![\"DeepSeek](\"images/logo.svg\"){kind=logo}
\n\n\nModel Download |\nEvaluation Results |\nQuick Start |\nLicense |\nCitation\n
\n\n\n\n![\"Awesome](\"images/Awesome_DeepSeek_Coder.png\")
\n
\n![\"DeepSeek](\"pictures/logo.png\"){kind=logo}
\n
[![](\"pictures/home.png\")
Homepage] | [๐ค Chat with DeepSeek Coder] | [๐ค Models Download] | [Discord] | [WeChat (ๅพฎไฟก)]
\n![\"result\"](\"pictures/result.png\")
\n
\n\n๐ฅ Model Download |\nโก Quick Start |\n๐ License |\n๐ Citation \nModel Download |\nQuick Start |\nEvaluation Results |\nLicense |\nCitation\n \nPaper Link๐๏ธ<"},{"ref":"P8","kind":"page","title":"deepseek-ai/awesome-deepseek-integration repository metadata","date":"2026-06-11T04:00:03.72998+00:00","date_source":null,"source_url":"https://github.com/deepseek-ai/awesome-deepseek-integration","signal_url":null,"signal_json_url":null,"text":"# deepseek-ai/awesome-deepseek-integration\n\nDescription: Integrate the DeepSeek API into popular software\n\nLicense: CC0-1.0\n\nStars: 37819\n\nForks: 4159\n\nOpen issues: 93\n\nCreated: 2024-01-11T02:43:17Z\n\nPushed: 2026-02-23T16:27:47Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n \n \nModel Download |\nEvaluation Resu"},{"ref":"P10","kind":"page","title":"deepseek-ai/DeepSeek-R1 repository metadata","date":"2026-06-11T04:00:02.688241+00:00","date_source":null,"source_url":"https://github.com/deepseek-ai/DeepSeek-R1","signal_url":null,"signal_json_url":null,"text":"# deepseek-ai/DeepSeek-R1\n\nLicense: MIT\n\nStars: 92010\n\nForks: 11712\n\nOpen issues: 49\n\nCreated: 2025-01-20T11:57:28Z\n\nPushed: 2025-06-27T08:35:54Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# DeepSeek-R1\n\n\n\n\n \n๐ฅ Model Download |\nโก Quick Start |\n๐ License |\n๐ Citation \n๐ฅ Model Download |\n\n\n\n\n \nModel Summary |\nProverBench |\n\n\n\n\n \n๐ฅ Model Download |\n๐ Paper Link |\n๐ Arxiv Paper Link |\n \nDeepSeek-OCR: Contexts Optical Compression\n \n
\n๐ P"},{"ref":"P6","kind":"page","title":"deepseek-ai/DeepSeek-V2 repository metadata","date":"2026-06-11T04:00:04.449305+00:00","date_source":null,"source_url":"https://github.com/deepseek-ai/DeepSeek-V2","signal_url":null,"signal_json_url":null,"text":"# deepseek-ai/DeepSeek-V2\n\nDescription: DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model\n\nLicense: MIT\n\nStars: 5011\n\nForks: 541\n\nOpen issues: 88\n\nCreated: 2024-04-22T06:55:47Z\n\nPushed: 2024-09-25T10:23:55Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n\n\n\n\n\n\n![\"DeepSeek-V2\"](\"https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/logo.svg?raw=true\"){kind=logo}
\n
\n\n\n\n\n![\"Discord\"](\"https://img.shields.io/badge/Discord-DeepSeek%20AI-7289da?logo=discord&logoColor=white&color=7289da\"){kind=icon}
![\"Wechat\"](\"https://img.shields.io/badge/WeChat-DeepSeek%20AI-brightgreen?logo=wechat&logoColor=white\"){kind=icon}
![\"Twitter](\"https://img.shields.io/badge/Twitter-deepseek_ai-white?logo=x&logoCol"},{"ref":"P7","kind":"page","title":"deepseek-ai/DeepSeek-LLM){kind=icon}
\n\n\n
\n\n\n\n\n\n\n\n\n\n\n## Table of Contents\n\n- [Awesome DeepSeek Integrations ](#awesome-deepseek-integrations-)\n- [Table of Contents](#table-of-contents)\n- [Project List](#project-list)\n- [Applications](#applications)\n- [AI Agent frameworks](#ai-agent-frameworks)\n- [Data AI Applications frameworks](#data-ai-applications-frameworks)\n- [RAG frameworks](#rag-frameworks)\n- [FHE (Fully Homomorphic Encryption) frameworks](#fhe-fully-homomorphic-encryption-frameworks)\n- [Solana frameworks](#solana-frameworks)\n- [Synthetic data curation](#synthetic-data-curation)\n- [IM Application Plugins](#im-application-plugins)\n- [Office Addin](#office-addin)\n- [Browser Extensions](#browser-extensions)\n- [VS Code Extensions](#vs-code-extensions)\n- [Visual Studio Extensions](#visual-studio-extensions)\n- [neovim Extensions](#neovim-extensions)\n- [JetBrains Extensions]("},{"ref":"P9","kind":"page","title":"deepseek-ai/DeepSeek-Math repository metadata","date":"2026-06-11T04:00:03.477856+00:00","date_source":null,"source_url":"https://github.com/deepseek-ai/DeepSeek-Math","signal_url":null,"signal_json_url":null,"text":"# deepseek-ai/DeepSeek-Math\n\nDescription: DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models\n\nLanguage: Python\n\nLicense: MIT\n\nStars: 3319\n\nForks: 582\n\nOpen issues: 43\n\nCreated: 2024-02-05T07:25:51Z\n\nPushed: 2024-04-15T07:55:37Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n\n\n\n\n\n
\n
\n\n\n
\n\n\n\n\n\n\n\n\n\n
\n\n\n## 1. Introduction\n\nWe introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1. \nDeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary"},{"ref":"P11","kind":"page","title":"deepseek-ai/ESFT repository metadata","date":"2026-06-11T04:00:02.53924+00:00","date_source":null,"source_url":"https://github.com/deepseek-ai/ESFT","signal_url":null,"signal_json_url":null,"text":"# deepseek-ai/ESFT\n\nDescription: Expert Specialized Fine-Tuning\n\nLanguage: Python\n\nLicense: MIT\n\nStars: 737\n\nForks: 263\n\nOpen issues: 7\n\nCreated: 2024-07-04T09:48:48Z\n\nPushed: 2025-05-22T07:47:15Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# Expert-Specialized Fine-Tuning\n\nOfficial Repo for paper [Let the Expert Stick to His Last: Expert-Specialized Fine-Tuning for Sparse Architectural Large Language Models](https://arxiv.org/abs/2407.01906) by \n[Zihan Wang](https://zihanwang314.github.io), [Deli Chen](https://victorchen96.github.io/chendeli.io/), [Damai Dai](https://scholar.google.com.hk/citations?user=8b-ysf0NWVoC&hl=zh-CN), [Runxin Xu](https://runxinxu.github.io/aboutme/), \n[Zhuoshu Li](http://www.idi.zju.edu.cn/member/3053.html) and\nY. Wu. \n\n**ESFT** aims to efficiently customize Large Language Models (LLMs) with Mixture-of-Experts (MoE) architecture by adjusting only task-relevant parts, improving efficiency and performance while using fewer resources and storage. \n\n## ๐ฐ News\n\n๐
**2024.9.20:** Glad to announce that ESFT has been accepted to the **EMNLP 2024 Main Conference**! \n\n๐
**2024.8.11:** We now release the **ESFT training code**! โจ You can now try it with your own models and dataset!\n\n## ๐ Quick Start \n### Installation and Setup\n```bash\ngit clone https://github.com/deepseek-ai/ESFT.git\ncd esft\n```\n\n### Install required dependencies\n```bash\npip install transformers torch safetensors accelerate\n```\n\n### Download necessary adapters\n```bash\nbash scripts/download_adapters.sh\n```\n\n## ๐งKey Scripts\n1. **eval_multigpu.py**\nThis script evaluates the performance of the model on various datasets. See **scripts/eval.sh** for detailed configs and explanations.\n\n**Usage:**\n```bash\npython eval_multigpu.py \\\n--eval_dataset=translation \\\n--base_model_path=deepseek-ai/ESFT-vanilla-lite \\\n--adapter_dir=all_models/adapters/token/translation \\\n--output_path=results/completions/token/translation.jsonl \\\n--openai_api_key=YOUR_OPENAI_API_KEY\n```\n\n2. **get_expert_scores.py**\nThis script calculates the scores for each expert based on the evaluation datasets.\n**Usage:**\n```bash\nexport PYTHONPATH=$PYTHONPATH:$(pwd)\npython scripts/expert/get_expert_scores.py \\\n--"},{"ref":"P12","kind":"page","title":"deepseek-ai/DeepSeek-Prover-V1.5 repository metadata","date":"2026-06-11T04:00:02.447661+00:00","date_source":null,"source_url":"https://github.com/deepseek-ai/DeepSeek-Prover-V1.5","signal_url":null,"signal_json_url":null,"text":"# deepseek-ai/DeepSeek-Prover-V1.5\n\nLanguage: Python\n\nLicense: MIT\n\nStars: 577\n\nForks: 239\n\nOpen issues: 9\n\nCreated: 2024-08-15T15:10:07Z\n\nPushed: 2024-08-16T03:33:41Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n\n\n\n\n\n\n\n
\n\n\n\n\n![\"Twitter](\"https://img.shields.io/badge/Twitter-deepseek_ai-white?logo=x&logoColor=white\"){kind=icon}
\n\n\n
\n\n\n\n\n![\"Twitter](\"https://img.shields.io/badge/Twitter-deepseek_ai-white?logo="},{"ref":"P14","kind":"page","title":"deepseek-ai/Janus){kind=icon}
\n\n\n
\n\n\n\n\n\n\n๐ Janus-Series: Unified Multimodal Understanding and Generation Models
\n\n\n\n\n\n\n\n\n\n\n\n
\n\n\n\n\n\n\n\n
\n\n\n## Table of Contents\n\n1. [Introduction](#1-introduction)\n2. [Model Summary](#2-model-summary)\n3. [Model Downloads](#3-model-downloads)\n4. [Evaluation Resul"},{"ref":"P17","kind":"page","title":"deepseek-ai/DeepSeek-VL2 repository metadata","date":"2026-06-11T04:00:01.460649+00:00","date_source":null,"source_url":"https://github.com/deepseek-ai/DeepSeek-VL2","signal_url":null,"signal_json_url":null,"text":"# deepseek-ai/DeepSeek-VL2\n\nDescription: DeepSeek-VL2: Mixture-of-Experts Vision-Language Models for Advanced Multimodal Understanding\n\nLanguage: Python\n\nLicense: MIT\n\nStars: 5298\n\nForks: 1810\n\nOpen issues: 121\n\nCreated: 2024-12-13T04:55:41Z\n\nPushed: 2025-02-26T05:03:42Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n\n\n\n\n\n\n\n
\n\n\n\n\n\n\n\n\n\n
\n\n# Hello, DeepSeek Open Infra!\n\n## 202505 Industry Track Paper (ISCA25) \n### Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Architectures\n[**๐ Arxiv Paper Link**](https://arxiv.org/abs/2505.09343)\n\n## 202504 [The Path to Open-Sourcing the DeepSeek Inference Engine](OpenSourcing_DeepSeek_Inference_Engine/README.md)\n\n## 202502 Open-Source Week\nWe're a tiny team @deepseek-ai pushing our limits in AGI exploration.\n\nStarting **this week** , Feb 24, 2025 we'll open-source 5 repos โ one daily drop โ not because we've made grand claims, \nbut simply as developers sharing our small-but-sincere progress with full transparency.\n\nThese are humble building blocks of our online service: documented, deployed, and battle-tested in production. \nNo vaporware, just sincere code that moved our tiny yet ambitious dream forward.\n\nWhy? Because every line shared becomes collective momentum that accelerates the journey.\nDaily unlocks begin soon. No ivory towers - just pure garage-energy and community-driven innovation ๐ง\n\nStay tuned โ let's geek out in the open together.\n\n### Day 1 - [FlashMLA](https://github.com/deepseek-ai/FlashMLA)\n\n**Efficient MLA Decoding Kernel for Hopper GPUs** \nOptimized for variable-length sequences, battle-tested in production \n\n๐ [**FlashMLA GitHub Repo**](https://github.com/deepseek-ai/FlashMLA) \nโ
BF16 support \nโ
Paged KV cache (block size 64) \nโก Performance: 3000 GB/s memory-bound | BF16 580 TFLOPS compute-bound on H800\n\n### Day 2 - [DeepEP](https://github.com/deepseek-ai/DeepEP)\n\nExcited to introduce **DeepEP** - t"},{"ref":"P20","kind":"page","title":"deepseek-ai/profile-data repository metadata","date":"2026-06-11T04:00:00.609086+00:00","date_source":null,"source_url":"https://github.com/deepseek-ai/profile-data","signal_url":null,"signal_json_url":null,"text":"# deepseek-ai/profile-data\n\nDescription: Analyze computation-communication overlap in V3/R1.\n\nStars: 1159\n\nForks: 147\n\nOpen issues: 13\n\nCreated: 2025-02-26T07:26:06Z\n\nPushed: 2025-03-21T02:23:51Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# Profiling Data in DeepSeek Infra\n\nHere, we publicly share profiling data from our training and inference framework to help the community better understand the communication-computation overlap strategies and low-level implementation details. The profiling data was captured using the PyTorch Profiler. After downloading, you can visualize it directly by navigating to chrome://tracing in the Chrome browser (or edge://tracing in the Edge browser). Notice that we simulate an absolutely balanced MoE routing strategy for profiling.\n\n## Training\n\n[[profile_data]](train.json)\n\n\n\nThe training profile data demonstrates our overlapping strategy for a pair of individual forward and backward chunks in [DualPipe](https://github.com/deepseek-ai/dualpipe). Each chunk contains 4 MoE (Mixture of Experts) layers.\nThe parallel configuration aligns with DeepSeek-V3 pretraining settings: EP64, TP1 with 4K sequence length. And the PP communication is not included during profiling for simplicity.\n\n## Inference\n\n### Prefilling\n\n[[profile_data]](prefill.json)\n\n\n\nFor prefilling, the profile employs EP32 and TP1 (in line with DeepSeek V3/R1 โs actual online deployment), with a prompt length set to 4K and a batch size of 16K tokens per GPU. In our prefilling stage, we utilize two micro-batches to overlap computation and all-to-all communication, while ensuring that the attention computation load is balanced across the two micro-batches โ meaning that the same prompt may be split between them.\n\n### Decoding\n\n[[profile_data]](decode.json)\n\n\n\nFor decoding, the profile employs EP128, TP1, and a prompt length of 4K (closely matching the actual online deployment configuration), with a batch size of 128 requests per GPU. Similar to prefilling, decoding also leverages two micro-batches for overlapping computation and all-to-all communication. However, unlike in prefill"},{"ref":"P21","kind":"page","title":"deepseek-ai/FlashMLA repository metadata","date":"2026-06-11T04:00:00.593416+00:00","date_source":null,"source_url":"https://github.com/deepseek-ai/FlashMLA","signal_url":null,"signal_json_url":null,"text":"# deepseek-ai/FlashMLA\n\nDescription: FlashMLA: Efficient Multi-head Latent Attention Kernels\n\nLanguage: C++\n\nLicense: MIT\n\nStars: 12698\n\nForks: 1057\n\nOpen issues: 105\n\nCreated: 2025-02-21T06:31:27Z\n\nPushed: 2026-04-30T09:04:01Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# FlashMLA\n\n## Introduction\n\nFlashMLA is DeepSeek's library of optimized attention kernels, powering the [DeepSeek-V3](https://github.com/deepseek-ai/DeepSeek-V3) and [DeepSeek-V3.2-Exp](https://github.com/deepseek-ai/DeepSeek-V3.2-Exp) models. This repository contains the following implementations:\n\n**Sparse Attention Kernels**\n\n*These kernels power DeepSeek Sparse Attention (DSA), as introduced in [this paper](https://github.com/deepseek-ai/DeepSeek-V3.2-Exp).*\n\n- Token-level sparse attention for the prefill stage\n- Token-level sparse attention for the decoding stage, with FP8 KV cache\n\n**Dense Attention Kernels**\n\n- Dense attention for the prefill stage\n- Dense attention for the decoding stage\n\n## News\n\n- **2025.09.29 Release of Sparse Attention Kernels**: With the launch of [DeepSeek-V3.2](https://github.com/deepseek-ai/DeepSeek-V3.2-Exp), we are releasing the corresponding token-level sparse attention kernels. These kernels power the model's DeepSeek Sparse Attention (DSA) and achieve up to 640 TFlops during prefilling and 410 TFlops during decoding. We also release a deep-dive blog for our new FP8 sparse decoding kernel. Check it out [here](docs/20250929-hopper-fp8-sparse-deep-dive.md).\n- **2025.08.01 Kernels for MHA on SM100**: Thanks to [NVIDIA's PR](https://github.com/deepseek-ai/FlashMLA/pull/76) for MHA forward / backward kernels on SM100!\n- **2025.04.22 Deep-Dive Blog**: We'd love to share the technical details behind the new FlashMLA kernel! Check out our deep-dive write-up [here](docs/20250422-new-kernel-deep-dive.md).\n- **2025.04.22 Performance Update**: We're excited to announce the new release of Flash MLA, which delivers 5% ~ 15% performance improvement for compute-bound workloads, achieving up to 660 TFlops on NVIDIA H800 SXM5 GPUs. The interface of the new version is fully compatible with the old one. Simply upgrade to the new version for an immediate performance b"},{"ref":"P22","kind":"page","title":"deepseek-ai/smallpond repository metadata","date":"2026-06-11T04:00:00.592386+00:00","date_source":null,"source_url":"https://github.com/deepseek-ai/smallpond","signal_url":null,"signal_json_url":null,"text":"# deepseek-ai/smallpond\n\nDescription: A lightweight data processing framework built on DuckDB and 3FS.\n\nLanguage: Python\n\nLicense: MIT\n\nStars: 4960\n\nForks: 444\n\nOpen issues: 32\n\nCreated: 2025-02-24T09:28:17Z\n\nPushed: 2025-03-05T18:23:54Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# smallpond\n\n[](https://github.com/deepseek-ai/smallpond/actions/workflows/ci.yml)\n[](https://pypi.org/project/smallpond/)\n[](https://deepseek-ai.github.io/smallpond/)\n[](LICENSE)\n\nA lightweight data processing framework built on [DuckDB] and [3FS].\n\n## Features\n\n- ๐ High-performance data processing powered by DuckDB\n- ๐ Scalable to handle PB-scale datasets\n- ๐ ๏ธ Easy operations with no long-running services\n\n## Installation\n\nPython 3.8 to 3.12 is supported.\n\n```bash\npip install smallpond\n```\n\n## Quick Start\n\n```bash\n# Download example data\nwget https://duckdb.org/data/prices.parquet\n```\n\n```python\nimport smallpond\n\n# Initialize session\nsp = smallpond.init()\n\n# Load data\ndf = sp.read_parquet(\"prices.parquet\")\n\n# Process data\ndf = df.repartition(3, hash_by=\"ticker\")\ndf = sp.partial_sql(\"SELECT ticker, min(price), max(price) FROM {0} GROUP BY ticker\", df)\n\n# Save results\ndf.write_parquet(\"output/\")\n# Show results\nprint(df.to_pandas())\n```\n\n## Documentation\n\nFor detailed guides and API reference:\n- [Getting Started](docs/source/getstarted.rst)\n- [API Reference](docs/source/api.rst)\n\n## Performance\n\nWe evaluated smallpond using the [GraySort benchmark] ([script]) on a cluster comprising 50 compute nodes and 25 storage nodes running [3FS]. The benchmark sorted 110.5TiB of data in 30 minutes and 14 seconds, achieving an average throughput of 3.66TiB/min.\n\nDetails can be found in [3FS - Gray Sort].\n\n[DuckDB]: https://duckdb.org/\n[3FS]: https://github.com/deepseek-ai/3FS\n[GraySort benchmark]: https://sortbenchmark.org/\n[script]: benchmarks/gray_sort_benchmark.py\n[3FS - Gray Sort]: https://github.com/deepseek-ai/3FS?ta"},{"ref":"P23","kind":"page","title":"deepseek-ai/EPLB repository metadata","date":"2026-06-11T04:00:00.404094+00:00","date_source":null,"source_url":"https://github.com/deepseek-ai/EPLB","signal_url":null,"signal_json_url":null,"text":"# deepseek-ai/EPLB\n\nDescription: Expert Parallelism Load Balancer\n\nLanguage: Python\n\nLicense: MIT\n\nStars: 1388\n\nForks: 202\n\nOpen issues: 10\n\nCreated: 2025-02-26T10:41:00Z\n\nPushed: 2025-03-24T09:17:10Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# Expert Parallelism Load Balancer (EPLB)\n\nWhen using expert parallelism (EP), different experts are assigned to different GPUs. Because the load of different \nexperts may vary depending on the current workload, it is important to keep the load of different GPUs balanced. \nAs described in the DeepSeek-V3 paper, we adopt a **redundant experts** strategy that duplicates heavy-loaded experts. \nThen, we heuristically pack the duplicated experts to GPUs to ensure load balancing across different GPUs. Moreover, \nthanks to the **group-limited expert routing** used in DeepSeek-V3, we also attempt to place the experts of the same \ngroup to the same node to reduce inter-node data traffic, whenever possible.\n\nTo facilitate reproduction and deployment, we open-source our deployed EP load balancing algorithm in `eplb.py`. \nThe algorithm computes a balanced expert replication and placement plan based on the estimated expert loads. Note \nthat the exact method to predict the loads of experts is out of this repo's scope. A common method is to use \nmoving average of historical statistics. \n\n## The Algorithm\n\nThe load balancing algorithm comes with two policies used for different cases.\n\n### Hierarchical Load Balancing\n\nWhen the number of server nodes divides the number of expert groups, we use the hierarchical load balancing policy to\nharness the group-limited expert routing. We first pack the expert groups to nodes evenly, ensuring the loads of \ndifferent nodes are balanced. Then, we replicate the experts within each node. Finally, we pack the replicated experts \nto individual GPUs to ensure different GPUs are load-balanced. The hierarchical load balancing policy can be used in \nprefilling stage with a smaller expert-parallel size.\n\n### Global Load Balancing\n\nIn other cases, we use the global load balancing policy that replicates the experts globally regardless of expert \ngroups, and pack the replicated experts to individual GP"},{"ref":"P24","kind":"page","title":"deepseek-ai/DualPipe repository metadata","date":"2026-06-11T04:00:00.001642+00:00","date_source":null,"source_url":"https://github.com/deepseek-ai/DualPipe","signal_url":null,"signal_json_url":null,"text":"# deepseek-ai/DualPipe\n\nDescription: A bidirectional pipeline parallelism algorithm for computation-communication overlap in DeepSeek V3/R1 training.\n\nLanguage: Python\n\nLicense: MIT\n\nStars: 2965\n\nForks: 326\n\nOpen issues: 5\n\nCreated: 2025-02-26T13:29:57Z\n\nPushed: 2026-01-14T06:34:59Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# DualPipe\n\nDualPipe is an innovative bidirectional pipeline parallelism algorithm introduced in the [DeepSeek-V3 Technical Report](https://arxiv.org/pdf/2412.19437). It achieves full overlap of forward and backward computation-communication phases, also reducing pipeline bubbles. For detailed information on computation-communication overlap, please refer to the [profile data](https://github.com/deepseek-ai/profile-data).\n\n### Schedules\n\n\n\nExample DualPipe scheduling for 8 PP ranks and 20 micro-batches in two directions.\nThe micro-batches in the reverse direction are symmetric to those in the forward direction, so\nwe omit their batch ID for illustration simplicity. Two cells enclosed by a shared black border\nhave mutually overlapped computation and communication\n\n## DualPipeV\n\nDualPipeV is a concise V-shape schedule derived from DualPipe using a \"cut-in-half\" procedure, introduced by Sea AI Lab as \"Cut-in-half\" in their [blog post](https://hackmd.io/@ufotalent/r1lVXsa9Jg). Thanks to them for this efficient schedule!\n\n### Schedules\n\n\n\nExample DualPipeV scheduling for 4 PP ranks (8 PP stages) and 10 micro-batches.\n\n## Pipeline Bubbles and Memory Usage Comparison (based on the same number of PP stages)\n\n| Method | Bubble | Parameter Per Device | Activation Per Device | #Devices |\n|-------------|---------------------------------|----------------------|-----------------------|----------|\n| 1F1B | (*PP*-1)(๐น+๐ต) | 1ร | *PP* | *PP* |\n| ZB1P | (*PP*-1)(๐น+๐ต-2๐) | 1ร | *PP* | *PP* |\n| DualPipe | (*PP*/2-1)(๐น&๐ต+๐ต-3๐) | 2ร | *PP*+1 | *PP* |\n| DualPipeV | (*PP*/2-1)(๐น&๐ต+๐ต-3๐) | 2ร | *PP*+1 | *PP*/2 |\n\n*PP* denotes the number of pp stages (even).\n๐น denotes the execution time of a forward chunk, ๐ต denotes the execution time of a\nfull backward chunk, ๐ denotes the e"},{"ref":"P25","kind":"page","title":"deepseek-ai/3FS repository metadata","date":"2026-06-11T03:59:59.934633+00:00","date_source":null,"source_url":"https://github.com/deepseek-ai/3FS","signal_url":null,"signal_json_url":null,"text":"# deepseek-ai/3FS\n\nDescription: A high-performance distributed file system designed to address the challenges of AI training and inference workloads. \n\nLanguage: C++\n\nLicense: MIT\n\nStars: 9955\n\nForks: 1056\n\nOpen issues: 153\n\nCreated: 2025-02-27T13:36:53Z\n\nPushed: 2026-05-07T09:31:54Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# Fire-Flyer File System\n\n[](https://github.com/deepseek-ai/3fs/actions/workflows/build.yml)\n[](LICENSE)\n\nThe Fire-Flyer File System (3FS) is a high-performance distributed file system designed to address the challenges of AI training and inference workloads. It leverages modern SSDs and RDMA networks to provide a shared storage layer that simplifies development of distributed applications. Key features and benefits of 3FS include:\n\n- Performance and Usability\n- **Disaggregated Architecture** Combines the throughput of thousands of SSDs and the network bandwidth of hundreds of storage nodes, enabling applications to access storage resource in a locality-oblivious manner.\n- **Strong Consistency** Implements Chain Replication with Apportioned Queries (CRAQ) for strong consistency, making application code simple and easy to reason about.\n- **File Interfaces** Develops stateless metadata services backed by a transactional key-value store (e.g., FoundationDB). The file interface is well known and used everywhere. There is no need to learn a new storage API.\n\n- Diverse Workloads\n- **Data Preparation** Organizes outputs of data analytics pipelines into hierarchical directory structures and manages a large volume of intermediate outputs efficiently.\n- **Dataloaders** Eliminates the need for prefetching or shuffling datasets by enabling random access to training samples across compute nodes.\n- **Checkpointing** Supports high-throughput parallel checkpointing for large-scale training.\n- **KVCache for Inference** Provides a cost-effective alternative to DRAM-based caching, offering high throughput and significantly larger capacity.\n\n## Documentation\n\n* [Design Notes](docs/design_notes.md)\n* [Setup G"},{"ref":"P26","kind":"page","title":"deepseek-ai/DeepSeek-Prover-V2 repository metadata","date":"2026-06-11T03:59:59.835689+00:00","date_source":null,"source_url":"https://github.com/deepseek-ai/DeepSeek-Prover-V2","signal_url":null,"signal_json_url":null,"text":"# deepseek-ai/DeepSeek-Prover-V2\n\nLicense: NOASSERTION\n\nStars: 1273\n\nForks: 102\n\nOpen issues: 13\n\nCreated: 2025-04-30T13:42:57Z\n\nPushed: 2025-07-18T08:11:37Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n\n\n\n\n\n\n\n
\n\n\n\n\n
\n\n\n\n\n\n\n![\"DeepSeek](\"assets/logo.svg\"){kind=logo}
\n
\n\n\n\n\n\n
![](\"assets/fig1.png\")
\n
![\"Homepage\"](\"images/badge.svg\"){kind=icon}
\n![\"Chat\"](\"https://img.shields.io/badge/๐ค%20Chat-DeepSeek%20LLM-536af5?color=536af5&logoColor=white\"){kind=icon}
\n![\"Hugging](\"https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-DeepSeek%20AI-ffc107?color=ffc107&logoColor=white\"){kind=icon}
\n![\"Wechat\"](\"images/qr.jpeg\")
\n![\"Code](\"https://img.shields.io/badge/Code_License-MIT-f5de53?&color=f5de53\"){kind=icon}
\n![\"Model](\"https://img.shields.io/badge/Model_License-Model_Agreement-f5de53?&color=f5de53\"){kind=icon}
\n![\"Chat\"](\"https://img.shields.io/badge/๐ค%20Chat-DeepSeek%20VL-536af5?color=536af5&logoColor=white\"){kind=icon}
\n![\"Homepage\"](\"https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/badge.svg?raw=true\"){kind=icon}
\n![\"Chat\"](\"https://img.shields.io/badge/๐ค%20Chat-DeepSeek%20V2-536af5?color=536af5&logoColor=white\"){kind=icon}
\n![\"Chat\"](\"https://img.shields.io/badge/๐ค%20Chat-DeepSeek%20V3-536af5?color=536af5&logoColor=white\"){kind=icon}
\n![\"Homepage\"](\"assets/badge.svg\"){kind=icon}
\n![[WeChat (ๅพฎไฟก)]](\"https://github.com/guoday/assert/blob/main/QR.png?raw=true\"){kind=link}