Skip to content
Computer vision

ParaThinker Qwen3-VL-8B SFT

финальный чекпоинт эксперимента

1 min readupdated

Это финальный checkpoint-153 SFT-эксперимента standard_sft_qwen3vl_plus_lr4e-5_gbs64_ep1, обученный от Qwen/Qwen3-VL-8B-Instruct на выровненных teacher-данных Qwen3VL-Plus Full-CG. Модель для image-text-to-text и мультираундовых видео-/VQA-сценариев ParaThinker. Веса сохранены в оригинальном FP32, без квантизации и конвертации.

Ключевые особенности

  • База — Qwen3-VL-8B-Instruct, зафиксированная revision; обучение с uniform CE.
  • SFT: LR 4e-5, global batch 64, 1 эпоха / 153 optimizer updates, 6 480 804 supervised tokens.
  • Обучение на 8×H100 80GB с FSDP1 full_shard; супервизия LLM и main visual merger.
  • Teacher-данные: corrected Qwen3VL-Plus Full-CG, 9 786 calls / 3 829 questions.
  • Сохранены все 30 файлов чекпоинта, 126.66 GiB: safetensors, processor/tokenizer, optimizer и FSDP state, scheduler, Trainer state и RNG.

Пример

import torch
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

repo = "Polly1231/ParaThinker-Qwen3-VL-8B-Instruct-SFT-Qwen3VLPlus-lr4e-5-gbs64-ep1-step153"
processor = AutoProcessor.from_pretrained(repo)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    repo, dtype=torch.bfloat16, device_map="auto"
)
ПодробнееLVBench: 385 correct / 1 549 questions, точность 24.8547%; 538 protocol failures засчитаны как incorrect. Это multi-round protocol, не one-shot VQA.
Для загрузки inference-весов оптимизаторные файлы не нужны.
LongVideoBench пока без финального аудированного результата в карточке.
Доступ к приватному архиву эксперимента управляется отдельно; teacher-данные и оригинальные изображения/видео не включены.
Discuss Telegram 0 20

Will it run?

Runs in BF16

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • F32weights 33 GB · ready files36 GB
  • BF16weights 2.0 GB2.7 GB
  • FP8 / INT8recommendedweights 1.0 GB1.6 GB
  • 4-bitweights 0.6 GB1.1 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

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.

    # Polly1231/ParaThinker-Qwen3-VL-8B-Instruct-SFT-Qwen3VLPlus-lr4e-5-gbs64-ep1-step153
    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 Polly1231/ParaThinker-Qwen3-VL-8B-Instruct-SFT-Qwen3VLPlus-lr4e-5-gbs64-ep1-step153 --local-dir ~/models/ParaThinker-Qwen3-VL-8B-Instruct-SFT-Qwen3VLPlus-lr4e-5-gbs64-ep1-step153
    # 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 Qwen3-VL-8B-Instruct (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 4All on Hugging Face
Merges 1All on Hugging Face

About the model

Polly1231/ParaThinker-Qwen3-VL-8B-Instruct-SFT-Qwen3VLPlus-lr4e-5-gbs64-ep1-step153
Author
Polly1231
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

200

Ask AI about this model

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

Comments