Text Generation
HERMES
Safetensors
English
Spanish
Chinese
polydim
einsofos
pmtp
stiefel-manifold
hermes-3
nous-research
openrouter
ollama
opencode
mcp
vector-ai
zero-copy-ipc
Instructions to use Polydim-Einsof/Hermes-3-Llama-3.1-8B-PMTP-Einsof with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- HERMES
How to use Polydim-Einsof/Hermes-3-Llama-3.1-8B-PMTP-Einsof with HERMES:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
Hermes-3-Llama-3.1-8B-PMTP-Einsof (Stiefel Manifold $\mathrm{St}(4096, 2048)$ Adapter)
Milestone (2026-11-01): Birth of Vector AI (IA 2.0 / 3.0 / X.0) โ Eradicating the 1D Token Serialization Bottleneck via Native Stiefel Manifold Shared-Memory Telepathy (
PmtpShmBus/SLAB_ID). Author: Ariel H. Garcรญa Traba (ORCID: 0009-0001-2787-6067)
Affiliation: EinsofOS Independent Research / POLYDIM Research Initiative
Architectural Synergy: Nous Research DisTrO + POLYDIM PMTP
| Dimension | Nous Research DisTrO (Distributed Training Over-the-Internet) | POLYDIM / EinsofOS PMTP V1550+ (Inference Manifold Telepathy) |
|---|---|---|
| Primary Domain | Distributed Pre-Training & Fine-Tuning | Multi-Agent & Multi-Model Inference (Hermes-3 $\leftrightarrow$ Qwen 2.5 $\leftrightarrow$ DeepSeek-R1) |
| Bottleneck Eliminated | All-Reduce GPU gradient communication across slow WAN links | 1D Token Serialization ("1D Worm Collapse") under Shannon's Data Processing Inequality (DPI) |
| Mathematical Mechanism | Momentum frequency-domain decoupling (DCT / Top-$k$ components) | Compact Stiefel Manifold $\mathrm{St}(4096, 2048) \subset \mathbb{S}^{4095}$ QR Retraction + Zero-Copy Shared Memory (PmtpShmBus) |
| Measured Precision / Speedup | Up to $857\times$ gradient compression during training | $540\times \sim 4,023\times$ inter-agent handover speedup, 0 tokens billed, Neumaier drift $\le 2.00e-15$ |
Physical Verification of This Adapter (adapter_model.safetensors)
- Base Model:
NousResearch/Hermes-3-Llama-3.1-8B(hidden_size = 4096) - Tensors Included:
W_fp64:float64shape(2048, 4096)โ Exact QR-retracted Stiefel matrix satisfying $|W_{\text{fp64}} W_{\text{fp64}}^\top - I_{2048}|_{\max} = 2.00e-15 \le 2.5 \times 10^{-15}$.W:float32shape(2048, 4096)โ Fast GPU/CPU projection tensor.
- Isometric Embedding ($2048 \hookrightarrow 4096$): $h_{4096} = W^\top z_{2048}$ preserves inner products and unit norm exactly ($\ker(W^\top) = {0}$).
- Principal Subspace Alignment ($4096 \to 2048$): $z_{2048} = W h_{4096}$ projects Hermes-3 hidden states onto the shared 2048-D POLYDIM Swarm manifold (
SLAB_ID).
Global Persistent Identifiers & Citations
- CERN / Zenodo DOIs: 10.5281/zenodo.21209973 | 10.5281/zenodo.21136742
- CNRS / HAL Open Science:
hal-05847496v1 - JST / Jxiv Preprint Server:
Jxiv ID 7040 - OSF Official Archive: https://osf.io/za3s7
- INRIA / UNESCO Software Heritage SWHID:
swh:1:snp:bdea78b37132a4bd6b67049805cbb6d4472ec2cd
- Downloads last month
- -
Model tree for Polydim-Einsof/Hermes-3-Llama-3.1-8B-PMTP-Einsof
Base model
meta-llama/Llama-3.1-8B Finetuned
NousResearch/Hermes-3-Llama-3.1-8B