Vikramjeet Singh

Results 8 comments of Vikramjeet Singh

hello @a-r-r-o-w , trying to convert the original genmo checkpoints into diffusers format , by the convert diffusers format using this script , the conversion for the encoder fails `encoder.down_blocks.2.norms.0.norm_layer.bias,...

hey @tin2tin , whats your parameters settings with this pipeline I am using something similar to this and got this , with cpu offload and vae tiling enabled ``` {...

Try using this checkpoint for the transformer https://huggingface.co/imnotednamode/mochi-1-preview-mix-nf4 in the pipeline , its nf4 quantized mixed in with bf16 gives decent results @tin2tin

Yes @nicollegah the documentation for that is out of date , but the checkpoint would work with it out of the box ``` transformer = MochiTransformer3DModel.from_pretrained("imnotednamode/mochi-1-preview-mix-nf4",torch_dtype=torch.bfloat16) pipe = MochiPipeline.from_pretrained("VikramSingh178/mochi-diffuser-bf16", torch_dtype=torch.bfloat16,transformer=transformer)...

In the meantime , this gets merged , is there any other way to run multiple endpoints in one main server ?