Downloads · 30 days
12
15% of all-time downloads
nemotron-diff-vlm-8b-fp8
yury-kapoyko/nemotron-diff-vlm-8b-fp8
nemotron-diff-vlm-8b-fp8 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.
FP8 weight-only (W8A16) quantization of nvidia/Nemotron-Labs-Diffusion-VLM-8B — FP8 (e4m3) weights with per-tensor weightscale, activations kept BF16. Roughly halves the checkpoint size vs BF16 with near-BF16 answer p…
All-time downloads
78
Public
Parameters
8.9B
19.8 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors10.4 GB · 100%
Parameter types
How the weights are stored.
F8_E4M37.4B · 83%
Base models
Model card
From the Hugging Face model README
Nemotron-Labs-Diffusion-VLM-8B — FP8 weight-only (W8A16)
FP8 weight-only (W8A16) quantization of
nvidia/Nemotron-Labs-Diffusion-VLM-8B
— FP8 (e4m3) weights with per-tensor weight_scale, activations kept
BF16. Roughly halves the checkpoint size vs BF16 with near-BF16 answer
parity.
- Made with NVIDIA Model Optimizer (modelopt) 0.45.0,
fp8recipe in weight-only mode, 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*. - Verified on a 48-image foot-scan set vs manual ground truth: 94.79% answer parity (BF16 baseline 95.31%), 48/48 valid JSON, one answer drift vs BF16 of 192.
- Runtime: on sm_89+ (Ada) and newer, FP8 scaled-matrix-multiply kernels run natively; on older GPUs (e.g. sm_86 Ampere) weights are decompressed to BF16 at load.
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 15, 2026
- Updated
- Sep 23, 2026