06,
+      "loss": 3.7467,
+      "step": 1180
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 8.819000000000001e-06,
+      "loss": 3.8078,
+      "step": 1185
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 8.814e-06,
+      "loss": 3.7465,
+      "step": 1190
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 8.809000000000002e-06,
+      "loss": 3.7955,
+      "step": 1195
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 8.804e-06,
+      "loss": 3.8281,
+      "step": 1200
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 8.799000000000002e-06,
+      "loss": 3.8035,
+      "step": 1205
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 8.794e-06,
+      "loss": 3.7963,
+      "step": 1210
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 8.789e-06,
+      "loss": 3.8061,
+      "step": 1215
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 8.784000000000001e-06,
+      "loss": 3.777,
+      "step": 1220
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 8.779e-06,
+      "loss": 3.7582,
+      "step": 1225
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 8.774000000000001e-06,
+      "loss": 3.7725,
+      "step": 1230
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 8.769e-06,
+      "loss": 3.7516,
+      "step": 1235
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 8.764e-06,
+      "loss": 3.8543,
+      "step": 1240
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 8.759e-06,
+      "loss": 3.8566,
+      "step": 1245
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 8.754e-06,
+      "loss": 3.7695,
+      "step": 1250
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 8.749e-06,
+      "loss": 3.8271,
+      "step": 1255
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 8.744e-06,
+      "loss": 3.773,
+      "step": 1260
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 8.739e-06,
+      "loss": 3.7283,
+      "step": 1265
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 8.734e-06,
+      "loss": 3.7822,
+      "step": 1270
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 8.729e-06,
+      "loss": 3.7816,
+      "step": 1275
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 8.724e-06,
+      "loss": 3.751,
+      "step": 1280
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 8.719e-06,
+      "loss": 3.8271,
+      "step": 1285
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 8.714e-06,
+      "loss": 3.7195,
+      "step": 1290
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 8.709e-06,
+      "loss": 3.7584,
+      "step": 1295
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 8.704e-06,
+      "loss": 3.7889,
+      "step": 1300
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 8.699000000000001e-06,
+      "loss": 3.8529,
+      "step": 1305
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 8.694e-06,
+      "loss": 3.8166,
+      "step": 1310
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 8.689000000000001e-06,
+      "loss": 3.7484,
+      "step": 1315
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 8.684e-06,
+      "loss": 3.8014,
+      "step": 1320
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 8.679000000000001e-06,
+      "loss": 3.7658,
+      "step": 1325
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 8.674e-06,
+      "loss": 3.7834,
+      "step": 1330
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 8.669000000000001e-06,
+      "loss": 3.7973,
+      "step": 1335
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 8.664e-06,
+      "loss": 3.7607,
+      "step": 1340
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 8.659000000000001e-06,
+      "loss": 3.7381,
+      "step": 1345
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 8.654e-06,
+      "loss": 3.751,
+      "step": 1350
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 8.649000000000001e-06,
+      "loss": 3.7201,
+      "step": 1355
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 8.644e-06,
+      "loss": 3.7969,
+      "step": 1360
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 8.639000000000001e-06,
+      "loss": 3.7773,
+      "step": 1365
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 8.634e-06,
+      "loss": 3.7752,
+      "step": 1370
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 8.629e-06,
+      "loss": 3.6992,
+      "step": 1375
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 8.624e-06,
+      "loss": 3.651,
+      "step": 1380
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 8.619e-06,
+      "loss": 3.7598,
+      "step": 1385
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 8.614000000000001e-06,
+      "loss": 3.7367,
+      "step": 1390
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 8.609e-06,
+      "loss": 3.6896,
+      "step": 1395
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 8.604000000000001e-06,
+      "loss": 3.7732,
+      "step": 1400
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 8.599e-06,
+      "loss": 3.7836,
+      "step": 1405
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 8.594000000000001e-06,
+      "loss": 3.7854,
+      "step": 1410
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 8.589e-06,
+      "loss": 3.701,
+      "step": 1415
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 8.584000000000001e-06,
+      "loss": 3.7652,
+      "step": 1420
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 8.579e-06,
+      "loss": 3.775,
+      "step": 1425
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 8.574000000000001e-06,
+      "loss": 3.7207,
+      "step": 1430
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 8.569e-06,
+      "loss": 3.71,
+      "step": 1435
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 8.564000000000001e-06,
+      "loss": 3.7359,
+      "step": 1440
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 8.559e-06,
+      "loss": 3.6854,
+      "step": 1445
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 8.554000000000001e-06,
+      "loss": 3.7342,
+      "step": 1450
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 8.549e-06,
+      "loss": 3.6707,
+      "step": 1455
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 8.544000000000002e-06,
+      "loss": 3.6596,
+      "step": 1460
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 8.539e-06,
+      "loss": 3.6711,
+      "step": 1465
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 8.534000000000002e-06,
+      "loss": 3.7279,
+      "step": 1470
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 8.529e-06,
+      "loss": 3.7115,
+      "step": 1475
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 8.524000000000002e-06,
+      "loss": 3.7139,
+      "step": 1480
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 8.519000000000001e-06,
+      "loss": 3.674,
+      "step": 1485
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 8.514000000000002e-06,
+      "loss": 3.6191,
+      "step": 1490
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 8.509000000000001e-06,
+      "loss": 3.6361,
+      "step": 1495
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 8.504000000000002e-06,
+      "loss": 3.7717,
+      "step": 1500
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 8.499000000000001e-06,
+      "loss": 3.6355,
+      "step": 1505
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 8.494e-06,
+      "loss": 3.8113,
+      "step": 1510
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 8.489000000000001e-06,
+      "loss": 3.7465,
+      "step": 1515
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 8.484e-06,
+      "loss": 3.8033,
+      "step": 1520
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 8.479000000000001e-06,
+      "loss": 3.6867,
+      "step": 1525
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 8.474e-06,
+      "loss": 3.7062,
+      "step": 1530
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 8.469e-06,
+      "loss": 3.726,
+      "step": 1535
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 8.464e-06,
+      "loss": 3.6432,
+      "step": 1540
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 8.459e-06,
+      "loss": 3.6943,
+      "step": 1545
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 8.454e-06,
+      "loss": 3.6127,
+      "step": 1550
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 8.449e-06,
+      "loss": 3.6529,
+      "step": 1555
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 8.444e-06,
+      "loss": 3.6063,
+      "step": 1560
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 8.439e-06,
+      "loss": 3.7633,
+      "step": 1565
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 8.434000000000001e-06,
+      "loss": 3.6211,
+      "step": 1570
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 8.429e-06,
+      "loss": 3.6895,
+      "step": 1575
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 8.424000000000001e-06,
+      "loss": 3.6152,
+      "step": 1580
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 8.419e-06,
+      "loss": 3.6549,
+      "step": 1585
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 8.414000000000001e-06,
+      "loss": 3.6502,
+      "step": 1590
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 8.409e-06,
+      "loss": 3.5689,
+      "step": 1595
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 8.404000000000001e-06,
+      "loss": 3.7002,
+      "step": 1600
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 8.399e-06,
+      "loss": 3.5998,
+      "step": 1605
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 8.394000000000001e-06,
+      "loss": 3.7164,
+      "step": 1610
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 8.389e-06,
+      "loss": 3.6006,
+      "step": 1615
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 8.384000000000001e-06,
+      "loss": 3.5586,
+      "step": 1620
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 8.379e-06,
+      "loss": 3.6801,
+      "step": 1625
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 8.374000000000001e-06,
+      "loss": 3.601,
+      "step": 1630
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 8.369e-06,
+      "loss": 3.6344,
+      "step": 1635
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 8.364000000000002e-06,
+      "loss": 3.6637,
+      "step": 1640
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 8.359e-06,
+      "loss": 3.6357,
+      "step": 1645
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 8.354000000000002e-06,
+      "loss": 3.652,
+      "step": 1650
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 8.349000000000001e-06,
+      "loss": 3.6439,
+      "step": 1655
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 8.344000000000002e-06,
+      "loss": 3.6051,
+      "step": 1660
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 8.339000000000001e-06,
+      "loss": 3.6207,
+      "step": 1665
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 8.334e-06,
+      "loss": 3.6059,
+      "step": 1670
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 8.329000000000001e-06,
+      "loss": 3.7102,
+      "step": 1675
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 8.324e-06,
+      "loss": 3.5629,
+      "step": 1680
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 8.319000000000001e-06,
+      "loss": 3.6357,
+      "step": 1685
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 8.314e-06,
+      "loss": 3.6416,
+      "step": 1690
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 8.309e-06,
+      "loss": 3.6572,
+      "step": 1695
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 8.304e-06,
+      "loss": 3.6244,
+      "step": 1700
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 8.299e-06,
+      "loss": 3.677,
+      "step": 1705
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 8.294e-06,
+      "loss": 3.6006,
+      "step": 1710
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 8.289e-06,
+      "loss": 3.7182,
+      "step": 1715
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 8.284e-06,
+      "loss": 3.6451,
+      "step": 1720
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 8.279e-06,
+      "loss": 3.508,
+      "step": 1725
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 8.274e-06,
+      "loss": 3.6182,
+      "step": 1730
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 8.269e-06,
+      "loss": 3.5447,
+      "step": 1735
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 8.264e-06,
+      "loss": 3.5941,
+      "step": 1740
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 8.259e-06,
+      "loss": 3.5094,
+      "step": 1745
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 8.254000000000001e-06,
+      "loss": 3.5988,
+      "step": 1750
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 8.249e-06,
+      "loss": 3.6652,
+      "step": 1755
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 8.244000000000001e-06,
+      "loss": 3.5957,
+      "step": 1760
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 8.239e-06,
+      "loss": 3.5326,
+      "step": 1765
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 8.234000000000001e-06,
+      "loss": 3.5537,
+      "step": 1770
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 8.229e-06,
+      "loss": 3.5834,
+      "step": 1775
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 8.224000000000001e-06,
+      "loss": 3.5666,
+      "step": 1780
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 8.219e-06,
+      "loss": 3.6174,
+      "step": 1785
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 8.214000000000001e-06,
+      "loss": 3.5148,
+      "step": 1790
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 8.209e-06,
+      "loss": 3.5037,
+      "step": 1795
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 8.204000000000001e-06,
+      "loss": 3.6,
+      "step": 1800
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 8.199e-06,
+      "loss": 3.5457,
+      "step": 1805
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 8.194000000000002e-06,
+      "loss": 3.5021,
+      "step": 1810
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 8.189e-06,
+      "loss": 3.509,
+      "step": 1815
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 8.184000000000002e-06,
+      "loss": 3.5457,
+      "step": 1820
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 8.179e-06,
+      "loss": 3.5449,
+      "step": 1825
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 8.174e-06,
+      "loss": 3.5832,
+      "step": 1830
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 8.169000000000001e-06,
+      "loss": 3.4852,
+      "step": 1835
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 8.164e-06,
+      "loss": 3.6166,
+      "step": 1840
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 8.159000000000001e-06,
+      "loss": 3.5248,
+      "step": 1845
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 8.154e-06,
+      "loss": 3.5617,
+      "step": 1850
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 8.149e-06,
+      "loss": 3.5119,
+      "step": 1855
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 8.144e-06,
+      "loss": 3.5475,
+      "step": 1860
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 8.139e-06,
+      "loss": 3.5646,
+      "step": 1865
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 8.134e-06,
+      "loss": 3.4521,
+      "step": 1870
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 8.129e-06,
+      "loss": 3.492,
+      "step": 1875
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 8.124e-06,
+      "loss": 3.6187,
+      "step": 1880
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 8.119e-06,
+      "loss": 3.4984,
+      "step": 1885
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 8.114e-06,
+      "loss": 3.5744,
+      "step": 1890
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 8.109e-06,
+      "loss": 3.5514,
+      "step": 1895
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 8.104e-06,
+      "loss": 3.4807,
+      "step": 1900
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 8.099e-06,
+      "loss": 3.5049,
+      "step": 1905
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 8.094e-06,
+      "loss": 3.5098,
+      "step": 1910
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 8.089e-06,
+      "loss": 3.4152,
+      "step": 1915
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 8.084000000000001e-06,
+      "loss": 3.4281,
+      "step": 1920
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 8.079e-06,
+      "loss": 3.5766,
+      "step": 1925
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 8.074000000000001e-06,
+      "loss": 3.4908,
+      "step": 1930
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 8.069e-06,
+      "loss": 3.5432,
+      "step": 1935
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 8.064000000000001e-06,
+      "loss": 3.5154,
+      "step": 1940
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 8.059e-06,
+      "loss": 3.4568,
+      "step": 1945
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 8.054000000000001e-06,
+      "loss": 3.5314,
+      "step": 1950
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 8.049e-06,
+      "loss": 3.5516,
+      "step": 1955
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 8.044000000000001e-06,
+      "loss": 3.4271,
+      "step": 1960
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 8.039e-06,
+      "loss": 3.4174,
+      "step": 1965
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 8.034000000000001e-06,
+      "loss": 3.5492,
+      "step": 1970
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 8.029e-06,
+      "loss": 3.568,
+      "step": 1975
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 8.024000000000001e-06,
+      "loss": 3.5455,
+      "step": 1980
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 8.019e-06,
+      "loss": 3.5598,
+      "step": 1985
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 8.014e-06,
+      "loss": 3.5848,
+      "step": 1990
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 8.009e-06,
+      "loss": 3.4631,
+      "step": 1995
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 8.004e-06,
+      "loss": 3.3873,
+      "step": 2000
+    }
+  ],
+  "max_steps": 10000,
+  "num_train_epochs": 3,
+  "total_flos": 1.5936160471711744e+18,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-2000/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f312be11981e62d6d20a9fdb31d7b06ad2b03abd4774fbd90a2c664ffe272e7
+size 4923

checkpoint-2000/zero_to_fp32.py

@@ -0,0 +1,578 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example: python zero_to_fp32.py . pytorch_model.bin
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage == 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+
+    total_files = len(files)
+    state_dicts = []
+    for f in files:
+        state_dicts.append(torch.load(f, map_location=device))
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage == 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    if zero_stage == 2:
+        fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    elif zero_stage == 3:
+        # if there is more than one param group, there will be multiple flattened tensors - one
+        # flattened tensor per group - for simplicity merge them into a single tensor
+        #
+        # XXX: could make the script more memory efficient for when there are multiple groups - it
+        # will require matching the sub-lists of param_shapes for each param group flattened tensor
+
+        fp32_flat_groups = [
+            torch.cat(state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key], 0) for i in range(len(state_dicts))
+        ]
+
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage == 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel()
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = fp32_flat_groups[0].numel() * world_size
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    for name, shape in param_shapes.items():
+
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # XXX: memory usage doubles here
+        state_dict[name] = torch.cat(
+            tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
+            0).narrow(0, 0, unpartitioned_numel).view(shape)
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    Note: this approach may not work if your application doesn't have sufficient free CPU memory and
+    you may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    return _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir, output_file, tag=None):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_file``: path to the pytorch fp32 state_dict output file (e.g. path/pytorch_model.bin)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+    """
+
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+    print(f"Saving fp32 state dict to {output_file}")
+    torch.save(state_dict, output_file)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument(
+        "output_file",
+        type=str,
+        help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir, args.output_file)