Downloads · 30 days
12
24% of all-time downloads
nemotron-diff-vlm-8b-w4a8_awq
yury-kapoyko/nemotron-diff-vlm-8b-w4a8_awq
nemotron-diff-vlm-8b-w4a8_awq is a machine learning model from yury-kapoyko. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.
INT4 + FP8-activations (W4A8) AWQ quantization of nvidia/Nemotron-Labs-Diffusion-VLM-8B: INT4 weights (group size 128, symmetric, per-group weightscale, activation smoothing via prequantscale) with FP8 per-token input…
All-time downloads
51
Public
Parameters
5.2B
7 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors6.9 GB · 100%
Parameter types
How the weights are stored.
U83.7B · 71%
Base models
Model card
From the Hugging Face model README
Nemotron-Labs-Diffusion-VLM-8B — INT4 W4A8 (AWQ)
INT4 + FP8-activations (W4A8) AWQ quantization of
nvidia/Nemotron-Labs-Diffusion-VLM-8B:
INT4 weights (group size 128, symmetric, per-group weight_scale,
activation smoothing via pre_quant_scale) with FP8 per-token input
activations. Smallest compressed variant of this model; needs FP8-capable
silicon to reach its full speed.
- Made with NVIDIA Model Optimizer (modelopt) 0.45.0 (
w4a8_awq), calibrated on 128 domain (foot-image, image-text) samples; exported throughcompressed-tensors. Excluded from quantization:diffusion_head,encoder.embed_tokens,encoder.multi_modal_projector*,encoder.vision_tower*. - Runtime: FP8 activation GEMMs require SM >= 8.9 (Ada Lovelace or newer). On older GPUs (e.g. sm_86 Ampere) the runtime decompresses to BF16 and answers can drift slightly vs BF16 on some inputs.
- Verified with an end-to-end golden (single-image JSON answers vs the BF16 reference) on an RTX 3090 Ti.
License
Derivative of the base model; use is governed by the NVIDIA Source Code License (NSCLv1) — see the base model card.
- Type
- nemotron_labs_diffusion_vlm
- License
- other
- Created
- Jul 24, 2026
- Updated
- Sep 23, 2026