servo/deepseek-infra-index
1
0
Fork 0
mirror of https://github.com/deepseek-ai/open-infra-index.git synced 2025-04-02 00:24:03 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

docs: optimize format

Browse source
This commit is contained in:
Nano 2025-02-27 11:09:09 +08:00 • committed by GitHub
parent bf880385d0
commit b3a029ae4c
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 19 additions and 18 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

37
README.md
View file

@ -24,10 +24,11 @@ Daily unlocks begin soon. No ivory towers - just pure garage-energy and communit
Stay tuned – let's geek out in the open together. Stay tuned – let's geek out in the open together.
### Day 1 - [FlashMLA](https://github.com/deepseek-ai/FlashMLA) ### Day 1 - [FlashMLA](https://github.com/deepseek-ai/FlashMLA)
**Efficient MLA Decoding Kernel for Hopper GPUs** **Efficient MLA Decoding Kernel for Hopper GPUs**
Optimized for variable-length sequences, battle-tested in production Optimized for variable-length sequences, battle-tested in production
🔗 <a href="https://github.com/deepseek-ai/FlashMLA"><b>FlashMLA GitHub Repo</b></a> 🔗 [**FlashMLA GitHub Repo**](https://github.com/deepseek-ai/FlashMLA)
✅ BF16 support ✅ BF16 support
✅ Paged KV cache (block size 64) ✅ Paged KV cache (block size 64)
⚡ Performance: 3000 GB/s memory-bound | BF16 580 TFLOPS compute-bound on H800 ⚡ Performance: 3000 GB/s memory-bound | BF16 580 TFLOPS compute-bound on H800
@ -36,7 +37,7 @@ Optimized for variable-length sequences, battle-tested in production
Excited to introduce **DeepEP** - the first open-source EP communication library for MoE model training and inference. Excited to introduce **DeepEP** - the first open-source EP communication library for MoE model training and inference.
🔗 <a href="https://github.com/deepseek-ai/DeepEP"><b>DeepEP GitHub Repo</b></a> 🔗 [**DeepEP GitHub Repo**](https://github.com/deepseek-ai/DeepEP)
✅ Efficient and optimized all-to-all communication ✅ Efficient and optimized all-to-all communication
✅ Both intranode and internode support with NVLink and RDMA ✅ Both intranode and internode support with NVLink and RDMA
✅ High-throughput kernels for training and inference prefilling ✅ High-throughput kernels for training and inference prefilling
@ -46,30 +47,30 @@ Excited to introduce **DeepEP** - the first open-source EP communication library
### Day 3 - [DeepGEMM](https://github.com/deepseek-ai/DeepGEMM) ### Day 3 - [DeepGEMM](https://github.com/deepseek-ai/DeepGEMM)
Introducing DeepGEMM - an FP8 GEMM library that supports both dense and MoE GEMMs, powering V3/R1 training and inference. Introducing **DeepGEMM** - an FP8 GEMM library that supports both dense and MoE GEMMs, powering V3/R1 training and inference.
⚡ Up to 1350+ FP8 TFLOPS on Hopper GPUs 🔗 [**DeepGEMM GitHub Repo**](https://github.com/deepseek-ai/DeepGEMM)
⚡ Up to 1350+ FP8 TFLOPS on Hopper GPUs
✅ No heavy dependency, as clean as a tutorial ✅ No heavy dependency, as clean as a tutorial
✅ Fully Just-In-Time compiled ✅ Fully Just-In-Time compiled
✅ Core logic at ~300 lines - yet outperforms expert-tuned kernels across most matrix sizes ✅ Core logic at ~300 lines - yet outperforms expert-tuned kernels across most matrix sizes
✅ Supports dense layout and two MoE layouts ✅ Supports dense layout and two MoE layouts
🔗 GitHub: https://github.com/deepseek-ai/DeepGEMM
### Day 4 - Optimized Parallelism Strategies ### Day 4 - Optimized Parallelism Strategies
✅ DualPipe - a bidirectional pipeline parallelism algorithm for computation-communication overlap in V3/R1 training.
🔗 https://github.com/deepseek-ai/DualPipe
✅ EPLB - an expert-parallel load balancer for V3/R1. ✅ **DualPipe** - a bidirectional pipeline parallelism algorithm for computation-communication overlap in V3/R1 training.
🔗 https://github.com/deepseek-ai/eplb 🔗 [**GitHub Repo**](https://github.com/deepseek-ai/DualPipe)
📊 Analyze computation-communication overlap in V3/R1. ✅ **EPLB** - an expert-parallel load balancer for V3/R1.
🔗 https://github.com/deepseek-ai/profile-data 🔗 [**GitHub Repo**](https://github.com/deepseek-ai/eplb)
📊 Analyze computation-communication overlap in V3/R1.
🔗 [**GitHub Repo**](https://github.com/deepseek-ai/profile-data)
### Ongoing Releases... ### Ongoing Releases...
## 2024 AI Infrastructure Paper (SC24) ## 2024 AI Infrastructure Paper (SC24)
### Fire-Flyer AI-HPC: A Cost-Effective Software-Hardware Co-Design for Deep Learning ### Fire-Flyer AI-HPC: A Cost-Effective Software-Hardware Co-Design for Deep Learning
<a href="https://dl.acm.org/doi/10.1109/SC41406.2024.00089"><b>📄 Paper Link</b></a> [**📄 Paper Link**](https://dl.acm.org/doi/10.1109/SC41406.2024.00089)
<a href="https://arxiv.org/abs/2408.14158"><b>📄 Arxiv Paper Link</b></a> [**📄 Arxiv Paper Link**](https://arxiv.org/abs/2408.14158)