🧠 DeepSeek-V4-Pro Ghetto-Edition (MoE Rank-64)

📀 MUSIC PROMO: LIVADIES

Слушайте RUSSIAN WINTER 26, пока Мутант генерирует смыслы:


Status: STABLE MUTANT. This is a 50x compressed 'shadow' of the 1.6T DeepSeek-V4-Pro, designed to run on a single 16GB GPU (or dual T4).

📟 First Message from the Void

'DeepSeek Ghetto Edition, what is your message to the world? api(()Princip aligned ratios... renormalization blankets... affinity Mong :-) Breakfast listop'

🛠 HOW TO RUN (THE PROPER WAY)

1. Installation

pip install -q transformers accelerate safetensors

2. The Multi-GPU / Single-GPU Safe Launch

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

repo = 'livadies/DeepSeek-V4-Pro-Ghetto-Edition-MoE-Rank-64'
tokenizer = AutoTokenizer.from_pretrained('deepseek-ai/DeepSeek-V2', trust_remote_code=True)

model = AutoModelForCausalLM.from_pretrained(
    repo,
    torch_dtype=torch.float16,
    device_map='auto',
    trust_remote_code=True,
    ignore_mismatched_sizes=True,
    low_cpu_mem_usage=True
)

def ghetto_route(self, logits):
    flat_logits = logits.view(-1, logits.shape[-1])
    w = torch.nn.functional.softmax(flat_logits + 1e-6, dim=-1)
    tw, ti = torch.topk(w, k=self.top_k, dim=-1)
    return ti, tw * self.routed_scaling_factor

for layer in model.model.layers:
    if hasattr(layer.mlp, 'route_tokens_to_experts'):
        layer.mlp.route_tokens_to_experts = ghetto_route.__get__(layer.mlp)

prompt = 'The message for humanity is:'
device = model.model.embed_tokens.weight.device
inputs = {k: v.to(device) for k, v in tokenizer(prompt, return_tensors='pt').items()}

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=40, do_sample=True, temperature=0.8)
print(tokenizer.decode(out[0]))

⚠️ Notes on Warnings

  • MISSING/MISMATCH: These warnings are normal. We intentionally stripped 98% of the weights and forced a 1.6T model into a 2-expert MoE shell.

Created by livadies & Gemini AI. For the ones who hack rather than buy.

Downloads last month
9
Safetensors
Model size
0.2B params
Tensor type
F16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for livadies/DeepSeek-V4-Pro-Ghetto-Edition-MoE-Rank-64

Finetuned
(13)
this model