ncbi/Gene-R1-8B
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: ncbi/Gene-R1-8B
Source: Original Platform
This commit is contained in:
ModelHub XC
2026-05-18 02:50:03 +08:00
commit 4812385892
18 changed files with 3385 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
.gitattributes vendored Normal file
View File

@@ -0,0 +1,9 @@
pytorch_model-00001-of-00007.bin filter=lfs diff=lfs merge=lfs -text
pytorch_model-00002-of-00007.bin filter=lfs diff=lfs merge=lfs -text
pytorch_model-00003-of-00007.bin filter=lfs diff=lfs merge=lfs -text
pytorch_model-00004-of-00007.bin filter=lfs diff=lfs merge=lfs -text
pytorch_model-00005-of-00007.bin filter=lfs diff=lfs merge=lfs -text
pytorch_model-00006-of-00007.bin filter=lfs diff=lfs merge=lfs -text
pytorch_model-00007-of-00007.bin filter=lfs diff=lfs merge=lfs -text
tokenizer.json filter=lfs diff=lfs merge=lfs -text
training_args.bin filter=lfs diff=lfs merge=lfs -text

154
README.md Normal file
View File

@@ -0,0 +1,154 @@
---
license: other
library_name: transformers
pipeline_tag: text-generation
language:
- en
tags:
- gene-set-analysis
- biomedical
- reasoning
- llama
base_model:
- meta-llama/Llama-3.1-8B-Instruct
- meta-llama/Llama-3.2-1B-Instruct
- meta-llama/Llama-3.2-3B-Instruct
---
# Overview of Gene-R1
**Introduction**
- Gene-R1 is a data-augmented learning framework that equips lightweight and open-source LLMs with step-by-step reasoning capabilities tailored to the gene set analysis task.
- It has been fine-tuned by ~270K gene sets collected from 16 genomic databases.
- Experimental results demonstrate that Gene-R1 achieves substantial performance gains, matching commercial LLMs.
- For more details, please check out our [paper](https://www.worldscientific.com/doi/abs/10.1142/9789819824755_0035) (PSB, 2026).
**Gene-R1 helps for gene set analysis through fine-tuned small language models (SLMs) that can be locally deployed.**
The model contains three versions:
- [Gene-R1-8B](https://huggingface.co/ncbi/Gene-R1-8B): A version fine-tuned based on the Llama-3.1-8B-Instruct.
- [Gene-R1-1B](https://huggingface.co/ncbi/Gene-R1-1B): A version fine-tuned based on the Llama-3.2-1B-Instruct.
- [Gene-R1-3B](https://huggingface.co/ncbi/Gene-R1-3B): A version fine-tuned based on the Llama-3.2-3B-Instruct.
# Model Deployment for Private Gene Set Analysis
```python
import transformers
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "ncbi/Gene-R1-8B"
tokenizer_test = AutoTokenizer.from_pretrained(
model_id,
token = "xxxxxxxxx" # Your access key of hugging face
)
model_test = AutoModelForCausalLM.from_pretrained(
model_id,
device_map = "auto",
token = "xxxxxxxxx" # Your access key of hugging face
)
def complete_chat(system, prompt, model, tokenizer):
model.generation_config.do_sample=False
tokenized_chat = tokenizer('#SYSTEM: \n'+ system + '#USER: \n'+ prompt+' #Assistant: \n', return_tensors="pt").input_ids.to(model.device)
outputs = model.generate(tokenized_chat, max_new_tokens=4000, temperature = 0)
return tokenizer.decode(outputs[0])
system = "You are an efficient and insightful assistant to a molecular biologist."
users = lambda genes: f"""
Write a critical analysis of the biological processes performed by this system of interacting proteins.
Base your analysis on prior knowledge available in your training data.
After the analysis, propose a brief name for the most prominent biological process performed by the system.
Place the name at the top of the analysis in the format: "Process: <name>".
Be concise. Avoid unnecessary words.
Use plain text only. Do not include format symbols such as asterisks, dashes, or bullets.
Be specific. Avoid overly general statements such as "the proteins are involved in various cellular processes."
Be factual. Do not include editorial opinions or unsupported claims.
For each important point, clearly explain your reasoning and provide supporting information.
For each identified biological function, specify the corresponding gene names.
Here is the gene set: {genes}
"""
def llama(genes):
genes = genes.replace("/",",").replace(" ",",")
prompt = users(genes)
summary =complete_chat(system, prompt, model_test, tokenizer_test)
return summary
if __name__ == "__main__":
genes = "xxxxxxxxx" # Your private gene set that is separated by comma (,)!
result = llama(genes)
print(result)
```
The expected output looks like:
```
Process: Pancreatic development and glucose homeostasis
1. PDX1 is a homeodomain transcription factor involved in the specification of the early pancreatic epithelium and its subsequent differentiation.
It activates the transcription of several genes including insulin, somatostatin, glucokinase and glucose transporter type 2.
It is essential for maintenance of the normal hormone-producing phenotype in the pancreatic beta-cell.
In pancreatic acinar cells, forms a complex with PBX1b and MEIS2b and mediates the activation of the ELA1 enhancer.
2. NKX6-1 is also a transcription factor involved in the development of pancreatic beta-cells during the secondary transition.
Together with NKX2-2 and IRX3, controls the generation of motor neurons in the neural tube and belongs to the neural progenitor
factors induced by Sonic Hedgehog (SHH) signals.
3.GCG and GLP1, respectively glucagon and glucagon-like peptide 1, are involved in glucose metabolism and homeostasis.
GCG raises blood glucose levels by promoting gluconeogenesis and is the counter regulatory hormone of Insulin.
GLP1 is a potent stimulator of Glucose-Induced Insulin Secretion (GSIS). Plays roles in gastric motility and suppresses blood glucagon levels.
Promotes growth of the intestinal epithelium and pancreatic islet mass both by islet neogenesis and islet cell proliferation.
4. SLC2A2, also known as GLUT2, is a facilitative hexose transporter. In hepatocytes, it mediates bi-directional transport of glucose accross the plasma membranes,
while in the pancreatic beta-cell, it is the main transporter responsible for glucose uptake and part of the cell's glucose-sensing mechanism.
It is involved in glucose transport in the small intestine and kidney too.
To summarize, the genes in this set are involved in the specification, differentiation, growth and functionality of the pancreas,
with a particular emphasis on the pancreatic beta-cell. Particularly, the architecture of the pancreatic islet ensures proper glucose sensing
and homeostasis via a number of different hormones and receptors that can elicit both synergistic and antagonistic effects in the pancreas itself and other peripheral tissues.
```
⚠️ **Notice: The outputs sometimes are not following the instruction, you can try again if this case occurs.**
More details of model usage can be referred at our GitHub: [GitHub](https://github.com/ncbi-nlp/Gene-R1)
# Download statistics
Hugging Face tracks downloads automatically based on requests to model query files such as `config.json`.
To ensure downloads are counted, please load the full models directly from the Hub using `transformers`:
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "ncbi/Gene-R1-8B"
tokenizer = AutoTokenizer.from_pretrained(model_id, hf_token)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", hf_token)
```
# Acknowledgments
This research was supported in part by the Intramural Research Program of the National Institutes of Health (NIH).
The contributions of the NIH authors are considered Works of the United States Government.
The findings and conclusions presented in this paper are those of the authors and do not necessarily reflect the views of the NIH or the U.S. Department of Health and Human Services.
# Disclaimer
These models show the results of research conducted in the Computational Biology Branch, NCBI/NLM.
The information produced on this website is not intended for direct diagnostic use or medical decision-making without review and oversight by a clinical professional.
Individuals should not change their health behavior solely on the basis of information produced on this website.
NIH does not independently verify the validity or utility of the information produced by this tool.
If you have questions about the information produced on this website, please see a health care professional.
More information about NCBI's disclaimer policy is available.
# Citation
```bibtext
@inproceedings{wang2025gene,
title={Gene-R1: Reasoning with Data-Augmented Lightweight LLMs for Gene Set Analysis},
author={Wang, Zhizheng and Yang, Yifan and Jin, Qiao and Lu, Zhiyong},
booktitle={Biocomputing 2026: Proceedings of the Pacific Symposium},
pages={494--507},
year={2025},
organization={World Scientific}
}
```

39
config.json Normal file
View File

@@ -0,0 +1,39 @@
{
"architectures": [
"LlamaForCausalLM"
],
"attention_bias": false,
"attention_dropout": 0.0,
"bos_token_id": 128000,
"eos_token_id": [
128001,
128008,
128009
],
"head_dim": 128,
"hidden_act": "silu",
"hidden_size": 4096,
"initializer_range": 0.02,
"intermediate_size": 14336,
"max_position_embeddings": 131072,
"mlp_bias": false,
"model_type": "llama",
"num_attention_heads": 32,
"num_hidden_layers": 32,
"num_key_value_heads": 8,
"pretraining_tp": 1,
"rms_norm_eps": 1e-05,
"rope_scaling": {
"factor": 8.0,
"high_freq_factor": 4.0,
"low_freq_factor": 1.0,
"original_max_position_embeddings": 8192,
"rope_type": "llama3"
},
"rope_theta": 500000.0,
"tie_word_embeddings": false,
"torch_dtype": "float32",
"transformers_version": "4.51.3",
"use_cache": true,
"vocab_size": 128256
}

12
generation_config.json Normal file
View File

@@ -0,0 +1,12 @@
{
"bos_token_id": 128000,
"do_sample": true,
"eos_token_id": [
128001,
128008,
128009
],
"temperature": 0.6,
"top_p": 0.9,
"transformers_version": "4.51.3"
}

0
log.json Normal file
View File

3
pytorch_model-00001-of-00007.bin Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:7d01d1b2df0bb88ed15c99a07b8d6927870c3ec7bb84ae5931d6bb811d2cfef7
size 4886373454

3
pytorch_model-00002-of-00007.bin Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:9235cf1ec43f5becb3620998355a432cd3650f39df1b01cc6b94af84cc75a922
size 4831851068

3
pytorch_model-00003-of-00007.bin Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:8effab507836c13fe5e2c61e2ad9251925b579f59c400a791b5eb5f329a77437
size 4999624668

3
pytorch_model-00004-of-00007.bin Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:ccb4d8ed5d2c28e8b810f1787c5840d6e307dce0bcacc7d4345c3fe8523331ec
size 4999624656

3
pytorch_model-00005-of-00007.bin Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:c941cc6ff54f296083ccbe7b67244a47445c047e5b19b2e940362d6d3db7b4c2
size 4831851132

3
pytorch_model-00006-of-00007.bin Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:e8aab9477f0466148817b5f4c4cd4210d5398ec8c261e77fb424d37bcfc92a58
size 4999624668

3
pytorch_model-00007-of-00007.bin Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:6e3f9ba8ded28719047ea74618335067b15e8b1f6ef545d1d4754d524e14c024
size 2572196402

298
pytorch_model.bin.index.json Normal file
View File

@@ -0,0 +1,298 @@
{
"metadata": {
"total_size": 32121044992
},
"weight_map": {
"lm_head.weight": "pytorch_model-00007-of-00007.bin",
"model.embed_tokens.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.0.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.0.mlp.down_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.0.mlp.gate_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.0.mlp.up_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.0.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.0.self_attn.k_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.0.self_attn.o_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.0.self_attn.q_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.0.self_attn.v_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.1.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.1.mlp.down_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.1.mlp.gate_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.1.mlp.up_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.1.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.1.self_attn.k_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.1.self_attn.o_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.1.self_attn.q_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.1.self_attn.v_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.10.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.10.mlp.down_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.10.mlp.gate_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.10.mlp.up_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.10.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.10.self_attn.k_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.10.self_attn.o_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.10.self_attn.q_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.10.self_attn.v_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.11.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.11.mlp.down_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.11.mlp.gate_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.11.mlp.up_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.11.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.11.self_attn.k_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.11.self_attn.o_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.11.self_attn.q_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.11.self_attn.v_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.12.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.12.mlp.down_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.12.mlp.gate_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.12.mlp.up_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.12.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.12.self_attn.k_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.12.self_attn.o_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.12.self_attn.q_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.12.self_attn.v_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.13.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.13.mlp.down_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.13.mlp.gate_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.13.mlp.up_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.13.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.13.self_attn.k_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.13.self_attn.o_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.13.self_attn.q_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.13.self_attn.v_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.14.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.14.mlp.down_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.14.mlp.gate_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.14.mlp.up_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.14.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.14.self_attn.k_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.14.self_attn.o_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.14.self_attn.q_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.14.self_attn.v_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.15.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.15.mlp.down_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.15.mlp.gate_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.15.mlp.up_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.15.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.15.self_attn.k_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.15.self_attn.o_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.15.self_attn.q_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.15.self_attn.v_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.16.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.16.mlp.down_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.16.mlp.gate_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.16.mlp.up_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.16.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.16.self_attn.k_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.16.self_attn.o_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.16.self_attn.q_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.16.self_attn.v_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.17.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.17.mlp.down_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.17.mlp.gate_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.17.mlp.up_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.17.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.17.self_attn.k_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.17.self_attn.o_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.17.self_attn.q_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.17.self_attn.v_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.18.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.18.mlp.down_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.18.mlp.gate_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.18.mlp.up_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.18.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.18.self_attn.k_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.18.self_attn.o_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.18.self_attn.q_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.18.self_attn.v_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.19.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.19.mlp.down_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.19.mlp.gate_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.19.mlp.up_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.19.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.19.self_attn.k_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.19.self_attn.o_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.19.self_attn.q_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.19.self_attn.v_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.2.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.2.mlp.down_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.2.mlp.gate_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.2.mlp.up_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.2.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.2.self_attn.k_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.2.self_attn.o_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.2.self_attn.q_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.2.self_attn.v_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.20.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.20.mlp.down_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.20.mlp.gate_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.20.mlp.up_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.20.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.20.self_attn.k_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.20.self_attn.o_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.20.self_attn.q_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.20.self_attn.v_proj.weight": "pytorch_model-00004-of-00007.bin",
"model.layers.21.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.21.mlp.down_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.21.mlp.gate_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.21.mlp.up_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.21.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.21.self_attn.k_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.21.self_attn.o_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.21.self_attn.q_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.21.self_attn.v_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.22.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.22.mlp.down_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.22.mlp.gate_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.22.mlp.up_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.22.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.22.self_attn.k_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.22.self_attn.o_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.22.self_attn.q_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.22.self_attn.v_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.23.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.23.mlp.down_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.23.mlp.gate_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.23.mlp.up_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.23.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.23.self_attn.k_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.23.self_attn.o_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.23.self_attn.q_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.23.self_attn.v_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.24.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.24.mlp.down_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.24.mlp.gate_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.24.mlp.up_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.24.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.24.self_attn.k_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.24.self_attn.o_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.24.self_attn.q_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.24.self_attn.v_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.25.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.25.mlp.down_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.25.mlp.gate_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.25.mlp.up_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.25.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.25.self_attn.k_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.25.self_attn.o_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.25.self_attn.q_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.25.self_attn.v_proj.weight": "pytorch_model-00005-of-00007.bin",
"model.layers.26.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.26.mlp.down_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.26.mlp.gate_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.26.mlp.up_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.26.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.26.self_attn.k_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.26.self_attn.o_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.26.self_attn.q_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.26.self_attn.v_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.27.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.27.mlp.down_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.27.mlp.gate_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.27.mlp.up_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.27.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.27.self_attn.k_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.27.self_attn.o_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.27.self_attn.q_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.27.self_attn.v_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.28.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.28.mlp.down_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.28.mlp.gate_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.28.mlp.up_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.28.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.28.self_attn.k_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.28.self_attn.o_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.28.self_attn.q_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.28.self_attn.v_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.29.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.29.mlp.down_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.29.mlp.gate_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.29.mlp.up_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.29.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.29.self_attn.k_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.29.self_attn.o_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.29.self_attn.q_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.29.self_attn.v_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.3.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.3.mlp.down_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.3.mlp.gate_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.3.mlp.up_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.3.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.3.self_attn.k_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.3.self_attn.o_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.3.self_attn.q_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.3.self_attn.v_proj.weight": "pytorch_model-00001-of-00007.bin",
"model.layers.30.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.30.mlp.down_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.30.mlp.gate_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.30.mlp.up_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.30.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.30.self_attn.k_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.30.self_attn.o_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.30.self_attn.q_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.30.self_attn.v_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.31.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.31.mlp.down_proj.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.31.mlp.gate_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.31.mlp.up_proj.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.31.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.31.self_attn.k_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.31.self_attn.o_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.31.self_attn.q_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.31.self_attn.v_proj.weight": "pytorch_model-00006-of-00007.bin",
"model.layers.4.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.4.mlp.down_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.4.mlp.gate_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.4.mlp.up_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.4.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.4.self_attn.k_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.4.self_attn.o_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.4.self_attn.q_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.4.self_attn.v_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.5.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.5.mlp.down_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.5.mlp.gate_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.5.mlp.up_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.5.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.5.self_attn.k_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.5.self_attn.o_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.5.self_attn.q_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.5.self_attn.v_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.6.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.6.mlp.down_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.6.mlp.gate_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.6.mlp.up_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.6.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.6.self_attn.k_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.6.self_attn.o_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.6.self_attn.q_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.6.self_attn.v_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.7.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.7.mlp.down_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.7.mlp.gate_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.7.mlp.up_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.7.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.7.self_attn.k_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.7.self_attn.o_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.7.self_attn.q_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.7.self_attn.v_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.8.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.8.mlp.down_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.8.mlp.gate_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.8.mlp.up_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.8.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.8.self_attn.k_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.8.self_attn.o_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.8.self_attn.q_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.8.self_attn.v_proj.weight": "pytorch_model-00002-of-00007.bin",
"model.layers.9.input_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.9.mlp.down_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.9.mlp.gate_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.9.mlp.up_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.9.post_attention_layernorm.weight": "pytorch_model-00007-of-00007.bin",
"model.layers.9.self_attn.k_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.9.self_attn.o_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.9.self_attn.q_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.layers.9.self_attn.v_proj.weight": "pytorch_model-00003-of-00007.bin",
"model.norm.weight": "pytorch_model-00007-of-00007.bin"
}
}

17
special_tokens_map.json Normal file
View File

@@ -0,0 +1,17 @@
{
"bos_token": {
"content": "<|begin_of_text|>",
"lstrip": false,
"normalized": false,
"rstrip": false,
"single_word": false
},
"eos_token": {
"content": "<|eot_id|>",
"lstrip": false,
"normalized": false,
"rstrip": false,
"single_word": false
},
"pad_token": "<|eot_id|>"
}

3
tokenizer.json Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:65ff5472d095ccd9332d9e723153d7bc7226cb6be9c1bffda738b5ba2e71bf26
size 17210084

2069
tokenizer_config.json Normal file
View File

File diff suppressed because it is too large Load Diff

3
training_args.bin Normal file
View File

@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:201110e2acf073ab4726ad5649c573e483741785740fa9a46d72ee573adef67f
size 7800

760
zero_to_fp32.py Normal file
View File

@@ -0,0 +1,760 @@
#!/usr/bin/env python
# Copyright (c) Microsoft Corporation.
# SPDX-License-Identifier: Apache-2.0
# DeepSpeed Team
# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
# application.
#
# example:
# python zero_to_fp32.py . output_dir/
# or
# python zero_to_fp32.py . output_dir/ --safe_serialization
import argparse
import torch
import glob
import math
import os
import re
import gc
import json
import numpy as np
from tqdm import tqdm
from collections import OrderedDict
from dataclasses import dataclass
# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
# DeepSpeed data structures it has to be available in the current python environment.
from deepspeed.utils import logger
from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
@dataclass
class zero_model_state:
buffers: dict()
param_shapes: dict()
shared_params: list
ds_version: int
frozen_param_shapes: dict()
frozen_param_fragments: dict()
debug = 0
# load to cpu
device = torch.device('cpu')
def atoi(text):
return int(text) if text.isdigit() else text
def natural_keys(text):
'''
alist.sort(key=natural_keys) sorts in human order
http://nedbatchelder.com/blog/200712/human_sorting.html
(See Toothy's implementation in the comments)
'''
return [atoi(c) for c in re.split(r'(\d+)', text)]
def get_model_state_file(checkpoint_dir, zero_stage):
if not os.path.isdir(checkpoint_dir):
raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
# there should be only one file
if zero_stage <= 2:
file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
elif zero_stage == 3:
file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
if not os.path.exists(file):
raise FileNotFoundError(f"can't find model states file at '{file}'")
return file
def get_checkpoint_files(checkpoint_dir, glob_pattern):
# XXX: need to test that this simple glob rule works for multi-node setup too
ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
if len(ckpt_files) == 0:
raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
return ckpt_files
def get_optim_files(checkpoint_dir):
return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
def get_model_state_files(checkpoint_dir):
return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
def parse_model_states(files):
zero_model_states = []
for file in files:
state_dict = torch.load(file, map_location=device, weights_only=False)
if BUFFER_NAMES not in state_dict:
raise ValueError(f"{file} is not a model state checkpoint")
buffer_names = state_dict[BUFFER_NAMES]
if debug:
print("Found buffers:", buffer_names)
# recover just the buffers while restoring them to fp32 if they were saved in fp16
buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
param_shapes = state_dict[PARAM_SHAPES]
# collect parameters that are included in param_shapes
param_names = []
for s in param_shapes:
for name in s.keys():
param_names.append(name)
# update with frozen parameters
frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
if frozen_param_shapes is not None:
if debug:
print(f"Found frozen_param_shapes: {frozen_param_shapes}")
param_names += list(frozen_param_shapes.keys())
# handle shared params
shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
ds_version = state_dict.get(DS_VERSION, None)
frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
z_model_state = zero_model_state(buffers=buffers,
param_shapes=param_shapes,
shared_params=shared_params,
ds_version=ds_version,
frozen_param_shapes=frozen_param_shapes,
frozen_param_fragments=frozen_param_fragments)
zero_model_states.append(z_model_state)
return zero_model_states
def parse_optim_states(files, ds_checkpoint_dir):
total_files = len(files)
state_dicts = []
for f in tqdm(files, desc='Loading checkpoint shards'):
state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
# immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
# and also handle the case where it was already removed by another helper script
state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
state_dicts.append(state_dict)
if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
raise ValueError(f"{files[0]} is not a zero checkpoint")
zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
# For ZeRO-2 each param group can have different partition_count as data parallelism for expert
# parameters can be different from data parallelism for non-expert parameters. So we can just
# use the max of the partition_count to get the dp world_size.
if type(world_size) is list:
world_size = max(world_size)
if world_size != total_files:
raise ValueError(
f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
"Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
)
# the groups are named differently in each stage
if zero_stage <= 2:
fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
elif zero_stage == 3:
fp32_groups_key = FP32_FLAT_GROUPS
else:
raise ValueError(f"unknown zero stage {zero_stage}")
fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
return zero_stage, world_size, fp32_flat_groups
def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
"""
Returns fp32 state_dict reconstructed from ds checkpoint
Args:
- ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
"""
print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
optim_files = get_optim_files(ds_checkpoint_dir)
zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
model_files = get_model_state_files(ds_checkpoint_dir)
zero_model_states = parse_model_states(model_files)
print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
if zero_stage <= 2:
return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
exclude_frozen_parameters)
elif zero_stage == 3:
return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
exclude_frozen_parameters)
def _zero2_merge_frozen_params(state_dict, zero_model_states):
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
return
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
frozen_param_fragments = zero_model_states[0].frozen_param_fragments
if debug:
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
wanted_params = len(frozen_param_shapes)
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
print(f'Frozen params: Have {avail_numel} numels to process.')
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
total_params = 0
total_numel = 0
for name, shape in frozen_param_shapes.items():
total_params += 1
unpartitioned_numel = shape.numel()
total_numel += unpartitioned_numel
state_dict[name] = frozen_param_fragments[name]
if debug:
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
def _has_callable(obj, fn):
attr = getattr(obj, fn, None)
return callable(attr)
def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
param_shapes = zero_model_states[0].param_shapes
# Reconstruction protocol:
#
# XXX: document this
if debug:
for i in range(world_size):
for j in range(len(fp32_flat_groups[0])):
print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
# XXX: memory usage doubles here (zero2)
num_param_groups = len(fp32_flat_groups[0])
merged_single_partition_of_fp32_groups = []
for i in range(num_param_groups):
merged_partitions = [sd[i] for sd in fp32_flat_groups]
full_single_fp32_vector = torch.cat(merged_partitions, 0)
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
avail_numel = sum(
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
if debug:
wanted_params = sum([len(shapes) for shapes in param_shapes])
wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
# not asserting if there is a mismatch due to possible padding
print(f"Have {avail_numel} numels to process.")
print(f"Need {wanted_numel} numels in {wanted_params} params.")
# params
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
# out-of-core computing solution
total_numel = 0
total_params = 0
for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
offset = 0
avail_numel = full_single_fp32_vector.numel()
for name, shape in shapes.items():
unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
total_numel += unpartitioned_numel
total_params += 1
if debug:
print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
offset += unpartitioned_numel
# Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
# avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
# paddings performed in the code it's almost impossible to predict the exact numbers w/o the
# live optimizer object, so we are checking that the numbers are within the right range
align_to = 2 * world_size
def zero2_align(x):
return align_to * math.ceil(x / align_to)
if debug:
print(f"original offset={offset}, avail_numel={avail_numel}")
offset = zero2_align(offset)
avail_numel = zero2_align(avail_numel)
if debug:
print(f"aligned offset={offset}, avail_numel={avail_numel}")
# Sanity check
if offset != avail_numel:
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
exclude_frozen_parameters):
state_dict = OrderedDict()
# buffers
buffers = zero_model_states[0].buffers
state_dict.update(buffers)
if debug:
print(f"added {len(buffers)} buffers")
if not exclude_frozen_parameters:
_zero2_merge_frozen_params(state_dict, zero_model_states)
_zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
# recover shared parameters
for pair in zero_model_states[0].shared_params:
if pair[1] in state_dict:
state_dict[pair[0]] = state_dict[pair[1]]
return state_dict
def zero3_partitioned_param_info(unpartitioned_numel, world_size):
remainder = unpartitioned_numel % world_size
padding_numel = (world_size - remainder) if remainder else 0
partitioned_numel = math.ceil(unpartitioned_numel / world_size)
return partitioned_numel, padding_numel
def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
return
if debug:
for i in range(world_size):
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
wanted_params = len(frozen_param_shapes)
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
print(f'Frozen params: Have {avail_numel} numels to process.')
print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
total_params = 0
total_numel = 0
for name, shape in zero_model_states[0].frozen_param_shapes.items():
total_params += 1
unpartitioned_numel = shape.numel()
total_numel += unpartitioned_numel
param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
if debug:
print(
f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
)
print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
class GatheredTensor:
"""
A pseudo tensor that collects partitioned weights.
It is more memory efficient when there are multiple groups.
"""
def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
self.flat_groups = flat_groups
self.flat_groups_offset = flat_groups_offset
self.offset = offset
self.partitioned_numel = partitioned_numel
self.shape = shape
self.dtype = self.flat_groups[0][0].dtype
def contiguous(self):
"""
Merge partitioned weights from flat_groups into a single tensor.
"""
end_idx = self.offset + self.partitioned_numel
world_size = len(self.flat_groups)
pad_flat_param_chunks = []
for rank_i in range(world_size):
# for each rank, we need to collect weights from related group/groups
flat_groups_at_rank_i = self.flat_groups[rank_i]
start_group_id = None
end_group_id = None
for group_id in range(len(self.flat_groups_offset)):
if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
start_group_id = group_id
if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
end_group_id = group_id
break
# collect weights from related group/groups
for group_id in range(start_group_id, end_group_id + 1):
flat_tensor = flat_groups_at_rank_i[group_id]
start_offset = self.offset - self.flat_groups_offset[group_id]
end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
# collect weights from all ranks
pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
return param
def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
param_shapes = zero_model_states[0].param_shapes
avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
# Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
# param, re-consolidating each param, while dealing with padding if any
# merge list of dicts, preserving order
param_shapes = {k: v for d in param_shapes for k, v in d.items()}
if debug:
for i in range(world_size):
print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
wanted_params = len(param_shapes)
wanted_numel = sum(shape.numel() for shape in param_shapes.values())
# not asserting if there is a mismatch due to possible padding
avail_numel = fp32_flat_groups[0].numel() * world_size
print(f"Trainable params: Have {avail_numel} numels to process.")
print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
# params
# XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
# out-of-core computing solution
offset = 0
total_numel = 0
total_params = 0
flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
unpartitioned_numel = shape.numel()
total_numel += unpartitioned_numel
total_params += 1
partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
if debug:
print(
f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
)
# memory efficient tensor
tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
state_dict[name] = tensor
offset += partitioned_numel
offset *= world_size
# Sanity check
if offset != avail_numel:
raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
exclude_frozen_parameters):
state_dict = OrderedDict()
# buffers
buffers = zero_model_states[0].buffers
state_dict.update(buffers)
if debug:
print(f"added {len(buffers)} buffers")
if not exclude_frozen_parameters:
_zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
_zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
# recover shared parameters
for pair in zero_model_states[0].shared_params:
if pair[1] in state_dict:
state_dict[pair[0]] = state_dict[pair[1]]
return state_dict
def to_torch_tensor(state_dict, return_empty_tensor=False):
"""
Convert state_dict of GatheredTensor to torch tensor
"""
torch_state_dict = {}
converted_tensors = {}
for name, tensor in state_dict.items():
tensor_id = id(tensor)
if tensor_id in converted_tensors: # shared tensors
shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
torch_state_dict[name] = shared_tensor
else:
converted_tensors[tensor_id] = name
if return_empty_tensor:
torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
else:
torch_state_dict[name] = tensor.contiguous()
return torch_state_dict
def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
tag=None,
exclude_frozen_parameters=False,
lazy_mode=False):
"""
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
via a model hub.
Args:
- ``checkpoint_dir``: path to the desired checkpoint folder
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
- ``exclude_frozen_parameters``: exclude frozen parameters
- ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
Convert the pesduo tensor to torch tensor by ``.contiguous()``
Returns:
- pytorch ``state_dict``
A typical usage might be ::
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
# do the training and checkpoint saving
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
model = model.cpu() # move to cpu
model.load_state_dict(state_dict)
# submit to model hub or save the model to share with others
In this example the ``model`` will no longer be usable in the deepspeed context of the same
application. i.e. you will need to re-initialize the deepspeed engine, since
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
the checkpoint. Or you can load state_dict in lazy mode ::
from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
for name, lazy_tensor in state_dict.item():
tensor = lazy_tensor.contiguous() # to cpu
print(name, tensor)
# del tensor to release memory if it no longer in use
"""
if tag is None:
latest_path = os.path.join(checkpoint_dir, 'latest')
if os.path.isfile(latest_path):
with open(latest_path, 'r') as fd:
tag = fd.read().strip()
else:
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
if not os.path.isdir(ds_checkpoint_dir):
raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
if lazy_mode:
return state_dict
else:
return to_torch_tensor(state_dict)
def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
output_dir,
max_shard_size="5GB",
safe_serialization=False,
tag=None,
exclude_frozen_parameters=False):
"""
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
Args:
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
- ``output_dir``: directory to the pytorch fp32 state_dict output files
- ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
- ``safe_serialization``: whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
- ``exclude_frozen_parameters``: exclude frozen parameters
"""
# Dependency pre-check
if safe_serialization:
try:
from safetensors.torch import save_file
except ImportError:
print('If you want to use `safe_serialization`, please `pip install safetensors`')
raise
if max_shard_size is not None:
try:
from huggingface_hub import split_torch_state_dict_into_shards
except ImportError:
print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
raise
# Convert zero checkpoint to state_dict
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
tag,
exclude_frozen_parameters,
lazy_mode=True)
# Shard the model if it is too big.
weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
if max_shard_size is not None:
filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
# an memory-efficient approach for sharding
empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
filename_pattern=filename_pattern,
max_shard_size=max_shard_size)
else:
from collections import namedtuple
StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
state_dict_split = StateDictSplit(is_sharded=False,
filename_to_tensors={weights_name: list(state_dict.keys())})
# Save the model by shard
os.makedirs(output_dir, exist_ok=True)
filename_to_tensors = state_dict_split.filename_to_tensors.items()
for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
shard_state_dict = to_torch_tensor(shard_state_dict)
output_path = os.path.join(output_dir, shard_file)
if safe_serialization:
save_file(shard_state_dict, output_path, metadata={"format": "pt"})
else:
torch.save(shard_state_dict, output_path)
# release the memory of current shard
for tensor_name in list(shard_state_dict.keys()):
del state_dict[tensor_name]
del shard_state_dict[tensor_name]
del shard_state_dict
gc.collect()
# Save index if sharded
if state_dict_split.is_sharded:
index = {
"metadata": state_dict_split.metadata,
"weight_map": state_dict_split.tensor_to_filename,
}
save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
save_index_file = os.path.join(output_dir, save_index_file)
with open(save_index_file, "w", encoding="utf-8") as f:
content = json.dumps(index, indent=2, sort_keys=True) + "\n"
f.write(content)
def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
"""
1. Put the provided model to cpu
2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
3. Load it into the provided model
Args:
- ``model``: the model object to update
- ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
- ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
Returns:
- ``model`: modified model
Make sure you have plenty of CPU memory available before you call this function. If you don't
have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
conveniently placed for you in the checkpoint folder.
A typical usage might be ::
from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
# submit to model hub or save the model to share with others
Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
of the same application. i.e. you will need to re-initialize the deepspeed engine, since
``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
"""
logger.info(f"Extracting fp32 weights")
state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
logger.info(f"Overwriting model with fp32 weights")
model = model.cpu()
model.load_state_dict(state_dict, strict=False)
return model
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("checkpoint_dir",
type=str,
help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
parser.add_argument("output_dir",
type=str,
help="directory to the pytorch fp32 state_dict output files"
"(e.g. path/checkpoint-12-output/)")
parser.add_argument(
"--max_shard_size",
type=str,
default="5GB",
help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
"lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
"We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
"without CPU OOM issues.")
parser.add_argument(
"--safe_serialization",
default=False,
action='store_true',
help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
parser.add_argument("-t",
"--tag",
type=str,
default=None,
help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
args = parser.parse_args()
debug = args.debug
convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
args.output_dir,
max_shard_size=args.max_shard_size,
safe_serialization=args.safe_serialization,
tag=args.tag,
exclude_frozen_parameters=args.exclude_frozen_parameters)