Add files using upload-large-folder tool

added_tokens.json

@@ -0,0 +1,13 @@
+{
+  "<|assistant|>": 32001,
+  "<|endoftext|>": 32000,
+  "<|end|>": 32007,
+  "<|placeholder1|>": 32002,
+  "<|placeholder2|>": 32003,
+  "<|placeholder3|>": 32004,
+  "<|placeholder4|>": 32005,
+  "<|placeholder5|>": 32008,
+  "<|placeholder6|>": 32009,
+  "<|system|>": 32006,
+  "<|user|>": 32010
+}

chat_template.jinja

@@ -0,0 +1,5 @@
+{% for message in messages %}{% if (message['role'] == 'user') %}{{'<|user|>' + '
+' + message['content'] + '<|end|>' + '
+' + '<|assistant|>' + '
+'}}{% elif (message['role'] == 'assistant') %}{{message['content'] + '<|end|>' + '
+'}}{% endif %}{% endfor %}

config.json

@@ -0,0 +1,170 @@
+{
+  "architectures": [
+    "Phi3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_phi3.Phi3Config",
+    "AutoModel": "modeling_phi3.Phi3ForCausalLM",
+    "AutoModelForCausalLM": "microsoft/Phi-3-medium-128k-instruct--modeling_phi3.Phi3ForCausalLM"
+  },
+  "bos_token_id": 1,
+  "embd_pdrop": 0.0,
+  "eos_token_id": 32000,
+  "hidden_act": "silu",
+  "hidden_size": 5120,
+  "initializer_range": 0.02,
+  "intermediate_size": 17920,
+  "max_position_embeddings": 131072,
+  "model_type": "phi3",
+  "num_attention_heads": 40,
+  "num_hidden_layers": 40,
+  "num_key_value_heads": 10,
+  "original_max_position_embeddings": 4096,
+  "pad_token_id": null,
+  "resid_pdrop": 0.0,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": {
+    "long_factor": [
+      1.0,
+      1.0,
+      1.0,
+      1.0,
+      1.0,
+      1.0,
+      1.0,
+      1.0,
+      1.0,
+      1.0,
+      1.0,
+      1.0,
+      1.0,
+      1.25,
+      1.25,
+      1.5,
+      2.0,
+      2.75,
+      5.75,
+      5.75,
+      6.5,
+      9.25,
+      11.0,
+      13.25,
+      19.25,
+      19.75,
+      19.75,
+      21.25,
+      21.5,
+      26.5,
+      30.0,
+      33.75,
+      35.25,
+      38.5,
+      42.0,
+      42.25,
+      46.0,
+      47.0,
+      50.0,
+      50.5,
+      51.0,
+      52.0,
+      52.75,
+      53.75,
+      54.75,
+      57.0,
+      57.25,
+      58.5,
+      59.25,
+      59.5,
+      62.0,
+      62.5,
+      62.75,
+      63.25,
+      63.25,
+      63.25,
+      63.75,
+      64.0,
+      64.0,
+      64.25,
+      64.5,
+      64.5,
+      65.0,
+      65.0
+    ],
+    "short_factor": [
+      1.0,
+      1.0,
+      1.0,
+      1.0,
+      1.0,
+      1.0,
+      1.01,
+      1.02,
+      1.02,
+      1.04,
+      1.04,
+      1.07,
+      1.07,
+      1.1,
+      1.3000000000000003,
+      1.3000000000000003,
+      1.5000000000000004,
+      1.5700000000000005,
+      1.9000000000000008,
+      2.3100000000000014,
+      2.759999999999992,
+      3.3899999999999784,
+      3.9399999999999666,
+      4.009999999999965,
+      4.289999999999959,
+      4.349999999999958,
+      5.349999999999937,
+      6.659999999999909,
+      7.029999999999901,
+      7.51999999999989,
+      8.00999999999988,
+      8.249999999999876,
+      8.279999999999875,
+      9.629999999999846,
+      9.89999999999984,
+      10.589999999999826,
+      11.049999999999816,
+      11.7899999999998,
+      12.189999999999792,
+      12.889999999999777,
+      13.129999999999772,
+      13.16999999999977,
+      13.20999999999977,
+      13.479999999999764,
+      13.539999999999763,
+      13.779999999999758,
+      13.929999999999755,
+      14.429999999999744,
+      14.759999999999737,
+      15.149999999999729,
+      15.419999999999723,
+      15.53999999999972,
+      15.659999999999718,
+      15.749999999999716,
+      15.759999999999716,
+      15.799999999999715,
+      16.05999999999971,
+      16.079999999999714,
+      16.11999999999972,
+      16.11999999999972,
+      16.18999999999973,
+      16.31999999999975,
+      16.539999999999786,
+      16.799999999999827
+    ],
+    "type": "su"
+  },
+  "rope_theta": 10000.0,
+  "sliding_window": 131072,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.4",
+  "use_cache": false,
+  "vocab_size": 32064
+}

configuration_phi3.py

@@ -0,0 +1,213 @@
+# coding=utf-8
+# Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+""" Phi-3 model configuration"""
+
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+PHI3_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "microsoft/Phi-3-mini-4k-instruct": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/resolve/main/config.json",
+    "microsoft/Phi-3-mini-128k-instruct": "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/resolve/main/config.json",
+}
+
+
+class Phi3Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Phi3Model`]. It is used to instantiate a Phi-3
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the
+    [microsoft/Phi-3-mini-4k-instruct](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct).
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 32064):
+            Vocabulary size of the Phi-3 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Phi3Model`].
+        hidden_size (`int`, *optional*, defaults to 3072):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 8192):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        num_key_value_heads (`int`, *optional*):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        resid_pdrop (`float`, *optional*, defaults to 0.0):
+            Dropout probability for mlp outputs.
+        embd_pdrop (`int`, *optional*, defaults to 0.0):
+            The dropout ratio for the embeddings.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio after computing the attention scores.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model might ever be used with.
+        original_max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model was trained with. This is used to determine the size of the
+            original RoPE embeddings when using long scaling.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon value used for the RMSNorm.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`. Whether to tie weight embeddings or not.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`dict`, *optional*):
+            The scaling strategy for the RoPE embeddings. If `None`, no scaling is applied. If a dictionary, it must
+            contain the following keys: `type`, `short_factor` and `long_factor`. The `type` must be either `su` or `yarn` and
+            the `short_factor` and `long_factor` must be lists of numbers with the same length as the hidden size
+            divided by the number of attention heads divided by 2.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            The id of the "beginning-of-sequence" token.
+        eos_token_id (`int`, *optional*, defaults to 32000):
+            The id of the "end-of-sequence" token.
+        pad_token_id (`int`, *optional*, defaults to 32000):
+            The id of the padding token.
+        sliding_window (`int`, *optional*):
+            Sliding window attention window size. If `None`, no sliding window is applied.
+
+    Example:
+
+    ```python
+    >>> from transformers import Phi3Model, Phi3Config
+
+    >>> # Initializing a Phi-3 style configuration
+    >>> configuration = Phi3Config.from_pretrained("microsoft/Phi-3-mini-4k-instruct")
+
+    >>> # Initializing a model from the configuration
+    >>> model = Phi3Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "phi3"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=32064,
+        hidden_size=3072,
+        intermediate_size=8192,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        resid_pdrop=0.0,
+        embd_pdrop=0.0,
+        attention_dropout=0.0,
+        hidden_act="silu",
+        max_position_embeddings=4096,
+        original_max_position_embeddings=4096,
+        initializer_range=0.02,
+        rms_norm_eps=1e-5,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        bos_token_id=1,
+        eos_token_id=32000,
+        pad_token_id=32000,
+        sliding_window=None,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.resid_pdrop = resid_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.attention_dropout = attention_dropout
+        self.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.original_max_position_embeddings = original_max_position_embeddings
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self._rope_scaling_validation()
+        self.sliding_window = sliding_window
+
+        super().__init__(
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            pad_token_id=pad_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+    def _rope_scaling_validation(self):
+        """
+        Validate the `rope_scaling` configuration.
+        """
+        if self.rope_scaling is None:
+            return
+
+        if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 3:
+            raise ValueError(
+                "`rope_scaling` must be a dictionary with three fields, `type`, `short_factor` and `long_factor`, "
+                f"got {self.rope_scaling}"
+            )
+        rope_scaling_type = self.rope_scaling.get("type", None)
+        rope_scaling_short_factor = self.rope_scaling.get("short_factor", None)
+        rope_scaling_long_factor = self.rope_scaling.get("long_factor", None)
+        if rope_scaling_type is None or rope_scaling_type not in ["su", "yarn"]:
+            raise ValueError(f"`rope_scaling`'s type field must be one of ['su', 'yarn'], got {rope_scaling_type}")
+        if not (
+            isinstance(rope_scaling_short_factor, list)
+            and all(isinstance(x, (int, float)) for x in rope_scaling_short_factor)
+        ):
+            raise ValueError(
+                f"`rope_scaling`'s short_factor field must be a list of numbers, got {rope_scaling_short_factor}"
+            )
+        if not len(rope_scaling_short_factor) == self.hidden_size // self.num_attention_heads // 2:
+            raise ValueError(
+                f"`rope_scaling`'s short_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_short_factor)}"
+            )
+        if not (
+            isinstance(rope_scaling_long_factor, list)
+            and all(isinstance(x, (int, float)) for x in rope_scaling_long_factor)
+        ):
+            raise ValueError(
+                f"`rope_scaling`'s long_factor field must be a list of numbers, got {rope_scaling_long_factor}"
+            )
+        if not len(rope_scaling_long_factor) == self.hidden_size // self.num_attention_heads // 2:
+            raise ValueError(
+                f"`rope_scaling`'s long_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_long_factor)}"
+            )

generation_config.json

@@ -0,0 +1,11 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": [
+    32000,
+    32001,
+    32007
+  ],
+  "pad_token_id": 32000,
+  "transformers_version": "4.52.4"
+}

latest

@@ -0,0 +1 @@
+global_step2996

model-00001-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:942344d8b3f73f82842cb76b7adc55649b3ac83edf25aa326d2769a541f05ab4
+size 4915982920

model-00002-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebd532cb8cc851af6a32373b6c4dc5304f909f738fcc90da43d510b9df990f35
+size 4954690720

model-00003-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18fe071183549f37d0efca4e39244a3e9e338086d9b47137e9c45b60f26018ac
+size 4902241352

model-00004-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:434d068c36c331f118756ba75d67a286da89c9e75febca3c778548b54f10e8a7
+size 4771169120

model-00005-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:613781e50192d7f185b42efacb456eca8ad00ac40857d42dc78f5516f5696750
+size 4771169120

model-00006-of-00006.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b422df5d8ed4b32ccf08cbd7c1334b23f9bd5804272bef85fa2aa6097fb74222
+size 3605251496

model.safetensors.index.json

@@ -0,0 +1,250 @@
+{
+  "metadata": {
+    "total_size": 27920476160
+  },
+  "weight_map": {
+    "lm_head.weight": "model-00006-of-00006.safetensors",
+    "model.embed_tokens.weight": "model-00001-of-00006.safetensors",
+    "model.layers.0.input_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.0.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.0.mlp.gate_up_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.0.self_attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.1.input_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.1.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.1.mlp.gate_up_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.1.self_attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.10.input_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.10.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.10.mlp.gate_up_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.10.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.10.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.10.self_attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.11.input_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.11.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.11.mlp.gate_up_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.11.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.11.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.11.self_attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.12.input_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.12.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.12.mlp.gate_up_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.12.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.12.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.12.self_attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.13.input_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.13.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.13.mlp.gate_up_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.13.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.13.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.13.self_attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.14.input_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.14.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.14.mlp.gate_up_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.14.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.14.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.14.self_attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.15.input_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.15.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.15.mlp.gate_up_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.15.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.15.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.15.self_attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.16.input_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.16.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.16.mlp.gate_up_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.16.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.16.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.16.self_attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.17.input_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.17.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.17.mlp.gate_up_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.17.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.17.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.17.self_attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.18.input_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.18.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.18.mlp.gate_up_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.18.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.18.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.18.self_attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.19.input_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.19.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.19.mlp.gate_up_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.19.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.19.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.19.self_attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.2.input_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.2.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.2.mlp.gate_up_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.2.self_attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.20.input_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.20.mlp.down_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.20.mlp.gate_up_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.20.post_attention_layernorm.weight": "model-00003-of-00006.safetensors",
+    "model.layers.20.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.20.self_attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.21.input_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.21.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.21.mlp.gate_up_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.21.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.21.self_attn.o_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.21.self_attn.qkv_proj.weight": "model-00003-of-00006.safetensors",
+    "model.layers.22.input_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.22.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.22.mlp.gate_up_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.22.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.22.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.22.self_attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.23.input_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.23.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.23.mlp.gate_up_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.23.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.23.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.23.self_attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.24.input_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.24.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.24.mlp.gate_up_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.24.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.24.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.24.self_attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.25.input_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.25.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.25.mlp.gate_up_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.25.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.25.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.25.self_attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.26.input_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.26.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.26.mlp.gate_up_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.26.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.26.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.26.self_attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.27.input_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.27.mlp.down_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.27.mlp.gate_up_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.27.post_attention_layernorm.weight": "model-00004-of-00006.safetensors",
+    "model.layers.27.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.27.self_attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.28.input_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.28.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.28.mlp.gate_up_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.28.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.28.self_attn.o_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.28.self_attn.qkv_proj.weight": "model-00004-of-00006.safetensors",
+    "model.layers.29.input_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.29.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.29.mlp.gate_up_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.29.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.29.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.29.self_attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.3.input_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.3.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.3.mlp.gate_up_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.3.self_attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.30.input_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.30.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.30.mlp.gate_up_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.30.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.30.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.30.self_attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.31.input_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.31.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.31.mlp.gate_up_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.31.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.31.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.31.self_attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.32.input_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.32.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.32.mlp.gate_up_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.32.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.32.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.32.self_attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.33.input_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.33.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.33.mlp.gate_up_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.33.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.33.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.33.self_attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.34.input_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.34.mlp.down_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.34.mlp.gate_up_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.34.post_attention_layernorm.weight": "model-00005-of-00006.safetensors",
+    "model.layers.34.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.34.self_attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.35.input_layernorm.weight": "model-00006-of-00006.safetensors",
+    "model.layers.35.mlp.down_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.35.mlp.gate_up_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.35.post_attention_layernorm.weight": "model-00006-of-00006.safetensors",
+    "model.layers.35.self_attn.o_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.35.self_attn.qkv_proj.weight": "model-00005-of-00006.safetensors",
+    "model.layers.36.input_layernorm.weight": "model-00006-of-00006.safetensors",
+    "model.layers.36.mlp.down_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.36.mlp.gate_up_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.36.post_attention_layernorm.weight": "model-00006-of-00006.safetensors",
+    "model.layers.36.self_attn.o_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.36.self_attn.qkv_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.37.input_layernorm.weight": "model-00006-of-00006.safetensors",
+    "model.layers.37.mlp.down_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.37.mlp.gate_up_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.37.post_attention_layernorm.weight": "model-00006-of-00006.safetensors",
+    "model.layers.37.self_attn.o_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.37.self_attn.qkv_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.38.input_layernorm.weight": "model-00006-of-00006.safetensors",
+    "model.layers.38.mlp.down_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.38.mlp.gate_up_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.38.post_attention_layernorm.weight": "model-00006-of-00006.safetensors",
+    "model.layers.38.self_attn.o_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.38.self_attn.qkv_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.39.input_layernorm.weight": "model-00006-of-00006.safetensors",
+    "model.layers.39.mlp.down_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.39.mlp.gate_up_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.39.post_attention_layernorm.weight": "model-00006-of-00006.safetensors",
+    "model.layers.39.self_attn.o_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.39.self_attn.qkv_proj.weight": "model-00006-of-00006.safetensors",
+    "model.layers.4.input_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.4.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.4.mlp.gate_up_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.4.self_attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.5.input_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.5.mlp.down_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.5.mlp.gate_up_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.5.post_attention_layernorm.weight": "model-00001-of-00006.safetensors",
+    "model.layers.5.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.5.self_attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.6.input_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.6.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.6.mlp.gate_up_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.6.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.6.self_attn.o_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.6.self_attn.qkv_proj.weight": "model-00001-of-00006.safetensors",
+    "model.layers.7.input_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.7.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.7.mlp.gate_up_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.7.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.7.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.7.self_attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.8.input_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.8.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.8.mlp.gate_up_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.8.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.8.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.8.self_attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.9.input_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.9.mlp.down_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.9.mlp.gate_up_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.9.post_attention_layernorm.weight": "model-00002-of-00006.safetensors",
+    "model.layers.9.self_attn.o_proj.weight": "model-00002-of-00006.safetensors",
+    "model.layers.9.self_attn.qkv_proj.weight": "model-00002-of-00006.safetensors",
+    "model.norm.weight": "model-00006-of-00006.safetensors"
+  }
+}

modeling_phi3.py

@@ -0,0 +1,1606 @@
+# coding=utf-8
+# Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+""" PyTorch Phi-3 model."""
+
+import inspect
+import math
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_phi3 import Phi3Config
+
+
+logger = logging.get_logger(__name__)
+
+# Transformers scans dependencies in the modeling file, causing issues on conditional loading. The regex only ignores try/catch blocks, but not if statements
+# if is_flash_attn_2_available():
+_flash_supports_window_size = False
+try:
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+    _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)
+except ImportError as error:
+    logger.warning(
+        f"`flash-attention` package not found, consider installing for better performance: {error}."
+    )
+    if not _flash_supports_window_size:
+        logger.warning(
+            "Current `flash-attenton` does not support `window_size`. Either upgrade or use `attn_implementation='eager'`."
+        )
+
+_CHECKPOINT_FOR_DOC = "microsoft/Phi-3-mini-4k-instruct"
+_CONFIG_FOR_DOC = "Phi3Config"
+
+PHI3_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "microsoft/Phi-3-mini-4k-instruct",
+    "microsoft/Phi-3-mini-128k-instruct",
+    # See all Phi-3 models at https://huggingface.co/models?filter=Phi-3
+]
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Phi3
+class Phi3RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Phi3RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+# Copied from transformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding with gemma->phi3, Gemma->Phi3
+class Phi3RotaryEmbedding(nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+        super().__init__()
+
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.base = base
+        self.register_buffer("inv_freq", None, persistent=False)
+
+    @torch.no_grad()
+    def forward(self, x, position_ids, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        if self.inv_freq is None:
+            self.inv_freq = 1.0 / (
+                self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim)
+            )
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 since bfloat16 loses precision on long contexts
+        # See https://github.com/huggingface/transformers/pull/29285
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+class Phi3SuScaledRotaryEmbedding(Phi3RotaryEmbedding):
+    def __init__(self, dim, config, device=None):
+        super().__init__(dim, config.max_position_embeddings, config.rope_theta, device)
+
+        self.short_factor = config.rope_scaling["short_factor"]
+        self.long_factor = config.rope_scaling["long_factor"]
+        self.original_max_position_embeddings = config.original_max_position_embeddings
+
+    @torch.no_grad()
+    def forward(self, x, position_ids, seq_len=None):
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.original_max_position_embeddings:
+            ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device)
+        else:
+            ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device)
+
+        inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim
+        self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape)
+
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        # Force float32 since bfloat16 loses precision on long contexts
+        # See https://github.com/huggingface/transformers/pull/29285
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+
+            scale = self.max_position_embeddings / self.original_max_position_embeddings
+            if scale <= 1.0:
+                scaling_factor = 1.0
+            else:
+                scaling_factor = math.sqrt(1 + math.log(scale) / math.log(self.original_max_position_embeddings))
+
+            cos = emb.cos() * scaling_factor
+            sin = emb.sin() * scaling_factor
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+class Phi3YarnScaledRotaryEmbedding(Phi3RotaryEmbedding):
+    def __init__(self, dim, config, device=None):
+        super().__init__(dim, config.max_position_embeddings, config.rope_theta, device)
+
+        self.short_factor = config.rope_scaling["short_factor"]
+        self.long_factor = config.rope_scaling["long_factor"]
+        self.original_max_position_embeddings = config.original_max_position_embeddings
+
+    @torch.no_grad()
+    def forward(self, x, position_ids, seq_len=None):
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.original_max_position_embeddings:
+            ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device)
+        else:
+            ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device)
+
+        inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim
+        self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape)
+
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        # Force float32 since bfloat16 loses precision on long contexts
+        # See https://github.com/huggingface/transformers/pull/29285
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+
+            scale = self.max_position_embeddings / self.original_max_position_embeddings
+            if scale <= 1.0:
+                scaling_factor = 1.0
+            else:
+                scaling_factor = 0.1 * math.log(scale) + 1.0
+
+            cos = emb.cos() * scaling_factor
+            sin = emb.sin() * scaling_factor
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class Phi3MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+
+        self.config = config
+        self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
+
+        self.activation_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        up_states = self.gate_up_proj(hidden_states)
+
+        gate, up_states = up_states.chunk(2, dim=-1)
+        up_states = up_states * self.activation_fn(gate)
+
+        return self.down_proj(up_states)
+
+
+# Copied from transformers.models.llama.modeling_llama.repeat_kv with llama->phi
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+class Phi3Attention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: Phi3Config, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
+                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.attention_dropout = config.attention_dropout
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.original_max_position_embeddings = config.original_max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.rope_scaling = config.rope_scaling
+        self.is_causal = True
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+
+        op_size = self.num_heads * self.head_dim + 2 * (self.num_key_value_heads * self.head_dim)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+        self.qkv_proj = nn.Linear(self.hidden_size, op_size, bias=False)
+        self._init_rope()
+
+    def _init_rope(self):
+        if self.rope_scaling is None:
+            self.rotary_emb = Phi3RotaryEmbedding(
+                self.head_dim,
+                max_position_embeddings=self.max_position_embeddings,
+                base=self.rope_theta,
+            )
+        else:
+            scaling_type = self.config.rope_scaling["type"]
+            if scaling_type == "su":
+                self.rotary_emb = Phi3SuScaledRotaryEmbedding(self.head_dim, self.config)
+            elif scaling_type == "yarn":
+                self.rotary_emb = Phi3YarnScaledRotaryEmbedding(self.head_dim, self.config)
+            else:
+                raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        logger.warning_once("You are not running the flash-attention implementation, expect numerical differences.")
+
+        bsz, q_len, _ = hidden_states.size()
+
+        qkv = self.qkv_proj(hidden_states)
+        query_pos = self.num_heads * self.head_dim
+        query_states = qkv[..., :query_pos]
+        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
+        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
+
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        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 Phi3FlashAttention2(Phi3Attention):
+    """
+    Phi-3 flash attention module. This module inherits from `Phi3Attention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        # Phi3FlashAttention2 attention does not support output_attentions
+
+        if not _flash_supports_window_size:
+            logger.warning_once(
+                "The current flash attention version does not support sliding window attention. Please use `attn_implementation='eager'` or upgrade flash-attn library."
+            )
+            raise ValueError("The current flash attention version does not support sliding window attention.")
+
+        output_attentions = False
+
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+
+            # overwrite attention_mask with padding_mask
+            attention_mask = kwargs.pop("padding_mask")
+
+        bsz, q_len, _ = hidden_states.size()
+
+        qkv = self.qkv_proj(hidden_states)
+        query_pos = self.num_heads * self.head_dim
+        query_states = qkv[..., :query_pos]
+        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
+        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+
+        # Because the input can be padded, the absolute sequence length depends on the max position id.
+        rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
+        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=rotary_seq_len)
+
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        use_sliding_windows = (
+            _flash_supports_window_size
+            and getattr(self.config, "sliding_window", None) is not None
+            and kv_seq_len > self.config.sliding_window
+        )
+
+        if past_key_value is not None:
+            # Activate slicing cache only if the config has a value `sliding_windows` attribute
+            cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
+            if (
+                getattr(self.config, "sliding_window", None) is not None
+                and kv_seq_len > self.config.sliding_window
+                and cache_has_contents
+            ):
+                slicing_tokens = 1 - self.config.sliding_window
+
+                past_key = past_key_value[self.layer_idx][0]
+                past_value = past_key_value[self.layer_idx][1]
+
+                past_key = past_key[:, :, slicing_tokens:, :].contiguous()
+                past_value = past_value[:, :, slicing_tokens:, :].contiguous()
+
+                if past_key.shape[-2] != self.config.sliding_window - 1:
+                    raise ValueError(
+                        f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
+                        f" {past_key.shape}"
+                    )
+
+                if attention_mask is not None:
+                    attention_mask = attention_mask[:, slicing_tokens:]
+                    attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
+
+            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_dropout = self.attention_dropout if self.training else 0.0
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
+        # in fp32.
+
+        if query_states.dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.qkv_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        # Reashape to the expected shape for Flash Attention
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        attn_output = self._flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            dropout=attn_dropout,
+            use_sliding_windows=use_sliding_windows,
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._flash_attention_forward
+    def _flash_attention_forward(
+        self,
+        query_states,
+        key_states,
+        value_states,
+        attention_mask,
+        query_length,
+        dropout=0.0,
+        softmax_scale=None,
+        use_sliding_windows=False,
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`float`):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+            use_sliding_windows (`bool`, *optional*):
+                Whether to activate sliding window attention.
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            if not use_sliding_windows:
+                attn_output_unpad = flash_attn_varlen_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    cu_seqlens_q=cu_seqlens_q,
+                    cu_seqlens_k=cu_seqlens_k,
+                    max_seqlen_q=max_seqlen_in_batch_q,
+                    max_seqlen_k=max_seqlen_in_batch_k,
+                    dropout_p=dropout,
+                    softmax_scale=softmax_scale,
+                    causal=causal,
+                )
+            else:
+                attn_output_unpad = flash_attn_varlen_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    cu_seqlens_q=cu_seqlens_q,
+                    cu_seqlens_k=cu_seqlens_k,
+                    max_seqlen_q=max_seqlen_in_batch_q,
+                    max_seqlen_k=max_seqlen_in_batch_k,
+                    dropout_p=dropout,
+                    softmax_scale=softmax_scale,
+                    causal=causal,
+                    window_size=(self.config.sliding_window, self.config.sliding_window),
+                )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            if not use_sliding_windows:
+                attn_output = flash_attn_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    dropout,
+                    softmax_scale=softmax_scale,
+                    causal=causal,
+                )
+            else:
+                attn_output = flash_attn_func(
+                    query_states,
+                    key_states,
+                    value_states,
+                    dropout,
+                    softmax_scale=softmax_scale,
+                    causal=causal,
+                    window_size=(self.config.sliding_window, self.config.sliding_window),
+                )
+
+        return attn_output
+
+    # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._upad_input
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
+
+        # On the first iteration we need to properly re-create the padding mask
+        # by slicing it on the proper place
+        if kv_seq_len != attention_mask.shape[-1]:
+            attention_mask_num_tokens = attention_mask.shape[-1]
+            attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]
+
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+
+        key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
+        value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
+
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+# copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Phi3
+# TODO @Arthur no longer copied from LLama after static cache
+class Phi3SdpaAttention(Phi3Attention):
+    """
+    Phi3 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `Phi3Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+
+    # Adapted from Phi3Attention.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "Phi3Model is using Phi3SdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+            )
+
+        bsz, q_len, _ = hidden_states.size()
+
+        qkv = self.qkv_proj(hidden_states)
+        query_pos = self.num_heads * self.head_dim
+        query_states = qkv[..., :query_pos]
+        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
+        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_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.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
+
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == "cuda" and attention_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
+            is_causal=self.is_causal and attention_mask is None and q_len > 1,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+PHI3_ATTENTION_CLASSES = {
+    "eager": Phi3Attention,
+    "flash_attention_2": Phi3FlashAttention2,
+    "sdpa": Phi3SdpaAttention,
+}
+
+
+class Phi3DecoderLayer(nn.Module):
+    def __init__(self, config: Phi3Config, layer_idx: int):
+        super().__init__()
+
+        self.config = config
+        self.self_attn = PHI3_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx)
+
+        self.mlp = Phi3MLP(config)
+        self.input_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.resid_attn_dropout = nn.Dropout(config.resid_pdrop)
+        self.resid_mlp_dropout = nn.Dropout(config.resid_pdrop)
+        self.post_attention_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+                Indices of positions of each input sequence tokens in the position embeddings. Selected in the range
+                `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        """
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        attn_outputs, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+        )
+
+        hidden_states = residual + self.resid_attn_dropout(attn_outputs)
+
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + self.resid_mlp_dropout(hidden_states)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+PHI3_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`Phi3Config`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare Phi-3 model outputting raw hidden-states without any specific head on top.",
+    PHI3_START_DOCSTRING,
+)
+class Phi3PreTrainedModel(PreTrainedModel):
+    config_class = Phi3Config
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Phi3DecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = False
+    _supports_cache_class = True
+
+    _version = "0.0.5"
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, 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_()
+
+
+PHI3_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance;
+            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+            cache format.
+
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare Phi-3 model outputting raw hidden-states without any specific head on top.",
+    PHI3_START_DOCSTRING,
+)
+class Phi3Model(Phi3PreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Phi3DecoderLayer`]
+
+    Args:
+        config: Phi3Config
+    """
+
+    def __init__(self, config: Phi3Config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.embed_dropout = nn.Dropout(config.embd_pdrop)
+        self.layers = nn.ModuleList(
+            [Phi3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self._attn_implementation = config._attn_implementation
+        self.norm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape[:2]
+        elif inputs_embeds is not None:
+            batch_size, seq_length = inputs_embeds.shape[:2]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        past_key_values_length = 0
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        if use_cache:
+            use_legacy_cache = not isinstance(past_key_values, Cache)
+            if use_legacy_cache:
+                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+            past_key_values_length = past_key_values.get_usable_length(seq_length)
+
+        if position_ids is None:
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(
+                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+        else:
+            position_ids = position_ids.view(-1, seq_length).long()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
+            is_padding_right = attention_mask[:, -1].sum().item() != batch_size
+            if is_padding_right:
+                raise ValueError(
+                    "You are attempting to perform batched generation with padding_side='right'"
+                    " this may lead to unexpected behaviour for Flash Attention version of Phi3. Make sure to "
+                    " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
+                )
+
+        if self._attn_implementation == "flash_attention_2":
+            # 2d mask is passed through the layers
+            attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+        else:
+            # 4d mask is passed through the layers
+            attention_mask = _prepare_4d_causal_attention_mask(
+                attention_mask,
+                (batch_size, seq_length),
+                inputs_embeds,
+                past_key_values_length,
+                sliding_window=self.config.sliding_window,
+            )
+
+        hidden_states = inputs_embeds
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    attention_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    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 = None
+        if use_cache:
+            next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
+        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 Phi3ForCausalLM(Phi3PreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.__init__ with Llama->Phi3
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Phi3Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_input_embeddings
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_input_embeddings
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_output_embeddings
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_output_embeddings
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_decoder
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_decoder
+    def get_decoder(self):
+        return self.model
+
+    # Ignore copy
+    @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = 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] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, Phi3ForCausalLM
+
+        >>> model = Phi3ForCausalLM.from_pretrained("microsoft/phi-3-mini-4k-instruct")
+        >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-3-mini-4k-instruct")
+
+        >>> prompt = "This is an example script ."
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum'
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        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,
+            position_ids=position_ids,
+            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)
+        logits = logits.float()
+
+        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,
+        )
+
+    # Copied from transformers.models.persimmon.modeling_persimmon.PersimmonForCausalLM.prepare_inputs_for_generation
+    def prepare_inputs_for_generation(
+        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        if past_key_values is not None:
+            if isinstance(past_key_values, Cache):
+                cache_length = past_key_values.get_seq_length()
+                past_length = past_key_values.seen_tokens
+                max_cache_length = past_key_values.get_max_length()
+            else:
+                cache_length = past_length = past_key_values[0][0].shape[2]
+                max_cache_length = None
+
+            # Keep only the unprocessed tokens:
+            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
+            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
+            # input)
+            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
+                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
+            # input_ids based on the past_length.
+            elif past_length < input_ids.shape[1]:
+                input_ids = input_ids[:, past_length:]
+            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
+
+            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
+            if (
+                max_cache_length is not None
+                and attention_mask is not None
+                and cache_length + input_ids.shape[1] > max_cache_length
+            ):
+                attention_mask = attention_mask[:, -max_cache_length:]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[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(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM._reorder_cache
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
+            )
+        return reordered_past
+
+
+@add_start_docstrings(
+    """
+    The [`Phi3Model`] with a sequence classification head on top (linear layer).
+
+    [`Phi3ForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+    (e.g. GPT-2) do.
+
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
+    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
+    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
+    each row of the batch).
+    """,
+    PHI3_START_DOCSTRING,
+)
+# Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->Phi3, LLAMA->PHI3, self.transformer->self.model, transformer_outputs->model_outputs
+class Phi3ForSequenceClassification(Phi3PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = Phi3Model(config)
+        self.score = nn.Linear(config.hidden_size, self.num_labels, 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
+
+    @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = 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] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        model_outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            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 = model_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size = input_ids.shape[0]
+        else:
+            batch_size = inputs_embeds.shape[0]
+
+        if self.config.pad_token_id is None and batch_size != 1:
+            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
+            else:
+                sequence_lengths = -1
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(pooled_logits, labels)
+        if not return_dict:
+            output = (pooled_logits,) + model_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=model_outputs.past_key_values,
+            hidden_states=model_outputs.hidden_states,
+            attentions=model_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    [`Phi3Model`] with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
+    Named-Entity-Recognition (NER) tasks.
+    """,
+    PHI3_START_DOCSTRING,
+)
+# Copied from transformers.models.mpt.modeling_mpt.MptForTokenClassification with Mpt->Phi3,MPT->PHI3,self.transformer->self.model,transformer_outputs->model_outputs
+class Phi3ForTokenClassification(Phi3PreTrainedModel):
+    def __init__(self, config: Phi3Config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.model = Phi3Model(config)
+        if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None:
+            classifier_dropout = config.classifier_dropout
+        elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
+            classifier_dropout = config.hidden_dropout
+        else:
+            classifier_dropout = 0.1
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **deprecated_arguments,
+    ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        model_outputs = self.model(
+            input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = model_outputs[0]
+        hidden_states = self.dropout(hidden_states)
+        logits = self.classifier(hidden_states)
+
+        loss = None
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(logits.device)
+            batch_size, seq_length = labels.shape
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(
+                logits.view(batch_size * seq_length, self.num_labels), labels.view(batch_size * seq_length)
+            )
+
+        if not return_dict:
+            output = (logits,) + model_outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=model_outputs.hidden_states,
+            attentions=model_outputs.attentions,
+        )

special_tokens_map.json

@@ -0,0 +1,30 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
+size 499723

tokenizer_config.json

@@ -0,0 +1,132 @@
+{
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "add_prefix_space": null,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": false
+    },
+    "32000": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32001": {
+      "content": "<|assistant|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    },
+    "32002": {
+      "content": "<|placeholder1|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    },
+    "32003": {
+      "content": "<|placeholder2|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    },
+    "32004": {
+      "content": "<|placeholder3|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    },
+    "32005": {
+      "content": "<|placeholder4|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    },
+    "32006": {
+      "content": "<|system|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    },
+    "32007": {
+      "content": "<|end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32008": {
+      "content": "<|placeholder5|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    },
+    "32009": {
+      "content": "<|placeholder6|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    },
+    "32010": {
+      "content": "<|user|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|end|>",
+  "extra_special_tokens": {},
+  "legacy": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "sp_model_kwargs": {},
+  "split_special_tokens": false,
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}

trainer_state.json

@@ -0,0 +1,2134 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 4.298368298368298,
+  "eval_steps": 500,
+  "global_step": 3000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.014344629729245113,
+      "grad_norm": 2.1563808263401123,
+      "learning_rate": 2.1479713603818614e-06,
+      "loss": 0.9832,
+      "step": 10
+    },
+    {
+      "epoch": 0.028689259458490227,
+      "grad_norm": 1.5376642197375787,
+      "learning_rate": 4.5346062052505965e-06,
+      "loss": 0.8531,
+      "step": 20
+    },
+    {
+      "epoch": 0.04303388918773534,
+      "grad_norm": 0.7108211564362369,
+      "learning_rate": 6.921241050119331e-06,
+      "loss": 0.7191,
+      "step": 30
+    },
+    {
+      "epoch": 0.05737851891698045,
+      "grad_norm": 0.5064006283904037,
+      "learning_rate": 9.307875894988068e-06,
+      "loss": 0.6532,
+      "step": 40
+    },
+    {
+      "epoch": 0.07172314864622557,
+      "grad_norm": 0.5016524926299113,
+      "learning_rate": 1.1694510739856803e-05,
+      "loss": 0.6245,
+      "step": 50
+    },
+    {
+      "epoch": 0.08606777837547068,
+      "grad_norm": 0.5064320977744877,
+      "learning_rate": 1.4081145584725539e-05,
+      "loss": 0.6017,
+      "step": 60
+    },
+    {
+      "epoch": 0.1004124081047158,
+      "grad_norm": 0.4525315758811991,
+      "learning_rate": 1.6467780429594274e-05,
+      "loss": 0.5788,
+      "step": 70
+    },
+    {
+      "epoch": 0.1147570378339609,
+      "grad_norm": 0.4891140282424216,
+      "learning_rate": 1.885441527446301e-05,
+      "loss": 0.5726,
+      "step": 80
+    },
+    {
+      "epoch": 0.129101667563206,
+      "grad_norm": 0.5816465774117703,
+      "learning_rate": 2.1241050119331742e-05,
+      "loss": 0.5669,
+      "step": 90
+    },
+    {
+      "epoch": 0.14344629729245115,
+      "grad_norm": 0.5416459422496235,
+      "learning_rate": 2.3627684964200477e-05,
+      "loss": 0.5621,
+      "step": 100
+    },
+    {
+      "epoch": 0.15779092702169625,
+      "grad_norm": 0.5501550374753971,
+      "learning_rate": 2.6014319809069216e-05,
+      "loss": 0.5543,
+      "step": 110
+    },
+    {
+      "epoch": 0.17213555675094136,
+      "grad_norm": 0.48897325487812016,
+      "learning_rate": 2.840095465393795e-05,
+      "loss": 0.5519,
+      "step": 120
+    },
+    {
+      "epoch": 0.1864801864801865,
+      "grad_norm": 0.5193769497689257,
+      "learning_rate": 3.0787589498806684e-05,
+      "loss": 0.546,
+      "step": 130
+    },
+    {
+      "epoch": 0.2008248162094316,
+      "grad_norm": 0.5100064220076413,
+      "learning_rate": 3.3174224343675416e-05,
+      "loss": 0.5454,
+      "step": 140
+    },
+    {
+      "epoch": 0.2151694459386767,
+      "grad_norm": 0.4992592686432826,
+      "learning_rate": 3.5560859188544155e-05,
+      "loss": 0.5437,
+      "step": 150
+    },
+    {
+      "epoch": 0.2295140756679218,
+      "grad_norm": 0.4838403686730783,
+      "learning_rate": 3.794749403341289e-05,
+      "loss": 0.5412,
+      "step": 160
+    },
+    {
+      "epoch": 0.24385870539716695,
+      "grad_norm": 0.45590237103117387,
+      "learning_rate": 4.0334128878281626e-05,
+      "loss": 0.541,
+      "step": 170
+    },
+    {
+      "epoch": 0.258203335126412,
+      "grad_norm": 0.44189108119445925,
+      "learning_rate": 4.272076372315036e-05,
+      "loss": 0.5328,
+      "step": 180
+    },
+    {
+      "epoch": 0.2725479648556572,
+      "grad_norm": 0.4596950199469469,
+      "learning_rate": 4.510739856801909e-05,
+      "loss": 0.5328,
+      "step": 190
+    },
+    {
+      "epoch": 0.2868925945849023,
+      "grad_norm": 0.4232582169854354,
+      "learning_rate": 4.749403341288783e-05,
+      "loss": 0.5331,
+      "step": 200
+    },
+    {
+      "epoch": 0.3012372243141474,
+      "grad_norm": 0.47183677702470816,
+      "learning_rate": 4.988066825775656e-05,
+      "loss": 0.5405,
+      "step": 210
+    },
+    {
+      "epoch": 0.3155818540433925,
+      "grad_norm": 0.510543623904468,
+      "learning_rate": 5.22673031026253e-05,
+      "loss": 0.5448,
+      "step": 220
+    },
+    {
+      "epoch": 0.3299264837726376,
+      "grad_norm": 0.49518213450267645,
+      "learning_rate": 5.465393794749404e-05,
+      "loss": 0.5441,
+      "step": 230
+    },
+    {
+      "epoch": 0.3442711135018827,
+      "grad_norm": 0.4081838992878798,
+      "learning_rate": 5.7040572792362765e-05,
+      "loss": 0.5342,
+      "step": 240
+    },
+    {
+      "epoch": 0.3586157432311278,
+      "grad_norm": 0.45201996051478616,
+      "learning_rate": 5.942720763723151e-05,
+      "loss": 0.5379,
+      "step": 250
+    },
+    {
+      "epoch": 0.372960372960373,
+      "grad_norm": 0.5163583891265898,
+      "learning_rate": 6.181384248210024e-05,
+      "loss": 0.5352,
+      "step": 260
+    },
+    {
+      "epoch": 0.3873050026896181,
+      "grad_norm": 0.43220360318869133,
+      "learning_rate": 6.420047732696898e-05,
+      "loss": 0.5319,
+      "step": 270
+    },
+    {
+      "epoch": 0.4016496324188632,
+      "grad_norm": 0.39341356230222857,
+      "learning_rate": 6.65871121718377e-05,
+      "loss": 0.531,
+      "step": 280
+    },
+    {
+      "epoch": 0.4159942621481083,
+      "grad_norm": 0.4843450935832954,
+      "learning_rate": 6.897374701670645e-05,
+      "loss": 0.5247,
+      "step": 290
+    },
+    {
+      "epoch": 0.4303388918773534,
+      "grad_norm": 0.43790304963108484,
+      "learning_rate": 7.136038186157519e-05,
+      "loss": 0.535,
+      "step": 300
+    },
+    {
+      "epoch": 0.4446835216065985,
+      "grad_norm": 0.3679883238347063,
+      "learning_rate": 7.374701670644391e-05,
+      "loss": 0.529,
+      "step": 310
+    },
+    {
+      "epoch": 0.4590281513358436,
+      "grad_norm": 0.38496528216858267,
+      "learning_rate": 7.613365155131266e-05,
+      "loss": 0.5358,
+      "step": 320
+    },
+    {
+      "epoch": 0.47337278106508873,
+      "grad_norm": 0.3554015988338894,
+      "learning_rate": 7.852028639618139e-05,
+      "loss": 0.5366,
+      "step": 330
+    },
+    {
+      "epoch": 0.4877174107943339,
+      "grad_norm": 0.3647316947863913,
+      "learning_rate": 8.090692124105012e-05,
+      "loss": 0.5305,
+      "step": 340
+    },
+    {
+      "epoch": 0.502062040523579,
+      "grad_norm": 0.40796953260473273,
+      "learning_rate": 8.329355608591885e-05,
+      "loss": 0.5361,
+      "step": 350
+    },
+    {
+      "epoch": 0.516406670252824,
+      "grad_norm": 0.345466577791704,
+      "learning_rate": 8.56801909307876e-05,
+      "loss": 0.5317,
+      "step": 360
+    },
+    {
+      "epoch": 0.5307512999820692,
+      "grad_norm": 0.3897824320306703,
+      "learning_rate": 8.806682577565633e-05,
+      "loss": 0.5384,
+      "step": 370
+    },
+    {
+      "epoch": 0.5450959297113144,
+      "grad_norm": 0.39750641265021325,
+      "learning_rate": 9.045346062052506e-05,
+      "loss": 0.5344,
+      "step": 380
+    },
+    {
+      "epoch": 0.5594405594405595,
+      "grad_norm": 0.3294501292354428,
+      "learning_rate": 9.28400954653938e-05,
+      "loss": 0.531,
+      "step": 390
+    },
+    {
+      "epoch": 0.5737851891698046,
+      "grad_norm": 0.3325704984940842,
+      "learning_rate": 9.522673031026254e-05,
+      "loss": 0.5282,
+      "step": 400
+    },
+    {
+      "epoch": 0.5881298188990497,
+      "grad_norm": 0.37577078172283135,
+      "learning_rate": 9.761336515513126e-05,
+      "loss": 0.5289,
+      "step": 410
+    },
+    {
+      "epoch": 0.6024744486282948,
+      "grad_norm": 0.32982285395946315,
+      "learning_rate": 0.0001,
+      "loss": 0.5326,
+      "step": 420
+    },
+    {
+      "epoch": 0.6168190783575399,
+      "grad_norm": 0.315570817514683,
+      "learning_rate": 9.999826305940802e-05,
+      "loss": 0.5276,
+      "step": 430
+    },
+    {
+      "epoch": 0.631163708086785,
+      "grad_norm": 0.32873479872603517,
+      "learning_rate": 9.99930523583106e-05,
+      "loss": 0.5329,
+      "step": 440
+    },
+    {
+      "epoch": 0.6455083378160301,
+      "grad_norm": 0.3754610016068392,
+      "learning_rate": 9.998436825873485e-05,
+      "loss": 0.5339,
+      "step": 450
+    },
+    {
+      "epoch": 0.6598529675452752,
+      "grad_norm": 0.2973537072730961,
+      "learning_rate": 9.997221136403139e-05,
+      "loss": 0.5249,
+      "step": 460
+    },
+    {
+      "epoch": 0.6741975972745203,
+      "grad_norm": 0.33421087994681,
+      "learning_rate": 9.995658251883237e-05,
+      "loss": 0.5196,
+      "step": 470
+    },
+    {
+      "epoch": 0.6885422270037654,
+      "grad_norm": 0.3120052809857615,
+      "learning_rate": 9.993748280899279e-05,
+      "loss": 0.5236,
+      "step": 480
+    },
+    {
+      "epoch": 0.7028868567330105,
+      "grad_norm": 0.28477505127058517,
+      "learning_rate": 9.991491356151515e-05,
+      "loss": 0.5166,
+      "step": 490
+    },
+    {
+      "epoch": 0.7172314864622557,
+      "grad_norm": 0.3282614273702267,
+      "learning_rate": 9.988887634445711e-05,
+      "loss": 0.5191,
+      "step": 500
+    },
+    {
+      "epoch": 0.7315761161915008,
+      "grad_norm": 0.3018575266984723,
+      "learning_rate": 9.985937296682264e-05,
+      "loss": 0.52,
+      "step": 510
+    },
+    {
+      "epoch": 0.745920745920746,
+      "grad_norm": 0.322642800391225,
+      "learning_rate": 9.982640547843628e-05,
+      "loss": 0.5193,
+      "step": 520
+    },
+    {
+      "epoch": 0.7602653756499911,
+      "grad_norm": 0.29334428049579575,
+      "learning_rate": 9.978997616980083e-05,
+      "loss": 0.5173,
+      "step": 530
+    },
+    {
+      "epoch": 0.7746100053792362,
+      "grad_norm": 0.30663608653187874,
+      "learning_rate": 9.975008757193805e-05,
+      "loss": 0.514,
+      "step": 540
+    },
+    {
+      "epoch": 0.7889546351084813,
+      "grad_norm": 0.2902575349921672,
+      "learning_rate": 9.970674245621296e-05,
+      "loss": 0.5173,
+      "step": 550
+    },
+    {
+      "epoch": 0.8032992648377264,
+      "grad_norm": 0.2750627268123107,
+      "learning_rate": 9.965994383414116e-05,
+      "loss": 0.5124,
+      "step": 560
+    },
+    {
+      "epoch": 0.8176438945669715,
+      "grad_norm": 0.31846444500772264,
+      "learning_rate": 9.960969495717975e-05,
+      "loss": 0.5105,
+      "step": 570
+    },
+    {
+      "epoch": 0.8319885242962166,
+      "grad_norm": 0.2850087845127407,
+      "learning_rate": 9.955599931650127e-05,
+      "loss": 0.505,
+      "step": 580
+    },
+    {
+      "epoch": 0.8463331540254617,
+      "grad_norm": 0.3011934822435399,
+      "learning_rate": 9.949886064275123e-05,
+      "loss": 0.4997,
+      "step": 590
+    },
+    {
+      "epoch": 0.8606777837547068,
+      "grad_norm": 0.28048546158040877,
+      "learning_rate": 9.943828290578892e-05,
+      "loss": 0.5039,
+      "step": 600
+    },
+    {
+      "epoch": 0.8750224134839519,
+      "grad_norm": 0.29482418586801606,
+      "learning_rate": 9.937427031441152e-05,
+      "loss": 0.5068,
+      "step": 610
+    },
+    {
+      "epoch": 0.889367043213197,
+      "grad_norm": 0.2868511054090736,
+      "learning_rate": 9.93068273160618e-05,
+      "loss": 0.5041,
+      "step": 620
+    },
+    {
+      "epoch": 0.9037116729424421,
+      "grad_norm": 0.26871712641833934,
+      "learning_rate": 9.9235958596519e-05,
+      "loss": 0.5031,
+      "step": 630
+    },
+    {
+      "epoch": 0.9180563026716873,
+      "grad_norm": 0.27176006324351715,
+      "learning_rate": 9.916166907957336e-05,
+      "loss": 0.4998,
+      "step": 640
+    },
+    {
+      "epoch": 0.9324009324009324,
+      "grad_norm": 0.2916123334399884,
+      "learning_rate": 9.908396392668397e-05,
+      "loss": 0.5045,
+      "step": 650
+    },
+    {
+      "epoch": 0.9467455621301775,
+      "grad_norm": 0.28657648009371867,
+      "learning_rate": 9.90028485366202e-05,
+      "loss": 0.5005,
+      "step": 660
+    },
+    {
+      "epoch": 0.9610901918594227,
+      "grad_norm": 0.2616587253845151,
+      "learning_rate": 9.891832854508661e-05,
+      "loss": 0.5017,
+      "step": 670
+    },
+    {
+      "epoch": 0.9754348215886678,
+      "grad_norm": 0.2642546302377749,
+      "learning_rate": 9.883040982433133e-05,
+      "loss": 0.492,
+      "step": 680
+    },
+    {
+      "epoch": 0.9897794513179129,
+      "grad_norm": 0.255936110537566,
+      "learning_rate": 9.87390984827382e-05,
+      "loss": 0.4934,
+      "step": 690
+    },
+    {
+      "epoch": 1.002868925945849,
+      "grad_norm": 0.3099905060480918,
+      "learning_rate": 9.864440086440223e-05,
+      "loss": 0.43,
+      "step": 700
+    },
+    {
+      "epoch": 1.017213555675094,
+      "grad_norm": 0.25984643430841164,
+      "learning_rate": 9.854632354868889e-05,
+      "loss": 0.3695,
+      "step": 710
+    },
+    {
+      "epoch": 1.0315581854043392,
+      "grad_norm": 0.28715953895875523,
+      "learning_rate": 9.844487334977705e-05,
+      "loss": 0.3792,
+      "step": 720
+    },
+    {
+      "epoch": 1.0459028151335843,
+      "grad_norm": 0.2649853690140653,
+      "learning_rate": 9.834005731618543e-05,
+      "loss": 0.3737,
+      "step": 730
+    },
+    {
+      "epoch": 1.0602474448628294,
+      "grad_norm": 0.26573177260306496,
+      "learning_rate": 9.823188273028297e-05,
+      "loss": 0.3771,
+      "step": 740
+    },
+    {
+      "epoch": 1.0745920745920745,
+      "grad_norm": 0.26574383245031147,
+      "learning_rate": 9.812035710778283e-05,
+      "loss": 0.3741,
+      "step": 750
+    },
+    {
+      "epoch": 1.0889367043213196,
+      "grad_norm": 0.28123208245100456,
+      "learning_rate": 9.800548819722026e-05,
+      "loss": 0.3731,
+      "step": 760
+    },
+    {
+      "epoch": 1.1032813340505647,
+      "grad_norm": 0.28349060376734364,
+      "learning_rate": 9.78872839794142e-05,
+      "loss": 0.3778,
+      "step": 770
+    },
+    {
+      "epoch": 1.11762596377981,
+      "grad_norm": 0.28893208285124705,
+      "learning_rate": 9.776575266691279e-05,
+      "loss": 0.3806,
+      "step": 780
+    },
+    {
+      "epoch": 1.1319705935090552,
+      "grad_norm": 0.2860386213893016,
+      "learning_rate": 9.764090270342286e-05,
+      "loss": 0.3799,
+      "step": 790
+    },
+    {
+      "epoch": 1.1463152232383003,
+      "grad_norm": 0.2460375162339276,
+      "learning_rate": 9.751274276322316e-05,
+      "loss": 0.3898,
+      "step": 800
+    },
+    {
+      "epoch": 1.1606598529675454,
+      "grad_norm": 0.251720247858439,
+      "learning_rate": 9.738128175056179e-05,
+      "loss": 0.3821,
+      "step": 810
+    },
+    {
+      "epoch": 1.1750044826967905,
+      "grad_norm": 0.25278081761302046,
+      "learning_rate": 9.724652879903751e-05,
+      "loss": 0.3798,
+      "step": 820
+    },
+    {
+      "epoch": 1.1893491124260356,
+      "grad_norm": 0.2599296471651565,
+      "learning_rate": 9.71084932709652e-05,
+      "loss": 0.3828,
+      "step": 830
+    },
+    {
+      "epoch": 1.2036937421552807,
+      "grad_norm": 0.2455565423628766,
+      "learning_rate": 9.696718475672532e-05,
+      "loss": 0.3743,
+      "step": 840
+    },
+    {
+      "epoch": 1.2180383718845258,
+      "grad_norm": 0.2888176539405626,
+      "learning_rate": 9.682261307409766e-05,
+      "loss": 0.381,
+      "step": 850
+    },
+    {
+      "epoch": 1.232383001613771,
+      "grad_norm": 0.25471629336116824,
+      "learning_rate": 9.667478826757916e-05,
+      "loss": 0.3832,
+      "step": 860
+    },
+    {
+      "epoch": 1.246727631343016,
+      "grad_norm": 0.2614351769400213,
+      "learning_rate": 9.652372060768608e-05,
+      "loss": 0.3848,
+      "step": 870
+    },
+    {
+      "epoch": 1.2610722610722611,
+      "grad_norm": 0.26412023986782596,
+      "learning_rate": 9.63694205902405e-05,
+      "loss": 0.3855,
+      "step": 880
+    },
+    {
+      "epoch": 1.2754168908015062,
+      "grad_norm": 0.2429603634039349,
+      "learning_rate": 9.621189893564092e-05,
+      "loss": 0.3819,
+      "step": 890
+    },
+    {
+      "epoch": 1.2897615205307513,
+      "grad_norm": 0.26829134565306034,
+      "learning_rate": 9.605116658811759e-05,
+      "loss": 0.3906,
+      "step": 900
+    },
+    {
+      "epoch": 1.3041061502599964,
+      "grad_norm": 0.27364921782590684,
+      "learning_rate": 9.588723471497208e-05,
+      "loss": 0.3848,
+      "step": 910
+    },
+    {
+      "epoch": 1.3184507799892415,
+      "grad_norm": 0.2515405826777948,
+      "learning_rate": 9.572011470580136e-05,
+      "loss": 0.3899,
+      "step": 920
+    },
+    {
+      "epoch": 1.3327954097184866,
+      "grad_norm": 0.2513235427794581,
+      "learning_rate": 9.554981817170655e-05,
+      "loss": 0.3912,
+      "step": 930
+    },
+    {
+      "epoch": 1.3471400394477318,
+      "grad_norm": 0.24955670786044107,
+      "learning_rate": 9.537635694448615e-05,
+      "loss": 0.3849,
+      "step": 940
+    },
+    {
+      "epoch": 1.3614846691769769,
+      "grad_norm": 0.24538389153554718,
+      "learning_rate": 9.519974307581404e-05,
+      "loss": 0.3867,
+      "step": 950
+    },
+    {
+      "epoch": 1.375829298906222,
+      "grad_norm": 0.25856641290644117,
+      "learning_rate": 9.50199888364021e-05,
+      "loss": 0.3899,
+      "step": 960
+    },
+    {
+      "epoch": 1.390173928635467,
+      "grad_norm": 0.26645029754281746,
+      "learning_rate": 9.483710671514777e-05,
+      "loss": 0.386,
+      "step": 970
+    },
+    {
+      "epoch": 1.4045185583647122,
+      "grad_norm": 0.25589965865278824,
+      "learning_rate": 9.465110941826622e-05,
+      "loss": 0.3856,
+      "step": 980
+    },
+    {
+      "epoch": 1.4188631880939573,
+      "grad_norm": 0.27815251555263987,
+      "learning_rate": 9.446200986840765e-05,
+      "loss": 0.3881,
+      "step": 990
+    },
+    {
+      "epoch": 1.4332078178232024,
+      "grad_norm": 0.26197920796578195,
+      "learning_rate": 9.426982120375943e-05,
+      "loss": 0.3878,
+      "step": 1000
+    },
+    {
+      "epoch": 1.4475524475524475,
+      "grad_norm": 0.2606524388643148,
+      "learning_rate": 9.407455677713328e-05,
+      "loss": 0.3883,
+      "step": 1010
+    },
+    {
+      "epoch": 1.4618970772816926,
+      "grad_norm": 0.2331343629588404,
+      "learning_rate": 9.387623015503753e-05,
+      "loss": 0.3848,
+      "step": 1020
+    },
+    {
+      "epoch": 1.4762417070109377,
+      "grad_norm": 0.25873984889230295,
+      "learning_rate": 9.367485511673462e-05,
+      "loss": 0.3895,
+      "step": 1030
+    },
+    {
+      "epoch": 1.4905863367401828,
+      "grad_norm": 0.2531065339256471,
+      "learning_rate": 9.347044565328367e-05,
+      "loss": 0.3937,
+      "step": 1040
+    },
+    {
+      "epoch": 1.504930966469428,
+      "grad_norm": 0.26767841128977615,
+      "learning_rate": 9.326301596656846e-05,
+      "loss": 0.3894,
+      "step": 1050
+    },
+    {
+      "epoch": 1.519275596198673,
+      "grad_norm": 0.2472066267101206,
+      "learning_rate": 9.30525804683107e-05,
+      "loss": 0.3889,
+      "step": 1060
+    },
+    {
+      "epoch": 1.5336202259279181,
+      "grad_norm": 0.24916114849718174,
+      "learning_rate": 9.283915377906875e-05,
+      "loss": 0.3874,
+      "step": 1070
+    },
+    {
+      "epoch": 1.5479648556571632,
+      "grad_norm": 0.24586564647507672,
+      "learning_rate": 9.262275072722181e-05,
+      "loss": 0.3899,
+      "step": 1080
+    },
+    {
+      "epoch": 1.5623094853864083,
+      "grad_norm": 0.24194199620674006,
+      "learning_rate": 9.240338634793969e-05,
+      "loss": 0.3867,
+      "step": 1090
+    },
+    {
+      "epoch": 1.5766541151156535,
+      "grad_norm": 0.24712089444412166,
+      "learning_rate": 9.218107588213813e-05,
+      "loss": 0.3902,
+      "step": 1100
+    },
+    {
+      "epoch": 1.5909987448448986,
+      "grad_norm": 0.24987071841082312,
+      "learning_rate": 9.195583477542009e-05,
+      "loss": 0.3851,
+      "step": 1110
+    },
+    {
+      "epoch": 1.6053433745741437,
+      "grad_norm": 0.2453499914136973,
+      "learning_rate": 9.172767867700236e-05,
+      "loss": 0.3906,
+      "step": 1120
+    },
+    {
+      "epoch": 1.6196880043033888,
+      "grad_norm": 0.2453362805106032,
+      "learning_rate": 9.149662343862851e-05,
+      "loss": 0.3905,
+      "step": 1130
+    },
+    {
+      "epoch": 1.6340326340326339,
+      "grad_norm": 0.23102523441076062,
+      "learning_rate": 9.126268511346744e-05,
+      "loss": 0.3903,
+      "step": 1140
+    },
+    {
+      "epoch": 1.648377263761879,
+      "grad_norm": 0.2542501707506408,
+      "learning_rate": 9.102587995499807e-05,
+      "loss": 0.3953,
+      "step": 1150
+    },
+    {
+      "epoch": 1.6627218934911243,
+      "grad_norm": 0.23285474972918488,
+      "learning_rate": 9.078622441588009e-05,
+      "loss": 0.391,
+      "step": 1160
+    },
+    {
+      "epoch": 1.6770665232203694,
+      "grad_norm": 0.24565957619352327,
+      "learning_rate": 9.054373514681085e-05,
+      "loss": 0.3923,
+      "step": 1170
+    },
+    {
+      "epoch": 1.6914111529496145,
+      "grad_norm": 0.2506739557436684,
+      "learning_rate": 9.029842899536853e-05,
+      "loss": 0.3909,
+      "step": 1180
+    },
+    {
+      "epoch": 1.7057557826788596,
+      "grad_norm": 0.2441587365940751,
+      "learning_rate": 9.005032300484162e-05,
+      "loss": 0.3915,
+      "step": 1190
+    },
+    {
+      "epoch": 1.7201004124081047,
+      "grad_norm": 0.26421110183322566,
+      "learning_rate": 8.979943441304473e-05,
+      "loss": 0.3904,
+      "step": 1200
+    },
+    {
+      "epoch": 1.7344450421373498,
+      "grad_norm": 0.24194171269463752,
+      "learning_rate": 8.954578065112107e-05,
+      "loss": 0.3892,
+      "step": 1210
+    },
+    {
+      "epoch": 1.748789671866595,
+      "grad_norm": 0.23054663152441476,
+      "learning_rate": 8.928937934233123e-05,
+      "loss": 0.3907,
+      "step": 1220
+    },
+    {
+      "epoch": 1.76313430159584,
+      "grad_norm": 0.2369813966398001,
+      "learning_rate": 8.903024830082887e-05,
+      "loss": 0.3849,
+      "step": 1230
+    },
+    {
+      "epoch": 1.7774789313250852,
+      "grad_norm": 0.24008220076352446,
+      "learning_rate": 8.876840553042296e-05,
+      "loss": 0.3904,
+      "step": 1240
+    },
+    {
+      "epoch": 1.7918235610543303,
+      "grad_norm": 0.23428608617416669,
+      "learning_rate": 8.850386922332696e-05,
+      "loss": 0.387,
+      "step": 1250
+    },
+    {
+      "epoch": 1.8061681907835754,
+      "grad_norm": 0.23331291847215246,
+      "learning_rate": 8.823665775889486e-05,
+      "loss": 0.3909,
+      "step": 1260
+    },
+    {
+      "epoch": 1.8205128205128205,
+      "grad_norm": 0.23850242149763548,
+      "learning_rate": 8.796678970234427e-05,
+      "loss": 0.3833,
+      "step": 1270
+    },
+    {
+      "epoch": 1.8348574502420656,
+      "grad_norm": 0.2219682422982644,
+      "learning_rate": 8.769428380346642e-05,
+      "loss": 0.3845,
+      "step": 1280
+    },
+    {
+      "epoch": 1.8492020799713107,
+      "grad_norm": 0.22216238994388604,
+      "learning_rate": 8.741915899532362e-05,
+      "loss": 0.3865,
+      "step": 1290
+    },
+    {
+      "epoch": 1.8635467097005558,
+      "grad_norm": 0.22824929990219722,
+      "learning_rate": 8.714143439293376e-05,
+      "loss": 0.3852,
+      "step": 1300
+    },
+    {
+      "epoch": 1.8778913394298011,
+      "grad_norm": 0.24280880504252028,
+      "learning_rate": 8.686112929194226e-05,
+      "loss": 0.3861,
+      "step": 1310
+    },
+    {
+      "epoch": 1.8922359691590462,
+      "grad_norm": 0.24001704017165865,
+      "learning_rate": 8.657826316728142e-05,
+      "loss": 0.3908,
+      "step": 1320
+    },
+    {
+      "epoch": 1.9065805988882913,
+      "grad_norm": 0.22100828767921185,
+      "learning_rate": 8.62928556718174e-05,
+      "loss": 0.3871,
+      "step": 1330
+    },
+    {
+      "epoch": 1.9209252286175365,
+      "grad_norm": 0.22202012109824457,
+      "learning_rate": 8.600492663498477e-05,
+      "loss": 0.3834,
+      "step": 1340
+    },
+    {
+      "epoch": 1.9352698583467816,
+      "grad_norm": 0.21529127705519435,
+      "learning_rate": 8.571449606140883e-05,
+      "loss": 0.388,
+      "step": 1350
+    },
+    {
+      "epoch": 1.9496144880760267,
+      "grad_norm": 0.23391440077905082,
+      "learning_rate": 8.542158412951563e-05,
+      "loss": 0.3844,
+      "step": 1360
+    },
+    {
+      "epoch": 1.9639591178052718,
+      "grad_norm": 0.2331711540562185,
+      "learning_rate": 8.512621119013013e-05,
+      "loss": 0.393,
+      "step": 1370
+    },
+    {
+      "epoch": 1.9783037475345169,
+      "grad_norm": 0.23305451446876246,
+      "learning_rate": 8.482839776506232e-05,
+      "loss": 0.3837,
+      "step": 1380
+    },
+    {
+      "epoch": 1.992648377263762,
+      "grad_norm": 0.24517430064973736,
+      "learning_rate": 8.452816454568124e-05,
+      "loss": 0.3852,
+      "step": 1390
+    },
+    {
+      "epoch": 2.005737851891698,
+      "grad_norm": 0.27916951560917247,
+      "learning_rate": 8.422553239147754e-05,
+      "loss": 0.2799,
+      "step": 1400
+    },
+    {
+      "epoch": 2.020082481620943,
+      "grad_norm": 0.23593724272097047,
+      "learning_rate": 8.392052232861411e-05,
+      "loss": 0.201,
+      "step": 1410
+    },
+    {
+      "epoch": 2.034427111350188,
+      "grad_norm": 0.23512809512185134,
+      "learning_rate": 8.361315554846534e-05,
+      "loss": 0.1983,
+      "step": 1420
+    },
+    {
+      "epoch": 2.0487717410794333,
+      "grad_norm": 0.2286549447255431,
+      "learning_rate": 8.330345340614471e-05,
+      "loss": 0.1942,
+      "step": 1430
+    },
+    {
+      "epoch": 2.0631163708086784,
+      "grad_norm": 0.24844506458021867,
+      "learning_rate": 8.299143741902111e-05,
+      "loss": 0.1943,
+      "step": 1440
+    },
+    {
+      "epoch": 2.0774610005379235,
+      "grad_norm": 0.25585326293058985,
+      "learning_rate": 8.267712926522389e-05,
+      "loss": 0.1993,
+      "step": 1450
+    },
+    {
+      "epoch": 2.0918056302671686,
+      "grad_norm": 0.2421387690266048,
+      "learning_rate": 8.236055078213666e-05,
+      "loss": 0.1965,
+      "step": 1460
+    },
+    {
+      "epoch": 2.1061502599964137,
+      "grad_norm": 0.24707973152402415,
+      "learning_rate": 8.204172396488013e-05,
+      "loss": 0.1992,
+      "step": 1470
+    },
+    {
+      "epoch": 2.120494889725659,
+      "grad_norm": 0.23509099311770917,
+      "learning_rate": 8.172067096478395e-05,
+      "loss": 0.2008,
+      "step": 1480
+    },
+    {
+      "epoch": 2.134839519454904,
+      "grad_norm": 0.23605150518346998,
+      "learning_rate": 8.139741408784764e-05,
+      "loss": 0.2019,
+      "step": 1490
+    },
+    {
+      "epoch": 2.149184149184149,
+      "grad_norm": 0.2546177191590111,
+      "learning_rate": 8.107197579319082e-05,
+      "loss": 0.2053,
+      "step": 1500
+    },
+    {
+      "epoch": 2.163528778913394,
+      "grad_norm": 0.2353676242323245,
+      "learning_rate": 8.074437869149288e-05,
+      "loss": 0.204,
+      "step": 1510
+    },
+    {
+      "epoch": 2.1778734086426392,
+      "grad_norm": 0.23401952893152606,
+      "learning_rate": 8.041464554342197e-05,
+      "loss": 0.2036,
+      "step": 1520
+    },
+    {
+      "epoch": 2.1922180383718843,
+      "grad_norm": 0.23141975512545726,
+      "learning_rate": 8.008279925805366e-05,
+      "loss": 0.2033,
+      "step": 1530
+    },
+    {
+      "epoch": 2.2065626681011294,
+      "grad_norm": 0.23587920943899052,
+      "learning_rate": 7.974886289127927e-05,
+      "loss": 0.2068,
+      "step": 1540
+    },
+    {
+      "epoch": 2.2209072978303745,
+      "grad_norm": 0.2394814609218661,
+      "learning_rate": 7.941285964420407e-05,
+      "loss": 0.2049,
+      "step": 1550
+    },
+    {
+      "epoch": 2.23525192755962,
+      "grad_norm": 0.2389110148718096,
+      "learning_rate": 7.907481286153516e-05,
+      "loss": 0.2116,
+      "step": 1560
+    },
+    {
+      "epoch": 2.249596557288865,
+      "grad_norm": 0.2282395291006806,
+      "learning_rate": 7.873474602995973e-05,
+      "loss": 0.2088,
+      "step": 1570
+    },
+    {
+      "epoch": 2.2639411870181103,
+      "grad_norm": 0.23275397540700887,
+      "learning_rate": 7.839268277651311e-05,
+      "loss": 0.2092,
+      "step": 1580
+    },
+    {
+      "epoch": 2.2782858167473554,
+      "grad_norm": 0.22624466416184327,
+      "learning_rate": 7.80486468669373e-05,
+      "loss": 0.2088,
+      "step": 1590
+    },
+    {
+      "epoch": 2.2926304464766005,
+      "grad_norm": 0.23126585599149074,
+      "learning_rate": 7.770266220402977e-05,
+      "loss": 0.2117,
+      "step": 1600
+    },
+    {
+      "epoch": 2.3069750762058456,
+      "grad_norm": 0.226948134461606,
+      "learning_rate": 7.735475282598271e-05,
+      "loss": 0.2097,
+      "step": 1610
+    },
+    {
+      "epoch": 2.3213197059350907,
+      "grad_norm": 0.22673465714008167,
+      "learning_rate": 7.700494290471296e-05,
+      "loss": 0.2104,
+      "step": 1620
+    },
+    {
+      "epoch": 2.335664335664336,
+      "grad_norm": 0.2556824784968339,
+      "learning_rate": 7.665325674418264e-05,
+      "loss": 0.2136,
+      "step": 1630
+    },
+    {
+      "epoch": 2.350008965393581,
+      "grad_norm": 0.25025658975825976,
+      "learning_rate": 7.629971877871039e-05,
+      "loss": 0.2084,
+      "step": 1640
+    },
+    {
+      "epoch": 2.364353595122826,
+      "grad_norm": 0.22536490579422702,
+      "learning_rate": 7.594435357127399e-05,
+      "loss": 0.2089,
+      "step": 1650
+    },
+    {
+      "epoch": 2.378698224852071,
+      "grad_norm": 0.2258065984765025,
+      "learning_rate": 7.558718581180355e-05,
+      "loss": 0.2067,
+      "step": 1660
+    },
+    {
+      "epoch": 2.3930428545813163,
+      "grad_norm": 0.2464593742203822,
+      "learning_rate": 7.522824031546629e-05,
+      "loss": 0.2137,
+      "step": 1670
+    },
+    {
+      "epoch": 2.4073874843105614,
+      "grad_norm": 0.24123071412945177,
+      "learning_rate": 7.486754202094229e-05,
+      "loss": 0.2115,
+      "step": 1680
+    },
+    {
+      "epoch": 2.4217321140398065,
+      "grad_norm": 0.23105649429700748,
+      "learning_rate": 7.450511598869194e-05,
+      "loss": 0.2138,
+      "step": 1690
+    },
+    {
+      "epoch": 2.4360767437690516,
+      "grad_norm": 0.22955721039077792,
+      "learning_rate": 7.414098739921471e-05,
+      "loss": 0.2125,
+      "step": 1700
+    },
+    {
+      "epoch": 2.4504213734982967,
+      "grad_norm": 0.23154193335740872,
+      "learning_rate": 7.377518155129973e-05,
+      "loss": 0.2183,
+      "step": 1710
+    },
+    {
+      "epoch": 2.464766003227542,
+      "grad_norm": 0.2340236121998045,
+      "learning_rate": 7.340772386026801e-05,
+      "loss": 0.2157,
+      "step": 1720
+    },
+    {
+      "epoch": 2.479110632956787,
+      "grad_norm": 0.2250255353665983,
+      "learning_rate": 7.303863985620676e-05,
+      "loss": 0.2123,
+      "step": 1730
+    },
+    {
+      "epoch": 2.493455262686032,
+      "grad_norm": 0.2283114308365594,
+      "learning_rate": 7.266795518219548e-05,
+      "loss": 0.2135,
+      "step": 1740
+    },
+    {
+      "epoch": 2.507799892415277,
+      "grad_norm": 0.23546636465212323,
+      "learning_rate": 7.22956955925245e-05,
+      "loss": 0.214,
+      "step": 1750
+    },
+    {
+      "epoch": 2.5221445221445222,
+      "grad_norm": 0.23275268765839288,
+      "learning_rate": 7.192188695090545e-05,
+      "loss": 0.2156,
+      "step": 1760
+    },
+    {
+      "epoch": 2.5364891518737673,
+      "grad_norm": 0.2457436947556184,
+      "learning_rate": 7.154655522867452e-05,
+      "loss": 0.2189,
+      "step": 1770
+    },
+    {
+      "epoch": 2.5508337816030124,
+      "grad_norm": 0.2385729628030818,
+      "learning_rate": 7.116972650298782e-05,
+      "loss": 0.2148,
+      "step": 1780
+    },
+    {
+      "epoch": 2.5651784113322575,
+      "grad_norm": 0.2382827317725779,
+      "learning_rate": 7.079142695500975e-05,
+      "loss": 0.2127,
+      "step": 1790
+    },
+    {
+      "epoch": 2.5795230410615027,
+      "grad_norm": 0.22496477508883403,
+      "learning_rate": 7.041168286809397e-05,
+      "loss": 0.2156,
+      "step": 1800
+    },
+    {
+      "epoch": 2.5938676707907478,
+      "grad_norm": 0.2337756123669142,
+      "learning_rate": 7.00305206259572e-05,
+      "loss": 0.2163,
+      "step": 1810
+    },
+    {
+      "epoch": 2.608212300519993,
+      "grad_norm": 0.23547675501490803,
+      "learning_rate": 6.964796671084631e-05,
+      "loss": 0.213,
+      "step": 1820
+    },
+    {
+      "epoch": 2.622556930249238,
+      "grad_norm": 0.236949625863052,
+      "learning_rate": 6.926404770169819e-05,
+      "loss": 0.2108,
+      "step": 1830
+    },
+    {
+      "epoch": 2.636901559978483,
+      "grad_norm": 0.22775808389184637,
+      "learning_rate": 6.887879027229332e-05,
+      "loss": 0.2131,
+      "step": 1840
+    },
+    {
+      "epoch": 2.651246189707728,
+      "grad_norm": 0.25558095929144115,
+      "learning_rate": 6.84922211894024e-05,
+      "loss": 0.2146,
+      "step": 1850
+    },
+    {
+      "epoch": 2.6655908194369733,
+      "grad_norm": 0.23865636643565702,
+      "learning_rate": 6.810436731092671e-05,
+      "loss": 0.2154,
+      "step": 1860
+    },
+    {
+      "epoch": 2.6799354491662184,
+      "grad_norm": 0.23347390914436725,
+      "learning_rate": 6.771525558403203e-05,
+      "loss": 0.2145,
+      "step": 1870
+    },
+    {
+      "epoch": 2.6942800788954635,
+      "grad_norm": 0.2311770851119529,
+      "learning_rate": 6.73249130432765e-05,
+      "loss": 0.2112,
+      "step": 1880
+    },
+    {
+      "epoch": 2.7086247086247086,
+      "grad_norm": 0.2326246785839781,
+      "learning_rate": 6.69333668087323e-05,
+      "loss": 0.2133,
+      "step": 1890
+    },
+    {
+      "epoch": 2.7229693383539537,
+      "grad_norm": 0.23563376415545254,
+      "learning_rate": 6.654064408410132e-05,
+      "loss": 0.2141,
+      "step": 1900
+    },
+    {
+      "epoch": 2.737313968083199,
+      "grad_norm": 0.2298522950109398,
+      "learning_rate": 6.614677215482527e-05,
+      "loss": 0.2142,
+      "step": 1910
+    },
+    {
+      "epoch": 2.751658597812444,
+      "grad_norm": 0.2364865163676101,
+      "learning_rate": 6.57517783861898e-05,
+      "loss": 0.2127,
+      "step": 1920
+    },
+    {
+      "epoch": 2.766003227541689,
+      "grad_norm": 0.22837021217881728,
+      "learning_rate": 6.535569022142335e-05,
+      "loss": 0.2145,
+      "step": 1930
+    },
+    {
+      "epoch": 2.780347857270934,
+      "grad_norm": 0.22749769763881308,
+      "learning_rate": 6.495853517979035e-05,
+      "loss": 0.2106,
+      "step": 1940
+    },
+    {
+      "epoch": 2.7946924870001792,
+      "grad_norm": 0.21764981978938533,
+      "learning_rate": 6.456034085467935e-05,
+      "loss": 0.2125,
+      "step": 1950
+    },
+    {
+      "epoch": 2.8090371167294244,
+      "grad_norm": 0.22774012921821585,
+      "learning_rate": 6.416113491168581e-05,
+      "loss": 0.213,
+      "step": 1960
+    },
+    {
+      "epoch": 2.8233817464586695,
+      "grad_norm": 0.22793686074861258,
+      "learning_rate": 6.376094508668999e-05,
+      "loss": 0.2116,
+      "step": 1970
+    },
+    {
+      "epoch": 2.8377263761879146,
+      "grad_norm": 0.24345345462191187,
+      "learning_rate": 6.335979918392999e-05,
+      "loss": 0.213,
+      "step": 1980
+    },
+    {
+      "epoch": 2.8520710059171597,
+      "grad_norm": 0.230566718186529,
+      "learning_rate": 6.295772507406982e-05,
+      "loss": 0.2123,
+      "step": 1990
+    },
+    {
+      "epoch": 2.866415635646405,
+      "grad_norm": 0.23922165240449358,
+      "learning_rate": 6.255475069226326e-05,
+      "loss": 0.211,
+      "step": 2000
+    },
+    {
+      "epoch": 2.88076026537565,
+      "grad_norm": 0.22058336484670613,
+      "learning_rate": 6.21509040362127e-05,
+      "loss": 0.2122,
+      "step": 2010
+    },
+    {
+      "epoch": 2.895104895104895,
+      "grad_norm": 0.2272702011851071,
+      "learning_rate": 6.174621316422417e-05,
+      "loss": 0.2147,
+      "step": 2020
+    },
+    {
+      "epoch": 2.90944952483414,
+      "grad_norm": 0.23799805104509125,
+      "learning_rate": 6.134070619325774e-05,
+      "loss": 0.212,
+      "step": 2030
+    },
+    {
+      "epoch": 2.923794154563385,
+      "grad_norm": 0.24608349983752625,
+      "learning_rate": 6.0934411296974184e-05,
+      "loss": 0.2122,
+      "step": 2040
+    },
+    {
+      "epoch": 2.9381387842926303,
+      "grad_norm": 0.23079480496683127,
+      "learning_rate": 6.052735670377736e-05,
+      "loss": 0.211,
+      "step": 2050
+    },
+    {
+      "epoch": 2.9524834140218754,
+      "grad_norm": 0.22680559271715478,
+      "learning_rate": 6.0119570694853155e-05,
+      "loss": 0.2102,
+      "step": 2060
+    },
+    {
+      "epoch": 2.9668280437511205,
+      "grad_norm": 0.22760761484882805,
+      "learning_rate": 5.97110816022044e-05,
+      "loss": 0.2113,
+      "step": 2070
+    },
+    {
+      "epoch": 2.9811726734803656,
+      "grad_norm": 0.23303799910976278,
+      "learning_rate": 5.930191780668258e-05,
+      "loss": 0.2088,
+      "step": 2080
+    },
+    {
+      "epoch": 2.9955173032096107,
+      "grad_norm": 0.22946738031807773,
+      "learning_rate": 5.88921077360159e-05,
+      "loss": 0.2097,
+      "step": 2090
+    },
+    {
+      "epoch": 3.008606777837547,
+      "grad_norm": 0.2697620900381124,
+      "learning_rate": 5.848167986283421e-05,
+      "loss": 0.1134,
+      "step": 2100
+    },
+    {
+      "epoch": 3.0229514075667923,
+      "grad_norm": 0.1885938841096422,
+      "learning_rate": 5.807066270269084e-05,
+      "loss": 0.0763,
+      "step": 2110
+    },
+    {
+      "epoch": 3.0372960372960374,
+      "grad_norm": 0.214693696805492,
+      "learning_rate": 5.765908481208139e-05,
+      "loss": 0.0756,
+      "step": 2120
+    },
+    {
+      "epoch": 3.0516406670252825,
+      "grad_norm": 0.2339101871402584,
+      "learning_rate": 5.724697478645963e-05,
+      "loss": 0.0744,
+      "step": 2130
+    },
+    {
+      "epoch": 3.0659852967545276,
+      "grad_norm": 0.1971755620952271,
+      "learning_rate": 5.6834361258250844e-05,
+      "loss": 0.072,
+      "step": 2140
+    },
+    {
+      "epoch": 3.0803299264837727,
+      "grad_norm": 0.1981153430750115,
+      "learning_rate": 5.642127289486246e-05,
+      "loss": 0.0748,
+      "step": 2150
+    },
+    {
+      "epoch": 3.094674556213018,
+      "grad_norm": 0.2172902671287561,
+      "learning_rate": 5.600773839669237e-05,
+      "loss": 0.0726,
+      "step": 2160
+    },
+    {
+      "epoch": 3.109019185942263,
+      "grad_norm": 0.19669334061877888,
+      "learning_rate": 5.559378649513478e-05,
+      "loss": 0.0733,
+      "step": 2170
+    },
+    {
+      "epoch": 3.123363815671508,
+      "grad_norm": 0.21027113699329436,
+      "learning_rate": 5.517944595058413e-05,
+      "loss": 0.0746,
+      "step": 2180
+    },
+    {
+      "epoch": 3.137708445400753,
+      "grad_norm": 0.20204087893287273,
+      "learning_rate": 5.476474555043688e-05,
+      "loss": 0.0748,
+      "step": 2190
+    },
+    {
+      "epoch": 3.152053075129998,
+      "grad_norm": 0.20638588917150788,
+      "learning_rate": 5.4349714107091335e-05,
+      "loss": 0.0744,
+      "step": 2200
+    },
+    {
+      "epoch": 3.1663977048592433,
+      "grad_norm": 0.20367882761147596,
+      "learning_rate": 5.393438045594595e-05,
+      "loss": 0.0755,
+      "step": 2210
+    },
+    {
+      "epoch": 3.1807423345884884,
+      "grad_norm": 0.20836979312681028,
+      "learning_rate": 5.351877345339583e-05,
+      "loss": 0.076,
+      "step": 2220
+    },
+    {
+      "epoch": 3.1950869643177335,
+      "grad_norm": 0.19643695987807314,
+      "learning_rate": 5.310292197482791e-05,
+      "loss": 0.0733,
+      "step": 2230
+    },
+    {
+      "epoch": 3.2094315940469786,
+      "grad_norm": 0.20621763947145422,
+      "learning_rate": 5.268685491261472e-05,
+      "loss": 0.075,
+      "step": 2240
+    },
+    {
+      "epoch": 3.2237762237762237,
+      "grad_norm": 0.20777873086593704,
+      "learning_rate": 5.227060117410702e-05,
+      "loss": 0.0746,
+      "step": 2250
+    },
+    {
+      "epoch": 3.238120853505469,
+      "grad_norm": 0.2021910407099938,
+      "learning_rate": 5.185418967962543e-05,
+      "loss": 0.0747,
+      "step": 2260
+    },
+    {
+      "epoch": 3.252465483234714,
+      "grad_norm": 0.2016612434414281,
+      "learning_rate": 5.143764936045106e-05,
+      "loss": 0.0743,
+      "step": 2270
+    },
+    {
+      "epoch": 3.266810112963959,
+      "grad_norm": 0.2180992659409795,
+      "learning_rate": 5.1021009156815414e-05,
+      "loss": 0.0744,
+      "step": 2280
+    },
+    {
+      "epoch": 3.281154742693204,
+      "grad_norm": 0.2056058145565962,
+      "learning_rate": 5.060429801588983e-05,
+      "loss": 0.0744,
+      "step": 2290
+    },
+    {
+      "epoch": 3.2954993724224493,
+      "grad_norm": 0.20164051829762908,
+      "learning_rate": 5.018754488977409e-05,
+      "loss": 0.0745,
+      "step": 2300
+    },
+    {
+      "epoch": 3.3098440021516944,
+      "grad_norm": 0.2026538165933443,
+      "learning_rate": 4.9770778733485065e-05,
+      "loss": 0.074,
+      "step": 2310
+    },
+    {
+      "epoch": 3.3241886318809395,
+      "grad_norm": 0.20427324673762595,
+      "learning_rate": 4.935402850294494e-05,
+      "loss": 0.0739,
+      "step": 2320
+    },
+    {
+      "epoch": 3.3385332616101846,
+      "grad_norm": 0.20831211218540635,
+      "learning_rate": 4.893732315296942e-05,
+      "loss": 0.0748,
+      "step": 2330
+    },
+    {
+      "epoch": 3.3528778913394297,
+      "grad_norm": 0.20740018500070947,
+      "learning_rate": 4.852069163525595e-05,
+      "loss": 0.0737,
+      "step": 2340
+    },
+    {
+      "epoch": 3.367222521068675,
+      "grad_norm": 0.20060155886370676,
+      "learning_rate": 4.810416289637234e-05,
+      "loss": 0.0729,
+      "step": 2350
+    },
+    {
+      "epoch": 3.38156715079792,
+      "grad_norm": 0.199826847154071,
+      "learning_rate": 4.7687765875745574e-05,
+      "loss": 0.0739,
+      "step": 2360
+    },
+    {
+      "epoch": 3.395911780527165,
+      "grad_norm": 0.20063705204581495,
+      "learning_rate": 4.727152950365117e-05,
+      "loss": 0.0737,
+      "step": 2370
+    },
+    {
+      "epoch": 3.41025641025641,
+      "grad_norm": 0.20947972363514977,
+      "learning_rate": 4.685548269920312e-05,
+      "loss": 0.0736,
+      "step": 2380
+    },
+    {
+      "epoch": 3.4246010399856552,
+      "grad_norm": 0.2006701925989043,
+      "learning_rate": 4.643965436834474e-05,
+      "loss": 0.075,
+      "step": 2390
+    },
+    {
+      "epoch": 3.4389456697149003,
+      "grad_norm": 0.20335025504735554,
+      "learning_rate": 4.6024073401840336e-05,
+      "loss": 0.0745,
+      "step": 2400
+    },
+    {
+      "epoch": 3.4532902994441455,
+      "grad_norm": 0.2192162565442083,
+      "learning_rate": 4.560876867326791e-05,
+      "loss": 0.0738,
+      "step": 2410
+    },
+    {
+      "epoch": 3.4676349291733906,
+      "grad_norm": 0.19858055523329815,
+      "learning_rate": 4.5193769037013066e-05,
+      "loss": 0.0732,
+      "step": 2420
+    },
+    {
+      "epoch": 3.4819795589026357,
+      "grad_norm": 0.20485303414115183,
+      "learning_rate": 4.477910332626438e-05,
+      "loss": 0.0728,
+      "step": 2430
+    },
+    {
+      "epoch": 3.4963241886318808,
+      "grad_norm": 0.19011594248287844,
+      "learning_rate": 4.4364800351010066e-05,
+      "loss": 0.0726,
+      "step": 2440
+    },
+    {
+      "epoch": 3.5106688183611263,
+      "grad_norm": 0.20410603253979742,
+      "learning_rate": 4.395088889603633e-05,
+      "loss": 0.0736,
+      "step": 2450
+    },
+    {
+      "epoch": 3.5250134480903714,
+      "grad_norm": 0.1994983957599032,
+      "learning_rate": 4.353739771892746e-05,
+      "loss": 0.073,
+      "step": 2460
+    },
+    {
+      "epoch": 3.5393580778196165,
+      "grad_norm": 0.20349060401414618,
+      "learning_rate": 4.312435554806787e-05,
+      "loss": 0.0736,
+      "step": 2470
+    },
+    {
+      "epoch": 3.5537027075488616,
+      "grad_norm": 0.20221336765718947,
+      "learning_rate": 4.271179108064605e-05,
+      "loss": 0.0713,
+      "step": 2480
+    },
+    {
+      "epoch": 3.5680473372781067,
+      "grad_norm": 0.1920539935100462,
+      "learning_rate": 4.229973298066083e-05,
+      "loss": 0.0714,
+      "step": 2490
+    },
+    {
+      "epoch": 3.582391967007352,
+      "grad_norm": 0.18514594535819984,
+      "learning_rate": 4.188820987692981e-05,
+      "loss": 0.0716,
+      "step": 2500
+    },
+    {
+      "epoch": 3.596736596736597,
+      "grad_norm": 0.19390555703637974,
+      "learning_rate": 4.1477250361100317e-05,
+      "loss": 0.072,
+      "step": 2510
+    },
+    {
+      "epoch": 3.611081226465842,
+      "grad_norm": 0.19881724163942532,
+      "learning_rate": 4.106688298566295e-05,
+      "loss": 0.0722,
+      "step": 2520
+    },
+    {
+      "epoch": 3.625425856195087,
+      "grad_norm": 0.19864848134388496,
+      "learning_rate": 4.065713626196778e-05,
+      "loss": 0.0697,
+      "step": 2530
+    },
+    {
+      "epoch": 3.6397704859243323,
+      "grad_norm": 0.20964033399772472,
+      "learning_rate": 4.0248038658243515e-05,
+      "loss": 0.0703,
+      "step": 2540
+    },
+    {
+      "epoch": 3.6541151156535774,
+      "grad_norm": 0.1887224816930325,
+      "learning_rate": 3.983961859761946e-05,
+      "loss": 0.071,
+      "step": 2550
+    },
+    {
+      "epoch": 3.6684597453828225,
+      "grad_norm": 0.1939910437911645,
+      "learning_rate": 3.9431904456150914e-05,
+      "loss": 0.0685,
+      "step": 2560
+    },
+    {
+      "epoch": 3.6828043751120676,
+      "grad_norm": 0.1905566250106664,
+      "learning_rate": 3.902492456084757e-05,
+      "loss": 0.0709,
+      "step": 2570
+    },
+    {
+      "epoch": 3.6971490048413127,
+      "grad_norm": 0.1954219594857734,
+      "learning_rate": 3.861870718770545e-05,
+      "loss": 0.0691,
+      "step": 2580
+    },
+    {
+      "epoch": 3.711493634570558,
+      "grad_norm": 0.20129771340548336,
+      "learning_rate": 3.821328055974231e-05,
+      "loss": 0.0688,
+      "step": 2590
+    },
+    {
+      "epoch": 3.725838264299803,
+      "grad_norm": 0.19424451985885532,
+      "learning_rate": 3.780867284503685e-05,
+      "loss": 0.0705,
+      "step": 2600
+    },
+    {
+      "epoch": 3.740182894029048,
+      "grad_norm": 0.19307288848206286,
+      "learning_rate": 3.7404912154771626e-05,
+      "loss": 0.069,
+      "step": 2610
+    },
+    {
+      "epoch": 3.754527523758293,
+      "grad_norm": 0.20224114631498458,
+      "learning_rate": 3.7002026541279905e-05,
+      "loss": 0.069,
+      "step": 2620
+    },
+    {
+      "epoch": 3.7688721534875382,
+      "grad_norm": 0.19645086260070405,
+      "learning_rate": 3.660004399609675e-05,
+      "loss": 0.0693,
+      "step": 2630
+    },
+    {
+      "epoch": 3.7832167832167833,
+      "grad_norm": 0.2009057118393354,
+      "learning_rate": 3.619899244801414e-05,
+      "loss": 0.0695,
+      "step": 2640
+    },
+    {
+      "epoch": 3.7975614129460284,
+      "grad_norm": 0.20154345565922616,
+      "learning_rate": 3.5798899761140626e-05,
+      "loss": 0.0688,
+      "step": 2650
+    },
+    {
+      "epoch": 3.8119060426752736,
+      "grad_norm": 0.19819908788727933,
+      "learning_rate": 3.5399793732965324e-05,
+      "loss": 0.0703,
+      "step": 2660
+    },
+    {
+      "epoch": 3.8262506724045187,
+      "grad_norm": 0.19579772630914064,
+      "learning_rate": 3.500170209242671e-05,
+      "loss": 0.0673,
+      "step": 2670
+    },
+    {
+      "epoch": 3.8405953021337638,
+      "grad_norm": 0.1930709905078437,
+      "learning_rate": 3.460465249798592e-05,
+      "loss": 0.068,
+      "step": 2680
+    },
+    {
+      "epoch": 3.854939931863009,
+      "grad_norm": 0.19375769656837338,
+      "learning_rate": 3.420867253570529e-05,
+      "loss": 0.0668,
+      "step": 2690
+    },
+    {
+      "epoch": 3.869284561592254,
+      "grad_norm": 0.19590535607906298,
+      "learning_rate": 3.381378971733161e-05,
+      "loss": 0.0658,
+      "step": 2700
+    },
+    {
+      "epoch": 3.883629191321499,
+      "grad_norm": 0.19485732453673113,
+      "learning_rate": 3.342003147838475e-05,
+      "loss": 0.0671,
+      "step": 2710
+    },
+    {
+      "epoch": 3.897973821050744,
+      "grad_norm": 0.19120048275674587,
+      "learning_rate": 3.302742517625144e-05,
+      "loss": 0.0665,
+      "step": 2720
+    },
+    {
+      "epoch": 3.9123184507799893,
+      "grad_norm": 0.19464302345990753,
+      "learning_rate": 3.2635998088284596e-05,
+      "loss": 0.0662,
+      "step": 2730
+    },
+    {
+      "epoch": 3.9266630805092344,
+      "grad_norm": 0.20017821890333443,
+      "learning_rate": 3.224577740990814e-05,
+      "loss": 0.0655,
+      "step": 2740
+    },
+    {
+      "epoch": 3.9410077102384795,
+      "grad_norm": 0.18866754533216776,
+      "learning_rate": 3.185679025272753e-05,
+      "loss": 0.0663,
+      "step": 2750
+    },
+    {
+      "epoch": 3.9553523399677246,
+      "grad_norm": 0.19243520237850759,
+      "learning_rate": 3.146906364264606e-05,
+      "loss": 0.0657,
+      "step": 2760
+    },
+    {
+      "epoch": 3.9696969696969697,
+      "grad_norm": 0.1924186499329835,
+      "learning_rate": 3.108262451798724e-05,
+      "loss": 0.0651,
+      "step": 2770
+    },
+    {
+      "epoch": 3.984041599426215,
+      "grad_norm": 0.21148279808210507,
+      "learning_rate": 3.069749972762316e-05,
+      "loss": 0.0648,
+      "step": 2780
+    },
+    {
+      "epoch": 3.99838622915546,
+      "grad_norm": 0.19991640585361364,
+      "learning_rate": 3.0313716029109064e-05,
+      "loss": 0.0645,
+      "step": 2790
+    },
+    {
+      "epoch": 4.011475703783396,
+      "grad_norm": 0.13064534629220334,
+      "learning_rate": 2.993130008682436e-05,
+      "loss": 0.0228,
+      "step": 2800
+    },
+    {
+      "epoch": 4.025820333512641,
+      "grad_norm": 0.14874535281957305,
+      "learning_rate": 2.955027847011993e-05,
+      "loss": 0.0176,
+      "step": 2810
+    },
+    {
+      "epoch": 4.040164963241886,
+      "grad_norm": 0.14336180228683498,
+      "learning_rate": 2.917067765147229e-05,
+      "loss": 0.0176,
+      "step": 2820
+    },
+    {
+      "epoch": 4.054509592971131,
+      "grad_norm": 0.12559441494646076,
+      "learning_rate": 2.8792524004644283e-05,
+      "loss": 0.0167,
+      "step": 2830
+    },
+    {
+      "epoch": 4.068854222700376,
+      "grad_norm": 0.12484448147694403,
+      "learning_rate": 2.8415843802852672e-05,
+      "loss": 0.0167,
+      "step": 2840
+    },
+    {
+      "epoch": 4.083198852429621,
+      "grad_norm": 0.1337296091314726,
+      "learning_rate": 2.8040663216942752e-05,
+      "loss": 0.0169,
+      "step": 2850
+    },
+    {
+      "epoch": 4.0975434821588665,
+      "grad_norm": 0.12242456577697475,
+      "learning_rate": 2.7667008313570076e-05,
+      "loss": 0.0161,
+      "step": 2860
+    },
+    {
+      "epoch": 4.111888111888112,
+      "grad_norm": 0.13243095768870966,
+      "learning_rate": 2.729490505338943e-05,
+      "loss": 0.0161,
+      "step": 2870
+    },
+    {
+      "epoch": 4.126232741617357,
+      "grad_norm": 0.12277718816926177,
+      "learning_rate": 2.692437928925109e-05,
+      "loss": 0.0157,
+      "step": 2880
+    },
+    {
+      "epoch": 4.140577371346602,
+      "grad_norm": 0.137540991678628,
+      "learning_rate": 2.655545676440464e-05,
+      "loss": 0.0159,
+      "step": 2890
+    },
+    {
+      "epoch": 4.154922001075847,
+      "grad_norm": 0.13131712471544715,
+      "learning_rate": 2.6188163110710435e-05,
+      "loss": 0.0161,
+      "step": 2900
+    },
+    {
+      "epoch": 4.169266630805092,
+      "grad_norm": 0.13640835620647865,
+      "learning_rate": 2.582252384685874e-05,
+      "loss": 0.0164,
+      "step": 2910
+    },
+    {
+      "epoch": 4.183611260534337,
+      "grad_norm": 0.12543818653508698,
+      "learning_rate": 2.5458564376596732e-05,
+      "loss": 0.0157,
+      "step": 2920
+    },
+    {
+      "epoch": 4.197955890263582,
+      "grad_norm": 0.11736167137678152,
+      "learning_rate": 2.509630998696349e-05,
+      "loss": 0.0154,
+      "step": 2930
+    },
+    {
+      "epoch": 4.212300519992827,
+      "grad_norm": 0.1245551329001544,
+      "learning_rate": 2.473578584653321e-05,
+      "loss": 0.0152,
+      "step": 2940
+    },
+    {
+      "epoch": 4.2266451497220725,
+      "grad_norm": 0.12763631881000323,
+      "learning_rate": 2.4377017003666413e-05,
+      "loss": 0.0155,
+      "step": 2950
+    },
+    {
+      "epoch": 4.240989779451318,
+      "grad_norm": 0.13313595943206588,
+      "learning_rate": 2.4020028384769795e-05,
+      "loss": 0.015,
+      "step": 2960
+    },
+    {
+      "epoch": 4.255334409180563,
+      "grad_norm": 0.125069284406997,
+      "learning_rate": 2.366484479256425e-05,
+      "loss": 0.015,
+      "step": 2970
+    },
+    {
+      "epoch": 4.269679038909808,
+      "grad_norm": 0.13167131954772826,
+      "learning_rate": 2.3311490904361738e-05,
+      "loss": 0.0159,
+      "step": 2980
+    },
+    {
+      "epoch": 4.284023668639053,
+      "grad_norm": 0.11344149792986571,
+      "learning_rate": 2.295999127035071e-05,
+      "loss": 0.0147,
+      "step": 2990
+    },
+    {
+      "epoch": 4.298368298368298,
+      "grad_norm": 0.1299095136285245,
+      "learning_rate": 2.26103703118905e-05,
+      "loss": 0.015,
+      "step": 3000
+    }
+  ],
+  "logging_steps": 10,
+  "max_steps": 4188,
+  "num_input_tokens_seen": 0,
+  "num_train_epochs": 6,
+  "save_steps": 500,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 940152769216512.0,
+  "train_batch_size": 1,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d984a7c189c70dbf5d61cb23d3712e05c9f7357204184deb2f3e29263430b342
+size 7953

zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/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 . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+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, weights_only=False)
+
+        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 tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # 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}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] 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, exclude_frozen_parameters):
+    """
+    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,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+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 _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+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() if _has_callable(shape, 'numel') else math.prod(shape)
+            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,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _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")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+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 = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * 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
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        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}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        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,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _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 to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    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.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    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")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    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_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``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``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+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_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    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("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    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_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)