From f0d22e926347b491f8a3ae4c19ca7df9b31c2541 Mon Sep 17 00:00:00 2001 From: ModelHub XC Date: Tue, 16 Jun 2026 05:50:16 +0800 Subject: [PATCH] =?UTF-8?q?=E5=88=9D=E5=A7=8B=E5=8C=96=E9=A1=B9=E7=9B=AE?= =?UTF-8?q?=EF=BC=8C=E7=94=B1ModelHub=20XC=E7=A4=BE=E5=8C=BA=E6=8F=90?= =?UTF-8?q?=E4=BE=9B=E6=A8=A1=E5=9E=8B?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Model: pa5haw/Phi-4-mini-instruct-mlx-fp16 Source: Original Platform --- .gitattributes | 36 + README.md | 67 ++ chat_template.jinja | 1 + config.json | 146 ++++ configuration_phi3.py | 226 ++++++ generation_config.json | 10 + model-00001-of-00002.safetensors | 3 + model-00002-of-00002.safetensors | 3 + model.safetensors.index.json | 202 +++++ modeling_phi3.py | 1180 ++++++++++++++++++++++++++++++ sample_finetune.py | 214 ++++++ tokenizer.json | 3 + tokenizer_config.json | 12 + 13 files changed, 2103 insertions(+) create mode 100644 .gitattributes create mode 100644 README.md create mode 100644 chat_template.jinja create mode 100644 config.json create mode 100644 configuration_phi3.py create mode 100644 generation_config.json create mode 100644 model-00001-of-00002.safetensors create mode 100644 model-00002-of-00002.safetensors create mode 100644 model.safetensors.index.json create mode 100644 modeling_phi3.py create mode 100644 sample_finetune.py create mode 100644 tokenizer.json create mode 100644 tokenizer_config.json diff --git a/.gitattributes b/.gitattributes new file mode 100644 index 0000000..52373fe --- /dev/null +++ b/.gitattributes @@ -0,0 +1,36 @@ +*.7z filter=lfs diff=lfs merge=lfs -text +*.arrow filter=lfs diff=lfs merge=lfs -text +*.bin filter=lfs diff=lfs merge=lfs -text +*.bz2 filter=lfs diff=lfs merge=lfs -text +*.ckpt filter=lfs diff=lfs merge=lfs -text +*.ftz filter=lfs diff=lfs merge=lfs -text +*.gz filter=lfs diff=lfs merge=lfs -text +*.h5 filter=lfs diff=lfs merge=lfs -text +*.joblib filter=lfs diff=lfs merge=lfs -text +*.lfs.* filter=lfs diff=lfs merge=lfs -text +*.mlmodel filter=lfs diff=lfs merge=lfs -text +*.model filter=lfs diff=lfs merge=lfs -text +*.msgpack filter=lfs diff=lfs merge=lfs -text +*.npy filter=lfs diff=lfs merge=lfs -text +*.npz filter=lfs diff=lfs merge=lfs -text +*.onnx filter=lfs diff=lfs merge=lfs -text +*.ot filter=lfs diff=lfs merge=lfs -text +*.parquet filter=lfs diff=lfs merge=lfs -text +*.pb filter=lfs diff=lfs merge=lfs -text +*.pickle filter=lfs diff=lfs merge=lfs -text +*.pkl filter=lfs diff=lfs merge=lfs -text +*.pt filter=lfs diff=lfs merge=lfs -text +*.pth filter=lfs diff=lfs merge=lfs -text +*.rar filter=lfs diff=lfs merge=lfs -text +*.safetensors filter=lfs diff=lfs merge=lfs -text +saved_model/**/* filter=lfs diff=lfs merge=lfs -text +*.tar.* filter=lfs diff=lfs merge=lfs -text +*.tar filter=lfs diff=lfs merge=lfs -text +*.tflite filter=lfs diff=lfs merge=lfs -text +*.tgz filter=lfs diff=lfs merge=lfs -text +*.wasm filter=lfs diff=lfs merge=lfs -text +*.xz filter=lfs diff=lfs merge=lfs -text +*.zip filter=lfs diff=lfs merge=lfs -text +*.zst filter=lfs diff=lfs merge=lfs -text +*tfevents* filter=lfs diff=lfs merge=lfs -text +tokenizer.json filter=lfs diff=lfs merge=lfs -text diff --git a/README.md b/README.md new file mode 100644 index 0000000..6334cd6 --- /dev/null +++ b/README.md @@ -0,0 +1,67 @@ +--- +language: +- multilingual +- ar +- zh +- cs +- da +- nl +- en +- fi +- fr +- de +- he +- hu +- it +- ja +- ko +- 'no' +- pl +- pt +- ru +- es +- sv +- th +- tr +- uk +library_name: transformers +license: mit +license_link: https://huggingface.co/microsoft/Phi-4-mini-instruct/resolve/main/LICENSE +pipeline_tag: text-generation +tags: +- nlp +- code +- mlx +- mlx-my-repo +widget: +- messages: + - role: user + content: Can you provide ways to eat combinations of bananas and dragonfruits? +base_model: microsoft/Phi-4-mini-instruct +--- + +# pa5haw/Phi-4-mini-instruct-mlx-fp16 + +The Model [pa5haw/Phi-4-mini-instruct-mlx-fp16](https://huggingface.co/pa5haw/Phi-4-mini-instruct-mlx-fp16) was converted to MLX format from [microsoft/Phi-4-mini-instruct](https://huggingface.co/microsoft/Phi-4-mini-instruct) using mlx-lm version **0.31.2**. + +## Use with mlx + +```bash +pip install mlx-lm +``` + +```python +from mlx_lm import load, generate + +model, tokenizer = load("pa5haw/Phi-4-mini-instruct-mlx-fp16") + +prompt="hello" + +if hasattr(tokenizer, "apply_chat_template") and tokenizer.chat_template is not None: + messages = [{"role": "user", "content": prompt}] + prompt = tokenizer.apply_chat_template( + messages, tokenize=False, add_generation_prompt=True + ) + +response = generate(model, tokenizer, prompt=prompt, verbose=True) +``` diff --git a/chat_template.jinja b/chat_template.jinja new file mode 100644 index 0000000..a9c00dd --- /dev/null +++ b/chat_template.jinja @@ -0,0 +1 @@ +{% for message in messages %}{% if message['role'] == 'system' and 'tools' in message and message['tools'] is not none %}{{ '<|' + message['role'] + '|>' + message['content'] + '<|tool|>' + message['tools'] + '<|/tool|>' + '<|end|>' }}{% else %}{{ '<|' + message['role'] + '|>' + message['content'] + '<|end|>' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|>' }}{% else %}{{ eos_token }}{% endif %} \ No newline at end of file diff --git a/config.json b/config.json new file mode 100644 index 0000000..79356be --- /dev/null +++ b/config.json @@ -0,0 +1,146 @@ +{ + "architectures": [ + "Phi3ForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "auto_map": { + "AutoConfig": "configuration_phi3.Phi3Config", + "AutoModelForCausalLM": "modeling_phi3.Phi3ForCausalLM", + "AutoTokenizer": "Xenova/gpt-4o" + }, + "bos_token_id": 199999, + "embd_pdrop": 0.0, + "eos_token_id": [ + 200020, + 199999 + ], + "full_attn_mod": 1, + "hidden_act": "silu", + "hidden_size": 3072, + "initializer_range": 0.02, + "intermediate_size": 8192, + "interpolate_factor": 1, + "lm_head_bias": false, + "max_position_embeddings": 131072, + "mlp_bias": false, + "model_type": "phi3", + "num_attention_heads": 24, + "num_hidden_layers": 32, + "num_key_value_heads": 8, + "original_max_position_embeddings": 4096, + "pad_token_id": 199999, + "partial_rotary_factor": 0.75, + "resid_pdrop": 0.0, + "rms_norm_eps": 1e-05, + "rope_scaling": { + "long_factor": [ + 1, + 1.118320672, + 1.250641126, + 1.398617824, + 1.564103225, + 1.74916897, + 1.956131817, + 2.187582649, + 2.446418898, + 2.735880826, + 3.059592084, + 3.421605075, + 3.826451687, + 4.279200023, + 4.785517845, + 5.351743533, + 5.984965424, + 6.693110555, + 7.485043894, + 8.370679318, + 9.36110372, + 10.4687158, + 11.70738129, + 13.09260651, + 14.64173252, + 16.37415215, + 18.31155283, + 20.47818807, + 22.90118105, + 25.61086418, + 28.64115884, + 32.03, + 32.1, + 32.13, + 32.23, + 32.6, + 32.61, + 32.64, + 32.66, + 32.7, + 32.71, + 32.93, + 32.97, + 33.28, + 33.49, + 33.5, + 44.16, + 47.77 + ], + "short_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, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "type": "longrope" + }, + "rope_theta": 10000.0, + "sliding_window": 262144, + "tie_word_embeddings": true, + "torch_dtype": "bfloat16", + "transformers_version": "4.45.0", + "use_cache": true, + "vocab_size": 200064 +} \ No newline at end of file diff --git a/configuration_phi3.py b/configuration_phi3.py new file mode 100644 index 0000000..01446e1 --- /dev/null +++ b/configuration_phi3.py @@ -0,0 +1,226 @@ +# 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__) + + +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. + partial_rotary_factor (`float`, *optional*, defaults to 1.0): + Percentage of the query and keys which will have rotary embedding. Must be between 0.0 and 1.0. + 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, + partial_rotary_factor=1.0, + 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.partial_rotary_factor = partial_rotary_factor + 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}" + ) + rotary_ndims = int(self.hidden_size // self.num_attention_heads * self.partial_rotary_factor) + if not len(rope_scaling_short_factor) == rotary_ndims // 2: + raise ValueError( + f"`rope_scaling`'s short_factor field must have length {rotary_ndims // 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) == rotary_ndims // 2: + raise ValueError( + f"`rope_scaling`'s long_factor field must have length {rotary_ndims // 2}, got {len(rope_scaling_long_factor)}" + ) diff --git a/generation_config.json b/generation_config.json new file mode 100644 index 0000000..43b32d0 --- /dev/null +++ b/generation_config.json @@ -0,0 +1,10 @@ +{ + "_from_model_config": true, + "bos_token_id": 199999, + "eos_token_id": [ + 200020, + 199999 + ], + "pad_token_id": 199999, + "transformers_version": "4.45.0" +} \ No newline at end of file diff --git a/model-00001-of-00002.safetensors b/model-00001-of-00002.safetensors new file mode 100644 index 0000000..acc2514 --- /dev/null +++ b/model-00001-of-00002.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d006053a54d48237488883d4cb5bf1ddb3bafcf5f35a1085b0583f20b0948464 +size 5306316717 diff --git a/model-00002-of-00002.safetensors b/model-00002-of-00002.safetensors new file mode 100644 index 0000000..e16c158 --- /dev/null +++ b/model-00002-of-00002.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aaa6502400a5576a4ad23869d87ef660b51ea3afe32420a285e32a7ea3d823ea +size 2365749254 diff --git a/model.safetensors.index.json b/model.safetensors.index.json new file mode 100644 index 0000000..973ac9c --- /dev/null +++ b/model.safetensors.index.json @@ -0,0 +1,202 @@ +{ + "metadata": { + "total_size": 7672043520, + "total_parameters": 3836021760 + }, + "weight_map": { + "model.embed_tokens.weight": "model-00001-of-00002.safetensors", + "model.layers.0.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.0.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.0.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.0.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.1.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.1.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.1.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.1.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.10.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.10.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.10.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.10.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.10.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.10.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.11.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.11.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.11.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.11.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.11.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.11.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.12.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.12.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.12.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.12.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.12.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.12.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.13.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.13.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.13.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.13.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.13.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.13.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.14.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.14.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.14.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.14.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.14.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.14.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.15.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.15.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.15.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.15.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.15.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.15.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.16.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.16.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.16.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.16.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.16.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.16.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.17.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.17.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.17.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.17.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.17.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.17.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.18.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.18.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.18.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.18.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.18.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.18.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.19.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.19.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.19.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.19.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.19.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.19.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.2.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.2.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.2.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.2.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.20.input_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.20.mlp.down_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.20.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.20.post_attention_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.20.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.20.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.21.input_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.21.mlp.down_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.21.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.21.post_attention_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.21.self_attn.o_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.21.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.22.input_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.22.mlp.down_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.22.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.22.post_attention_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.22.self_attn.o_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.22.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.23.input_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.23.mlp.down_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.23.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.23.post_attention_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.23.self_attn.o_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.23.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.24.input_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.24.mlp.down_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.24.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.24.post_attention_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.24.self_attn.o_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.24.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.25.input_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.25.mlp.down_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.25.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.25.post_attention_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.25.self_attn.o_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.25.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.26.input_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.26.mlp.down_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.26.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.26.post_attention_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.26.self_attn.o_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.26.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.27.input_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.27.mlp.down_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.27.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.27.post_attention_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.27.self_attn.o_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.27.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.28.input_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.28.mlp.down_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.28.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.28.post_attention_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.28.self_attn.o_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.28.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.29.input_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.29.mlp.down_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.29.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.29.post_attention_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.29.self_attn.o_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.29.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.3.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.3.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.3.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.3.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.30.input_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.30.mlp.down_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.30.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.30.post_attention_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.30.self_attn.o_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.30.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.31.input_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.31.mlp.down_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.31.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.31.post_attention_layernorm.weight": "model-00002-of-00002.safetensors", + "model.layers.31.self_attn.o_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.31.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors", + "model.layers.4.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.4.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.4.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.4.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.5.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.5.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.5.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.5.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.5.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.5.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.6.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.6.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.6.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.6.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.6.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.6.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.7.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.7.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.7.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.7.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.7.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.7.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.8.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.8.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.8.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.8.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.8.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.8.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.9.input_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.9.mlp.down_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.9.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.9.post_attention_layernorm.weight": "model-00001-of-00002.safetensors", + "model.layers.9.self_attn.o_proj.weight": "model-00001-of-00002.safetensors", + "model.layers.9.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors", + "model.norm.weight": "model-00002-of-00002.safetensors" + } +} \ No newline at end of file diff --git a/modeling_phi3.py b/modeling_phi3.py new file mode 100644 index 0000000..71597d1 --- /dev/null +++ b/modeling_phi3.py @@ -0,0 +1,1180 @@ +# 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.""" + +from typing import Callable, List, Optional, Tuple, Union + +import torch +from torch import nn + +from transformers.activations import ACT2FN +from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache +from transformers.generation import GenerationMixin +from transformers.modeling_attn_mask_utils import AttentionMaskConverter +from transformers.modeling_flash_attention_utils import FlashAttentionKwargs +from transformers.modeling_outputs import ( + BaseModelOutputWithPast, + CausalLMOutputWithPast, + SequenceClassifierOutputWithPast, + TokenClassifierOutput, +) +from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS +from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel +from transformers.processing_utils import Unpack +from transformers.utils import ( + LossKwargs, + add_code_sample_docstrings, + add_start_docstrings, + add_start_docstrings_to_model_forward, + logging, + replace_return_docstrings, +) +from transformers.utils.deprecation import deprecate_kwarg +from .configuration_phi3 import Phi3Config + + +logger = logging.get_logger(__name__) + +_CHECKPOINT_FOR_DOC = "microsoft/Phi-3-mini-4k-instruct" +_CONFIG_FOR_DOC = "Phi3Config" + + +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) + + +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) + + +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) + + +def eager_attention_forward( + module: nn.Module, + query: torch.Tensor, + key: torch.Tensor, + value: torch.Tensor, + attention_mask: Optional[torch.Tensor], + scaling: float, + dropout: float = 0.0, + **kwargs, +): + key_states = repeat_kv(key, module.num_key_value_groups) + value_states = repeat_kv(value, module.num_key_value_groups) + + attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling + if attention_mask is not None: + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] + attn_weights = attn_weights + causal_mask + + attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype) + attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training) + attn_output = torch.matmul(attn_weights, value_states) + attn_output = attn_output.transpose(1, 2).contiguous() + + return attn_output, attn_weights + + +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) + + rotary_dim = cos.shape[-1] + q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:] + k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:] + + q_embed = torch.cat([(q_rot * cos) + (rotate_half(q_rot) * sin), q_pass], dim=-1) + k_embed = torch.cat([(k_rot * cos) + (rotate_half(k_rot) * sin), k_pass], dim=-1) + return q_embed, k_embed + + +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 + self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads) + self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads + self.num_key_value_heads = config.num_key_value_heads + self.scaling = self.head_dim**-0.5 + self.attention_dropout = config.attention_dropout + self.is_causal = True + + op_size = config.num_attention_heads * self.head_dim + 2 * (config.num_key_value_heads * self.head_dim) + self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False) + self.qkv_proj = nn.Linear(config.hidden_size, op_size, bias=False) + + def forward( + self, + hidden_states: torch.Tensor, + position_embeddings: Tuple[torch.Tensor, torch.Tensor], + attention_mask: Optional[torch.Tensor], + past_key_value: Optional[Cache] = None, + cache_position: Optional[torch.LongTensor] = None, + **kwargs: Unpack[FlashAttentionKwargs], + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + input_shape = hidden_states.shape[:-1] + hidden_shape = (*input_shape, -1, self.head_dim) + + qkv = self.qkv_proj(hidden_states) + query_pos = self.config.num_attention_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(hidden_shape).transpose(1, 2) + key_states = key_states.view(hidden_shape).transpose(1, 2) + value_states = value_states.view(hidden_shape).transpose(1, 2) + + cos, sin = position_embeddings + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) + + if past_key_value is not None: + # sin and cos are specific to RoPE models; cache_position needed for the static cache + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} + key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) + + attention_interface: Callable = eager_attention_forward + if self.config._attn_implementation != "eager": + if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False): + logger.warning_once( + "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to " + 'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.' + ) + else: + attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation] + + attn_output, attn_weights = attention_interface( + self, + query_states, + key_states, + value_states, + attention_mask, + dropout=0.0 if not self.training else self.attention_dropout, + scaling=self.scaling, + sliding_window=getattr(self.config, "sliding_window", None), + **kwargs, + ) + + attn_output = attn_output.reshape(*input_shape, -1).contiguous() + attn_output = self.o_proj(attn_output) + return attn_output, attn_weights + + +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) + + def extra_repr(self): + return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}" + + +class Phi3DecoderLayer(nn.Module): + def __init__(self, config: Phi3Config, layer_idx: int): + super().__init__() + self.hidden_size = config.hidden_size + self.self_attn = Phi3Attention(config=config, layer_idx=layer_idx) + self.mlp = Phi3MLP(config) + self.input_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_attention_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.config = config + self.resid_attn_dropout = nn.Dropout(config.resid_pdrop) + self.resid_mlp_dropout = nn.Dropout(config.resid_pdrop) + + 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: Optional[bool] = False, + use_cache: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC + **kwargs: Unpack[FlashAttentionKwargs], + ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: + """ + 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) + past_key_value (`Cache`, *optional*): cached past key and value projection states + 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`). + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence + kwargs (`dict`, *optional*): + Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code + into the model + """ + residual = hidden_states + + hidden_states = self.input_layernorm(hidden_states) + + # Self Attention + hidden_states, self_attn_weights = 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, + cache_position=cache_position, + position_embeddings=position_embeddings, + **kwargs, + ) + hidden_states = residual + self.resid_attn_dropout(hidden_states) # main diff with Llama + + 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) # main diff with Llama + + outputs = (hidden_states,) + if output_attentions: + outputs += (self_attn_weights,) + + return outputs + + +class Phi3RotaryEmbedding(nn.Module): + def __init__(self, config: Phi3Config, device=None): + super().__init__() + # BC: "rope_type" was originally "type" + if hasattr(config, "rope_scaling") and config.rope_scaling is not None: + self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type")) + else: + self.rope_type = "default" + self.max_seq_len_cached = config.max_position_embeddings + self.original_max_seq_len = config.max_position_embeddings + + self.config = config + self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type] + + inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self.original_inv_freq = self.inv_freq + + def _dynamic_frequency_update(self, position_ids, device): + """ + dynamic RoPE layers should recompute `inv_freq` in the following situations: + 1 - growing beyond the cached sequence length (allow scaling) + 2 - the current sequence length is in the original scale (avoid losing precision with small sequences) + """ + seq_len = torch.max(position_ids) + 1 + if seq_len > self.max_seq_len_cached: # growth + inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, seq_len=seq_len) + self.register_buffer("inv_freq", inv_freq, persistent=False) # TODO joao: may break with compilation + self.max_seq_len_cached = seq_len + + if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len: # reset + # This .to() is needed if the model has been moved to a device after being initialized (because + # the buffer is automatically moved, but not the original copy) + self.original_inv_freq = self.original_inv_freq.to(device) + self.register_buffer("inv_freq", self.original_inv_freq, persistent=False) + self.max_seq_len_cached = self.original_max_seq_len + + @torch.no_grad() + def forward(self, x, position_ids): + if "dynamic" in self.rope_type: + self._dynamic_frequency_update(position_ids, device=x.device) + elif self.rope_type == "longrope": + self._longrope_frequency_update(position_ids, device=x.device) + + # Core RoPE block + 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 (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() + + # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention + cos = cos * self.attention_scaling + sin = sin * self.attention_scaling + + return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) + + def _longrope_frequency_update(self, position_ids, device): + """Longrope uses long factor if sequence is larger than original pretraining length, short otherwise.""" + seq_len = torch.max(position_ids) + 1 + if hasattr(self.config, "original_max_position_embeddings"): + original_max_position_embeddings = self.config.original_max_position_embeddings + else: + original_max_position_embeddings = self.config.max_position_embeddings + if seq_len > original_max_position_embeddings: + if not hasattr(self, "long_inv_freq"): + self.long_inv_freq, _ = self.rope_init_fn( + self.config, device, seq_len=original_max_position_embeddings + 1 + ) + self.register_buffer("inv_freq", self.long_inv_freq, persistent=False) + else: + # This .to() is needed if the model has been moved to a device after being initialized (because + # the buffer is automatically moved, but not the original copy) + self.original_inv_freq = self.original_inv_freq.to(device) + self.register_buffer("inv_freq", self.original_inv_freq, persistent=False) + + +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 Phi3 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 = True + _supports_flex_attn = True + _supports_cache_class = True + _supports_quantized_cache = True + _supports_static_cache = True + _supports_attention_backend = 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, see our + [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache); + - 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. + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, + this tensor is not affected by padding. It is used to update the cache in the correct position and to infer + the complete sequence length. +""" + + +@add_start_docstrings( + "The bare Phi3 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.layers = nn.ModuleList( + [Phi3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] + ) + self.norm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.rotary_emb = Phi3RotaryEmbedding(config=config) + 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[Cache] = 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, + cache_position: Optional[torch.LongTensor] = None, + **flash_attn_kwargs: Unpack[FlashAttentionKwargs], + ) -> 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 + + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError("You must specify exactly one of input_ids or inputs_embeds") + + if self.gradient_checkpointing and self.training and use_cache: + logger.warning_once( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`." + ) + use_cache = False + + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) + + if use_cache and past_key_values is None: + past_key_values = DynamicCache() + + if cache_position is None: + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + cache_position = torch.arange( + past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device + ) + + if position_ids is None: + position_ids = cache_position.unsqueeze(0) + + causal_mask = self._update_causal_mask( + attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions + ) + + hidden_states = inputs_embeds + + # create position embeddings to be shared across the decoder layers + position_embeddings = self.rotary_emb(hidden_states, position_ids) + + # decoder layers + all_hidden_states = () if output_hidden_states else None + all_self_attns = () if output_attentions else None + + for decoder_layer in self.layers[: self.config.num_hidden_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, + causal_mask, + position_ids, + past_key_values, + output_attentions, + use_cache, + cache_position, + position_embeddings, + ) + else: + layer_outputs = decoder_layer( + hidden_states, + attention_mask=causal_mask, + position_ids=position_ids, + past_key_value=past_key_values, + output_attentions=output_attentions, + use_cache=use_cache, + cache_position=cache_position, + position_embeddings=position_embeddings, + **flash_attn_kwargs, + ) + + hidden_states = layer_outputs[0] + + 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,) + + output = BaseModelOutputWithPast( + last_hidden_state=hidden_states, + past_key_values=past_key_values if use_cache else None, + hidden_states=all_hidden_states, + attentions=all_self_attns, + ) + return output if return_dict else output.to_tuple() + + def _update_causal_mask( + self, + attention_mask: torch.Tensor, + input_tensor: torch.Tensor, + cache_position: torch.Tensor, + past_key_values: Cache, + output_attentions: bool, + ): + if self.config._attn_implementation == "flash_attention_2": + if attention_mask is not None and past_key_values is not None: + is_padding_right = attention_mask[:, -1].sum().item() != input_tensor.size()[0] + 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 attention_mask is not None and 0.0 in attention_mask: + return attention_mask + return None + + # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in + # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail + # to infer the attention mask. + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + using_static_cache = isinstance(past_key_values, StaticCache) + using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache) + + # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward + if ( + self.config._attn_implementation == "sdpa" + and not (using_static_cache or using_sliding_window_cache) + and not output_attentions + ): + if AttentionMaskConverter._ignore_causal_mask_sdpa( + attention_mask, + inputs_embeds=input_tensor, + past_key_values_length=past_seen_tokens, + sliding_window=self.config.sliding_window, + is_training=self.training, + ): + return None + + dtype, device = input_tensor.dtype, input_tensor.device + min_dtype = torch.finfo(dtype).min + sequence_length = input_tensor.shape[1] + # SlidingWindowCache or StaticCache + if using_sliding_window_cache or using_static_cache: + target_length = past_key_values.get_max_cache_shape() + # DynamicCache or no cache + else: + target_length = ( + attention_mask.shape[-1] + if isinstance(attention_mask, torch.Tensor) + else past_seen_tokens + sequence_length + 1 + ) + + # In case the provided `attention` mask is 2D, we generate a causal mask here (4D). + causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position( + attention_mask, + sequence_length=sequence_length, + target_length=target_length, + dtype=dtype, + device=device, + cache_position=cache_position, + batch_size=input_tensor.shape[0], + config=self.config, + past_key_values=past_key_values, + ) + + if ( + self.config._attn_implementation == "sdpa" + and attention_mask is not None + and attention_mask.device.type in ["cuda", "xpu"] + and not output_attentions + ): + # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when + # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. + # Details: https://github.com/pytorch/pytorch/issues/110213 + causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) + + return causal_mask + + @staticmethod + def _prepare_4d_causal_attention_mask_with_cache_position( + attention_mask: torch.Tensor, + sequence_length: int, + target_length: int, + dtype: torch.dtype, + device: torch.device, + cache_position: torch.Tensor, + batch_size: int, + config: Phi3Config, + past_key_values: Cache, + ): + """ + Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape + `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing. + + Args: + attention_mask (`torch.Tensor`): + A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`. + sequence_length (`int`): + The sequence length being processed. + target_length (`int`): + The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet. + dtype (`torch.dtype`): + The dtype to use for the 4D attention mask. + device (`torch.device`): + The device to plcae the 4D attention mask on. + cache_position (`torch.Tensor`): + Indices depicting the position of the input sequence tokens in the sequence. + batch_size (`torch.Tensor`): + Batch size. + config (`Phi3Config`): + The model's configuration class + past_key_values (`Cache`): + The cache class that is being used currently to generate + """ + if attention_mask is not None and attention_mask.dim() == 4: + # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing. + causal_mask = attention_mask + else: + min_dtype = torch.finfo(dtype).min + causal_mask = torch.full( + (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device + ) + diagonal_attend_mask = torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) + if config.sliding_window is not None: + # if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also + # the check is needed to verify is current checkpoint was trained with sliding window or not + if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length: + sliding_attend_mask = torch.arange(target_length, device=device) <= ( + cache_position.reshape(-1, 1) - config.sliding_window + ) + diagonal_attend_mask.bitwise_or_(sliding_attend_mask) + causal_mask *= diagonal_attend_mask + causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1) + if attention_mask is not None: + causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit + if attention_mask.shape[-1] > target_length: + attention_mask = attention_mask[:, :target_length] + mask_length = attention_mask.shape[-1] + padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to( + causal_mask.device + ) + padding_mask = padding_mask == 0 + causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( + padding_mask, min_dtype + ) + return causal_mask + + +class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ... + + +class Phi3ForCausalLM(Phi3PreTrainedModel, GenerationMixin): + _tied_weights_keys = ["lm_head.weight"] + _tp_plan = {"lm_head": "colwise_rep"} + _pp_plan = {"lm_head": (["hidden_states"], ["logits"])} + + 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() + + def get_input_embeddings(self): + return self.model.embed_tokens + + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + def get_output_embeddings(self): + return self.lm_head + + def set_output_embeddings(self, new_embeddings): + self.lm_head = new_embeddings + + def set_decoder(self, decoder): + self.model = decoder + + def get_decoder(self): + return self.model + + @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep") + @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[Union[Cache, 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, + cache_position: Optional[torch.LongTensor] = None, + logits_to_keep: Union[int, torch.Tensor] = 0, + **kwargs: Unpack[KwargsForCausalLM], + ) -> 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]`. + + logits_to_keep (`int` or `torch.Tensor`, *optional*): + If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all + `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that + token can save memory, which becomes pretty significant for long sequences or large vocabulary size. + If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension. + This is useful when using packed tensor format (single dimension for batch and sequence length). + + Returns: + + Example: + + ```python + >>> from transformers import AutoTokenizer, Phi3ForCausalLM + + >>> model = Phi3ForCausalLM.from_pretrained("meta-phi3/Phi3-2-7b-hf") + >>> tokenizer = AutoTokenizer.from_pretrained("meta-phi3/Phi3-2-7b-hf") + + >>> prompt = "Hey, are you conscious? Can you talk to me?" + >>> 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] + "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you." + ```""" + 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, + cache_position=cache_position, + **kwargs, + ) + + hidden_states = outputs[0] + # Only compute necessary logits, and do not upcast them to float if we are not computing the loss + slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep + logits = self.lm_head(hidden_states[:, slice_indices, :]) + + loss = None + if labels is not None: + loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs) + + if not return_dict: + output = (logits,) + outputs[1:] + return (loss,) + output if loss is not None else output + + return CausalLMOutputWithPast( + loss=loss, + logits=logits, + past_key_values=outputs.past_key_values, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) + + def prepare_inputs_for_generation( + self, + input_ids, + past_key_values=None, + attention_mask=None, + inputs_embeds=None, + cache_position=None, + position_ids=None, + use_cache=True, + logits_to_keep=None, + **kwargs, + ): + # Overwritten -- this model may need to switch between short and long rope, invalidating the cache in the + # process + + # When the first time input length reached long and short factor switching point, enforce re-compute cache + # It will cause downside of slower at this single token position, however, better than current failure. + if ( + past_key_values + and self.config.rope_scaling + and input_ids.shape[1] >= self.config.original_max_position_embeddings + 1 + ): + past_length = cache_position[0] + if past_length <= self.config.original_max_position_embeddings: + past_key_values = None + + model_inputs = super().prepare_inputs_for_generation( + input_ids=input_ids, + past_key_values=past_key_values, + attention_mask=attention_mask, + inputs_embeds=inputs_embeds, + cache_position=cache_position, + position_ids=position_ids, + use_cache=use_cache, + logits_to_keep=logits_to_keep, + **kwargs, + ) + return model_inputs + + +@add_start_docstrings( + """ + The Phi3 Model transformer 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, +) +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: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Union[Cache, 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 + + transformer_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 = transformer_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: + last_non_pad_token = -1 + elif input_ids is not None: + # To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id + non_pad_mask = (input_ids != self.config.pad_token_id).to(logits.device, torch.int32) + token_indices = torch.arange(input_ids.shape[-1], device=logits.device) + last_non_pad_token = (token_indices * non_pad_mask).argmax(-1) + else: + last_non_pad_token = -1 + logger.warning_once( + f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be " + "unexpected if using padding tokens in conjunction with `inputs_embeds.`" + ) + + pooled_logits = logits[torch.arange(batch_size, device=logits.device), last_non_pad_token] + + loss = None + if labels is not None: + loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config) + + if not return_dict: + output = (pooled_logits,) + transformer_outputs[1:] + return ((loss,) + output) if loss is not None else output + + return SequenceClassifierOutputWithPast( + loss=loss, + logits=pooled_logits, + past_key_values=transformer_outputs.past_key_values, + hidden_states=transformer_outputs.hidden_states, + attentions=transformer_outputs.attentions, + ) + + +@add_start_docstrings( + """ + The Phi3 Model transformer 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, +) +class Phi3ForTokenClassification(Phi3PreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.model = Phi3Model(config) + if getattr(config, "classifier_dropout", None) is not None: + classifier_dropout = config.classifier_dropout + elif getattr(config, "hidden_dropout", None) is not None: + classifier_dropout = config.hidden_dropout + else: + classifier_dropout = 0.1 + self.dropout = nn.Dropout(classifier_dropout) + self.score = nn.Linear(config.hidden_size, config.num_labels) + + # 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) + @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, + 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, 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 + + 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, + ) + sequence_output = outputs[0] + sequence_output = self.dropout(sequence_output) + logits = self.score(sequence_output) + + loss = None + if labels is not None: + loss = self.loss_function(logits, labels, self.config) + + if not return_dict: + output = (logits,) + outputs[2:] + return ((loss,) + output) if loss is not None else output + + return TokenClassifierOutput( + loss=loss, + logits=logits, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) diff --git a/sample_finetune.py b/sample_finetune.py new file mode 100644 index 0000000..499f5f0 --- /dev/null +++ b/sample_finetune.py @@ -0,0 +1,214 @@ +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-4-Mini-Instruct model. 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=1.3.0 + pip3 install -i https://pypi.org/simple/ bitsandbytes + pip3 install peft==0.14.0 + pip3 install transformers==4.48.1 + pip3 install 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-4-mini-instruct" +model_kwargs = dict( + use_cache=False, + trust_remote_code=True, + attn_implementation="flash_attention_2", # loading the model with flash-attention 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 + + +train_dataset, test_dataset = load_dataset("HuggingFaceH4/ultrachat_200k", split=["train_sft", "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) diff --git a/tokenizer.json b/tokenizer.json new file mode 100644 index 0000000..8655fee --- /dev/null +++ b/tokenizer.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ea8bdf68c3e7549a3fb4342523288ce628f6ab56a618f9a4dfb234a0b4d46a8 +size 15524476 diff --git a/tokenizer_config.json b/tokenizer_config.json new file mode 100644 index 0000000..eb46300 --- /dev/null +++ b/tokenizer_config.json @@ -0,0 +1,12 @@ +{ + "add_prefix_space": false, + "backend": "tokenizers", + "bos_token": "<|endoftext|>", + "clean_up_tokenization_spaces": false, + "eos_token": "<|endoftext|>", + "is_local": true, + "model_max_length": 131072, + "pad_token": "<|endoftext|>", + "tokenizer_class": "TokenizersBackend", + "unk_token": "<|endoftext|>" +}