Skip to content
Other

ice40GPT

чат-модель на ~1,07 млн параметров для FPGA

2 min readupdated

ice40GPT — decoder-only трансформер, который целиком помещается в память FPGA iCE40UP5K. Хранится 1 069 280 параметров: один блок повторяется 6 раз, на каждом проходе свои нормы. Ширина 96, 4 головы, 1 KV-голова, FFN 256; словарь 2048 BPE-токенов, контекст 32 токена. Эмбеддинги — 64-широкая таблица токенов, проецируется до 96 и связана с выходной головой.

Ключевые особенности

  • Веса, KV-кэш и таблицы занимают 134KB — это укладывается в 128KB SPRAM + 15KB EBR чипа iCE40UP5K.
  • int4-блоки весов и int4-эмбеддинги с per-row scales; обучение с QAT. В forward pass применяется int4/int4-округление (fake_quant), логиты уже отражают квантованные веса.
  • Формат промпта: [INST] текст [/INST]. Текст приводится к нижнему регистру, делится на слова и пунктуацию, затем BPE; конец слова — </w>.
  • Генерация идёт скользящим окном 32 токена и без KV-кэша; останавливается на <eos>, [INST] или [/INST].
  • Обучение на TinyChat (CC BY-NC 4.0, данные сгенерированы GPT-4o mini); веса non-commercial.

Пример

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained('.', trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained('.', trust_remote_code=True)
prompt = tok('[INST] hello, how are you today? [/INST]', return_tensors='pt').input_ids
out = model.generate(prompt, max_new_tokens=40)
print(tok.decode(out[0, prompt.shape[1]:], skip_special_tokens=True))
ПодробнееВалидация: 2.631 nats/word, ppl 10.3, top-1 47.6%.
Веса, KV-кэш и таблицы — 134KB; чип: 128KB SPRAM + 15KB EBR.
Контекст — 32 токена; язык — английский; обучение на TinyChat (CC BY-NC 4.0), веса non-commercial.
Отвечает на бытовой и эмоциональный small talk; фактические вопросы не решает.
Discuss Telegram 0 24

Will it run?

Runs in F32

Fits without quantisation.

  • F32weights 0.0 GB · ready files0.5 GB
  • BF16weights 0.0 GB0.5 GB
  • FP8 / INT8recommendedweights 0.0 GB0.5 GB
  • 4-bitweights 0.0 GB0.5 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 2060

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.

    # nishtahir/ice40GPT
    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 nishtahir/ice40GPT --local-dir ~/models/ice40GPT
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to ice40GPT
  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 ice40GPT (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

nishtahir/ice40GPT
Author
nishtahir
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

240

Ask AI about this model

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

Comments