diff --git a/cogvideox_fun/pipeline_cogvideox_inpaint.py b/cogvideox_fun/pipeline_cogvideox_inpaint.py
index 7b9d8e7..a6f0e9e 100644
--- a/cogvideox_fun/pipeline_cogvideox_inpaint.py
+++ b/cogvideox_fun/pipeline_cogvideox_inpaint.py
@@ -739,6 +739,8 @@ class CogVideoX_Fun_Pipeline_Inpaint(DiffusionPipeline):
         num_channels_transformer = self.transformer.config.in_channels
         return_image_latents = num_channels_transformer == num_channels_latents
 
+        self.vae.to(device)
+
         latents_outputs = self.prepare_latents(
             batch_size * num_videos_per_prompt,
             num_channels_latents,
@@ -840,6 +842,9 @@ class CogVideoX_Fun_Pipeline_Inpaint(DiffusionPipeline):
                 mask = rearrange(mask, "b c f h w -> b f c h w")
 
                 inpaint_latents = None
+
+        self.vae.to(torch.device("cpu"))
+
         if comfyui_progressbar:
             pbar.update(1)
 
diff --git a/model_loading.py b/model_loading.py
index d1482d3..e627351 100644
--- a/model_loading.py
+++ b/model_loading.py
@@ -121,7 +121,7 @@ class DownloadAndLoadCogVideoModel:
                 "precision": (["fp16", "fp32", "bf16"],
                     {"default": "bf16", "tooltip": "official recommendation is that 2b model should be fp16, 5b model should be bf16"}
                 ),
-                "fp8_transformer": (['disabled', 'enabled', 'fastmode', 'torchao_fp8dq', "torchao_fp6"], {"default": 'disabled', "tooltip": "enabled casts the transformer to torch.float8_e4m3fn, fastmode is only for latest nvidia GPUs and requires torch 2.4.0 and cu124 minimum"}),
+                "fp8_transformer": (['disabled', 'enabled', 'fastmode', 'torchao_fp8dq', "torchao_fp8dqrow", "torchao_int8dq", "torchao_fp6"], {"default": 'disabled', "tooltip": "enabled casts the transformer to torch.float8_e4m3fn, fastmode is only for latest nvidia GPUs and requires torch 2.4.0 and cu124 minimum"}),
                 "compile": (["disabled","onediff","torch"], {"tooltip": "compile the model for faster inference, these are advanced options only available on Linux, see readme for more info"}),
                 "enable_sequential_cpu_offload": ("BOOLEAN", {"default": False, "tooltip": "significantly reducing memory usage and slows down the inference"}),
                 "block_edit": ("TRANSFORMERBLOCKS", {"default": None}),
@@ -301,7 +301,8 @@ class DownloadAndLoadCogVideoModel:
                 from torchao.quantization import (
                 quantize_,
                 fpx_weight_only,
-                float8_dynamic_activation_float8_weight
+                float8_dynamic_activation_float8_weight,
+                int8_dynamic_activation_int8_weight
             )
             except:
                 raise ImportError("torchao is not installed, please install torchao to use fp8dq")
@@ -316,11 +317,16 @@ class DownloadAndLoadCogVideoModel:
                 quant_func = fpx_weight_only(3, 2)
             elif "fp8dq" in fp8_transformer: #very fast on 4090 when compiled
                 quant_func = float8_dynamic_activation_float8_weight()
+            elif 'fp8dqrow' in fp8_transformer:
+                from torchao.quantization.quant_api import PerRow
+                quant_func = float8_dynamic_activation_float8_weight(granularity=PerRow())
+            elif 'int8dq' in fp8_transformer:
+                quant_func = int8_dynamic_activation_int8_weight()
         
             for i, block in enumerate(pipe.transformer.transformer_blocks):
                 if "CogVideoXBlock" in str(block):
                     quantize_(block, quant_func, filter_fn=filter_fn)
-                
+                        
             manual_offloading = False # to disable manual .to(device) calls
         
         if enable_sequential_cpu_offload:
diff --git a/nodes.py b/nodes.py
index dd9589a..b18a978 100644
--- a/nodes.py
+++ b/nodes.py
@@ -1100,10 +1100,6 @@ class CogVideoXFunSampler:
         base_path = pipeline["base_path"]
         assert "fun" in base_path.lower(), "'Unfun' models not supported in 'CogVideoXFunSampler', use the 'CogVideoSampler'"
         assert "pose" not in base_path.lower(), "'Pose' models not supported in 'CogVideoXFunSampler', use the 'CogVideoXFunControlSampler'"
-        
-
-        if not pipeline["cpu_offloading"]:
-            pipe.enable_model_cpu_offload(device=device)
 
         mm.soft_empty_cache()
 
@@ -1123,8 +1119,8 @@ class CogVideoXFunSampler:
         else:
             raise ValueError(f"Unknown scheduler: {scheduler}")
 
-        #if not pipeline["cpu_offloading"]:
-        #    pipe.transformer.to(device)
+        if not pipeline["cpu_offloading"] and pipeline["manual_offloading"]:
+            pipe.transformer.to(device)
 
         if context_options is not None:
             context_frames = context_options["context_frames"] // 4
@@ -1184,8 +1180,8 @@ class CogVideoXFunSampler:
                 noise_aug_strength = noise_aug_strength,
                 strength = vid2vid_denoise,
             )
-        #if not pipeline["cpu_offloading"]:
-        #     pipe.transformer.to(offload_device)
+        if not pipeline["cpu_offloading"] and pipeline["manual_offloading"]:
+            pipe.transformer.to(offload_device)
         #clear FasterCache
         if fastercache is not None:
             for block in pipe.transformer.transformer_blocks: