update

added_tokens.json

@@ -0,0 +1,9 @@
+{
+  "<eot>": 64005,
+  "<|assistant|>": 64001,
+  "<|inner_thoughts|>": 64004,
+  "<|prefix_begin|>": 64002,
+  "<|prefix_end|>": 64000,
+  "<|prompter|>": 64006,
+  "<|system|>": 64003
+}

config.json

@@ -0,0 +1,28 @@
+{
+  "_from_model_config": true,
+  "_name_or_path": "/mnt/petrelfs/chenqin.p/dyh/educhat-base-002-13b-baichuan/",
+  "architectures": [
+    "BaichuanForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_baichuan.BaichuanConfig",
+    "AutoModelForCausalLM": "modeling_baichuan.BaichuanForCausalLM"
+  },
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "initializer_range": 0.02,
+  "intermediate_size": 13696,
+  "model_max_length": 4096,
+  "model_type": "baichuan",
+  "num_attention_heads": 40,
+  "num_hidden_layers": 40,
+  "pad_token_id": 0,
+  "rms_norm_eps": 1e-06,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.31.0",
+  "use_cache": true,
+  "vocab_size": 64016
+}

configuration_baichuan.py

@@ -0,0 +1,46 @@
+# Copyright (c) 2023, Baichuan Intelligent Technology. All rights reserved.
+
+from transformers.configuration_utils import PretrainedConfig
+
+class BaichuanConfig(PretrainedConfig):
+    model_type = "baichuan"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=64000,
+        hidden_size=5120,
+        intermediate_size=13696,
+        num_hidden_layers=40,
+        num_attention_heads=40,
+        hidden_act="silu",
+        model_max_length=4096,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        tie_word_embeddings=False,
+        gradient_checkpointing=False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.model_max_length = model_max_length
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.gradient_checkpointing = gradient_checkpointing,
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "pad_token_id": 0,
+  "transformers_version": "4.31.0"
+}

generation_utils.py

@@ -0,0 +1,82 @@
+from typing import List
+from queue import Queue
+
+import torch
+
+
+def build_chat_input(model, tokenizer, messages: List[dict], max_new_tokens: int=0):
+    def _parse_messages(messages, split_role="user"):
+        system, rounds = "", []
+        round = []
+        for i, message in enumerate(messages):
+            if message["role"] == "system":
+                assert i == 0
+                system = message["content"]
+                continue
+            if message["role"] == split_role and round:
+                rounds.append(round)
+                round = []
+            round.append(message)
+        if round:
+            rounds.append(round)
+        return system, rounds
+
+    max_new_tokens = max_new_tokens or model.generation_config.max_new_tokens
+    max_input_tokens = model.config.model_max_length - max_new_tokens
+    system, rounds = _parse_messages(messages, split_role="user")
+    system_tokens = tokenizer.encode(system)
+    max_history_tokens = max_input_tokens - len(system_tokens)
+
+    history_tokens = []
+    for round in rounds[::-1]:
+        round_tokens = []
+        for message in round:
+            if message["role"] == "user":
+                round_tokens.append(model.generation_config.user_token_id)
+            else:
+                round_tokens.append(model.generation_config.assistant_token_id)
+            round_tokens.extend(tokenizer.encode(message["content"]))
+        if len(history_tokens) == 0 or len(history_tokens) + len(round_tokens) <= max_history_tokens:
+            history_tokens = round_tokens + history_tokens  # concat left
+            if len(history_tokens) < max_history_tokens:
+                continue
+        break
+
+    input_tokens = system_tokens + history_tokens
+    if messages[-1]["role"] != "assistant":
+        input_tokens.append(model.generation_config.assistant_token_id)
+    input_tokens = input_tokens[-max_input_tokens:]  # truncate left
+    return torch.LongTensor([input_tokens]).to(model.device)
+
+
+class TextIterStreamer:
+    def __init__(self, tokenizer, skip_prompt=False, skip_special_tokens=False):
+        self.tokenizer = tokenizer
+        self.skip_prompt = skip_prompt
+        self.skip_special_tokens = skip_special_tokens
+        self.tokens = []
+        self.text_queue = Queue()
+        self.next_tokens_are_prompt = True
+
+    def put(self, value):
+        if self.skip_prompt and self.next_tokens_are_prompt:
+            self.next_tokens_are_prompt = False
+        else:
+            if len(value.shape) > 1:
+                value = value[0]
+            self.tokens.extend(value.tolist())
+            self.text_queue.put(
+                self.tokenizer.decode(self.tokens, skip_special_tokens=self.skip_special_tokens))
+
+    def end(self):
+        self.text_queue.put(None)
+
+    def __iter__(self):
+        return self
+
+    def __next__(self):
+        value = self.text_queue.get()
+        if value is None:
+            raise StopIteration()
+        else:
+            return value

modeling_baichuan.py

@@ -0,0 +1,572 @@
+# Copyright (c) 2023, Baichuan Intelligent Technology. All rights reserved.
+
+import math
+from threading import Thread
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch.nn import CrossEntropyLoss
+from transformers import PreTrainedModel
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.utils import logging
+from transformers.generation.utils import GenerationConfig
+
+from .configuration_baichuan import BaichuanConfig
+from .generation_utils import build_chat_input, TextIterStreamer
+
+logger = logging.get_logger(__name__)
+
+
+def _get_interleave(n):
+    def _get_interleave_power_of_2(n):
+        start = (2 ** (-2 ** -(math.log2(n) - 3)))
+        ratio = start
+        return [start * ratio ** i for i in range(n)]
+
+    if math.log2(n).is_integer():
+        return _get_interleave_power_of_2(n)
+    else:
+        closest_power_of_2 = 2 ** math.floor(math.log2(n))
+        return _get_interleave_power_of_2(closest_power_of_2) + \
+               _get_interleave(2 * closest_power_of_2)[0::2][:n - closest_power_of_2]
+
+def _fill_with_neg_inf(t):
+    """FP16-compatible function that fills a tensor with -inf."""
+    return t.float().fill_(float("-inf")).type_as(t)
+
+def _gen_alibi_mask(n_head, max_pos):
+    """used in inference only"""
+    slopes = torch.Tensor(_get_interleave(n_head))
+    alibi = slopes.unsqueeze(1).unsqueeze(1) * torch.arange(max_pos).unsqueeze(0).unsqueeze(0).expand(
+        n_head, -1, -1)
+    alibi = alibi.view(n_head, 1, max_pos)
+    alibi_mask = torch.triu(
+        _fill_with_neg_inf(torch.zeros([max_pos, max_pos])), 1
+    )
+    alibi_mask = alibi_mask.unsqueeze(0) + alibi
+    return alibi_mask
+
+def _buffered_future_mask(tensor, maxpos, alibi, attn_heads):
+    """used in training only"""
+    dim = tensor.size(1)
+    _future_mask = torch.triu(
+        _fill_with_neg_inf(torch.zeros([maxpos, maxpos])), 1
+    )   
+    _future_mask = _future_mask.unsqueeze(0) + alibi
+    _future_mask = _future_mask.to(tensor)
+    return _future_mask[:tensor.shape[0] * attn_heads, :maxpos, :maxpos]
+
+
+class RMSNorm(torch.nn.Module):
+    def __init__(self, hidden_size, epsilon=1e-6):
+        super().__init__()
+        self.weight = torch.nn.Parameter(torch.empty(hidden_size))
+        self.epsilon = epsilon
+
+    def forward(self, hidden_states):
+        variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.epsilon)
+
+        # convert into half-precision
+        if self.weight.dtype in [torch.float16, torch.bfloat16]:
+            hidden_states = hidden_states.to(self.weight.dtype)
+
+        return self.weight * hidden_states
+
+
+class MLP(torch.nn.Module):
+    def __init__(
+            self,
+            hidden_size: int,
+            intermediate_size: int,
+            hidden_act: str,
+    ):
+        super().__init__()
+        self.gate_proj = torch.nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.down_proj = torch.nn.Linear(intermediate_size, hidden_size, bias=False)
+        self.up_proj = torch.nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.act_fn = ACT2FN[hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class BaichuanAttention(torch.nn.Module):
+    def __init__(self, config: BaichuanConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.max_position_embeddings = config.model_max_length
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size {self.hidden_size} is not divisible by num_heads {self.num_heads}"
+            )
+        self.W_pack = torch.nn.Linear(self.hidden_size, 3 * self.hidden_size, bias=False)
+        self.o_proj = torch.nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_value: Optional[Tuple[torch.Tensor]] = None,
+            output_attentions: bool = False,
+            use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+
+        bsz, q_len, _ = hidden_states.size()
+
+        proj = self.W_pack(hidden_states)
+        proj = proj.unflatten(-1, (3, self.hidden_size)).unsqueeze(0).transpose(0, -2).squeeze(-2)
+        query_states = proj[0].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = proj[1].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = proj[2].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value[0].shape[-2]
+
+        if past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+        past_key_value = (key_states, value_states) if use_cache else None
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        if attention_mask is not None:
+            if q_len == 1: # inference with cache
+                if len(attention_mask.size()) == 4:
+                    attention_mask = attention_mask[:, :, -1:, :]   
+                else:
+                    attention_mask = attention_mask[:, -1:, :]    
+            attn_weights = attn_weights + attention_mask
+            attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
+
+        attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1)
+
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        attn_output = attn_output.transpose(1, 2)
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class BaichuanLayer(torch.nn.Module):
+    def __init__(self, config: BaichuanConfig):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = BaichuanAttention(config=config)
+        self.mlp = MLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+        )
+        self.input_layernorm = RMSNorm(config.hidden_size, epsilon=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(config.hidden_size, epsilon=config.rms_norm_eps)
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_value: Optional[Tuple[torch.Tensor]] = None,
+            output_attentions: Optional[bool] = False,
+            use_cache: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+class BaichuanPreTrainedModel(PreTrainedModel):
+    config_class = BaichuanConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["BaichuanLayer"]
+    _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, torch.nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, torch.nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, BaichuanModel):
+            module.gradient_checkpointing = value
+
+
+class BaichuanModel(BaichuanPreTrainedModel):
+    def __init__(self, config: BaichuanConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        self.n_head = config.num_attention_heads
+        self.embed_tokens = torch.nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = torch.nn.ModuleList([BaichuanLayer(config) for _ in range(config.num_hidden_layers)])
+        self.norm = RMSNorm(config.hidden_size, epsilon=config.rms_norm_eps)
+
+        self.gradient_checkpointing = config.gradient_checkpointing
+        self.post_init()
+        self.max_cache_pos = config.model_max_length
+        self.first_run = True
+        self.alibi_mask = None
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def get_alibi_mask(self, tensor, seq_length_with_past):
+        if self.training:
+            slopes = torch.Tensor(_get_interleave(self.n_head))
+            alibi = slopes.unsqueeze(1).unsqueeze(1) * torch.arange(seq_length_with_past).unsqueeze(0).unsqueeze(0).expand(
+                self.n_head,
+                -1, -1) 
+            alibi = alibi.view(self.n_head, 1, seq_length_with_past)
+            mask = _buffered_future_mask(tensor, seq_length_with_past, alibi, self.n_head)
+        else:
+            if self.first_run:
+                self.first_run = False
+                self.register_buffer("future_mask", _gen_alibi_mask(self.n_head, self.max_cache_pos).to(tensor), persistent=False)
+            if seq_length_with_past > self.max_cache_pos:
+                self.max_cache_pos = seq_length_with_past
+                self.register_buffer("future_mask", _gen_alibi_mask(self.n_head, self.max_cache_pos).to(tensor), persistent=False)
+            mask = self.future_mask[:self.n_head, :seq_length_with_past, :seq_length_with_past] 
+        return mask
+
+    def forward(
+            self,
+            input_ids: torch.LongTensor = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_values: Optional[List[torch.FloatTensor]] = None,
+            inputs_embeds: Optional[torch.FloatTensor] = None,
+            use_cache: Optional[bool] = False,
+            output_attentions: Optional[bool] = False,
+            output_hidden_states: Optional[bool] = False,
+            return_dict: Optional[bool] = True,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot provide both input_ids and inputs_embeds simultaneously")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+        else:
+            raise ValueError("You need to provide input_ids or inputs_embeds")
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        seq_length_with_past = seq_length
+
+        if past_key_values is not None:
+            past_key_values_length = past_key_values[0][0].shape[2]
+            seq_length_with_past = seq_length_with_past + past_key_values_length
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if self.training:
+            if self.alibi_mask is None or self.alibi_mask.shape[-1] != seq_length_with_past:
+                self.alibi_mask = self.get_alibi_mask(inputs_embeds, seq_length_with_past)
+            alibi_mask = self.alibi_mask
+        else:
+            alibi_mask = self.get_alibi_mask(inputs_embeds, seq_length_with_past)
+
+        if attention_mask is not None:
+            if len(attention_mask.shape) == 2:
+                expanded_mask = attention_mask.to(alibi_mask.dtype)
+                expanded_mask = torch.tril(torch.gt(expanded_mask[:, :, None] * expanded_mask[:, None, :], 0)
+                                ) * torch.eq(expanded_mask[:, :, None] - expanded_mask[:, None, :], 0)
+            else:
+                expanded_mask = attention_mask 
+            bsz = inputs_embeds.size(0)
+            src_len, tgt_len = alibi_mask.size()[-2:]
+            expanded_mask = expanded_mask.unsqueeze(1).expand(bsz, 1, src_len, tgt_len).to(alibi_mask.dtype)
+            inverted_mask = 1.0 - expanded_mask
+            inverted_mask = inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(alibi_mask.dtype).min)
+            attention_mask = inverted_mask + alibi_mask.unsqueeze(0)
+        else:
+            attention_mask = alibi_mask
+
+        hidden_states = inputs_embeds
+
+        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
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = () if use_cache else None
+
+        for idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # None for past_key_value
+                        return module(*inputs, output_attentions, None)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(decoder_layer),
+                    hidden_states,
+                    attention_mask,
+                    None,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    past_key_value=past_key_value,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class BaichuanForCausalLM(BaichuanPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = BaichuanModel(config)
+        self.lm_head = torch.nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def forward(
+            self,
+            input_ids: torch.LongTensor = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_values: Optional[List[torch.FloatTensor]] = None,
+            inputs_embeds: Optional[torch.FloatTensor] = None,
+            labels: Optional[torch.LongTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = False,
+            output_hidden_states: Optional[bool] = False,
+            return_dict: Optional[bool] = True,
+            **kwargs
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_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 = outputs[0]
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+            self,
+            input_ids: torch.LongTensor,
+            past_key_values: Optional[torch.Tensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            inputs_embeds: Optional[torch.Tensor] = None,
+            **kwargs
+    ):
+        if past_key_values:
+            input_ids = input_ids[:, -1:]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        return tuple(
+            tuple(past_state.index_select(0, beam_idx) for past_state in layer_past)
+            for layer_past in past_key_values
+        )
+
+    def quantize(self, bits: int):
+        try:
+            from .quantizer import QLinear
+        except ImportError:
+            raise ImportError(
+                f"Needs QLinear to run quantize."
+            )
+
+        for layer in self.model.layers:
+            layer.self_attn.W_pack = QLinear(
+                bits=bits,
+                weight=layer.self_attn.W_pack.weight,
+                bias = None,
+            )
+            layer.self_attn.o_proj = QLinear(
+                bits=bits,
+                weight=layer.self_attn.o_proj.weight,
+                bias = None,
+            )
+            layer.mlp.gate_proj = QLinear(
+                bits=bits,
+                weight=layer.mlp.gate_proj.weight,
+                bias = None,
+            )
+            layer.mlp.down_proj = QLinear(
+                bits=bits,
+                weight=layer.mlp.down_proj.weight,
+                bias = None,
+            )
+            layer.mlp.up_proj = QLinear(
+                bits=bits,
+                weight=layer.mlp.up_proj.weight,
+                bias = None,
+            )
+        return self
+
+    @torch.no_grad()
+    def chat(self, tokenizer, messages: List[dict], stream=False,
+             generation_config: Optional[GenerationConfig]=None):
+        generation_config = generation_config or self.generation_config
+        input_ids = build_chat_input(self, tokenizer, messages, generation_config.max_new_tokens)
+        if stream:
+            streamer = TextIterStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
+            Thread(target=self.generate, kwargs=dict(
+                inputs=input_ids, streamer=streamer,
+                generation_config=generation_config,
+            )).start()
+            return streamer
+        else:
+            outputs = self.generate(input_ids, generation_config=generation_config)
+            response = tokenizer.decode(outputs[0][len(input_ids[0]):], skip_special_tokens=True)
+            return response

pytorch_model-00001-of-00003.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:99b028dee5b892a9fdb2bb1b03ee50614f32f26c00a4b82af6305c2c4f6d14da
+size 9972443620

pytorch_model-00002-of-00003.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:953a36711f31ff068cd600ae1f74152806affcf32f617b5669e382531ca783b1
+size 9947419824

pytorch_model-00003-of-00003.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:85137388811d9ab10ac6de22832b3e5bfe9d562260bb174175ccfeb23defa23f
+size 6610363169

pytorch_model.bin.index.json

@@ -0,0 +1,290 @@
+{
+  "metadata": {
+    "total_size": 26530129920
+  },
+  "weight_map": {
+    "lm_head.weight": "pytorch_model-00003-of-00003.bin",
+    "model.embed_tokens.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.0.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.0.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.0.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.0.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.0.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.0.self_attn.W_pack.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.0.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.1.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.1.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.1.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.1.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.1.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.1.self_attn.W_pack.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.1.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.10.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.10.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.10.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.10.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.10.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.10.self_attn.W_pack.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.10.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.11.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.11.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.11.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.11.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.11.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.11.self_attn.W_pack.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.11.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.12.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.12.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.12.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.12.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.12.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.12.self_attn.W_pack.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.12.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.13.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.13.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.13.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.13.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.13.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.13.self_attn.W_pack.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.13.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.14.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.14.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.14.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.14.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.14.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.14.self_attn.W_pack.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.14.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.15.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.15.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.15.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.15.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.15.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.15.self_attn.W_pack.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.15.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.16.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.16.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.16.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.16.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.16.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.16.self_attn.W_pack.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.16.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.17.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.17.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.17.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.17.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.17.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.17.self_attn.W_pack.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.17.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.18.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.18.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.18.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.18.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.18.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.18.self_attn.W_pack.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.18.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.19.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.19.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.19.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.19.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.19.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.19.self_attn.W_pack.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.19.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.2.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.2.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.2.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.2.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.2.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.2.self_attn.W_pack.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.2.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.20.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.20.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.20.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.20.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.20.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.20.self_attn.W_pack.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.20.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.21.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.21.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.21.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.21.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.21.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.21.self_attn.W_pack.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.21.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.22.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.22.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.22.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.22.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.22.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.22.self_attn.W_pack.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.22.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.23.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.23.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.23.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.23.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.23.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.23.self_attn.W_pack.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.23.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.24.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.24.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.24.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.24.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.24.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.24.self_attn.W_pack.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.24.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.25.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.25.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.25.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.25.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.25.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.25.self_attn.W_pack.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.25.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.26.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.26.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.26.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.26.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.26.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.26.self_attn.W_pack.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.26.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.27.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.27.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.27.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.27.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.27.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.27.self_attn.W_pack.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.27.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.28.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.28.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.28.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.28.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.28.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.28.self_attn.W_pack.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.28.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.29.input_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.29.mlp.down_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.29.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.29.mlp.up_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.29.post_attention_layernorm.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.29.self_attn.W_pack.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.29.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.3.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.3.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.3.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.3.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.3.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.3.self_attn.W_pack.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.3.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.30.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.30.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.30.mlp.gate_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.30.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.30.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.30.self_attn.W_pack.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.30.self_attn.o_proj.weight": "pytorch_model-00002-of-00003.bin",
+    "model.layers.31.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.31.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.31.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.31.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.31.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.31.self_attn.W_pack.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.31.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.32.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.32.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.32.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.32.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.32.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.32.self_attn.W_pack.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.32.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.33.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.33.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.33.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.33.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.33.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.33.self_attn.W_pack.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.33.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.34.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.34.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.34.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.34.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.34.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.34.self_attn.W_pack.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.34.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.35.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.35.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.35.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.35.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.35.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.35.self_attn.W_pack.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.35.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.36.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.36.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.36.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.36.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.36.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.36.self_attn.W_pack.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.36.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.37.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.37.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.37.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.37.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.37.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.37.self_attn.W_pack.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.37.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.38.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.38.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.38.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.38.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.38.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.38.self_attn.W_pack.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.38.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.39.input_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.39.mlp.down_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.39.mlp.gate_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.39.mlp.up_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.39.post_attention_layernorm.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.39.self_attn.W_pack.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.39.self_attn.o_proj.weight": "pytorch_model-00003-of-00003.bin",
+    "model.layers.4.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.4.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.4.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.4.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.4.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.4.self_attn.W_pack.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.4.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.5.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.5.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.5.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.5.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.5.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.5.self_attn.W_pack.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.5.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.6.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.6.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.6.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.6.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.6.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.6.self_attn.W_pack.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.6.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.7.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.7.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.7.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.7.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.7.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.7.self_attn.W_pack.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.7.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.8.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.8.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.8.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.8.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.8.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.8.self_attn.W_pack.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.8.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.9.input_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.9.mlp.down_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.9.mlp.gate_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.9.mlp.up_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.9.post_attention_layernorm.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.9.self_attn.W_pack.weight": "pytorch_model-00001-of-00003.bin",
+    "model.layers.9.self_attn.o_proj.weight": "pytorch_model-00001-of-00003.bin",
+    "model.norm.weight": "pytorch_model-00003-of-00003.bin"
+  }
+}

quantizer.py

@@ -0,0 +1,123 @@
+# Copyright (c) 2023, Baichuan Intelligent Technology. All rights reserved.
+
+import torch
+from typing import List
+import bz2
+import base64
+import ctypes
+from transformers.utils import logging
+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)),
+        [
+            "int4_to_fp16",
+            "fp16_to_int4",
+            "int8_to_fp16",
+            "fp16_to_int8",
+            "int4_to_bf16",
+            "bf16_to_int4",
+            "int8_to_bf16",
+            "bf16_to_int8",
+        ],
+    )
+except Exception as exception:
+    kernels = None
+    logger.warning("Failed to load kernels:" + str(exception))
+
+def quant4(weight: torch.Tensor, scale: torch.Tensor):
+    stream = torch.cuda.current_stream()
+    num_row = weight.size(0)
+    num_chan_fp16 = weight.size(1)
+    # 4bit
+    num_chan_int = num_chan_fp16 // 8
+    qweight = torch.zeros((num_row, num_chan_int), dtype=torch.int32, device=weight.device)
+    intweight = torch.empty(num_row, num_chan_fp16, dtype = torch.int32)
+    intweight = torch.clip(torch.round(weight.to(scale.dtype) / scale[:, None]),-16, 15).to(dtype=torch.int32)
+
+    for j in range(num_chan_int):
+        qweight[:, j] = ((intweight[:, j*8+7] & 0x0f) << 28) \
+            | ((intweight[:, j*8+6] & 0x0f) << 24) \
+            | ((intweight[:, j*8+5] & 0x0f) << 20) \
+            | ((intweight[:, j*8+4] & 0x0f) << 16) \
+            | ((intweight[:, j*8+3] & 0x0f) << 12) \
+            | ((intweight[:, j*8+2] & 0x0f) << 8) \
+            | ((intweight[:, j*8+1] & 0x0f) << 4) \
+            | ((intweight[:, j*8] & 0x0f))
+    return qweight
+
+def dequant4(qweight: torch.Tensor, scale: torch.Tensor, input: torch.Tensor):
+    stream = torch.cuda.current_stream()
+    num_row = qweight.size(0)
+    num_chan_int = qweight.size(1)
+    # 4bit
+    num_chan_fp16 = num_chan_int * 8
+    
+    out = torch.empty((num_row, num_chan_fp16), dtype=input.dtype, device=qweight.device)
+
+    blockDim = (128, 1, 1)
+    gridDim = ((num_chan_int + blockDim[0] - 1) // blockDim[0], num_row, 1)
+    if input.dtype == torch.bfloat16:
+    	kernels.int4_to_bf16(
+            gridDim,
+       	    blockDim,
+            0,
+            stream,
+            [ctypes.c_void_p(out.data_ptr()), ctypes.c_void_p(qweight.data_ptr()), 
+            ctypes.c_void_p(scale.data_ptr()), ctypes.c_int32(num_row), ctypes.c_int32(num_chan_int), ctypes.c_int32(num_chan_fp16)],
+        )
+    elif input.dtype == torch.float16:
+        kernels.int4_to_fp16(
+            gridDim,
+            blockDim,
+            0,
+            stream,
+            [ctypes.c_void_p(out.data_ptr()), ctypes.c_void_p(qweight.data_ptr()),
+            ctypes.c_void_p(scale.data_ptr()), ctypes.c_int32(num_row), ctypes.c_int32(num_chan_int), ctypes.c_int32(num_chan_fp16)],
+        )
+    return out
+
+class QLinear(torch.nn.Module):
+    def __init__(self, bits: int, weight: torch.Tensor, bias=None):
+        super().__init__()
+        self.quant_bits = bits
+        self.scale = weight.abs().max(dim=-1).values / ((2 ** (bits - 1)) - 1)
+        self.scale = self.scale.to(torch.float32)
+        if self.quant_bits == 4:
+            self.weight = quant4(weight, self.scale)
+        elif self.quant_bits == 8:
+            self.weight = torch.round(weight.to(self.scale.dtype) / self.scale[:, None]).to(torch.int8)
+        if self.quant_bits == 8:
+            self.weight = self.weight.T
+        self.bias = None
+
+    def forward(self, input):
+        if self.quant_bits == 4:
+            assert(input.dtype == torch.bfloat16 or input.dtype == torch.float16)            
+
+        if self.weight.device != input.device:
+            self.weight = self.weight.to(input.device)
+            self.scale = self.scale.to(input.device)
+        
+        if self.quant_bits == 4:
+            self.scale = self.scale.to(input.dtype)
+            rweight = dequant4(self.weight, self.scale, input).T
+            output = torch.matmul(input, rweight)
+        elif self.quant_bits == 8:
+            rweight = self.weight.to(input.dtype) * self.scale.to(input.dtype)
+            output = torch.matmul(input, rweight)
+        if self.bias is not None:
+            output = output + self.bias
+        return output

special_tokens_map.json

@@ -0,0 +1,28 @@
+{
+  "additional_special_tokens": [
+    "<|inner_thoughts|>",
+    "<|system|>",
+    "<|assistant|>",
+    "<|prefix_begin|>",
+    "<|prefix_end|>",
+    "<eot>",
+    "<|prompter|>"
+  ],
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": true
+  },
+  "eos_token": "</s>",
+  "pad_token": "</s>",
+  "sep_token": "<s>",
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": true
+  }
+}

tokenization_baichuan.py

@@ -0,0 +1,232 @@
+# Copyright (c) 2023, Baichuan Intelligent Technology. All rights reserved.
+
+import os
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple
+
+import sentencepiece as spm
+from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model"}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {},
+    "tokenizer_file": {},
+}
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {}
+
+
+class BaichuanTokenizer(PreTrainedTokenizer):
+    """
+    Construct a Baichuan tokenizer. Based on byte-level Byte-Pair-Encoding.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        unk_token="<unk>",
+        bos_token="<s>",
+        eos_token="</s>",
+        pad_token=None,
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
+        add_bos_token=True,
+        add_eos_token=False,
+        clean_up_tokenization_spaces=False,
+        **kwargs,
+    ):
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+        bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
+        unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
+        pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
+        super().__init__(
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            add_bos_token=add_bos_token,
+            add_eos_token=add_eos_token,
+            sp_model_kwargs=self.sp_model_kwargs,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+        self.vocab_file = vocab_file
+        self.add_bos_token = add_bos_token
+        self.add_eos_token = add_eos_token
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(vocab_file)
+
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["sp_model"] = None
+        return state
+
+    def __setstate__(self, d):
+        self.__dict__ = d
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(self.vocab_file)
+
+    @property
+    def vocab_size(self):
+        """Returns vocab size"""
+        return self.sp_model.get_piece_size()
+
+    def get_vocab(self):
+        """Returns vocab as a dict"""
+        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def _tokenize(self, text):
+        """Returns a tokenized string."""
+        return self.sp_model.encode(text, out_type=str)
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.sp_model.piece_to_id(token)
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        token = self.sp_model.IdToPiece(index)
+        return token
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        current_sub_tokens = []
+        out_string = ""
+        prev_is_special = False
+        for i, token in enumerate(tokens):
+            # make sure that special tokens are not decoded using sentencepiece model
+            if token in self.all_special_tokens:
+                if not prev_is_special and i != 0:
+                    out_string += " "
+                out_string += self.sp_model.decode(current_sub_tokens) + token
+                prev_is_special = True
+                current_sub_tokens = []
+            else:
+                current_sub_tokens.append(token)
+                prev_is_special = False
+        out_string += self.sp_model.decode(current_sub_tokens)
+        return out_string
+
+    def save_vocabulary(self, save_directory, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        """
+        Save the vocabulary and special tokens file to a directory.
+
+        Args:
+            save_directory (`str`):
+                The directory in which to save the vocabulary.
+
+        Returns:
+            `Tuple(str)`: Paths to the files saved.
+        """
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+        elif not os.path.isfile(self.vocab_file):
+            with open(out_vocab_file, "wb") as fi:
+                content_spiece_model = self.sp_model.serialized_model_proto()
+                fi.write(content_spiece_model)
+
+        return (out_vocab_file,)
+
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+        eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+        output = bos_token_id + token_ids_0 + eos_token_id
+
+        if token_ids_1 is not None:
+            output = output + bos_token_id + token_ids_1 + eos_token_id
+
+        return output
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        bos_token_id = [1] if self.add_bos_token else []
+        eos_token_id = [1] if self.add_eos_token else []
+
+        if token_ids_1 is None:
+            return bos_token_id + ([0] * len(token_ids_0)) + eos_token_id
+        return (
+            bos_token_id
+            + ([0] * len(token_ids_0))
+            + eos_token_id
+            + bos_token_id
+            + ([0] * len(token_ids_1))
+            + eos_token_id
+        )
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Creates a mask from the two sequences passed to be used in a sequence-pair classification task. An ALBERT
+        sequence pair mask has the following format:
+
+        ```
+        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+        | first sequence    | second sequence |
+        ```
+
+        if token_ids_1 is None, only returns the first portion of the mask (0s).
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of ids.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
+        """
+        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+        eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+        output = [0] * len(bos_token_id + token_ids_0 + eos_token_id)
+
+        if token_ids_1 is not None:
+            output += [1] * len(bos_token_id + token_ids_1 + eos_token_id)
+
+        return output
+

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7d1ab69d25c74644af5c5e4dcd1cc6e96d33783dbd257b6bdea55b643c72813
+size 1136765

tokenizer_config.json

@@ -0,0 +1,46 @@
+{
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_baichuan.BaichuanTokenizer",
+      null
+    ]
+  },
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": true
+  },
+  "clean_up_tokenization_spaces": false,
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": true
+  },
+  "model_max_length": 4096,
+  "pad_token": {
+    "__type": "AddedToken",
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": true
+  },
+  "sp_model_kwargs": {},
+  "tokenizer_class": "BaichuanTokenizer",
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": true
+  }
+}

trainer_state.json

@@ -0,0 +1,3793 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.9999180932099271,
+  "global_step": 6104,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.0,
+      "learning_rate": 1.8e-07,
+      "loss": 1.4588,
+      "step": 10
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 3.8e-07,
+      "loss": 1.5163,
+      "step": 20
+    },
+    {
+      "epoch": 0.0,
+      "learning_rate": 5.800000000000001e-07,
+      "loss": 1.5383,
+      "step": 30
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 7.8e-07,
+      "loss": 1.4794,
+      "step": 40
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 9.800000000000001e-07,
+      "loss": 1.4675,
+      "step": 50
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 1.1800000000000001e-06,
+      "loss": 1.392,
+      "step": 60
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 1.3800000000000001e-06,
+      "loss": 1.4392,
+      "step": 70
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 1.5800000000000001e-06,
+      "loss": 1.3785,
+      "step": 80
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 1.7800000000000001e-06,
+      "loss": 1.3863,
+      "step": 90
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 1.98e-06,
+      "loss": 1.4214,
+      "step": 100
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 2.1800000000000003e-06,
+      "loss": 1.4741,
+      "step": 110
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 2.38e-06,
+      "loss": 1.3113,
+      "step": 120
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 2.5800000000000003e-06,
+      "loss": 1.3297,
+      "step": 130
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 2.7800000000000005e-06,
+      "loss": 1.3179,
+      "step": 140
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 2.9800000000000003e-06,
+      "loss": 1.3273,
+      "step": 150
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.1800000000000005e-06,
+      "loss": 1.3637,
+      "step": 160
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.3800000000000007e-06,
+      "loss": 1.4357,
+      "step": 170
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.58e-06,
+      "loss": 1.3675,
+      "step": 180
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.7800000000000002e-06,
+      "loss": 1.3398,
+      "step": 190
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 3.980000000000001e-06,
+      "loss": 1.3694,
+      "step": 200
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 4.18e-06,
+      "loss": 1.3456,
+      "step": 210
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 4.38e-06,
+      "loss": 1.3203,
+      "step": 220
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 4.58e-06,
+      "loss": 1.3678,
+      "step": 230
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 4.78e-06,
+      "loss": 1.3973,
+      "step": 240
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 4.980000000000001e-06,
+      "loss": 1.3017,
+      "step": 250
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 5.18e-06,
+      "loss": 1.2614,
+      "step": 260
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 5.380000000000001e-06,
+      "loss": 1.3756,
+      "step": 270
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 5.580000000000001e-06,
+      "loss": 1.4691,
+      "step": 280
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 5.78e-06,
+      "loss": 1.2958,
+      "step": 290
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 5.98e-06,
+      "loss": 1.3148,
+      "step": 300
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 6.18e-06,
+      "loss": 1.418,
+      "step": 310
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 6.380000000000001e-06,
+      "loss": 1.3144,
+      "step": 320
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 6.5800000000000005e-06,
+      "loss": 1.3724,
+      "step": 330
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 6.780000000000001e-06,
+      "loss": 1.363,
+      "step": 340
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 6.98e-06,
+      "loss": 1.2373,
+      "step": 350
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 7.180000000000001e-06,
+      "loss": 1.3572,
+      "step": 360
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 7.3800000000000005e-06,
+      "loss": 1.3494,
+      "step": 370
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 7.58e-06,
+      "loss": 1.3055,
+      "step": 380
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 7.78e-06,
+      "loss": 1.3626,
+      "step": 390
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 7.980000000000002e-06,
+      "loss": 1.3851,
+      "step": 400
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 8.18e-06,
+      "loss": 1.2608,
+      "step": 410
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 8.380000000000001e-06,
+      "loss": 1.3282,
+      "step": 420
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 8.580000000000001e-06,
+      "loss": 1.3424,
+      "step": 430
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 8.78e-06,
+      "loss": 1.3204,
+      "step": 440
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 8.98e-06,
+      "loss": 1.3753,
+      "step": 450
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 9.180000000000002e-06,
+      "loss": 1.2741,
+      "step": 460
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 9.38e-06,
+      "loss": 1.4024,
+      "step": 470
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 9.58e-06,
+      "loss": 1.2724,
+      "step": 480
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 9.780000000000001e-06,
+      "loss": 1.4165,
+      "step": 490
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 9.980000000000001e-06,
+      "loss": 1.3768,
+      "step": 500
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 1e-05,
+      "loss": 1.3327,
+      "step": 510
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 1e-05,
+      "loss": 1.3627,
+      "step": 520
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 1e-05,
+      "loss": 1.3674,
+      "step": 530
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 1e-05,
+      "loss": 1.4293,
+      "step": 540
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 1e-05,
+      "loss": 1.2829,
+      "step": 550
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 1e-05,
+      "loss": 1.4258,
+      "step": 560
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 1e-05,
+      "loss": 1.3552,
+      "step": 570
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 1e-05,
+      "loss": 1.4267,
+      "step": 580
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 1e-05,
+      "loss": 1.3482,
+      "step": 590
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 1e-05,
+      "loss": 1.3768,
+      "step": 600
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 1e-05,
+      "loss": 1.4167,
+      "step": 610
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 1e-05,
+      "loss": 1.3562,
+      "step": 620
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 1e-05,
+      "loss": 1.3819,
+      "step": 630
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 1e-05,
+      "loss": 1.3362,
+      "step": 640
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1e-05,
+      "loss": 1.3477,
+      "step": 650
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1e-05,
+      "loss": 1.3238,
+      "step": 660
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1e-05,
+      "loss": 1.3467,
+      "step": 670
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1e-05,
+      "loss": 1.3679,
+      "step": 680
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1e-05,
+      "loss": 1.325,
+      "step": 690
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 1e-05,
+      "loss": 1.2288,
+      "step": 700
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 1e-05,
+      "loss": 1.2394,
+      "step": 710
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 1e-05,
+      "loss": 1.2945,
+      "step": 720
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 1e-05,
+      "loss": 1.2176,
+      "step": 730
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 1e-05,
+      "loss": 1.3438,
+      "step": 740
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 1e-05,
+      "loss": 1.3514,
+      "step": 750
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 1e-05,
+      "loss": 1.3976,
+      "step": 760
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1e-05,
+      "loss": 1.3542,
+      "step": 770
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1e-05,
+      "loss": 1.3002,
+      "step": 780
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1e-05,
+      "loss": 1.2712,
+      "step": 790
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1e-05,
+      "loss": 1.427,
+      "step": 800
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1e-05,
+      "loss": 1.329,
+      "step": 810
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 1e-05,
+      "loss": 1.3053,
+      "step": 820
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.3724,
+      "step": 830
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.2535,
+      "step": 840
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.3213,
+      "step": 850
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.23,
+      "step": 860
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.2782,
+      "step": 870
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 1e-05,
+      "loss": 1.3225,
+      "step": 880
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1e-05,
+      "loss": 1.2838,
+      "step": 890
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1e-05,
+      "loss": 1.3434,
+      "step": 900
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1e-05,
+      "loss": 1.2737,
+      "step": 910
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1e-05,
+      "loss": 1.3334,
+      "step": 920
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1e-05,
+      "loss": 1.2737,
+      "step": 930
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 1e-05,
+      "loss": 1.3841,
+      "step": 940
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 1e-05,
+      "loss": 1.3208,
+      "step": 950
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 1e-05,
+      "loss": 1.2985,
+      "step": 960
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 1e-05,
+      "loss": 1.3755,
+      "step": 970
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 1e-05,
+      "loss": 1.2616,
+      "step": 980
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 1e-05,
+      "loss": 1.3529,
+      "step": 990
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 1e-05,
+      "loss": 1.3163,
+      "step": 1000
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.2914,
+      "step": 1010
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.4547,
+      "step": 1020
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.2758,
+      "step": 1030
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.3591,
+      "step": 1040
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.3162,
+      "step": 1050
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 1e-05,
+      "loss": 1.2589,
+      "step": 1060
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1e-05,
+      "loss": 1.3123,
+      "step": 1070
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1e-05,
+      "loss": 1.3089,
+      "step": 1080
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1e-05,
+      "loss": 1.2332,
+      "step": 1090
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1e-05,
+      "loss": 1.3301,
+      "step": 1100
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1e-05,
+      "loss": 1.3039,
+      "step": 1110
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 1e-05,
+      "loss": 1.3132,
+      "step": 1120
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 1.3389,
+      "step": 1130
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 1.2938,
+      "step": 1140
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 1.2528,
+      "step": 1150
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 1.3062,
+      "step": 1160
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 1.2864,
+      "step": 1170
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 1.2356,
+      "step": 1180
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 1e-05,
+      "loss": 1.2604,
+      "step": 1190
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 1e-05,
+      "loss": 1.2845,
+      "step": 1200
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 1e-05,
+      "loss": 1.3801,
+      "step": 1210
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 1e-05,
+      "loss": 1.3173,
+      "step": 1220
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 1e-05,
+      "loss": 1.272,
+      "step": 1230
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 1e-05,
+      "loss": 1.3123,
+      "step": 1240
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 1e-05,
+      "loss": 1.3632,
+      "step": 1250
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 1e-05,
+      "loss": 1.3333,
+      "step": 1260
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 1e-05,
+      "loss": 1.3122,
+      "step": 1270
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 1e-05,
+      "loss": 1.317,
+      "step": 1280
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 1e-05,
+      "loss": 1.2972,
+      "step": 1290
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 1e-05,
+      "loss": 1.2496,
+      "step": 1300
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 1e-05,
+      "loss": 1.2843,
+      "step": 1310
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1e-05,
+      "loss": 1.3373,
+      "step": 1320
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1e-05,
+      "loss": 1.2508,
+      "step": 1330
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1e-05,
+      "loss": 1.4132,
+      "step": 1340
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1e-05,
+      "loss": 1.2656,
+      "step": 1350
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1e-05,
+      "loss": 1.3105,
+      "step": 1360
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 1e-05,
+      "loss": 1.3734,
+      "step": 1370
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1e-05,
+      "loss": 1.3538,
+      "step": 1380
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1e-05,
+      "loss": 1.2767,
+      "step": 1390
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1e-05,
+      "loss": 1.4437,
+      "step": 1400
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1e-05,
+      "loss": 1.2782,
+      "step": 1410
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1e-05,
+      "loss": 1.2588,
+      "step": 1420
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 1e-05,
+      "loss": 1.3658,
+      "step": 1430
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1e-05,
+      "loss": 1.3161,
+      "step": 1440
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1e-05,
+      "loss": 1.3481,
+      "step": 1450
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1e-05,
+      "loss": 1.3329,
+      "step": 1460
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1e-05,
+      "loss": 1.2643,
+      "step": 1470
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1e-05,
+      "loss": 1.3022,
+      "step": 1480
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 1e-05,
+      "loss": 1.2482,
+      "step": 1490
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 1e-05,
+      "loss": 1.3112,
+      "step": 1500
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 1e-05,
+      "loss": 1.335,
+      "step": 1510
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 1e-05,
+      "loss": 1.2738,
+      "step": 1520
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 1e-05,
+      "loss": 1.3219,
+      "step": 1530
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 1e-05,
+      "loss": 1.3446,
+      "step": 1540
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 1e-05,
+      "loss": 1.2233,
+      "step": 1550
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1e-05,
+      "loss": 1.2723,
+      "step": 1560
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1e-05,
+      "loss": 1.3748,
+      "step": 1570
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1e-05,
+      "loss": 1.3433,
+      "step": 1580
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1e-05,
+      "loss": 1.2581,
+      "step": 1590
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1e-05,
+      "loss": 1.2938,
+      "step": 1600
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 1e-05,
+      "loss": 1.3463,
+      "step": 1610
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1e-05,
+      "loss": 1.3424,
+      "step": 1620
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1e-05,
+      "loss": 1.2639,
+      "step": 1630
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1e-05,
+      "loss": 1.3822,
+      "step": 1640
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1e-05,
+      "loss": 1.3105,
+      "step": 1650
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1e-05,
+      "loss": 1.3273,
+      "step": 1660
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 1e-05,
+      "loss": 1.4105,
+      "step": 1670
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1e-05,
+      "loss": 1.3346,
+      "step": 1680
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1e-05,
+      "loss": 1.3361,
+      "step": 1690
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1e-05,
+      "loss": 1.3085,
+      "step": 1700
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1e-05,
+      "loss": 1.3822,
+      "step": 1710
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1e-05,
+      "loss": 1.2784,
+      "step": 1720
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 1e-05,
+      "loss": 1.3267,
+      "step": 1730
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 1.2102,
+      "step": 1740
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 1.2318,
+      "step": 1750
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 1.2865,
+      "step": 1760
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 1.4117,
+      "step": 1770
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 1.2947,
+      "step": 1780
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 1.3536,
+      "step": 1790
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 1e-05,
+      "loss": 1.3179,
+      "step": 1800
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 1e-05,
+      "loss": 1.3403,
+      "step": 1810
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 1e-05,
+      "loss": 1.3984,
+      "step": 1820
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 1e-05,
+      "loss": 1.2817,
+      "step": 1830
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 1e-05,
+      "loss": 1.3239,
+      "step": 1840
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 1e-05,
+      "loss": 1.3182,
+      "step": 1850
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 1e-05,
+      "loss": 1.3002,
+      "step": 1860
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1e-05,
+      "loss": 1.2652,
+      "step": 1870
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1e-05,
+      "loss": 1.2937,
+      "step": 1880
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1e-05,
+      "loss": 1.3046,
+      "step": 1890
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1e-05,
+      "loss": 1.3152,
+      "step": 1900
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1e-05,
+      "loss": 1.3017,
+      "step": 1910
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 1e-05,
+      "loss": 1.2774,
+      "step": 1920
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1e-05,
+      "loss": 1.2409,
+      "step": 1930
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1e-05,
+      "loss": 1.2773,
+      "step": 1940
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1e-05,
+      "loss": 1.3688,
+      "step": 1950
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1e-05,
+      "loss": 1.3041,
+      "step": 1960
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1e-05,
+      "loss": 1.3089,
+      "step": 1970
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 1e-05,
+      "loss": 1.2131,
+      "step": 1980
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1e-05,
+      "loss": 1.3107,
+      "step": 1990
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1e-05,
+      "loss": 1.2902,
+      "step": 2000
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1e-05,
+      "loss": 1.2635,
+      "step": 2010
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1e-05,
+      "loss": 1.3283,
+      "step": 2020
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1e-05,
+      "loss": 1.37,
+      "step": 2030
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 1e-05,
+      "loss": 1.3471,
+      "step": 2040
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 1e-05,
+      "loss": 1.2582,
+      "step": 2050
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 1e-05,
+      "loss": 1.2883,
+      "step": 2060
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 1e-05,
+      "loss": 1.3454,
+      "step": 2070
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 1e-05,
+      "loss": 1.2927,
+      "step": 2080
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 1e-05,
+      "loss": 1.337,
+      "step": 2090
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 1e-05,
+      "loss": 1.2725,
+      "step": 2100
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1e-05,
+      "loss": 1.3667,
+      "step": 2110
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1e-05,
+      "loss": 1.2807,
+      "step": 2120
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1e-05,
+      "loss": 1.283,
+      "step": 2130
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1e-05,
+      "loss": 1.2922,
+      "step": 2140
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1e-05,
+      "loss": 1.2685,
+      "step": 2150
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 1e-05,
+      "loss": 1.2355,
+      "step": 2160
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1e-05,
+      "loss": 1.2418,
+      "step": 2170
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1e-05,
+      "loss": 1.2759,
+      "step": 2180
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1e-05,
+      "loss": 1.2873,
+      "step": 2190
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1e-05,
+      "loss": 1.2961,
+      "step": 2200
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1e-05,
+      "loss": 1.3602,
+      "step": 2210
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 1e-05,
+      "loss": 1.463,
+      "step": 2220
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1e-05,
+      "loss": 1.2493,
+      "step": 2230
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1e-05,
+      "loss": 1.2928,
+      "step": 2240
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1e-05,
+      "loss": 1.2797,
+      "step": 2250
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1e-05,
+      "loss": 1.2727,
+      "step": 2260
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1e-05,
+      "loss": 1.3153,
+      "step": 2270
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 1e-05,
+      "loss": 1.3328,
+      "step": 2280
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 1.312,
+      "step": 2290
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 1.3117,
+      "step": 2300
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 1.3142,
+      "step": 2310
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 1.3553,
+      "step": 2320
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 1.3259,
+      "step": 2330
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 1.2908,
+      "step": 2340
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 1e-05,
+      "loss": 1.3421,
+      "step": 2350
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 1e-05,
+      "loss": 1.2947,
+      "step": 2360
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 1e-05,
+      "loss": 1.2784,
+      "step": 2370
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 1e-05,
+      "loss": 1.2878,
+      "step": 2380
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 1e-05,
+      "loss": 1.2465,
+      "step": 2390
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 1e-05,
+      "loss": 1.2567,
+      "step": 2400
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 1e-05,
+      "loss": 1.3259,
+      "step": 2410
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 1e-05,
+      "loss": 1.3074,
+      "step": 2420
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 1e-05,
+      "loss": 1.2679,
+      "step": 2430
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 1e-05,
+      "loss": 1.2891,
+      "step": 2440
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 1e-05,
+      "loss": 1.2942,
+      "step": 2450
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 1e-05,
+      "loss": 1.309,
+      "step": 2460
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 1e-05,
+      "loss": 1.2833,
+      "step": 2470
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 1e-05,
+      "loss": 1.2876,
+      "step": 2480
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 1e-05,
+      "loss": 1.3055,
+      "step": 2490
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 1e-05,
+      "loss": 1.2866,
+      "step": 2500
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 1e-05,
+      "loss": 1.2896,
+      "step": 2510
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 1e-05,
+      "loss": 1.2554,
+      "step": 2520
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 1e-05,
+      "loss": 1.3605,
+      "step": 2530
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 1e-05,
+      "loss": 1.2808,
+      "step": 2540
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 1e-05,
+      "loss": 1.3809,
+      "step": 2550
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 1e-05,
+      "loss": 1.2405,
+      "step": 2560
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 1e-05,
+      "loss": 1.3486,
+      "step": 2570
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 1e-05,
+      "loss": 1.2717,
+      "step": 2580
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 1e-05,
+      "loss": 1.2951,
+      "step": 2590
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 1e-05,
+      "loss": 1.3206,
+      "step": 2600
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 1e-05,
+      "loss": 1.2649,
+      "step": 2610
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 1e-05,
+      "loss": 1.3111,
+      "step": 2620
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 1e-05,
+      "loss": 1.3412,
+      "step": 2630
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 1e-05,
+      "loss": 1.3971,
+      "step": 2640
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 1e-05,
+      "loss": 1.2913,
+      "step": 2650
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 1e-05,
+      "loss": 1.3284,
+      "step": 2660
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 1e-05,
+      "loss": 1.233,
+      "step": 2670
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 1e-05,
+      "loss": 1.2013,
+      "step": 2680
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 1e-05,
+      "loss": 1.3606,
+      "step": 2690
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 1e-05,
+      "loss": 1.3042,
+      "step": 2700
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 1e-05,
+      "loss": 1.3331,
+      "step": 2710
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 1e-05,
+      "loss": 1.3148,
+      "step": 2720
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 1e-05,
+      "loss": 1.2421,
+      "step": 2730
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 1e-05,
+      "loss": 1.3807,
+      "step": 2740
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 1e-05,
+      "loss": 1.3048,
+      "step": 2750
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 1e-05,
+      "loss": 1.3048,
+      "step": 2760
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 1e-05,
+      "loss": 1.3163,
+      "step": 2770
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 1e-05,
+      "loss": 1.2634,
+      "step": 2780
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 1e-05,
+      "loss": 1.36,
+      "step": 2790
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 1e-05,
+      "loss": 1.2809,
+      "step": 2800
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 1e-05,
+      "loss": 1.283,
+      "step": 2810
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 1e-05,
+      "loss": 1.3317,
+      "step": 2820
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 1e-05,
+      "loss": 1.3165,
+      "step": 2830
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 1e-05,
+      "loss": 1.3155,
+      "step": 2840
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 1e-05,
+      "loss": 1.1975,
+      "step": 2850
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 1e-05,
+      "loss": 1.1896,
+      "step": 2860
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 1e-05,
+      "loss": 1.208,
+      "step": 2870
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 1e-05,
+      "loss": 1.2704,
+      "step": 2880
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 1e-05,
+      "loss": 1.2951,
+      "step": 2890
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 1e-05,
+      "loss": 1.1931,
+      "step": 2900
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 1e-05,
+      "loss": 1.2505,
+      "step": 2910
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 1e-05,
+      "loss": 1.2654,
+      "step": 2920
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 1e-05,
+      "loss": 1.378,
+      "step": 2930
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 1e-05,
+      "loss": 1.3106,
+      "step": 2940
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 1e-05,
+      "loss": 1.2436,
+      "step": 2950
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 1e-05,
+      "loss": 1.3263,
+      "step": 2960
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 1e-05,
+      "loss": 1.315,
+      "step": 2970
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 1e-05,
+      "loss": 1.3096,
+      "step": 2980
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 1e-05,
+      "loss": 1.3328,
+      "step": 2990
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 1e-05,
+      "loss": 1.273,
+      "step": 3000
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 1e-05,
+      "loss": 1.2463,
+      "step": 3010
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 1e-05,
+      "loss": 1.313,
+      "step": 3020
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 1e-05,
+      "loss": 1.2886,
+      "step": 3030
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 1e-05,
+      "loss": 1.2971,
+      "step": 3040
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 1e-05,
+      "loss": 1.3871,
+      "step": 3050
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 1e-05,
+      "loss": 1.2677,
+      "step": 3060
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 1e-05,
+      "loss": 1.3094,
+      "step": 3070
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 1e-05,
+      "loss": 1.2652,
+      "step": 3080
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 1e-05,
+      "loss": 1.2308,
+      "step": 3090
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 1e-05,
+      "loss": 1.3286,
+      "step": 3100
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 1e-05,
+      "loss": 1.2933,
+      "step": 3110
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 1e-05,
+      "loss": 1.2355,
+      "step": 3120
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 1e-05,
+      "loss": 1.3138,
+      "step": 3130
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 1e-05,
+      "loss": 1.2825,
+      "step": 3140
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 1e-05,
+      "loss": 1.258,
+      "step": 3150
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 1e-05,
+      "loss": 1.2926,
+      "step": 3160
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 1e-05,
+      "loss": 1.3267,
+      "step": 3170
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 1e-05,
+      "loss": 1.334,
+      "step": 3180
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 1e-05,
+      "loss": 1.2268,
+      "step": 3190
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 1e-05,
+      "loss": 1.3248,
+      "step": 3200
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 1e-05,
+      "loss": 1.2997,
+      "step": 3210
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 1e-05,
+      "loss": 1.259,
+      "step": 3220
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 1e-05,
+      "loss": 1.3282,
+      "step": 3230
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 1e-05,
+      "loss": 1.2567,
+      "step": 3240
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 1e-05,
+      "loss": 1.3052,
+      "step": 3250
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 1e-05,
+      "loss": 1.2964,
+      "step": 3260
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 1e-05,
+      "loss": 1.2399,
+      "step": 3270
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 1e-05,
+      "loss": 1.2876,
+      "step": 3280
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 1e-05,
+      "loss": 1.2853,
+      "step": 3290
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 1e-05,
+      "loss": 1.3239,
+      "step": 3300
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 1e-05,
+      "loss": 1.2694,
+      "step": 3310
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 1e-05,
+      "loss": 1.3172,
+      "step": 3320
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 1e-05,
+      "loss": 1.2988,
+      "step": 3330
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 1e-05,
+      "loss": 1.246,
+      "step": 3340
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 1e-05,
+      "loss": 1.302,
+      "step": 3350
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 1e-05,
+      "loss": 1.2601,
+      "step": 3360
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 1e-05,
+      "loss": 1.3388,
+      "step": 3370
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 1e-05,
+      "loss": 1.1941,
+      "step": 3380
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 1e-05,
+      "loss": 1.1759,
+      "step": 3390
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 1e-05,
+      "loss": 1.2403,
+      "step": 3400
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 1e-05,
+      "loss": 1.2024,
+      "step": 3410
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 1e-05,
+      "loss": 1.2752,
+      "step": 3420
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 1e-05,
+      "loss": 1.2273,
+      "step": 3430
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 1e-05,
+      "loss": 1.2261,
+      "step": 3440
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 1e-05,
+      "loss": 1.2652,
+      "step": 3450
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 1e-05,
+      "loss": 1.3601,
+      "step": 3460
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 1e-05,
+      "loss": 1.3543,
+      "step": 3470
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 1e-05,
+      "loss": 1.3679,
+      "step": 3480
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 1e-05,
+      "loss": 1.2009,
+      "step": 3490
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 1e-05,
+      "loss": 1.2414,
+      "step": 3500
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 1e-05,
+      "loss": 1.1929,
+      "step": 3510
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 1e-05,
+      "loss": 1.2947,
+      "step": 3520
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 1e-05,
+      "loss": 1.2843,
+      "step": 3530
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 1e-05,
+      "loss": 1.2223,
+      "step": 3540
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 1e-05,
+      "loss": 1.2569,
+      "step": 3550
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 1e-05,
+      "loss": 1.1927,
+      "step": 3560
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 1e-05,
+      "loss": 1.3239,
+      "step": 3570
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 1e-05,
+      "loss": 1.3273,
+      "step": 3580
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 1e-05,
+      "loss": 1.2522,
+      "step": 3590
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 1e-05,
+      "loss": 1.3084,
+      "step": 3600
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 1e-05,
+      "loss": 1.2368,
+      "step": 3610
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 1e-05,
+      "loss": 1.2644,
+      "step": 3620
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 1e-05,
+      "loss": 1.2636,
+      "step": 3630
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 1e-05,
+      "loss": 1.2676,
+      "step": 3640
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 1e-05,
+      "loss": 1.2844,
+      "step": 3650
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 1e-05,
+      "loss": 1.3151,
+      "step": 3660
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 1e-05,
+      "loss": 1.2261,
+      "step": 3670
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 1e-05,
+      "loss": 1.2124,
+      "step": 3680
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 1e-05,
+      "loss": 1.2555,
+      "step": 3690
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 1e-05,
+      "loss": 1.2169,
+      "step": 3700
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 1e-05,
+      "loss": 1.3497,
+      "step": 3710
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 1e-05,
+      "loss": 1.2615,
+      "step": 3720
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 1e-05,
+      "loss": 1.28,
+      "step": 3730
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 1e-05,
+      "loss": 1.2943,
+      "step": 3740
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 1e-05,
+      "loss": 1.2596,
+      "step": 3750
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 1e-05,
+      "loss": 1.2416,
+      "step": 3760
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 1e-05,
+      "loss": 1.2909,
+      "step": 3770
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 1e-05,
+      "loss": 1.2281,
+      "step": 3780
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 1e-05,
+      "loss": 1.2071,
+      "step": 3790
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 1e-05,
+      "loss": 1.279,
+      "step": 3800
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 1e-05,
+      "loss": 1.2921,
+      "step": 3810
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 1e-05,
+      "loss": 1.2398,
+      "step": 3820
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 1e-05,
+      "loss": 1.2424,
+      "step": 3830
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 1e-05,
+      "loss": 1.2615,
+      "step": 3840
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 1e-05,
+      "loss": 1.3171,
+      "step": 3850
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 1e-05,
+      "loss": 1.3122,
+      "step": 3860
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 1e-05,
+      "loss": 1.2333,
+      "step": 3870
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 1e-05,
+      "loss": 1.3382,
+      "step": 3880
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 1e-05,
+      "loss": 1.2945,
+      "step": 3890
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 1e-05,
+      "loss": 1.2718,
+      "step": 3900
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 1e-05,
+      "loss": 1.3351,
+      "step": 3910
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 1e-05,
+      "loss": 1.2759,
+      "step": 3920
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 1e-05,
+      "loss": 1.3349,
+      "step": 3930
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 1e-05,
+      "loss": 1.2619,
+      "step": 3940
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 1e-05,
+      "loss": 1.2832,
+      "step": 3950
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 1e-05,
+      "loss": 1.2646,
+      "step": 3960
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 1e-05,
+      "loss": 1.3354,
+      "step": 3970
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 1e-05,
+      "loss": 1.3257,
+      "step": 3980
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 1e-05,
+      "loss": 1.2637,
+      "step": 3990
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 1e-05,
+      "loss": 1.3233,
+      "step": 4000
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 1e-05,
+      "loss": 1.2305,
+      "step": 4010
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 1e-05,
+      "loss": 1.3121,
+      "step": 4020
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 1e-05,
+      "loss": 1.2502,
+      "step": 4030
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 1e-05,
+      "loss": 1.3405,
+      "step": 4040
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 1e-05,
+      "loss": 1.265,
+      "step": 4050
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 1e-05,
+      "loss": 1.3124,
+      "step": 4060
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 1e-05,
+      "loss": 1.2664,
+      "step": 4070
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 1e-05,
+      "loss": 1.3427,
+      "step": 4080
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 1e-05,
+      "loss": 1.2858,
+      "step": 4090
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 1e-05,
+      "loss": 1.2414,
+      "step": 4100
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 1e-05,
+      "loss": 1.1726,
+      "step": 4110
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 1e-05,
+      "loss": 1.2786,
+      "step": 4120
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 1e-05,
+      "loss": 1.2386,
+      "step": 4130
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 1e-05,
+      "loss": 1.2194,
+      "step": 4140
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 1e-05,
+      "loss": 1.3317,
+      "step": 4150
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 1e-05,
+      "loss": 1.3069,
+      "step": 4160
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 1e-05,
+      "loss": 1.182,
+      "step": 4170
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 1e-05,
+      "loss": 1.2819,
+      "step": 4180
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 1e-05,
+      "loss": 1.2253,
+      "step": 4190
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 1e-05,
+      "loss": 1.2657,
+      "step": 4200
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 1e-05,
+      "loss": 1.2418,
+      "step": 4210
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 1e-05,
+      "loss": 1.2285,
+      "step": 4220
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 1e-05,
+      "loss": 1.3214,
+      "step": 4230
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 1e-05,
+      "loss": 1.2696,
+      "step": 4240
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 1e-05,
+      "loss": 1.3138,
+      "step": 4250
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 1e-05,
+      "loss": 1.2085,
+      "step": 4260
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 1e-05,
+      "loss": 1.2584,
+      "step": 4270
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 1e-05,
+      "loss": 1.2333,
+      "step": 4280
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 1e-05,
+      "loss": 1.3303,
+      "step": 4290
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 1e-05,
+      "loss": 1.2612,
+      "step": 4300
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 1e-05,
+      "loss": 1.1736,
+      "step": 4310
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 1e-05,
+      "loss": 1.2627,
+      "step": 4320
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 1e-05,
+      "loss": 1.3033,
+      "step": 4330
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 1e-05,
+      "loss": 1.2415,
+      "step": 4340
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 1e-05,
+      "loss": 1.2739,
+      "step": 4350
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 1e-05,
+      "loss": 1.2591,
+      "step": 4360
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 1e-05,
+      "loss": 1.2699,
+      "step": 4370
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 1e-05,
+      "loss": 1.3777,
+      "step": 4380
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 1e-05,
+      "loss": 1.2144,
+      "step": 4390
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 1e-05,
+      "loss": 1.3081,
+      "step": 4400
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 1e-05,
+      "loss": 1.2445,
+      "step": 4410
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 1e-05,
+      "loss": 1.1593,
+      "step": 4420
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 1e-05,
+      "loss": 1.2987,
+      "step": 4430
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 1e-05,
+      "loss": 1.2467,
+      "step": 4440
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 1e-05,
+      "loss": 1.2139,
+      "step": 4450
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 1e-05,
+      "loss": 1.2999,
+      "step": 4460
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 1e-05,
+      "loss": 1.2681,
+      "step": 4470
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 1e-05,
+      "loss": 1.3139,
+      "step": 4480
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 1e-05,
+      "loss": 1.2685,
+      "step": 4490
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 1e-05,
+      "loss": 1.3077,
+      "step": 4500
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 1e-05,
+      "loss": 1.2559,
+      "step": 4510
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 1e-05,
+      "loss": 1.2181,
+      "step": 4520
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 1e-05,
+      "loss": 1.3011,
+      "step": 4530
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 1e-05,
+      "loss": 1.3051,
+      "step": 4540
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 1e-05,
+      "loss": 1.2037,
+      "step": 4550
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 1e-05,
+      "loss": 1.2504,
+      "step": 4560
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 1e-05,
+      "loss": 1.212,
+      "step": 4570
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 1e-05,
+      "loss": 1.2554,
+      "step": 4580
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 1e-05,
+      "loss": 1.2767,
+      "step": 4590
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 1e-05,
+      "loss": 1.2845,
+      "step": 4600
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 1e-05,
+      "loss": 1.2142,
+      "step": 4610
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 1e-05,
+      "loss": 1.2353,
+      "step": 4620
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 1e-05,
+      "loss": 1.2294,
+      "step": 4630
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 1e-05,
+      "loss": 1.2984,
+      "step": 4640
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 1e-05,
+      "loss": 1.279,
+      "step": 4650
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 1e-05,
+      "loss": 1.2119,
+      "step": 4660
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 1e-05,
+      "loss": 1.2493,
+      "step": 4670
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 1e-05,
+      "loss": 1.1979,
+      "step": 4680
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 1e-05,
+      "loss": 1.2422,
+      "step": 4690
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 1e-05,
+      "loss": 1.2443,
+      "step": 4700
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 1e-05,
+      "loss": 1.3392,
+      "step": 4710
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 1e-05,
+      "loss": 1.315,
+      "step": 4720
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 1e-05,
+      "loss": 1.2761,
+      "step": 4730
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 1e-05,
+      "loss": 1.2308,
+      "step": 4740
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 1e-05,
+      "loss": 1.2628,
+      "step": 4750
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 1e-05,
+      "loss": 1.2887,
+      "step": 4760
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 1e-05,
+      "loss": 1.2942,
+      "step": 4770
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 1e-05,
+      "loss": 1.2921,
+      "step": 4780
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 1e-05,
+      "loss": 1.2017,
+      "step": 4790
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 1e-05,
+      "loss": 1.2328,
+      "step": 4800
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 1e-05,
+      "loss": 1.2817,
+      "step": 4810
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 1e-05,
+      "loss": 1.2046,
+      "step": 4820
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 1e-05,
+      "loss": 1.2962,
+      "step": 4830
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 1e-05,
+      "loss": 1.2223,
+      "step": 4840
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 1e-05,
+      "loss": 1.3326,
+      "step": 4850
+    },
+    {
+      "epoch": 0.8,
+      "learning_rate": 1e-05,
+      "loss": 1.2584,
+      "step": 4860
+    },
+    {
+      "epoch": 0.8,
+      "learning_rate": 1e-05,
+      "loss": 1.1835,
+      "step": 4870
+    },
+    {
+      "epoch": 0.8,
+      "learning_rate": 1e-05,
+      "loss": 1.2967,
+      "step": 4880
+    },
+    {
+      "epoch": 0.8,
+      "learning_rate": 1e-05,
+      "loss": 1.2358,
+      "step": 4890
+    },
+    {
+      "epoch": 0.8,
+      "learning_rate": 1e-05,
+      "loss": 1.2331,
+      "step": 4900
+    },
+    {
+      "epoch": 0.8,
+      "learning_rate": 1e-05,
+      "loss": 1.2126,
+      "step": 4910
+    },
+    {
+      "epoch": 0.81,
+      "learning_rate": 1e-05,
+      "loss": 1.3323,
+      "step": 4920
+    },
+    {
+      "epoch": 0.81,
+      "learning_rate": 1e-05,
+      "loss": 1.2067,
+      "step": 4930
+    },
+    {
+      "epoch": 0.81,
+      "learning_rate": 1e-05,
+      "loss": 1.2419,
+      "step": 4940
+    },
+    {
+      "epoch": 0.81,
+      "learning_rate": 1e-05,
+      "loss": 1.2713,
+      "step": 4950
+    },
+    {
+      "epoch": 0.81,
+      "learning_rate": 1e-05,
+      "loss": 1.1894,
+      "step": 4960
+    },
+    {
+      "epoch": 0.81,
+      "learning_rate": 1e-05,
+      "loss": 1.2695,
+      "step": 4970
+    },
+    {
+      "epoch": 0.82,
+      "learning_rate": 1e-05,
+      "loss": 1.2709,
+      "step": 4980
+    },
+    {
+      "epoch": 0.82,
+      "learning_rate": 1e-05,
+      "loss": 1.2718,
+      "step": 4990
+    },
+    {
+      "epoch": 0.82,
+      "learning_rate": 1e-05,
+      "loss": 1.31,
+      "step": 5000
+    },
+    {
+      "epoch": 0.82,
+      "learning_rate": 1e-05,
+      "loss": 1.2691,
+      "step": 5010
+    },
+    {
+      "epoch": 0.82,
+      "learning_rate": 1e-05,
+      "loss": 1.2704,
+      "step": 5020
+    },
+    {
+      "epoch": 0.82,
+      "learning_rate": 1e-05,
+      "loss": 1.2374,
+      "step": 5030
+    },
+    {
+      "epoch": 0.83,
+      "learning_rate": 1e-05,
+      "loss": 1.3116,
+      "step": 5040
+    },
+    {
+      "epoch": 0.83,
+      "learning_rate": 1e-05,
+      "loss": 1.3039,
+      "step": 5050
+    },
+    {
+      "epoch": 0.83,
+      "learning_rate": 1e-05,
+      "loss": 1.223,
+      "step": 5060
+    },
+    {
+      "epoch": 0.83,
+      "learning_rate": 1e-05,
+      "loss": 1.2047,
+      "step": 5070
+    },
+    {
+      "epoch": 0.83,
+      "learning_rate": 1e-05,
+      "loss": 1.2721,
+      "step": 5080
+    },
+    {
+      "epoch": 0.83,
+      "learning_rate": 1e-05,
+      "loss": 1.292,
+      "step": 5090
+    },
+    {
+      "epoch": 0.84,
+      "learning_rate": 1e-05,
+      "loss": 1.2201,
+      "step": 5100
+    },
+    {
+      "epoch": 0.84,
+      "learning_rate": 1e-05,
+      "loss": 1.2874,
+      "step": 5110
+    },
+    {
+      "epoch": 0.84,
+      "learning_rate": 1e-05,
+      "loss": 1.2849,
+      "step": 5120
+    },
+    {
+      "epoch": 0.84,
+      "learning_rate": 1e-05,
+      "loss": 1.2056,
+      "step": 5130
+    },
+    {
+      "epoch": 0.84,
+      "learning_rate": 1e-05,
+      "loss": 1.2918,
+      "step": 5140
+    },
+    {
+      "epoch": 0.84,
+      "learning_rate": 1e-05,
+      "loss": 1.2618,
+      "step": 5150
+    },
+    {
+      "epoch": 0.85,
+      "learning_rate": 1e-05,
+      "loss": 1.2592,
+      "step": 5160
+    },
+    {
+      "epoch": 0.85,
+      "learning_rate": 1e-05,
+      "loss": 1.221,
+      "step": 5170
+    },
+    {
+      "epoch": 0.85,
+      "learning_rate": 1e-05,
+      "loss": 1.2401,
+      "step": 5180
+    },
+    {
+      "epoch": 0.85,
+      "learning_rate": 1e-05,
+      "loss": 1.2268,
+      "step": 5190
+    },
+    {
+      "epoch": 0.85,
+      "learning_rate": 1e-05,
+      "loss": 1.2221,
+      "step": 5200
+    },
+    {
+      "epoch": 0.85,
+      "learning_rate": 1e-05,
+      "loss": 1.1223,
+      "step": 5210
+    },
+    {
+      "epoch": 0.86,
+      "learning_rate": 1e-05,
+      "loss": 1.1579,
+      "step": 5220
+    },
+    {
+      "epoch": 0.86,
+      "learning_rate": 1e-05,
+      "loss": 1.2795,
+      "step": 5230
+    },
+    {
+      "epoch": 0.86,
+      "learning_rate": 1e-05,
+      "loss": 1.2462,
+      "step": 5240
+    },
+    {
+      "epoch": 0.86,
+      "learning_rate": 1e-05,
+      "loss": 1.2158,
+      "step": 5250
+    },
+    {
+      "epoch": 0.86,
+      "learning_rate": 1e-05,
+      "loss": 1.1765,
+      "step": 5260
+    },
+    {
+      "epoch": 0.86,
+      "learning_rate": 1e-05,
+      "loss": 1.2117,
+      "step": 5270
+    },
+    {
+      "epoch": 0.86,
+      "learning_rate": 1e-05,
+      "loss": 1.2117,
+      "step": 5280
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 1e-05,
+      "loss": 1.204,
+      "step": 5290
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 1e-05,
+      "loss": 1.2754,
+      "step": 5300
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 1e-05,
+      "loss": 1.2453,
+      "step": 5310
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 1e-05,
+      "loss": 1.2434,
+      "step": 5320
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 1e-05,
+      "loss": 1.2598,
+      "step": 5330
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 1e-05,
+      "loss": 1.2409,
+      "step": 5340
+    },
+    {
+      "epoch": 0.88,
+      "learning_rate": 1e-05,
+      "loss": 1.2302,
+      "step": 5350
+    },
+    {
+      "epoch": 0.88,
+      "learning_rate": 1e-05,
+      "loss": 1.2618,
+      "step": 5360
+    },
+    {
+      "epoch": 0.88,
+      "learning_rate": 1e-05,
+      "loss": 1.3057,
+      "step": 5370
+    },
+    {
+      "epoch": 0.88,
+      "learning_rate": 1e-05,
+      "loss": 1.2942,
+      "step": 5380
+    },
+    {
+      "epoch": 0.88,
+      "learning_rate": 1e-05,
+      "loss": 1.1696,
+      "step": 5390
+    },
+    {
+      "epoch": 0.88,
+      "learning_rate": 1e-05,
+      "loss": 1.2329,
+      "step": 5400
+    },
+    {
+      "epoch": 0.89,
+      "learning_rate": 1e-05,
+      "loss": 1.2805,
+      "step": 5410
+    },
+    {
+      "epoch": 0.89,
+      "learning_rate": 1e-05,
+      "loss": 1.2741,
+      "step": 5420
+    },
+    {
+      "epoch": 0.89,
+      "learning_rate": 1e-05,
+      "loss": 1.2481,
+      "step": 5430
+    },
+    {
+      "epoch": 0.89,
+      "learning_rate": 1e-05,
+      "loss": 1.2315,
+      "step": 5440
+    },
+    {
+      "epoch": 0.89,
+      "learning_rate": 1e-05,
+      "loss": 1.3099,
+      "step": 5450
+    },
+    {
+      "epoch": 0.89,
+      "learning_rate": 1e-05,
+      "loss": 1.2454,
+      "step": 5460
+    },
+    {
+      "epoch": 0.9,
+      "learning_rate": 1e-05,
+      "loss": 1.2796,
+      "step": 5470
+    },
+    {
+      "epoch": 0.9,
+      "learning_rate": 1e-05,
+      "loss": 1.2591,
+      "step": 5480
+    },
+    {
+      "epoch": 0.9,
+      "learning_rate": 1e-05,
+      "loss": 1.2764,
+      "step": 5490
+    },
+    {
+      "epoch": 0.9,
+      "learning_rate": 1e-05,
+      "loss": 1.2662,
+      "step": 5500
+    },
+    {
+      "epoch": 0.9,
+      "learning_rate": 1e-05,
+      "loss": 1.2407,
+      "step": 5510
+    },
+    {
+      "epoch": 0.9,
+      "learning_rate": 1e-05,
+      "loss": 1.219,
+      "step": 5520
+    },
+    {
+      "epoch": 0.91,
+      "learning_rate": 1e-05,
+      "loss": 1.2577,
+      "step": 5530
+    },
+    {
+      "epoch": 0.91,
+      "learning_rate": 1e-05,
+      "loss": 1.2551,
+      "step": 5540
+    },
+    {
+      "epoch": 0.91,
+      "learning_rate": 1e-05,
+      "loss": 1.1574,
+      "step": 5550
+    },
+    {
+      "epoch": 0.91,
+      "learning_rate": 1e-05,
+      "loss": 1.2744,
+      "step": 5560
+    },
+    {
+      "epoch": 0.91,
+      "learning_rate": 1e-05,
+      "loss": 1.2122,
+      "step": 5570
+    },
+    {
+      "epoch": 0.91,
+      "learning_rate": 1e-05,
+      "loss": 1.2706,
+      "step": 5580
+    },
+    {
+      "epoch": 0.92,
+      "learning_rate": 1e-05,
+      "loss": 1.2703,
+      "step": 5590
+    },
+    {
+      "epoch": 0.92,
+      "learning_rate": 1e-05,
+      "loss": 1.2772,
+      "step": 5600
+    },
+    {
+      "epoch": 0.92,
+      "learning_rate": 1e-05,
+      "loss": 1.2145,
+      "step": 5610
+    },
+    {
+      "epoch": 0.92,
+      "learning_rate": 1e-05,
+      "loss": 1.3134,
+      "step": 5620
+    },
+    {
+      "epoch": 0.92,
+      "learning_rate": 1e-05,
+      "loss": 1.234,
+      "step": 5630
+    },
+    {
+      "epoch": 0.92,
+      "learning_rate": 1e-05,
+      "loss": 1.2484,
+      "step": 5640
+    },
+    {
+      "epoch": 0.93,
+      "learning_rate": 1e-05,
+      "loss": 1.1774,
+      "step": 5650
+    },
+    {
+      "epoch": 0.93,
+      "learning_rate": 1e-05,
+      "loss": 1.3045,
+      "step": 5660
+    },
+    {
+      "epoch": 0.93,
+      "learning_rate": 1e-05,
+      "loss": 1.2948,
+      "step": 5670
+    },
+    {
+      "epoch": 0.93,
+      "learning_rate": 1e-05,
+      "loss": 1.25,
+      "step": 5680
+    },
+    {
+      "epoch": 0.93,
+      "learning_rate": 1e-05,
+      "loss": 1.2265,
+      "step": 5690
+    },
+    {
+      "epoch": 0.93,
+      "learning_rate": 1e-05,
+      "loss": 1.3164,
+      "step": 5700
+    },
+    {
+      "epoch": 0.94,
+      "learning_rate": 1e-05,
+      "loss": 1.2688,
+      "step": 5710
+    },
+    {
+      "epoch": 0.94,
+      "learning_rate": 1e-05,
+      "loss": 1.2615,
+      "step": 5720
+    },
+    {
+      "epoch": 0.94,
+      "learning_rate": 1e-05,
+      "loss": 1.2785,
+      "step": 5730
+    },
+    {
+      "epoch": 0.94,
+      "learning_rate": 1e-05,
+      "loss": 1.2813,
+      "step": 5740
+    },
+    {
+      "epoch": 0.94,
+      "learning_rate": 1e-05,
+      "loss": 1.2287,
+      "step": 5750
+    },
+    {
+      "epoch": 0.94,
+      "learning_rate": 1e-05,
+      "loss": 1.2914,
+      "step": 5760
+    },
+    {
+      "epoch": 0.95,
+      "learning_rate": 1e-05,
+      "loss": 1.2133,
+      "step": 5770
+    },
+    {
+      "epoch": 0.95,
+      "learning_rate": 1e-05,
+      "loss": 1.2405,
+      "step": 5780
+    },
+    {
+      "epoch": 0.95,
+      "learning_rate": 1e-05,
+      "loss": 1.3153,
+      "step": 5790
+    },
+    {
+      "epoch": 0.95,
+      "learning_rate": 1e-05,
+      "loss": 1.2094,
+      "step": 5800
+    },
+    {
+      "epoch": 0.95,
+      "learning_rate": 1e-05,
+      "loss": 1.2593,
+      "step": 5810
+    },
+    {
+      "epoch": 0.95,
+      "learning_rate": 1e-05,
+      "loss": 1.3192,
+      "step": 5820
+    },
+    {
+      "epoch": 0.96,
+      "learning_rate": 1e-05,
+      "loss": 1.2264,
+      "step": 5830
+    },
+    {
+      "epoch": 0.96,
+      "learning_rate": 1e-05,
+      "loss": 1.267,
+      "step": 5840
+    },
+    {
+      "epoch": 0.96,
+      "learning_rate": 1e-05,
+      "loss": 1.1878,
+      "step": 5850
+    },
+    {
+      "epoch": 0.96,
+      "learning_rate": 1e-05,
+      "loss": 1.2058,
+      "step": 5860
+    },
+    {
+      "epoch": 0.96,
+      "learning_rate": 1e-05,
+      "loss": 1.1589,
+      "step": 5870
+    },
+    {
+      "epoch": 0.96,
+      "learning_rate": 1e-05,
+      "loss": 1.1993,
+      "step": 5880
+    },
+    {
+      "epoch": 0.96,
+      "learning_rate": 1e-05,
+      "loss": 1.2312,
+      "step": 5890
+    },
+    {
+      "epoch": 0.97,
+      "learning_rate": 1e-05,
+      "loss": 1.2511,
+      "step": 5900
+    },
+    {
+      "epoch": 0.97,
+      "learning_rate": 1e-05,
+      "loss": 1.3163,
+      "step": 5910
+    },
+    {
+      "epoch": 0.97,
+      "learning_rate": 1e-05,
+      "loss": 1.2523,
+      "step": 5920
+    },
+    {
+      "epoch": 0.97,
+      "learning_rate": 1e-05,
+      "loss": 1.3137,
+      "step": 5930
+    },
+    {
+      "epoch": 0.97,
+      "learning_rate": 1e-05,
+      "loss": 1.2225,
+      "step": 5940
+    },
+    {
+      "epoch": 0.97,
+      "learning_rate": 1e-05,
+      "loss": 1.2642,
+      "step": 5950
+    },
+    {
+      "epoch": 0.98,
+      "learning_rate": 1e-05,
+      "loss": 1.2146,
+      "step": 5960
+    },
+    {
+      "epoch": 0.98,
+      "learning_rate": 1e-05,
+      "loss": 1.3105,
+      "step": 5970
+    },
+    {
+      "epoch": 0.98,
+      "learning_rate": 1e-05,
+      "loss": 1.2762,
+      "step": 5980
+    },
+    {
+      "epoch": 0.98,
+      "learning_rate": 1e-05,
+      "loss": 1.2242,
+      "step": 5990
+    },
+    {
+      "epoch": 0.98,
+      "learning_rate": 1e-05,
+      "loss": 1.2385,
+      "step": 6000
+    },
+    {
+      "epoch": 0.98,
+      "learning_rate": 1e-05,
+      "loss": 1.2467,
+      "step": 6010
+    },
+    {
+      "epoch": 0.99,
+      "learning_rate": 1e-05,
+      "loss": 1.2547,
+      "step": 6020
+    },
+    {
+      "epoch": 0.99,
+      "learning_rate": 1e-05,
+      "loss": 1.2799,
+      "step": 6030
+    },
+    {
+      "epoch": 0.99,
+      "learning_rate": 1e-05,
+      "loss": 1.2916,
+      "step": 6040
+    },
+    {
+      "epoch": 0.99,
+      "learning_rate": 1e-05,
+      "loss": 1.2668,
+      "step": 6050
+    },
+    {
+      "epoch": 0.99,
+      "learning_rate": 1e-05,
+      "loss": 1.1854,
+      "step": 6060
+    },
+    {
+      "epoch": 0.99,
+      "learning_rate": 1e-05,
+      "loss": 1.2529,
+      "step": 6070
+    },
+    {
+      "epoch": 1.0,
+      "learning_rate": 1e-05,
+      "loss": 1.2653,
+      "step": 6080
+    },
+    {
+      "epoch": 1.0,
+      "learning_rate": 1e-05,
+      "loss": 1.2418,
+      "step": 6090
+    },
+    {
+      "epoch": 1.0,
+      "learning_rate": 1e-05,
+      "loss": 1.246,
+      "step": 6100
+    },
+    {
+      "epoch": 1.0,
+      "eval_oasst_export_accuracy": 0.6878104734550841,
+      "eval_oasst_export_loss": 1.349609375,
+      "eval_oasst_export_runtime": 19.3636,
+      "eval_oasst_export_samples_per_second": 15.493,
+      "eval_oasst_export_steps_per_second": 1.962,
+      "step": 6104
+    },
+    {
+      "epoch": 1.0,
+      "eval_dolly15k_accuracy": 0.6627859972511925,
+      "eval_dolly15k_loss": 1.2138671875,
+      "eval_dolly15k_runtime": 15.5602,
+      "eval_dolly15k_samples_per_second": 19.28,
+      "eval_dolly15k_steps_per_second": 2.442,
+      "step": 6104
+    },
+    {
+      "epoch": 1.0,
+      "eval__ruozhiba_accuracy": 0.5154994259471871,
+      "eval__ruozhiba_loss": 1.7412109375,
+      "eval__ruozhiba_runtime": 0.6619,
+      "eval__ruozhiba_samples_per_second": 10.576,
+      "eval__ruozhiba_steps_per_second": 1.511,
+      "step": 6104
+    },
+    {
+      "epoch": 1.0,
+      "eval__sharegpt_format_accuracy": 0.8137771622392778,
+      "eval__sharegpt_format_loss": 1.0361328125,
+      "eval__sharegpt_format_runtime": 32.5298,
+      "eval__sharegpt_format_samples_per_second": 15.371,
+      "eval__sharegpt_format_steps_per_second": 1.937,
+      "step": 6104
+    },
+    {
+      "epoch": 1.0,
+      "eval__similar_question_accuracy": 0.577728776185226,
+      "eval__similar_question_loss": 1.6435546875,
+      "eval__similar_question_runtime": 0.7886,
+      "eval__similar_question_samples_per_second": 8.877,
+      "eval__similar_question_steps_per_second": 1.268,
+      "step": 6104
+    },
+    {
+      "epoch": 1.0,
+      "eval__open_domain_subject_accuracy": 0.6082298514730947,
+      "eval__open_domain_subject_loss": 1.451171875,
+      "eval__open_domain_subject_runtime": 2.7365,
+      "eval__open_domain_subject_samples_per_second": 16.81,
+      "eval__open_domain_subject_steps_per_second": 2.193,
+      "step": 6104
+    },
+    {
+      "epoch": 1.0,
+      "eval__empdia_ly_07_17_accuracy": 0.6261699064074874,
+      "eval__empdia_ly_07_17_loss": 1.33984375,
+      "eval__empdia_ly_07_17_runtime": 5.1144,
+      "eval__empdia_ly_07_17_samples_per_second": 17.988,
+      "eval__empdia_ly_07_17_steps_per_second": 2.346,
+      "step": 6104
+    },
+    {
+      "epoch": 1.0,
+      "eval__zuowen_accuracy": 0.9144507380228243,
+      "eval__zuowen_loss": 0.299560546875,
+      "eval__zuowen_runtime": 6.9808,
+      "eval__zuowen_samples_per_second": 14.325,
+      "eval__zuowen_steps_per_second": 1.862,
+      "step": 6104
+    },
+    {
+      "epoch": 1.0,
+      "eval__psychat_6_29_and_11_15_mix_accuracy": 0.5055779183438758,
+      "eval__psychat_6_29_and_11_15_mix_loss": 1.9296875,
+      "eval__psychat_6_29_and_11_15_mix_runtime": 6.0911,
+      "eval__psychat_6_29_and_11_15_mix_samples_per_second": 16.417,
+      "eval__psychat_6_29_and_11_15_mix_steps_per_second": 2.134,
+      "step": 6104
+    },
+    {
+      "epoch": 1.0,
+      "eval__psy_qa_1208_accuracy": 0.49772520473157417,
+      "eval__psy_qa_1208_loss": 2.26171875,
+      "eval__psy_qa_1208_runtime": 2.7612,
+      "eval__psy_qa_1208_samples_per_second": 16.297,
+      "eval__psy_qa_1208_steps_per_second": 2.173,
+      "step": 6104
+    },
+    {
+      "epoch": 1.0,
+      "eval__socrates_teaching_accuracy": 0.6412340032115226,
+      "eval__socrates_teaching_loss": 1.3828125,
+      "eval__socrates_teaching_runtime": 5.1769,
+      "eval__socrates_teaching_samples_per_second": 16.612,
+      "eval__socrates_teaching_steps_per_second": 2.125,
+      "step": 6104
+    },
+    {
+      "epoch": 1.0,
+      "eval__socrates_teaching_math1213_accuracy": 0.9996854356715948,
+      "eval__socrates_teaching_math1213_loss": 0.0015439987182617188,
+      "eval__socrates_teaching_math1213_runtime": 1.6723,
+      "eval__socrates_teaching_math1213_samples_per_second": 14.949,
+      "eval__socrates_teaching_math1213_steps_per_second": 2.392,
+      "step": 6104
+    },
+    {
+      "epoch": 1.0,
+      "eval__search_accuracy": 0.6996122053270742,
+      "eval__search_loss": 1.099609375,
+      "eval__search_runtime": 5.9225,
+      "eval__search_samples_per_second": 16.885,
+      "eval__search_steps_per_second": 2.195,
+      "step": 6104
+    }
+  ],
+  "max_steps": 12208,
+  "num_train_epochs": 2,
+  "total_flos": 5281029455085568.0,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:535ba74a271c3cad7b287704e9bf2a2d5382df0492157794efcf2367722a2425
+size 6203

zero_to_fp32.py

@@ -0,0 +1,587 @@
+#!/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 1, 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_dict = torch.load(f, map_location=device)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    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("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    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, tag=args.tag)