Text Generation
Transformers
Safetensors
qwen2
coder
code
agent
conversational
text-generation-inference
Instructions to use AdminReal/NexusCoder with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use AdminReal/NexusCoder with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="AdminReal/NexusCoder") messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("AdminReal/NexusCoder") model = AutoModelForCausalLM.from_pretrained("AdminReal/NexusCoder", device_map="auto") messages = [ {"role": "user", "content": "Who are you?"}, ] inputs = tokenizer.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=40) print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use AdminReal/NexusCoder with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "AdminReal/NexusCoder" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "AdminReal/NexusCoder", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/AdminReal/NexusCoder
- SGLang
How to use AdminReal/NexusCoder with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "AdminReal/NexusCoder" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "AdminReal/NexusCoder", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "AdminReal/NexusCoder" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "AdminReal/NexusCoder", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use AdminReal/NexusCoder with Docker Model Runner:
docker model run hf.co/AdminReal/NexusCoder
File size: 3,214 Bytes
eca5751 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 | # Kiến trúc Nexus Coder / Nexus Coder Architecture
## Tổng quan / Overview
Nexus Coder v0.1 sử dụng kiến trúc **Mixture of Experts (MoE) Transformer** tương tự Mixtral 8x7B và DeepSeek-V3.
Nexus Coder v0.1 uses a **Mixture of Experts (MoE) Transformer** architecture similar to Mixtral 8x7B and DeepSeek-V3.
## Các thành phần / Components
### 1. Token Embedding
- Vocab size: 32,000
- Hidden size: 2,048
- Tokens được nhúng thành vector 2048 chiều
### 2. Grouped Query Attention (GQA)
- 16 query heads
- 4 KV heads (ratio 4:1)
- Head dimension: 128
- Giảm 4x memory cho KV cache so với MHA truyền thống
### 3. Rotary Position Embedding (RoPE)
- Base: 10,000
- Hỗ trợ tối đa 50,000 positions
- Cho phép model hiểu vị trí tương đối giữa các tokens
### 4. RMSNorm
- Thay thế LayerNorm truyền thống
- Không có bias, không trừ mean
- Nhanh hơn ~10-20%
### 5. SwiGLU Activation
- `SiLU(gate(x)) * up(x)`
- Hiệu quả hơn ReLU/GELU
- Có 3 ma trận: gate, up, down (3 * hidden * intermediate params)
### 6. Mixture of Experts (MoE) - Cốt lõi
- **24 experts** tổng cộng (mỗi expert là một SwiGLU FFN)
- **3 active experts** mỗi token (top-3 routing)
- Router: linear layer (hidden_size → num_experts)
- Load balancing loss: auxiliary loss để tránh expert collapse
#### Routing Algorithm
```
1. Router tính gate_logits = W_router @ x
2. routing_weights = softmax(gate_logits)
3. top_k_weights, top_k_indices = topk(routing_weights, k=3)
4. Normalize top_k_weights
5. Mỗi token đi qua 3 expert được chọn
6. Output = sum(weight_i * expert_i(x))
```
## Tính toán tham số / Parameter Math
```
Embedding: vocab_size × hidden = 32000 × 2048 = 65.5M
Per layer attn: 2048² + 2×(2048×512) + 2048² = 10.5M (Q, K, V, O with GQA)
Per expert: 3 × 2048 × 5632 = 34.6M (gate + up + down)
Per layer MoE: 24 × 34.6M = 830M (total)
3 × 34.6M = 104M (active)
Per layer total: 10.5M + 830M = 840.5M
12 layers: 10,086M
LM head: 65.5M
────────────────────────────────────
TOTAL: 10,223M ≈ 10.22B ✓
ACTIVE: 65.5 + 12×(10.5 + 104) + 65.5 = 1,503M ≈ 1.50B ✓
```
## Workflow
### Training Workflow
1. Tokenize input text → token IDs
2. Embed tokens → hidden states [B, L, H]
3. For each layer:
- Pre-norm → Attention → residual
- Pre-norm → MoE (router + experts) → residual
4. Final norm → LM head → logits
5. Compute cross-entropy loss + aux loss
6. Backpropagation
### Inference Workflow
1. Tokenize prompt
2. Forward pass through all layers
3. Get logits for last position
4. Apply temperature, top-k, top-p
5. Sample next token
6. Append to sequence, repeat
## Tối ưu / Optimizations
- **KV Cache**: Cache K, V từ các token trước để tăng tốc generation
- **GQA**: Giảm memory và computation cho attention
- **Pre-norm**: Ổn định hơn post-norm trong training
- **Mixed Precision**: Hỗ trợ fp16/bf16 để tiết kiệm memory
- **Gradient Checkpointing**: Đánh đổi compute lấy memory (chưa implement trong v0.1)
|