Spaces:

bibibi12345
/

llada-image-base

Running on Zero

App Files Files Community

llada-image-base

135 kB

Ctrl+K

Ctrl+K

1 contributor

History: 17 commits

bibibi12345's picture

revert to bf16 original + drop nvidia-modelopt (xlarge 96GB holds 49GB bf16 natively, no quant layer needed)

4ee3e63 about 2 months ago

repo
revert vendored pipeline hacks: original self.transformer.device is correct for full-resident bf16 on xlarge 96GB (offload-era patches caused this mess) about 2 months ago
.gitattributes

1.52 kB
init: LLaDA-Image Base (50 steps) on ZeroGPU, src vendored (binary assets dropped) about 2 months ago
.gitignore

32 Bytes
init: LLaDA-Image Base (50 steps) on ZeroGPU, src vendored (binary assets dropped) about 2 months ago
README.md

194 Bytes
init: LLaDA-Image Base (50 steps) on ZeroGPU, src vendored (binary assets dropped) about 2 months ago
app.py

3.41 kB
fix: remove cpu offload entirely - xlarge 96GB holds all bf16 weights (49.2G) resident; cuda generator; offload was causing scheduler.step device mismatch about 2 months ago
requirements.txt

105 Bytes
revert to bf16 original + drop nvidia-modelopt (xlarge 96GB holds 49GB bf16 natively, no quant layer needed) about 2 months ago