microsoft/Phi-tiny-MoE-instruct
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

初始化项目，由ModelHub XC社区提供模型

Browse Source 
Model: microsoft/Phi-tiny-MoE-instruct
Source: Original Platform
This commit is contained in:
ModelHub XC
2026-06-16 16:20:13 +08:00
commit 5d613f4da4
20 changed files with 4891 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
.gitattributes vendored Normal file
View File

@@ -0,0 +1,49 @@
*.7z filter=lfs diff=lfs merge=lfs -text
*.arrow filter=lfs diff=lfs merge=lfs -text
*.bin filter=lfs diff=lfs merge=lfs -text
*.bin.* filter=lfs diff=lfs merge=lfs -text
*.bz2 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
*.model filter=lfs diff=lfs merge=lfs -text
*.msgpack 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
*.pt filter=lfs diff=lfs merge=lfs -text
*.pth filter=lfs diff=lfs merge=lfs -text
*.rar filter=lfs diff=lfs merge=lfs -text
saved_model/**/* 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
*.xz filter=lfs diff=lfs merge=lfs -text
*.zip filter=lfs diff=lfs merge=lfs -text
*.zstandard filter=lfs diff=lfs merge=lfs -text
*.tfevents* filter=lfs diff=lfs merge=lfs -text
*.db* filter=lfs diff=lfs merge=lfs -text
*.ark* filter=lfs diff=lfs merge=lfs -text
**/*ckpt*data* filter=lfs diff=lfs merge=lfs -text
**/*ckpt*.meta filter=lfs diff=lfs merge=lfs -text
**/*ckpt*.index filter=lfs diff=lfs merge=lfs -text
*.safetensors filter=lfs diff=lfs merge=lfs -text
*.ckpt filter=lfs diff=lfs merge=lfs -text
*.gguf* filter=lfs diff=lfs merge=lfs -text
*.ggml filter=lfs diff=lfs merge=lfs -text
*.llamafile* filter=lfs diff=lfs merge=lfs -text
*.pt2 filter=lfs diff=lfs merge=lfs -text
*.mlmodel filter=lfs diff=lfs merge=lfs -text
*.npy filter=lfs diff=lfs merge=lfs -text
*.npz filter=lfs diff=lfs merge=lfs -text
*.pickle filter=lfs diff=lfs merge=lfs -text
*.pkl filter=lfs diff=lfs merge=lfs -text
*.tar filter=lfs diff=lfs merge=lfs -text
*.wasm 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

9
CODE_OF_CONDUCT.md Normal file
View File

@@ -0,0 +1,9 @@
# Microsoft Open Source Code of Conduct
This project has adopted the [Microsoft Open Source Code of Conduct](https://opensource.microsoft.com/codeofconduct/).
Resources:
- [Microsoft Open Source Code of Conduct](https://opensource.microsoft.com/codeofconduct/)
- [Microsoft Code of Conduct FAQ](https://opensource.microsoft.com/codeofconduct/faq/)
- Contact [opencode@microsoft.com](mailto:opencode@microsoft.com) with questions or concerns

22
LICENSE Normal file
View File

@@ -0,0 +1,22 @@
Microsoft.
Copyright (c) Microsoft Corporation.
MIT License
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED *AS IS*, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

38
NOTICE.md Normal file
View File

@@ -0,0 +1,38 @@
NOTICES AND INFORMATION
Do Not Translate or Localize
This software incorporates material from third parties.
**Component.** https://github.com/Dao-AILab/flash-attention
**Open Source License/Copyright Notice.**
BSD 3-Clause License
Copyright (c) 2022, the respective contributors, as shown by the AUTHORS file.
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
* Redistributions of source code must retain the above copyright notice, this
list of conditions and the following disclaimer.
* Redistributions in binary form must reproduce the above copyright notice,
this list of conditions and the following disclaimer in the documentation
and/or other materials provided with the distribution.
* Neither the name of the copyright holder nor the names of its
contributors may be used to endorse or promote products derived from
this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

172
README.md Normal file
View File

@@ -0,0 +1,172 @@
---
language:
- en
license: mit
pipeline_tag: text-generation
context_length:
- 4k
library_name: transformers
---
## Model Summary
Phi-tiny-MoE is a lightweight Mixture of Experts (MoE) model with 3.8B total parameters and 1.1B activated parameters. It is compressed and distilled from the base model shared by [Phi-3.5-MoE](https://huggingface.co/microsoft/Phi-3.5-MoE-instruct) and [GRIN-MoE](https://huggingface.co/microsoft/GRIN-MoE) using the [SlimMoE](https://arxiv.org/pdf/2506.18349) approach, then post-trained via supervised fine-tuning and direct preference optimization for instruction following and safety. The model is trained on Phi-3 synthetic data and filtered public documents, with a focus on high-quality, reasoning-dense content. It is part of the SlimMoE series, which includes a larger variant, [Phi-mini-MoE](https://huggingface.co/microsoft/Phi-mini-MoE-instruct), with 7.6B total and 2.4B activated parameters.
References: <br>
📖 [SlimMoE Paper](https://arxiv.org/pdf/2506.18349) <br>
📖 [Phi-3 Technical Report](https://arxiv.org/abs/2404.14219) <br>
📖 [GRIN-MoE](https://arxiv.org/abs/2409.12136) <br>
## Intended Uses
### Primary Use Cases
The model is intended for commercial and research use in English. The model provides uses for general purpose AI systems and applications which require memory/compute constrained environments and latency bound scenarios.
### Use Case Considerations
Our models are not specifically designed or evaluated for all downstream purposes. Developers should consider common limitations of language models as they select use cases, and evaluate and mitigate for accuracy, safety, and fariness before using within a specific downstream use case, particularly for high risk scenarios. Developers should be aware of and adhere to applicable laws or regulations (including privacy, trade compliance laws, etc.) that are relevant to their use case.
***Nothing contained in this Model Card should be interpreted as or deemed a restriction or modification to the license the model is released under.***
## Usage
### Input Formats
Given the nature of the training data, the Phi-tiny-MoE model is best suited for prompts using the chat format as follows:
```
<|system|>
You are a helpful assistant.<|end|>
<|user|>
How to explain Internet for a medieval knight?<|end|>
<|assistant|>
```
### Loading the model locally
After obtaining the Phi-tiny-MoE model checkpoints, users can use this sample code for inference.
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
torch.random.manual_seed(0)
model = AutoModelForCausalLM.from_pretrained(
"microsoft/Phi-tiny-MoE-instruct",
device_map="cuda",
torch_dtype="auto",
trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-tiny-MoE-instruct")
messages = [
{"role": "system", "content": "You are a helpful AI assistant."},
{"role": "user", "content": "Can you provide ways to eat combinations of bananas and dragonfruits?"},
{"role": "assistant", "content": "Sure! Here are some ways to eat bananas and dragonfruits together: 1. Banana and dragonfruit smoothie: Blend bananas and dragonfruits together with some milk and honey. 2. Banana and dragonfruit salad: Mix sliced bananas and dragonfruits together with some lemon juice and honey."},
{"role": "user", "content": "What about solving an 2x + 3 = 7 equation?"},
]
pipe = pipeline(
"text-generation",
model=model,
tokenizer=tokenizer,
)
generation_args = {
"max_new_tokens": 500,
"return_full_text": False,
"temperature": 0.0,
"do_sample": False,
}
output = pipe(messages, **generation_args)
print(output[0]['generated_text'])
```
## Benchmarks
To understand the capabilities, we compare Phi-tiny-MoE with a set of models over a variety of benchmarks using [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness). Detailed evaluation settings can be found in the SlimMoE paper.
| Model | # Total param | # Act. param | MMLU | MMLU pro | BBH | Arc-C (chat) | Human-eval | GSM8K | MT-bench |
|----------------------|---------------|--------------|-------|----------|-------|---------------|-------------|--------|----------|
| **MoE Models** |||||||||||
| Phi 3.5-MoE | 42B | 6.6B | 78.36 | 59.38 | 63.93 | 91.38 | 81.70 | 87.87 | 8.34 |
| Qwen 1.5 MoE | 14B | 2.7B | 60.73 | 26.49 | 42.65 | 67.24 | 46.30 | 53.07 | 6.55 |
| DeepSeek V2 Lite | 16B | 2.4B | 56.69 | 17.89 | 36.30 | 61.09 | 54.40 | 63.23 | 6.82 |
| OL-MoE | 7B | 1.3B | 54.27 | 20.87 | 38.00 | 55.63 | 37.80 | 71.49 | 6.60 |
| Granite 3.0 MoE | 3.4B | 0.8B | 50.06 | 4.82 | 39.65 | 56.06 | 51.80 | 60.12 | 6.91 |
| **Dense Models** |||||||||||
| LLaMA 3.1 8B | 8B | 8B | 68.71 | 45.28 | 50.86 | 82.42 | 69.50 | 84.84 | 8.03 |
| Qwen 2.5 7B | 7.6B | 7.6B | 73.47 | 56.24 | 53.74 | 88.82 | 81.70 | 84.84 | 8.34 |
| Phi 3 small | 7.4B | 7.4B | 75.35 | 52.06 | 62.07 | 84.30 | 70.10 | 84.84 | 8.03 |
| Gemma 3 4B | 4B | 4B | 59.49 | 40.13 | 49.45 | 75.85 | 67.10 | 78.92 | 8.28 |
| Phi 3 mini | 3.8B | 3.8B | 69.94 | 45.65 | 54.94 | 85.58 | 72.60 | 84.61 | 7.46 |
| LLaMA 3.2 3B | 3.2B | 3.2B | 61.73 | 36.70 | 45.46 | 75.77 | 52.40 | 77.41 | 7.46 |
| Qwen 2.5 3B | 3B | 3B | 65.06 | 41.00 | 46.61 | 80.20 | 73.80 | 76.57 | 7.60 |
| Gemma 3 1B | 1B | 1B | 40.80 | 14.70 | 34.80 | 37.46 | 41.50 | 41.77 | 6.67 |
| LLaMA 3.2 1B | 1B | 1B | 46.30 | 18.67 | 35.18 | 49.91 | 35.40 | 44.96 | 5.23 |
| **Our (SlimMoE) Models** |||||||||||
| Phi-mini-MoE | 7.6B | 2.4B | 70.68 | 49.68 | 55.27 | 84.91 | 73.80 | 84.89 | 7.59 |
| Phi-tiny-MoE | 3.8B | 1.1B | 60.83 | 36.34 | 45.58 | 76.37 | 58.50 | 78.47 | 7.05 |
## Training
### Model
**Architecture:** Phi-tiny-MoE has 3.8 total parameters with **1.1B active parameters**. The model is a mixture-of-expert decoder-only Transformer model using the tokenizer with vocabulary size of 32,064.<br>
**Inputs:** Text. It is best suited for prompts using chat format.<br>
**Context length:** 4k tokens<br>
**GPUs:** 64 A100-80G<br>
**Training time:** 11 days<br>
**Training data:** 400B tokens<br>
**Outputs:** Generated text in response to the input<br>
**Dates:** Trained between September 2024 and March 2025<br>
**Status:** This is a static model trained on an offline dataset with cutoff date October 2023 for publicly available data.<br>
### Training Datasets
Our training data is a subset with 400B tokens of Phi-3 datasets, which includes a wide variety of sources and is a combination of
1) publicly available documents filtered rigorously for quality, selected high-quality educational data, and code;
2) newly created synthetic, “textbook-like” data for the purpose of teaching math, coding, common sense reasoning, general knowledge of the world (science, daily activities, theory of mind, etc.);
3) high quality chat format supervised data covering various topics to reflect human preferences on different aspects such as instruct-following, truthfulness, honesty and helpfulness.
More details about data can be found in the [Phi-3 Technical Report](https://arxiv.org/pdf/2404.14219).
## Responsible AI Considerations
Like other language models, Phi-tiny-MoE can potentially behave in ways that are unfair, unreliable, or offensive. Some of the limiting behaviors to be aware of include:
* Quality of Service: The models are trained primarily on English text and some additional multilingual text. Languages other than English will experience worse performance as well as performance disparities across non-English. English language varieties with less representation in the training data might experience worse performance than standard American English.
* Representation of Harms & Perpetuation of Stereotypes: These models can over- or under-represent groups of people, erase representation of some groups, or reinforce demeaning or negative stereotypes. Despite safety post-training, these limitations may still be present due to differing levels of representation of different groups, cultural contexts, or prevalence of examples of negative stereotypes in training data that reflect real-world patterns and societal biases.
* Inappropriate or Offensive Content: These models may produce other types of inappropriate or offensive content, which may make it inappropriate to deploy for sensitive contexts without additional mitigations that are specific to the use case.
* Information Reliability: Language models can generate nonsensical content or fabricate content that might sound reasonable but is inaccurate or outdated.
* Limited Scope for Code: Majority of Phi-3 training data is based in Python and use common packages such as "typing, math, random, collections, datetime, itertools". If the model generates Python scripts that utilize other packages or scripts in other languages, we strongly recommend users manually verify all API uses.
* The High ECI: The model has an elevated defect rate when responding to election-critical queries, which may result in incorrect or unauthoritative election critical information being presented. Users should verify information related to elections with the election authority in their region.
* Long Conversation: Phi-3 models, like other models, can in some cases generate responses that are repetitive, unhelpful, or inconsistent in very long chat sessions in both English and non-English languages. Developers are encouraged to place appropriate mitigations, like limiting conversation turns to account for the possible conversational drift
Developers should apply responsible AI best practices, including mapping, measuring, and mitigating risks associated with their specific use case and cultural, linguistic context. Important areas for consideration include:
* Allocation: Models may not be suitable for scenarios that could have consequential impact on legal status or the allocation of resources or life opportunities (ex: housing, employment, credit, etc.) without further assessments and additional debiasing techniques.
* High-Risk Scenarios: Developers should assess the suitability of using models in high-risk scenarios where unfair, unreliable or offensive outputs might be extremely costly or lead to harm. This includes providing advice in sensitive or expert domains where accuracy and reliability are critical (ex: legal or health advice). Additional safeguards should be implemented at the application level according to the deployment context.
* Misinformation: Models may produce inaccurate information. Developers should follow transparency best practices and inform end-users they are interacting with an AI system. At the application level, developers can build feedback mechanisms and pipelines to ground responses in use-case specific, contextual information, a technique known as Retrieval Augmented Generation (RAG).
* Generation of Harmful Content: Developers should assess outputs for their context and use available safety classifiers or custom solutions appropriate for their use case.
* Misuse: Other forms of misuse such as fraud, spam, or malware production may be possible, and developers should ensure that their applications do not violate applicable laws and regulations.
## Software
* [PyTorch](https://github.com/pytorch/pytorch)
* [Transformers](https://github.com/huggingface/transformers)
* [Flash-Attention](https://github.com/HazyResearch/flash-attention)
## Hardware
Note that by default, the Phi-tiny-MoE model uses flash attention, which requires certain types of GPU hardware to run. We have tested on the following GPU types:
* NVIDIA A100
* NVIDIA A6000
* NVIDIA H100
## License
The model is licensed under the [MIT license](./LICENSE).
## Trademarks
This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow [Microsoft’s Trademark & Brand Guidelines](https://www.microsoft.com/en-us/legal/intellectualproperty/trademarks). Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party’s policies.
## Data Summary
https://huggingface.co/microsoft/Phi-tiny-MoE-instruct/blob/main/data_summary_card.md

41
SECURITY.md Normal file
View File

@@ -0,0 +1,41 @@
<!-- BEGIN MICROSOFT SECURITY.MD V0.0.9 BLOCK -->
## Security
Microsoft takes the security of our software products and services seriously, which includes all source code repositories managed through our GitHub organizations, which include [Microsoft](https://github.com/Microsoft), [Azure](https://github.com/Azure), [DotNet](https://github.com/dotnet), [AspNet](https://github.com/aspnet) and [Xamarin](https://github.com/xamarin).
If you believe you have found a security vulnerability in any Microsoft-owned repository that meets [Microsoft's definition of a security vulnerability](https://aka.ms/security.md/definition), please report it to us as described below.
## Reporting Security Issues
**Please do not report security vulnerabilities through public GitHub issues.**
Instead, please report them to the Microsoft Security Response Center (MSRC) at [https://msrc.microsoft.com/create-report](https://aka.ms/security.md/msrc/create-report).
If you prefer to submit without logging in, send email to [secure@microsoft.com](mailto:secure@microsoft.com). If possible, encrypt your message with our PGP key; please download it from the [Microsoft Security Response Center PGP Key page](https://aka.ms/security.md/msrc/pgp).
You should receive a response within 24 hours. If for some reason you do not, please follow up via email to ensure we received your original message. Additional information can be found at [microsoft.com/msrc](https://www.microsoft.com/msrc).
Please include the requested information listed below (as much as you can provide) to help us better understand the nature and scope of the possible issue:
* Type of issue (e.g. buffer overflow, SQL injection, cross-site scripting, etc.)
* Full paths of source file(s) related to the manifestation of the issue
* The location of the affected source code (tag/branch/commit or direct URL)
* Any special configuration required to reproduce the issue
* Step-by-step instructions to reproduce the issue
* Proof-of-concept or exploit code (if possible)
* Impact of the issue, including how an attacker might exploit the issue
This information will help us triage your report more quickly.
If you are reporting for a bug bounty, more complete reports can contribute to a higher bounty award. Please visit our [Microsoft Bug Bounty Program](https://aka.ms/security.md/msrc/bounty) page for more details about our active programs.
## Preferred Languages
We prefer all communications to be in English.
## Policy
Microsoft follows the principle of [Coordinated Vulnerability Disclosure](https://aka.ms/security.md/cvd).
<!-- END MICROSOFT SECURITY.MD BLOCK -->

13
added_tokens.json Normal file
View File

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

41
config.json Normal file
View File

@@ -0,0 +1,41 @@
{
"_name_or_path": "Phi-tiny-MoE",
"architectures": [
"PhiMoEForCausalLM"
],
"attention_bias": true,
"attention_dropout": 0.0,
"auto_map": {
"AutoConfig": "configuration_slimmoe.PhiMoEConfig",
"AutoModelForCausalLM": "modeling_slimmoe.PhiMoEForCausalLM"
},
"bos_token_id": 1,
"eos_token_id": 32000,
"expert_dropout": 0.0,
"head_dim": 128,
"hidden_act": "silu",
"hidden_dropout": 0.0,
"hidden_size": 4096,
"initializer_range": 0.02,
"input_jitter_noise": 0.01,
"intermediate_size": 448,
"lm_head_bias": true,
"max_position_embeddings": 4096,
"model_type": "phimoe",
"num_attention_heads": 16,
"num_experts_per_tok": 2,
"num_hidden_layers": 32,
"num_key_value_heads": 4,
"num_local_experts": 16,
"output_router_logits": false,
"rms_norm_eps": 1e-05,
"rope_theta": 10000.0,
"router_aux_loss_coef": 0.0,
"router_jitter_noise": 0.01,
"sliding_window": 2047,
"tie_word_embeddings": false,
"torch_dtype": "bfloat16",
"transformers_version": "4.43.3",
"use_cache": true,
"vocab_size": 32064
}

1
configuration.json Normal file
View File

@@ -0,0 +1 @@
{"framework": "pytorch", "task": "text-generation", "allow_remote": true}

244
configuration_slimmoe.py Normal file
View File

@@ -0,0 +1,244 @@
# 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-MoE model."""
from transformers.configuration_utils import PretrainedConfig
from transformers.utils import logging
logger = logging.get_logger(__name__)
class PhiMoEConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`PhiMoEModel`]. It is used to instantiate a Phi-MoE
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-3.5-MoE-instruct](https://huggingface.co/microsoft/Phi-3.5-MoE-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 PhiMoE model. Defines the number of different tokens that can be represented by the
`inputs_ids` passed when calling [`PhiMoEModel`]
hidden_size (`int`, *optional*, defaults to 4096):
Dimension of the hidden representations.
intermediate_size (`int`, *optional*, defaults to 6400):
Dimension of the MLP representations.
num_hidden_layers (`int`, *optional*, defaults to 32):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 32):
Number of attention heads for each attention layer in the Transformer encoder.
num_key_value_heads (`int`, *optional*, defaults to 8):
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 `8`.
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*32`):
The maximum sequence length that this model might ever be used with. Mixtral's sliding window attention
allows sequence of up to 4096*32 tokens.
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 used by the rms normalization layers.
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`.
pad_token_id (`int`, *optional*):
The id of the padding token.
bos_token_id (`int`, *optional*, defaults to 1):
The id of the "beginning-of-sequence" token.
eos_token_id (`int`, *optional*, defaults to 2):
The id of the "end-of-sequence" token.
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
Whether the model's input and output word embeddings should be tied.
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`, `long_factor`, `short_mscale`, `long_mscale` and
`original_max_position_embeddings`. The `type` must be `longrope`, the `short_mscale` and `long_scale` must
be numbers, the `short_factor` and `long_factor` must be lists of numbers with the same length as half of
the attention head size and the `original_max_position_embeddings` must be an integer.
sliding_window (`int`, *optional*):
Sliding window attention window size. If not specified, will default to `262144`.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
num_experts_per_tok (`int`, *optional*, defaults to 2):
The number of experts to root per-token, can be also interpreted as the `top-p` routing
parameter
num_local_experts (`int`, *optional*, defaults to 16):
Number of experts per Sparse MLP layer.
output_router_logits (`bool`, *optional*, defaults to `False`):
Whether or not the router logits should be returned by the model. Enabeling this will also
allow the model to output the auxiliary loss. See [here]() for more details
router_aux_loss_coef (`float`, *optional*, defaults to 0.0):
The aux loss factor for the total loss.
router_jitter_noise (`float`, *optional*, defaults to 0.01):
Amount of noise to add to the router.
```python
>>> from transformers import PhiMoEModel, PhiMoEConfig
>>> # Initializing a Phi-3 style configuration
>>> configuration = PhiMoEConfig.from_pretrained("microsoft/Phi-3.5-MoE-instruct")
>>> # Initializing a model from the configuration
>>> model = PhiMoEModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "phimoe"
keys_to_ignore_at_inference = ["past_key_values"]
def __init__(
self,
vocab_size=32064,
hidden_size=4096,
intermediate_size=6400,
num_hidden_layers=32,
num_attention_heads=32,
num_key_value_heads=8,
head_dim=None, # added to control head dimension
hidden_act="silu",
max_position_embeddings=4096 * 32,
initializer_range=0.02,
rms_norm_eps=1e-5,
use_cache=True,
pad_token_id=None,
bos_token_id=1,
eos_token_id=2,
tie_word_embeddings=False,
rope_theta=1e6,
rope_scaling=None,
sliding_window=None,
attention_dropout=0.0,
num_experts_per_tok=2,
num_local_experts=16,
output_router_logits=False,
router_aux_loss_coef=0.001,
router_jitter_noise=0.01,
input_jitter_noise=0.0,
attention_bias = False,
lm_head_bias = False,
**kwargs,
):
self.vocab_size = vocab_size
self.max_position_embeddings = max_position_embeddings
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.sliding_window = sliding_window
self.attention_bias = attention_bias
self.lm_head_bias = lm_head_bias
# for backward compatibility
if num_key_value_heads is None:
num_key_value_heads = num_attention_heads
if head_dim is None:
head_dim = hidden_size // num_attention_heads
self.head_dim = head_dim
self.num_key_value_heads = num_key_value_heads
self.hidden_act = hidden_act
self.initializer_range = initializer_range
self.rms_norm_eps = rms_norm_eps
self.use_cache = use_cache
self.rope_theta = rope_theta
self.attention_dropout = attention_dropout
self.num_experts_per_tok = num_experts_per_tok
self.num_local_experts = num_local_experts
self.output_router_logits = output_router_logits
self.router_aux_loss_coef = router_aux_loss_coef
self.router_jitter_noise = router_jitter_noise
self.input_jitter_noise = input_jitter_noise
self.rope_scaling = rope_scaling
self._rope_scaling_validation()
super().__init__(
pad_token_id=pad_token_id,
bos_token_id=bos_token_id,
eos_token_id=eos_token_id,
tie_word_embeddings=tie_word_embeddings,
**kwargs,
)
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) != 6:
raise ValueError(
"`rope_scaling` must be a dictionary with three fields, `type`, `short_factor`, `long_factor`, "
f"`short_mscale`, `long_mscale` and `original_max_position_embeddings`, 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)
rope_scaling_short_mscale = self.rope_scaling.get("short_mscale", None)
rope_scaling_long_mscale = self.rope_scaling.get("long_mscale", None)
original_max_position_embeddings = self.rope_scaling.get("original_max_position_embeddings", 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)}"
)
if not isinstance(rope_scaling_short_mscale, (int, float)):
raise ValueError(
f"`rope_scaling`'s short_mscale field must be a number, got {rope_scaling_short_mscale}"
)
if not isinstance(rope_scaling_long_mscale, (int, float)):
raise ValueError(
f"`rope_scaling`'s long_mscale field must be a number, got {rope_scaling_long_mscale}"
)
if not isinstance(original_max_position_embeddings, int):
raise ValueError(
f"`rope_scaling`'s original_max_position_embeddings field must be an integer, got {original_max_position_embeddings}"
)

147
data_summary_card.md Normal file
View File

@@ -0,0 +1,147 @@
# Data Summary for microsoft_GRIN-MoE, phi-tiny-MoE-instruct, phi-mini-MoE-instruct
## 1. General information
**1.0.1 Version of the Summary:** 1.0
**1.0.2 Last update:** 10-Dec-2025
## 1.1 Model Developer Identification
**1.1.1 Model Developer name and contact details:** Microsoft Corporation at One Microsoft Way, Redmond, WA 98052. Tel: 425-882-8080
## 1.2 Model Identification
**1.2.1 Versioned model name(s):** GRIN MoE 16x3.8B
**1.2.2 Model release date:** 18-Sept-2024
## 1.3 Overall training data size and characteristics
### 1.3.1 Size of dataset and characteristics
**1.3.1.A Text training data size:** 1 billion to 10 trillion tokens
**1.3.1.B Text training data content:** Our training data includes a wide variety of sources and is a combination of publicly available documents selected for quality, educational data, and code; newly created synthetic, “textbook-like” data for the purpose of teaching math, coding, common sense reasoning, general knowledge of the world (science, daily activities, theory of mind, etc.); chat format supervised data covering various topics to reflect preferences on different aspects such as instruct-following, truthfulness, honesty and helpfulness.
**1.3.1.C Image training data size:** Not applicable. Images are not part of the training data
**1.3.1.D Image training data content:** Not applicable
**1.3.1.E Audio training data size:** Not applicable. Audio data is not part of the training data
**1.3.1.F Audio training data content:** Not applicable
**1.3.1.G Video training data size:** Not applicable. Video data is not part of the training data
**1.3.1.H Video training data content:** Not applicable
**1.3.1.I Other training data size:** Not applicable
**1.3.1.J Other training data content:** Not applicable
**1.3.2 Latest date of data acquisition/collection for model training:** 03-Jun-2024
**1.3.3 Is data collection ongoing to update the model with new data collection after deployment?** No
**1.3.4 Date the training dataset was first used to train the model:** 01-Apr-2024
**1.3.5 Rationale or purpose of data selection:** Datasets were selected to maximize coverage of reasoning, math, coding, and conversational domains, supporting strong performance on benchmarks like MMLU, HumanEval, MBPP, and MATH. The corpus combines publicly available data, curated educational content, and synthetic textbook-like data to teach math, coding, and general knowledge.
## 2. List of data sources
### 2.1 Publicly available datasets
**2.1.1 Have you used publicly available datasets to train the model?** Yes
## 2.2 Private non-publicly available datasets obtained from third parties
### 2.2.1 Datasets commercially licensed by rights holders or their representatives
**2.2.1.A Have you concluded transactional commercial licensing agreement(s) with rights holder(s) or with their representatives?** No
### 2.2.2 Private datasets obtained from other third-parties
**2.2.2.A Have you obtained private datasets from third parties that are not licensed as described in Section 2.2.1, such as data obtained from providers of private databases, or data intermediaries?** No
## 2.3 Personal Information
**2.3.1 Was personal data used to train the model?** Microsoft follows all relevant laws and regulations pertaining to personal information
## 2.4 Synthetic data
**2.4.1 Was any synthetic AI-generated data used to train the model?** Yes
## 3. Data processing aspects
### 3.1 Respect of reservation of rights from text and data mining exception or limitation
**3.1.1 Does this dataset include any data protected by copyright, trademark, or patent?** Microsoft follows all required regulations and laws for processing data protected by copyright, trademark, or patent
## 3.2 Other information
**3.2.1 Does the dataset include information about consumer groups without revealing individual consumer identities?** Microsoft follows all required regulations and laws for protecting consumer identities
**3.2.2 Was the dataset cleaned or modified before model training?** Yes

11
generation_config.json Normal file
View File

@@ -0,0 +1,11 @@
{
"_from_model_config": true,
"bos_token_id": 1,
"eos_token_id": [
32000,
32001,
32007
],
"transformers_version": "4.43.3",
"pad_token_id": 32000
}

3
model-00001-of-00002.safetensors Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:dcb4621a07a01e28cc2161d6a2a5ce3a15f343057722da2169e2c6907bfdb2ae
size 4998945744

3
model-00002-of-00002.safetensors Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:1bc35ff74f5dfc65ed0cb968ba4c707ffcdc30c84d799ff4dfe2e9525a38f12c
size 2511742008

1964
model.safetensors.index.json Normal file
View File

File diff suppressed because it is too large Load Diff

1800
modeling_slimmoe.py Normal file
View File

File diff suppressed because it is too large Load Diff

170
sample_finetune.py Normal file
View File

@@ -0,0 +1,170 @@
import sys
import logging
import datasets
from datasets import load_dataset
import torch
import transformers
from trl import SFTTrainer, SFTConfig
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
"""
A simple example on using SFTTrainer to finetune SlimMoE models. For
a more advanced example, please follow HF alignment-handbook/scripts/run_sft.py.
The script can be run on a single 80GB A100 or later generation GPU. Here are some suggestions on
further reducing memory consumption:
- use deepspeed zero3
- use gradient checkpointing
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 trl transformers datasets
pip3 install einops flash_attn torchao
2. Run the code:
python sample_finetune.py
"""
logger = logging.getLogger(__name__)
###################
# Hyper-parameters
###################
training_config = {
"bf16": True,
"do_eval": False,
"optim": "adamw_torch_8bit",
"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": 1,
"per_device_train_batch_size": 1,
"remove_unused_columns": True,
"save_steps": 100,
"save_total_limit": 1,
"seed": 0,
"gradient_checkpointing": False,
# "gradient_checkpointing_kwargs":{"use_reentrant": False},
"gradient_accumulation_steps": 1,
"warmup_ratio": 0.2,
"max_length": 4096,
"dataset_text_field": "text",
"packing": True,
}
train_conf = SFTConfig(**training_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}")
################
# Model Loading
################
checkpoint_path = "microsoft/Phi-tiny-MoE-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,
train_dataset=processed_train_dataset,
eval_dataset=processed_test_dataset,
processing_class=tokenizer,
)
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)

30
special_tokens_map.json Normal file
View File

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

3
tokenizer.json Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:d0f067e1e15cd0a36ebef3668024882cb67a80b86fb4b7b4b128481f0d474db7
size 1844436

130
tokenizer_config.json Normal file
View File

@@ -0,0 +1,130 @@
{
"add_bos_token": false,
"add_eos_token": false,
"added_tokens_decoder": {
"0": {
"content": "<unk>",
"lstrip": false,
"normalized": false,
"rstrip": false,
"single_word": false,
"special": true
},
"1": {
"content": "<s>",
"lstrip": false,
"normalized": false,
"rstrip": false,
"single_word": false,
"special": true
},
"2": {
"content": "</s>",
"lstrip": false,
"normalized": false,
"rstrip": true,
"single_word": false,
"special": false
},
"32000": {
"content": "<|endoftext|>",
"lstrip": false,
"normalized": false,
"rstrip": false,
"single_word": false,
"special": true
},
"32001": {
"content": "<|assistant|>",
"lstrip": false,
"normalized": false,
"rstrip": true,
"single_word": false,
"special": true
},
"32002": {
"content": "<|placeholder1|>",
"lstrip": false,
"normalized": false,
"rstrip": true,
"single_word": false,
"special": true
},
"32003": {
"content": "<|placeholder2|>",
"lstrip": false,
"normalized": false,
"rstrip": true,
"single_word": false,
"special": true
},
"32004": {
"content": "<|placeholder3|>",
"lstrip": false,
"normalized": false,
"rstrip": true,
"single_word": false,
"special": true
},
"32005": {
"content": "<|placeholder4|>",
"lstrip": false,
"normalized": false,
"rstrip": true,
"single_word": false,
"special": true
},
"32006": {
"content": "<|system|>",
"lstrip": false,
"normalized": false,
"rstrip": true,
"single_word": false,
"special": true
},
"32007": {
"content": "<|end|>",
"lstrip": false,
"normalized": false,
"rstrip": true,
"single_word": false,
"special": true
},
"32008": {
"content": "<|placeholder5|>",
"lstrip": false,
"normalized": false,
"rstrip": true,
"single_word": false,
"special": true
},
"32009": {
"content": "<|placeholder6|>",
"lstrip": false,
"normalized": false,
"rstrip": true,
"single_word": false,
"special": true
},
"32010": {
"content": "<|user|>",
"lstrip": false,
"normalized": false,
"rstrip": true,
"single_word": false,
"special": true
}
},
"bos_token": "<s>",
"chat_template": "{% for message in messages %}{{'<|' + message['role'] + '|>' + '\n' + message['content'] + '<|end|>\n' }}{% 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": "<|endoftext|>",
"padding_side": "left",
"sp_model_kwargs": {},
"tokenizer_class": "LlamaTokenizer",
"unk_token": "<unk>",
"use_default_system_prompt": false
}