# deepseek-ai/profile-data

Analyze computation-communication overlap in V3/R1.

Repository: https://github.com/deepseek-ai/profile-data
Canonical: https://ross.abutalabs.com/products/profile-data
License Family: other
Last push: 2025-03-21T02:23:51+00:00

## Health v2 (maintenance only)
Score: 25/100 (v2, computed 2026-09-03T02:39:23.370411+00:00)
- activity 12, release rhythm 35, longevity 39
- inputs: {"age_days": 553, "days_push": 531, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1182, forks 152 (observed 2026-08-28T04:03:54.007532+00:00)

## What it is
A public release of PyTorch Profiler traces from DeepSeek's V3/R1 training and inference infrastructure, showing computation-communication overlap strategies for MoE workloads. The traces can be visualized in Chrome or Edge via chrome://tracing.

## Use cases
- study how DeepSeek overlaps computation and communication in MoE training
- understand DualPipe forward-backward chunk overlap strategies
- analyze prefill and decode all-to-all communication patterns
- learn EP/TP parallelism configurations used in DeepSeek-V3 deployment
- visualize GPU kernel timelines for large-scale LLM training
- research low-level implementation details of MoE routing and RDMA communication

## When to choose
- you are researching distributed LLM training or inference performance
- you want real-world profiling traces of MoE computation-communication overlap
- you are building or optimizing systems like DualPipe or DeepEP
- you need reference parallelism configurations (EP/TP) for large-scale training

## When to avoid
- you need runnable code rather than profiling data
- you want general-purpose GPU profiling tools rather than example traces
- you need profiling of non-MoE or small-scale workloads

## Facets
- artifact type: dataset
- maturity: active
- function: benchmarking, monitoring, gpu-computing, llm-training, llm-inference
- domain: large-language-models, gpu-computing, performance, microservices, deep-learning
- platform: python, browser
- tags: profiling-data, mixture-of-experts, computation-communication-overlap, pytorch-profiler, dualpipe, deepep, chrome-tracing, rdma, gpu

## Member repositories
- deepseek-ai/profile-data (main) score 25

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:54.007532+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T06:24:47.338999+00:00, confidence not recorded.
  - readme: https://github.com/deepseek-ai/profile-data (fetched 2026-08-28T04:03:54.007532+00:00, sha c6e52994ad28)
- Data as of 2026-08-30T08:39:29.467469+00:00.