Skip to content
Other

Transformer-1.3B EHWS: A 90% Compressed Sparse Language Model

2 min readupdated

The Transformer-1.3B-semi-structured-EHWS-90pct model is an experimental 1.3B-parameter language model with its linear layers artificially pruned down to 90% sparsity. It is built upon the base fla-hub/transformer-1.3B-100B model, using a two-phase ADMM method and C4 data calibration for pruning. The architecture features semi-structured sparsity: every row in the pruned linear layer retains an identical weight count (top 10%), while the embeddings and final lm_head remain dense. This model appeals to researchers and engineers exploring optimization methods, inference acceleration, and LLM compression techniques.

Key features

  • Extreme compression: 90% sparsity applied to the linear layers of the model blocks.
  • Balanced load: each row in the pruned layer maintains an equal number of non-zero weights (exactly 10%).
  • Training method: two-phase optimization using ADMM and C4 calibration sequences.
  • Integration: fully compatible with the flash-linear-attention (fla) library and the AutoModelForCausalLM standard.

What you can do

  • Inference testing: run the model locally using the flash-linear-attention library to experiment with sparse computations.
  • Quality analysis: compare accuracy and perplexity metrics between the sparse version and its dense base counterpart on text generation tasks.
  • Optimization research: inspect the configuration files and results.json data to reproduce the ADMM method on other architectures.

Example

import torch
from transformers import AutoModelForCausalLM

# Загрузка модели требует установленной библиотеки flash-linear-attention (import fla)
model_id = "ikimyaii/Transformer-1.3B-semi-structured-EHWS-90pct"
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")
print(model)
Read moreDetails
• Perplexity (WikiText-2): Dense — 9.77, this model — 42.59
• Perplexity (C4): Dense — 14.65, this model — 38.86
• Average zero-shot accuracy (lm-evaluation-harness): 0.4390 (Dense) vs 0.3663 (this model)
• Training time: 2.76 hours on a single A100 GPU (40GB)
• Limitations: embeddings and lm_head remain dense, accuracy drops due to heavy pruning
Discuss Telegram 0 10

Will it run?

Runs in F32

Fits without quantisation.

  • F32weights 5.1 GB · ready files6.1 GB
  • BF16weights 2.5 GB3.3 GB
  • FP8 / INT8recommendedweights 1.3 GB1.9 GB
  • 4-bitweights 0.7 GB1.3 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 2060 · F32 — from RTX 5060 Ti 8GB

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.

    # ikimyaii/Transformer-1.3B-semi-structured-EHWS-90pct
    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 ikimyaii/Transformer-1.3B-semi-structured-EHWS-90pct --local-dir ~/models/Transformer-1.3B-semi-structured-EHWS-90pct
    # 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 transformer-1.3B-100B (quants, finetunes, LoRAs) will show up under the bell.

Fine-tunes 9All on Hugging Face

About the model

ikimyaii/Transformer-1.3B-semi-structured-EHWS-90pct
Author
ikimyaii
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

100

Ask AI about this model

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

Comments