Skip to content
Other

DM-GDNGQA-9.3B: A hybrid language model with FP8 Delta-Matching attention

2 min readupdated

The DM-GDNGQA-9.3B-LR1.2e-3-DM model, featuring 9.27 billion parameters, is an experimental development with a hybrid architecture combining GatedDeltaNet layers and GQA (Grouped-Query Attention, a variant of the attention mechanism with a reduced number of key and value heads). It was trained from scratch on 30 billion tokens using the Delta-Matching method for FP8 attention. Created as part of research into training precision, this model delivers results fully comparable to standard FP32/BF16 training. It will be useful for language model architecture researchers and developers exploring efficient quantization methods.

Key features

  • Hybrid architecture: GatedDeltaNet layers are combined with GQA (GatedDeltaNet, GatedDeltaNet, GatedDeltaNet, GQA, repeating 9 times).
  • Trained on 30B tokens: Pretrained with a peak learning rate of 1.2e-3 on the Nemotron-CC dataset.
  • Exported to bf16: Weights are saved in bf16 safetensors format with support for standard SDPA attention.
  • High metrics: The model achieves a val CE of 1.2985 and a score of 44.11 on the MMLU benchmark.

What you can do

  • Run text generation: Use the model for text completion and generation via standard Hugging Face tools.
  • Analyze the architecture: Explore the features of the hybrid GatedDeltaNet layers and short convolution in flash-linear-attention.
  • Fine-tune: Perform fine-tuning using the standard Trainer class from the transformers library.

Example

Example code for running model inference in Python:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "haok1402/DM-GDNGQA-9.3B-LR1.2e-3-DM"
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, trust_remote_code=True).cuda()
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)

inputs = tokenizer("The capital of France is", return_tensors="pt").to("cuda")
print(tokenizer.decode(model.generate(**inputs, max_new_tokens=32)[0], skip_special_tokens=True))
Read moreDetails
Metrics (seed 11785): val CE 1.2985, RULER 62.84, MMLU 0.4411, TriviaQA 0.2564.
Hardware requirements: Weights require 18.6 GB of VRAM in bf16.
Limitations: Batched prompts must have the same length due to the design of GatedDeltaNet layers without padding masking.
Stack: torch, transformers >= 5, flash-linear-attention 0.5.2.
Discuss Telegram 0 22

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 17 GB · ready files20 GB
  • FP8 / INT8recommendedweights 8.6 GB10 GB
  • 4-bitweights 4.9 GB5.8 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2080 Ti · BF16 — from RTX 4090

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # haok1402/DM-GDNGQA-9.3B-LR1.2e-3-DM
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download haok1402/DM-GDNGQA-9.3B-LR1.2e-3-DM --local-dir ~/models/DM-GDNGQA-9.3B-LR1.2e-3-DM
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of DM-GDNGQA-9.3B-LR1.2e-3-DM (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

haok1402/DM-GDNGQA-9.3B-LR1.2e-3-DM
Author
haok1402
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

220

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments