Update model

Microsoft [released a new version](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/tree/main) of this model, which I've converted and quantized.

config.json

@@ -2,6 +2,7 @@
     "architectures": [
         "Phi3ForCausalLM"
     ],
+    "attention_bias": false,
     "attention_dropout": 0.0,
     "auto_map": {
         "AutoConfig": "configuration_phi3.Phi3Config",
@@ -32,7 +33,7 @@
     "sliding_window": 2047,
     "tie_word_embeddings": false,
     "torch_dtype": "bfloat16",
-    "transformers_version": "4.39.3",
+    "transformers_version": "4.40.2",
     "use_cache": true,
     "vocab_size": 32064
 }

configuration_phi3.py

@@ -1,200 +1,227 @@
-# 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 factor 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 `longrope` 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.
-        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,
-        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.sliding_window = sliding_window
-
-        super().__init__(
-            eos_token_id=eos_token_id,
-            pad_token_id=pad_token_id,
-            tie_word_embeddings=tie_word_embeddings,
-            **kwargs,
-        )
-
-    def _rope_scaling_validation(self):
-        if self.rope_scaling is None:
-            return
-
-        assert (
-            (isinstance(self.rope_scaling, dict))
-            and ("type" in self.rope_scaling)
-            and ("short_factor" in self.rope_scaling)
-            and ("long_factor" in self.rope_scaling)
-        ), (
-            "`rope_scaling` must be a dictionary with three keys: `type`, `short_factor` and `long_factor`, "
-            f"got {self.rope_scaling}."
-        )
-
-        assert self.rope_scaling["type"].lower() == "longrope", "RoPE scaling type must be `longrope`."
-
-        short_factor = self.rope_scaling["short_factor"]
-        assert isinstance(short_factor, list) and all(
-            isinstance(x, (int, float)) for x in short_factor
-        ), f"RoPE scaling factor must be a list of numbers, got {short_factor}."
-        assert (
-            len(short_factor) == self.hidden_size // self.num_attention_heads // 2
-        ), f"Length of RoPE scaling factor must be half of the attention head, got {short_factor}."
-
-        long_factor = self.rope_scaling["long_factor"]
-        assert isinstance(long_factor, list) and all(
-            isinstance(x, (int, float)) for x in long_factor
-        ), f"RoPE scaling factor must be a list of numbers, got {long_factor}."
-        assert (
-            len(long_factor) == self.hidden_size // self.num_attention_heads // 2
-        ), f"Length of RoPE scaling factor must be half of the attention head, got {long_factor}."
+# 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 `longrope` 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_adjustment()
+        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_adjustment(self):
+        """
+        Adjust the `type` of the `rope_scaling` configuration for backward compatibility.
+        """
+        if self.rope_scaling is None:
+            return
+
+        rope_scaling_type = self.rope_scaling.get("type", None)
+
+        # For backward compatibility if previous version used "su" or "yarn"
+        if rope_scaling_type is not None and rope_scaling_type in ["su", "yarn"]:
+            self.rope_scaling["type"] = "longrope"
+
+    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 ["longrope"]:
+            raise ValueError(f"`rope_scaling`'s type field must be one of ['longrope'], 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)}"
+            )

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:eea61bcc55335594f54c275e0149c12c9e0818bde5d9b6c3419420f663059bcf
-size 2149696133
+oid sha256:8d75680621a09474f6601e9176f2f61f92a5e4c079d68d583901f51699fda50a
+size 2149696167

sample_finetune.py

@@ -1,129 +1,214 @@
-import torch
-from datasets import load_dataset
-from trl import SFTTrainer
-from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
-
-"""
-A simple example on using SFTTrainer and Accelerate to finetune Phi-3 models. For
-a more advanced example, please follow HF alignment-handbook/scripts/run_sft.py
-
-1. Install accelerate: 
-    conda install -c conda-forge accelerate
-2. Setup accelerate config:
-    accelerate config
-to simply use all the GPUs available:
-    python -c "from accelerate.utils import write_basic_config; write_basic_config(mixed_precision='bf16')"
-check accelerate config:
-    accelerate env
-3. Run the code:
-    accelerate launch sample_finetune.py
-"""
-
-###################
-# Hyper-parameters
-###################
-args = {
-    "bf16": True,
-    "do_eval": False,
-    "eval_strategy": "no",
-    "learning_rate": 5.0e-06,
-    "log_level": "info",
-    "logging_steps": 20,
-    "logging_strategy": "steps",
-    "lr_scheduler_type": "cosine",
-    "num_train_epochs": 1,
-    "max_steps": -1,
-    "output_dir": "./checkpoint_dir",
-    "overwrite_output_dir": True,
-    "per_device_eval_batch_size": 4,
-    "per_device_train_batch_size": 8,
-    "remove_unused_columns": True,
-    "save_steps": 100,
-    "save_total_limit": 1,
-    "seed": 0,
-    "gradient_checkpointing": True,
-    "gradient_checkpointing_kwargs":{"use_reentrant": False},
-    "gradient_accumulation_steps": 1,
-    "warmup_ratio": 0.2,
-    }
-    
-training_args = TrainingArguments(**args)
-
-################
-# Modle Loading
-################
-checkpoint_path = "microsoft/Phi-3-mini-4k-instruct"
-# checkpoint_path = "microsoft/Phi-3-mini-128k-instruct"
-model_kwargs = dict(
-    use_cache=False,
-    trust_remote_code=True,
-    attn_implementation="flash_attention_2",  # loading the model with flash-attenstion support
-    torch_dtype=torch.bfloat16,
-    device_map="cuda",
-)
-model = AutoModelForCausalLM.from_pretrained(checkpoint_path, **model_kwargs)
-tokenizer = AutoTokenizer.from_pretrained(checkpoint_path)
-tokenizer.pad_token = tokenizer.unk_token  # use unk rather than eos token to prevent endless generation
-tokenizer.pad_token_id = tokenizer.convert_tokens_to_ids(tokenizer.pad_token)
-tokenizer.padding_side = 'right'
-
-##################
-# Data Processing
-##################
-def apply_chat_template(
-    example,
-    tokenizer,
-):
-    messages = example["messages"]
-    # Add an empty system message if there is none
-    if messages[0]["role"] != "system":
-        messages.insert(0, {"role": "system", "content": ""})
-    example["text"] = tokenizer.apply_chat_template(
-        messages, tokenize=False, add_generation_prompt=False)
-    return example
-
-raw_dataset = load_dataset("HuggingFaceH4/ultrachat_200k")
-column_names = list(raw_dataset["train_sft"].features)
-
-processed_dataset = raw_dataset.map(
-    apply_chat_template,
-    fn_kwargs={"tokenizer": tokenizer},
-    num_proc=12,
-    remove_columns=column_names,
-    desc="Applying chat template",
-)
-train_dataset = processed_dataset["train_sft"]
-eval_dataset = processed_dataset["test_sft"]
-
-###########
-# Training
-###########
-trainer = SFTTrainer(
-    model=model,
-    args=training_args,
-    train_dataset=train_dataset,
-    eval_dataset=eval_dataset,
-    max_seq_length=2048,
-    dataset_text_field="text",
-    tokenizer=tokenizer,
-    packing=True
-)
-train_result = trainer.train()
-metrics = train_result.metrics
-trainer.log_metrics("train", metrics)
-trainer.save_metrics("train", metrics)
-trainer.save_state()
-
-#############
-# Evaluation
-#############
-tokenizer.padding_side = 'left'
-metrics = trainer.evaluate()
-metrics["eval_samples"] = len(eval_dataset)
-trainer.log_metrics("eval", metrics)
-trainer.save_metrics("eval", metrics)
-
-############
-# Save model
-############
-trainer.save_model(training_args.output_dir)
+import sys
+import logging
+
+import datasets
+from datasets import load_dataset
+from peft import LoraConfig
+import torch
+import transformers
+from trl import SFTTrainer
+from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, BitsAndBytesConfig
+
+"""
+A simple example on using SFTTrainer and Accelerate to finetune Phi-3 models. For
+a more advanced example, please follow HF alignment-handbook/scripts/run_sft.py.
+This example has utilized DeepSpeed ZeRO3 offload to reduce the memory usage. The
+script can be run on V100 or later generation GPUs. Here are some suggestions on 
+futher reducing memory consumption:
+    - reduce batch size
+    - decrease lora dimension
+    - restrict lora target modules
+Please follow these steps to run the script:
+1. Install dependencies: 
+    conda install -c conda-forge accelerate
+    pip3 install -i https://pypi.org/simple/ bitsandbytes
+    pip3 install peft transformers trl datasets
+    pip3 install deepspeed
+2. Setup accelerate and deepspeed config based on the machine used:
+    accelerate config
+Here is a sample config for deepspeed zero3:
+    compute_environment: LOCAL_MACHINE
+    debug: false
+    deepspeed_config:
+      gradient_accumulation_steps: 1
+      offload_optimizer_device: none
+      offload_param_device: none
+      zero3_init_flag: true
+      zero3_save_16bit_model: true
+      zero_stage: 3
+    distributed_type: DEEPSPEED
+    downcast_bf16: 'no'
+    enable_cpu_affinity: false
+    machine_rank: 0
+    main_training_function: main
+    mixed_precision: bf16
+    num_machines: 1
+    num_processes: 4
+    rdzv_backend: static
+    same_network: true
+    tpu_env: []
+    tpu_use_cluster: false
+    tpu_use_sudo: false
+    use_cpu: false
+3. check accelerate config:
+    accelerate env
+4. Run the code:
+    accelerate launch sample_finetune.py
+"""
+
+logger = logging.getLogger(__name__)
+
+
+###################
+# Hyper-parameters
+###################
+training_config = {
+    "bf16": True,
+    "do_eval": False,
+    "learning_rate": 5.0e-06,
+    "log_level": "info",
+    "logging_steps": 20,
+    "logging_strategy": "steps",
+    "lr_scheduler_type": "cosine",
+    "num_train_epochs": 1,
+    "max_steps": -1,
+    "output_dir": "./checkpoint_dir",
+    "overwrite_output_dir": True,
+    "per_device_eval_batch_size": 4,
+    "per_device_train_batch_size": 4,
+    "remove_unused_columns": True,
+    "save_steps": 100,
+    "save_total_limit": 1,
+    "seed": 0,
+    "gradient_checkpointing": True,
+    "gradient_checkpointing_kwargs":{"use_reentrant": False},
+    "gradient_accumulation_steps": 1,
+    "warmup_ratio": 0.2,
+    }
+
+peft_config = {
+    "r": 16,
+    "lora_alpha": 32,
+    "lora_dropout": 0.05,
+    "bias": "none",
+    "task_type": "CAUSAL_LM",
+    "target_modules": "all-linear",
+    "modules_to_save": None,
+}
+train_conf = TrainingArguments(**training_config)
+peft_conf = LoraConfig(**peft_config)
+
+
+###############
+# Setup logging
+###############
+logging.basicConfig(
+    format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+    datefmt="%Y-%m-%d %H:%M:%S",
+    handlers=[logging.StreamHandler(sys.stdout)],
+)
+log_level = train_conf.get_process_log_level()
+logger.setLevel(log_level)
+datasets.utils.logging.set_verbosity(log_level)
+transformers.utils.logging.set_verbosity(log_level)
+transformers.utils.logging.enable_default_handler()
+transformers.utils.logging.enable_explicit_format()
+
+# Log on each process a small summary
+logger.warning(
+    f"Process rank: {train_conf.local_rank}, device: {train_conf.device}, n_gpu: {train_conf.n_gpu}"
+    + f" distributed training: {bool(train_conf.local_rank != -1)}, 16-bits training: {train_conf.fp16}"
+)
+logger.info(f"Training/evaluation parameters {train_conf}")
+logger.info(f"PEFT parameters {peft_conf}")
+
+
+################
+# Model Loading
+################
+checkpoint_path = "microsoft/Phi-3-mini-4k-instruct"
+# checkpoint_path = "microsoft/Phi-3-mini-128k-instruct"
+model_kwargs = dict(
+    use_cache=False,
+    trust_remote_code=True,
+    attn_implementation="flash_attention_2",  # loading the model with flash-attenstion support
+    torch_dtype=torch.bfloat16,
+    device_map=None
+)
+model = AutoModelForCausalLM.from_pretrained(checkpoint_path, **model_kwargs)
+tokenizer = AutoTokenizer.from_pretrained(checkpoint_path)
+tokenizer.model_max_length = 2048
+tokenizer.pad_token = tokenizer.unk_token  # use unk rather than eos token to prevent endless generation
+tokenizer.pad_token_id = tokenizer.convert_tokens_to_ids(tokenizer.pad_token)
+tokenizer.padding_side = 'right'
+
+
+##################
+# Data Processing
+##################
+def apply_chat_template(
+    example,
+    tokenizer,
+):
+    messages = example["messages"]
+    example["text"] = tokenizer.apply_chat_template(
+        messages, tokenize=False, add_generation_prompt=False)
+    return example
+
+raw_dataset = load_dataset("HuggingFaceH4/ultrachat_200k")
+train_dataset = raw_dataset["train_sft"]
+test_dataset = raw_dataset["test_sft"]
+column_names = list(train_dataset.features)
+
+processed_train_dataset = train_dataset.map(
+    apply_chat_template,
+    fn_kwargs={"tokenizer": tokenizer},
+    num_proc=10,
+    remove_columns=column_names,
+    desc="Applying chat template to train_sft",
+)
+
+processed_test_dataset = test_dataset.map(
+    apply_chat_template,
+    fn_kwargs={"tokenizer": tokenizer},
+    num_proc=10,
+    remove_columns=column_names,
+    desc="Applying chat template to test_sft",
+)
+
+
+###########
+# Training
+###########
+trainer = SFTTrainer(
+    model=model,
+    args=train_conf,
+    peft_config=peft_conf,
+    train_dataset=processed_train_dataset,
+    eval_dataset=processed_test_dataset,
+    max_seq_length=2048,
+    dataset_text_field="text",
+    tokenizer=tokenizer,
+    packing=True
+)
+train_result = trainer.train()
+metrics = train_result.metrics
+trainer.log_metrics("train", metrics)
+trainer.save_metrics("train", metrics)
+trainer.save_state()
+
+
+#############
+# Evaluation
+#############
+tokenizer.padding_side = 'left'
+metrics = trainer.evaluate()
+metrics["eval_samples"] = len(processed_test_dataset)
+trainer.log_metrics("eval", metrics)
+trainer.save_metrics("eval", metrics)
+
+
+# ############
+# # Save model
+# ############
+trainer.save_model(train_conf.output_dir)

special_tokens_map.json

@@ -1,7 +1,4 @@
 {
-  "additional_special_tokens": [
-    "<|/inst|>"
-  ],
   "bos_token": {
     "content": "<s>",
     "lstrip": false,
@@ -17,7 +14,7 @@
     "single_word": false
   },
   "pad_token": {
-    "content": "<|end|>",
+    "content": "<|endoftext|>",
     "lstrip": false,
     "normalized": false,
     "rstrip": false,

tokenizer.json

@@ -50,7 +50,7 @@
     },
     {
       "id": 32002,
-      "content": "<|step|>",
+      "content": "<|placeholder1|>",
       "single_word": false,
       "lstrip": false,
       "rstrip": true,
@@ -59,7 +59,7 @@
     },
     {
       "id": 32003,
-      "content": "<|function_output|>",
+      "content": "<|placeholder2|>",
       "single_word": false,
       "lstrip": false,
       "rstrip": true,
@@ -68,7 +68,7 @@
     },
     {
       "id": 32004,
-      "content": "<|tag|>",
+      "content": "<|placeholder3|>",
       "single_word": false,
       "lstrip": false,
       "rstrip": true,
@@ -77,7 +77,7 @@
     },
     {
       "id": 32005,
-      "content": "<|function_call|>",
+      "content": "<|placeholder4|>",
       "single_word": false,
       "lstrip": false,
       "rstrip": true,
@@ -104,7 +104,7 @@
     },
     {
       "id": 32008,
-      "content": "<|raw|>",
+      "content": "<|placeholder5|>",
       "single_word": false,
       "lstrip": false,
       "rstrip": true,
@@ -113,7 +113,7 @@
     },
     {
       "id": 32009,
-      "content": "<|continue|>",
+      "content": "<|placeholder6|>",
       "single_word": false,
       "lstrip": false,
       "rstrip": true,
@@ -128,249 +128,6 @@
       "rstrip": true,
       "normalized": false,
       "special": true
-    },
-    {
-      "id": 32011,
-      "content": "<|function_list|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32012,
-      "content": "<|calc|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32013,
-      "content": "<|code|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32014,
-      "content": "<|/code|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32015,
-      "content": "<|summary|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32016,
-      "content": "<|resource|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32017,
-      "content": "<|assistant_mask|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32018,
-      "content": "<|start|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32019,
-      "content": "<|message|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32020,
-      "content": "<|fim_prefix|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32021,
-      "content": "<|fim_middle|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32022,
-      "content": "<|fim_suffix|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32023,
-      "content": "<|meta_start|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32024,
-      "content": "<|ipynb_marker|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32025,
-      "content": "<|diff_marker|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32026,
-      "content": "<|ghissue|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32027,
-      "content": "<|ghreview|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32028,
-      "content": "<|disc_start|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32029,
-      "content": "<|disc_sep|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32030,
-      "content": "<|disc_thread|><|query|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32031,
-      "content": "<|/query|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32032,
-      "content": "<|data|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32033,
-      "content": "<|/data|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32034,
-      "content": "<|sys|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32035,
-      "content": "<|/sys|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32036,
-      "content": "<|inst|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
-    },
-    {
-      "id": 32037,
-      "content": "<|/inst|>",
-      "single_word": false,
-      "lstrip": false,
-      "rstrip": true,
-      "normalized": false,
-      "special": true
     }
   ],
   "normalizer": {
@@ -393,12 +150,6 @@
   "post_processor": {
     "type": "TemplateProcessing",
     "single": [
-      {
-        "SpecialToken": {
-          "id": "<s>",
-          "type_id": 0
-        }
-      },
       {
         "Sequence": {
           "id": "A",
@@ -407,24 +158,12 @@
       }
     ],
     "pair": [
-      {
-        "SpecialToken": {
-          "id": "<s>",
-          "type_id": 0
-        }
-      },
       {
         "Sequence": {
           "id": "A",
           "type_id": 0
         }
       },
-      {
-        "SpecialToken": {
-          "id": "<s>",
-          "type_id": 1
-        }
-      },
       {
         "Sequence": {
           "id": "B",
@@ -432,17 +171,7 @@
         }
       }
     ],
-    "special_tokens": {
-      "<s>": {
-        "id": "<s>",
-        "ids": [
-          1
-        ],
-        "tokens": [
-          "<s>"
-        ]
-      }
-    }
+    "special_tokens": {}
   },
   "decoder": {
     "type": "Sequence",

tokenizer_config.json

@@ -1,6 +1,7 @@
 {
-  "add_bos_token": true,
+  "add_bos_token": false,
   "add_eos_token": false,
+  "add_prefix_space": null,
   "added_tokens_decoder": {
     "0": {
       "content": "<unk>",
@@ -43,7 +44,7 @@
       "special": true
     },
     "32002": {
-      "content": "<|step|>",
+      "content": "<|placeholder1|>",
       "lstrip": false,
       "normalized": false,
       "rstrip": true,
@@ -51,7 +52,7 @@
       "special": true
     },
     "32003": {
-      "content": "<|function_output|>",
+      "content": "<|placeholder2|>",
       "lstrip": false,
       "normalized": false,
       "rstrip": true,
@@ -59,7 +60,7 @@
       "special": true
     },
     "32004": {
-      "content": "<|tag|>",
+      "content": "<|placeholder3|>",
       "lstrip": false,
       "normalized": false,
       "rstrip": true,
@@ -67,7 +68,7 @@
       "special": true
     },
     "32005": {
-      "content": "<|function_call|>",
+      "content": "<|placeholder4|>",
       "lstrip": false,
       "normalized": false,
       "rstrip": true,
@@ -91,7 +92,7 @@
       "special": true
     },
     "32008": {
-      "content": "<|raw|>",
+      "content": "<|placeholder5|>",
       "lstrip": false,
       "normalized": false,
       "rstrip": true,
@@ -99,7 +100,7 @@
       "special": true
     },
     "32009": {
-      "content": "<|continue|>",
+      "content": "<|placeholder6|>",
       "lstrip": false,
       "normalized": false,
       "rstrip": true,
@@ -113,234 +114,15 @@
       "rstrip": true,
       "single_word": false,
       "special": true
-    },
-    "32011": {
-      "content": "<|function_list|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32012": {
-      "content": "<|calc|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32013": {
-      "content": "<|code|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32014": {
-      "content": "<|/code|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32015": {
-      "content": "<|summary|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32016": {
-      "content": "<|resource|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32017": {
-      "content": "<|assistant_mask|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32018": {
-      "content": "<|start|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32019": {
-      "content": "<|message|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32020": {
-      "content": "<|fim_prefix|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32021": {
-      "content": "<|fim_middle|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32022": {
-      "content": "<|fim_suffix|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32023": {
-      "content": "<|meta_start|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32024": {
-      "content": "<|ipynb_marker|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32025": {
-      "content": "<|diff_marker|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32026": {
-      "content": "<|ghissue|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32027": {
-      "content": "<|ghreview|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32028": {
-      "content": "<|disc_start|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32029": {
-      "content": "<|disc_sep|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32030": {
-      "content": "<|disc_thread|><|query|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32031": {
-      "content": "<|/query|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32032": {
-      "content": "<|data|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32033": {
-      "content": "<|/data|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32034": {
-      "content": "<|sys|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32035": {
-      "content": "<|/sys|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32036": {
-      "content": "<|inst|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
-    },
-    "32037": {
-      "content": "<|/inst|>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": true,
-      "single_word": false,
-      "special": true
     }
   },
-  "additional_special_tokens": [
-    "<|/inst|>"
-  ],
   "bos_token": "<s>",
-  "chat_template": "{{ bos_token }}{% for message in messages %}{{'<|' + message['role'] + '|>' + '\n' + message['content'] + '<|end|>\n' }}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|>\n' }}{% else %}{{ eos_token }}{% endif %}",
+  "chat_template": "{% for message in messages %}{% if message['role'] == 'system' %}{{'<|system|>\n' + message['content'] + '<|end|>\n'}}{% elif message['role'] == 'user' %}{{'<|user|>\n' + message['content'] + '<|end|>\n'}}{% elif message['role'] == 'assistant' %}{{'<|assistant|>\n' + message['content'] + '<|end|>\n'}}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|>\n' }}{% else %}{{ eos_token }}{% endif %}",
   "clean_up_tokenization_spaces": false,
   "eos_token": "<|endoftext|>",
   "legacy": false,
   "model_max_length": 4096,
-  "pad_token": "<|end|>",
+  "pad_token": "<|endoftext|>",
   "padding_side": "left",
   "sp_model_kwargs": {},
   "tokenizer_class": "LlamaTokenizer",