Skip to content
Computer vision

InternVL3-78B-WSP-SFT-Compressed

78B VLM, сжатая до 8,68 ГиБ

1 min readupdated

Это структурно сжатая, WSP- и мультимодально дообученная версия OpenGVLab/InternVL3-78B. Исходные 78,4B параметров и 146,05 ГиБ упакованы в один checkpoint model.safetensors на 8,68 ГиБ — 16,8-кратное сокращение. Модель работает через Hugging Face transformers на CPU и GPU, язык обучения — английский.

Ключевые особенности

  • Сжатие InternViT-6B: 45→8 слоёв, MLP 12800→3200; LLM Qwen2.5-72B: 80→6 слоёв, SwiGLU 29568→4096.
  • Сохранены 3200-мерные динамические patch embeddings для изображений, видео, UI/UX-схем, сетевых топологий и плотного OCR.
  • 8 эпох joint multimodal SFT на 16 учебных направлениях: image/video, UI→React/Tailwind, аудит сетей, software engineering, CEH, pentest и web security.
  • Один portable checkpoint 8,68 ГиБ; запуск нативно через transformers.
  • Карточка и обучение ориентированы на английский язык.

Пример

from transformers import AutoModel, AutoTokenizer
model_id = "EverythingIsForGod/InternVL3-78B-WSP-SFT-Compressed"
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True, use_fast=False)
model = AutoModel.from_pretrained(model_id, torch_dtype="bfloat16", low_cpu_mem_usage=True, trust_remote_code=True).eval()
response = model.chat(tok, pixel_values=None, question="Explain TOCTOU prevention in PostgreSQL.", generation_config=dict(max_new_tokens=256, do_sample=False))
print(response)
ПодробнееСжатие: 146,05 ГиБ → 8,68 ГиБ (16,8x); по метаданным — 4,66B параметров.
Запуск: CPU/GPU, transformers, trust_remote_code=True, use_flash_attn=False.
Ограничения: язык en, бенчмарки в карточке не указаны.
Discuss Telegram 1 37

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 8.7 GB · ready files10 GB
  • FP8 / INT8recommendedweights 4.3 GB5.3 GB
  • 4-bitweights 2.4 GB3.2 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 · BF16 — from RTX 2080 Ti

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.

    # EverythingIsForGod/InternVL3-78B-WSP-SFT-Compressed
    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 EverythingIsForGod/InternVL3-78B-WSP-SFT-Compressed --local-dir ~/models/InternVL3-78B-WSP-SFT-Compressed
    # 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

  1. InternViT-6B-448px-V1-5family root
  2. InternViT-6B-448px-V2_5
  3. InternVL3-78B-Pretrained
  4. InternVL3-78B-Instruct
  5. InternVL3-78B
  6. InternVL3-78B-WSP-SFT-Compressed

New posts about versions of InternViT-6B-448px-V1-5 (quants, finetunes, LoRAs) will show up under the bell.

Fine-tunes 1All on Hugging Face
Merges 6All on Hugging Face

About the model

EverythingIsForGod/InternVL3-78B-WSP-SFT-Compressed
Author
EverythingIsForGod
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

371

Ask AI about this model

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

Comments