Instructions to use MD-Mushfiqur123/m-droplychee with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use MD-Mushfiqur123/m-droplychee with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="MD-Mushfiqur123/m-droplychee", trust_remote_code=True)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("MD-Mushfiqur123/m-droplychee", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use MD-Mushfiqur123/m-droplychee with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "MD-Mushfiqur123/m-droplychee"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MD-Mushfiqur123/m-droplychee",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/MD-Mushfiqur123/m-droplychee

SGLang

How to use MD-Mushfiqur123/m-droplychee 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 "MD-Mushfiqur123/m-droplychee" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MD-Mushfiqur123/m-droplychee",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "MD-Mushfiqur123/m-droplychee" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MD-Mushfiqur123/m-droplychee",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use MD-Mushfiqur123/m-droplychee with Docker Model Runner:
```
docker model run hf.co/MD-Mushfiqur123/m-droplychee
```

🫐 m-droplychee (`xe_droplychee` Architecture)

Next-Generation Frontier MoE & Latent Attention Architecture

Optimized for Pre-Training on 94GB/96GB VRAM (NVIDIA RTX PRO 6000)

📌 Model Summary

m-droplychee is powered by the custom xe_droplychee architecture—an ultra-efficient, highly scalable design (~6.5B Total, ~450M Active parameters per token) optimized for single-node multi-gigabyte workstations.

Core Architectural Upgrades:

Multi-Head Latent Attention (MLA) with QK-Head Normalization: Slashes KV-cache memory bandwidth consumption by ~90% via low-rank compression ($d_{c}^{KV} = 256$) coupled with decoupled 64-dimensional Rotary Position Embeddings (RoPE). RMSNorm on Q and K heads prevents attention entropy collapse and stabilizes FP8/BF16 training.
Accelerated Attention Kernels: Direct PyTorch SDPA integration for maximum throughput and memory efficiency.
Fine-Grained MoE (64 Routed + 1 Dedicated Shared Expert): Top-4 dynamic expert routing with 1 dedicated shared expert (5 active experts per token).
Auxiliary-Loss-Free Bias Balancing with Exponential Decay: Dynamic parameter updates ($\pm \gamma$) with exponential decay factor ($0.999$) to eliminate expert collapse without sacrificing routing representation entropy.
Multi-Token Prediction (MTP) Depth-1 Module: Simultaneous 2-token prediction during pre-training, doubling token representation density and enabling speculative decoding speedups.
Native GRPO Reinforcement Learning Engine: Native policy ratio calculation with clipped surrogate objectives and KL divergence regularizers directly in the model class—completely bypassing the need for a separate critic model.

💻 Quickstart & Usage

import torch
from transformers import AutoConfig, AutoModelForCausalLM

# Load model directly from Hugging Face Hub (or local directory)
model_id = "MD-Mushfiqur123/m-droplychee"

config = AutoConfig.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16
).cuda()

# Autoregressive generation utilizing MLA compressed KV cache
prompt = torch.tensor([[1, 100, 200, 300]]).cuda()
output = model.generate(prompt, max_new_tokens=64, temperature=0.7)
print("Generated Sequence:", output.tolist())

🏋️‍♂️ Continued Pre-Training on 94GB/96GB VRAM

python continue_training.py --model_name_or_path MD-Mushfiqur123/m-droplychee --dataset_name "your_bangla_dataset"

📜 License

Apache-2.0

Downloads last month: 78