Should INT8 work on RDNA3?

[AlfredSartan572]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Expected Behavior

Running a Krea-2 Turbo workflow using an INT8 model works without crashing (and is faster than when using a FP8 model?)

Actual Behavior

The INT8 model I first downloaded causes a crash in comfyUI : Unknown quantization format for layer blocks.0.attn.wq
I later downloaded other INT8 models and these seem to work.
However, they are 2 to 3 times slower than when using FP8 or GGUF Q8.

As the Radeon RX 7900 XTX supports INT8 natively, but not FP8, I expected INT8 to be faster than FP8 (but maybe my expectations are wrong?)

Q1: is the crash with that INT8 model normal?
Q2: is the 2- to 3-times slowdown normal?

Steps to Reproduce

Model:
download the Krea2-Turbo-int8-ConvRot.safetensors model from https://huggingface.co/lilcheaty/Krea2-INT8-ConvRot/tree/main and use it in a standard Krea-2 text-to-image workflow.

Environment Variables:
export HSA_OVERRIDE_GFX_VERSION=11.0.0
export TORCH_BLAS_PREFER_HIPBLASLT=1
export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
export TORCH_ROCM_AOTRITON_ENABLE=1
export FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE"
COMFYUI_ENABLE_MIOPEN=0
MIOPEN_ENABLE_LOGGING=0
MIOPEN_FIND_MODE=DYNAMIC_HYBRID
MIOPEN_DEBUG_CONV_DIRECT=0
MIOPEN_DEBUG_CONV_DIRECT_NAIVE_CONV_FWD=0

output image resolution:
1024 x 1024

Lora's :
none

Command line arguments:
--use-flash-attention --enable-dynamic-vram
(other command line arguments all give the same result)

This causes a crash. (see debug log below)

Additionally:
Now run the same flow, with the same settings and arguments, but use krea2turbo_INT8_comfyfixed.safetensors from https://huggingface.co/tsolful/Krea2_Turbo_Raw_INT8/tree/main
This works correctly, but the execution time is 2 to 3 times slower. (see "other" section below for timing results)

comfyui_INT8_v4[OK].log

Krea2_FP8_INT8.json

Debug Logs

## ComfyUI-Manager: installing dependencies done.
[2026-07-05 12:04:20.095] ** ComfyUI startup time: 2026-07-05 12:04:20.095
[2026-07-05 12:04:20.095] ** Platform: Linux
[2026-07-05 12:04:20.095] ** Python version: 3.12.3 (main, Mar 23 2026, 19:04:32) [GCC 13.3.0]
[2026-07-05 12:04:20.095] ** Python executable: ~/comfyui-venv/bin/python3
[2026-07-05 12:04:20.095] ** ComfyUI Path: ~/ComfyUI
[2026-07-05 12:04:20.095] ** ComfyUI Base Folder Path: ~/ComfyUI
[2026-07-05 12:04:20.095] ** User directory: ~/ComfyUI/user
[2026-07-05 12:04:20.095] ** ComfyUI-Manager config path: ~/ComfyUI/user/__manager/config.ini
[2026-07-05 12:04:20.095] ** Log path: ~/ComfyUI/user/comfyui.log

Prestartup times for custom nodes:
[2026-07-05 12:04:20.170]    0.0 seconds: ~/ComfyUI/custom_nodes/rgthree-comfy
[2026-07-05 12:04:20.170]    0.0 seconds: ~/ComfyUI/custom_nodes/comfyui-easy-use
[2026-07-05 12:04:20.170]    0.4 seconds: ~/ComfyUI/custom_nodes/ComfyUI-Manager
[2026-07-05 12:04:20.170] 
[2026-07-05 12:04:21.475] Found comfy_kitchen backend cuda: {'available': True, 'disabled': True, 'unavailable_reason': None, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'dequantize_int8_convrot_weight', 'dequantize_int8_convrot_weight_dtype', 'dequantize_int8_simple', 'dequantize_int8_simple_dtype', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'gemv_awq_w4a16', 'quantize_and_rotate_rowwise', 'quantize_int8_convrot_weight', 'quantize_int8_rowwise', 'quantize_int8_tensorwise', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8']}
[2026-07-05 12:04:21.475] Found comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'dequantize_int8_convrot_weight', 'dequantize_int8_convrot_weight_dtype', 'dequantize_int8_simple', 'dequantize_int8_simple_dtype', 'dequantize_mxfp8', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'gemv_awq_w4a16', 'int8_linear', 'quantize_and_rotate_rowwise', 'quantize_int8_convrot_weight', 'quantize_int8_rowwise', 'quantize_int8_tensorwise', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'scaled_mm_mxfp8', 'scaled_mm_nvfp4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8']}
[2026-07-05 12:04:21.475] Found comfy_kitchen backend triton: {'available': True, 'disabled': True, 'unavailable_reason': None, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'int8_linear', 'quantize_and_rotate_rowwise', 'quantize_int8_rowwise', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8']}
[2026-07-05 12:04:21.478] Checkpoint files will always be loaded safely.
[2026-07-05 12:04:21.484] Total VRAM 24560 MB, total RAM 63938 MB
[2026-07-05 12:04:21.484] pytorch version: 2.9.1+rocm7.2.4.git39497456
[2026-07-05 12:04:21.484] Set: torch.backends.cudnn.enabled = False for better AMD performance.
[2026-07-05 12:04:21.485] AMD arch: gfx1100
[2026-07-05 12:04:21.485] ROCm version: (7, 2)
[2026-07-05 12:04:21.485] Set vram state to: NORMAL_VRAM
[2026-07-05 12:04:21.485] Device: cuda:0 AMD Radeon RX 7900 XTX : native
[2026-07-05 12:04:21.485] Using async weight offloading with 2 streams
[2026-07-05 12:04:21.485] Enabled pinned memory 57544.0
[2026-07-05 12:04:21.940] [aiter] import [module_aiter_core] under ~/comfyui-venv/lib/python3.12/site-packages/aiter/jit/module_aiter_core.so
[2026-07-05 12:04:22.513] Using Flash Attention
[2026-07-05 12:04:23.724] DynamicVRAM support detected and enabled
[2026-07-05 12:04:23.724] Python version: 3.12.3 (main, Mar 23 2026, 19:04:32) [GCC 13.3.0]
[2026-07-05 12:04:23.724] ComfyUI version: 0.27.0
[2026-07-05 12:04:23.724] comfy-aimdo version: 0.4.10
[2026-07-05 12:04:23.724] comfy-kitchen version: 0.2.16
[2026-07-05 12:04:23.726] comfyui-frontend-package version: 1.45.20
[2026-07-05 12:04:23.726] comfyui-workflow-templates version: 0.11.2
[2026-07-05 12:04:23.726] comfyui-embedded-docs version: 0.5.6
[2026-07-05 12:04:23.726] comfy-kitchen version: 0.2.16
[2026-07-05 12:04:23.726] comfy-aimdo version: 0.4.10
[2026-07-05 12:04:23.726] [Prompt Server] web root: ~/comfyui-venv/lib/python3.12/site-packages/comfyui_frontend_package/static
[2026-07-05 12:04:23.727] Asset seeder disabled
[2026-07-05 12:04:23.913] No OpenGL_accelerate module loaded: Acceleration disabled
[2026-07-05 12:04:24.276] 
[2026-07-05 12:04:24.276] �[92m[rgthree-comfy] Loaded 48 exciting nodes. 🎉�[0m
[2026-07-05 12:04:24.276] 
[2026-07-05 12:04:24.276] �[33m[rgthree-comfy] ComfyUI's new Node 2.0 rendering may be incompatible with some rgthree-comfy nodes and features, breaking some rendering as well as losing the ability to access a node's properties (a vital part of many nodes). It also appears to run MUCH more slowly spiking CPU usage and causing jankiness and unresponsiveness, especially with large workflows. Personally I am not planning to use the new Nodes 2.0 and, unfortunately, am not able to invest the time to investigate and overhaul rgthree-comfy where needed. If you have issues when Nodes 2.0 is enabled, I'd urge you to switch it off as well and join me in hoping ComfyUI is not planning to deprecate the existing, stable canvas rendering all together.
�[0m
[2026-07-05 12:04:24.292] ### Loading: ComfyUI-Impact-Pack (V8.28.3)
[2026-07-05 12:04:24.566] [Impact Pack] Wildcard total size (0.00 MB) is within cache limit (50.00 MB). Using full cache mode.
[2026-07-05 12:04:24.566] [Impact Pack] Wildcards loading done.
[2026-07-05 12:04:24.568] ### Loading: ComfyUI-Impact-Subpack (V1.3.5)
[2026-07-05 12:04:24.569] [Impact Pack/Subpack] Using folder_paths to determine whitelist path: ~/ComfyUI/user/default/ComfyUI-Impact-Subpack/model-whitelist.txt
[2026-07-05 12:04:24.569] [Impact Pack/Subpack] Ensured whitelist directory exists: ~/ComfyUI/user/default/ComfyUI-Impact-Subpack
[2026-07-05 12:04:24.569] [Impact Pack/Subpack] Loaded 0 model(s) from whitelist: ~/ComfyUI/user/default/ComfyUI-Impact-Subpack/model-whitelist.txt
[2026-07-05 12:04:24.749] [Impact Subpack] ultralytics_bbox: ~/ComfyUI/models/ultralytics/bbox
[2026-07-05 12:04:24.750] [Impact Subpack] ultralytics_segm: ~/ComfyUI/models/ultralytics/segm
[2026-07-05 12:04:24.760] ### Loading: ComfyUI-Manager (V3.41)
[2026-07-05 12:04:24.760] [ComfyUI-Manager] network_mode: offline
[2026-07-05 12:04:24.761] [ComfyUI-Manager] ComfyUI per-queue preview override detected (PR #11261). Manager's preview method feature is disabled. Use ComfyUI's --preview-method CLI option or 'Settings > Execution > Live preview method'.
[2026-07-05 12:04:24.797] ### ComfyUI Version: v0.27.0-15-g6c62ca0b | Released on '2026-07-04'
[2026-07-05 12:04:24.800] [ComfyUI-Manager] All startup tasks have been completed.
[2026-07-05 12:04:25.336] �[34m[ComfyUI-Easy-Use] server: �[0mv1.3.6 �[92mLoaded�[0m
[2026-07-05 12:04:25.336] �[34m[ComfyUI-Easy-Use] web root: �[0m~/ComfyUI/custom_nodes/comfyui-easy-use/web_version/v2 �[92mLoaded�[0m
[2026-07-05 12:04:25.338] ComfyUI-GGUF: Allowing full torch compile
[2026-07-05 12:04:25.340] 
Import times for custom nodes:
[2026-07-05 12:04:25.340]    0.0 seconds: ~/ComfyUI/custom_nodes/websocket_image_save.py
[2026-07-05 12:04:25.340]    0.0 seconds: ~/ComfyUI/custom_nodes/ComfyUI-ConditioningKrea2Rebalance
[2026-07-05 12:04:25.340]    0.0 seconds: ~/ComfyUI/custom_nodes/ComfyUI-GGUF
[2026-07-05 12:04:25.340]    0.0 seconds: ~/ComfyUI/custom_nodes/ComfyUI-PromptRelay
[2026-07-05 12:04:25.340]    0.0 seconds: ~/ComfyUI/custom_nodes/rgthree-comfy
[2026-07-05 12:04:25.340]    0.0 seconds: ~/ComfyUI/custom_nodes/comfyui_essentials
[2026-07-05 12:04:25.340]    0.0 seconds: ~/ComfyUI/custom_nodes/ComfyUI-KJNodes
[2026-07-05 12:04:25.340]    0.0 seconds: ~/ComfyUI/custom_nodes/ComfyUI-Manager
[2026-07-05 12:04:25.340]    0.2 seconds: ~/ComfyUI/custom_nodes/comfyui-impact-subpack
[2026-07-05 12:04:25.340]    0.3 seconds: ~/ComfyUI/custom_nodes/comfyui-impact-pack
[2026-07-05 12:04:25.340]    0.5 seconds: ~/ComfyUI/custom_nodes/comfyui-easy-use
[2026-07-05 12:04:25.340] 
[2026-07-05 12:04:25.341] Context impl SQLiteImpl.
[2026-07-05 12:04:25.341] Will assume non-transactional DDL.
[2026-07-05 12:04:25.370] Using RAM pressure cache.
[2026-07-05 12:04:25.370] Starting server

[2026-07-05 12:04:25.371] To see the GUI go to: http://127.0.0.1:8188
[2026-07-05 12:04:31.026] [DEPRECATION WARNING] Detected import of deprecated legacy API: /scripts/ui.js. This is likely caused by a custom node extension using outdated APIs. Please update your extensions or contact the extension author for an updated version.
[2026-07-05 12:04:31.026] [DEPRECATION WARNING] Detected import of deprecated legacy API: /extensions/core/clipspace.js. This is likely caused by a custom node extension using outdated APIs. Please update your extensions or contact the extension author for an updated version.
[2026-07-05 12:04:31.027] [DEPRECATION WARNING] Detected import of deprecated legacy API: /extensions/core/groupNode.js. This is likely caused by a custom node extension using outdated APIs. Please update your extensions or contact the extension author for an updated version.
[2026-07-05 12:04:31.623] [DEPRECATION WARNING] Detected import of deprecated legacy API: /scripts/ui/components/buttonGroup.js. This is likely caused by a custom node extension using outdated APIs. Please update your extensions or contact the extension author for an updated version.
[2026-07-05 12:04:31.633] [DEPRECATION WARNING] Detected import of deprecated legacy API: /scripts/ui/components/button.js. This is likely caused by a custom node extension using outdated APIs. Please update your extensions or contact the extension author for an updated version.
[2026-07-05 12:04:36.991] got prompt
[2026-07-05 12:04:37.000] Using split attention in VAE
[2026-07-05 12:04:37.001] Using split attention in VAE
[2026-07-05 12:04:37.075] VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
[2026-07-05 12:04:37.378] Found quantization metadata version 1
[2026-07-05 12:04:37.378] Using MixedPrecisionOps for text encoder
[2026-07-05 12:04:37.606] CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
[2026-07-05 12:04:37.615] Found quantization metadata version 1
[2026-07-05 12:04:37.615] Detected mixed precision quantization
[2026-07-05 12:04:37.615] Using mixed precision operations
[2026-07-05 12:04:37.615] Native ops: int8_tensorwise , emulated ops: float8_e5m2, float8_e4m3fn, mxfp8, nvfp4
[2026-07-05 12:04:37.622] model weight dtype torch.bfloat16, manual cast: torch.bfloat16
[2026-07-05 12:04:37.623] model_type FLUX
[2026-07-05 12:04:37.625] !!! Exception during processing !!! Unknown quantization format for layer blocks.0.attn.wq
[2026-07-05 12:04:37.628] Traceback (most recent call last):
  File "~/ComfyUI/execution.py", line 542, in execute
    output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
                                                              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "~/ComfyUI/execution.py", line 341, in get_output_data
    return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "~/ComfyUI/execution.py", line 315, in _async_map_node_over_list
    await process_inputs(input_dict, i)
  File "~/ComfyUI/execution.py", line 303, in process_inputs
    result = f(**inputs)
             ^^^^^^^^^^^
  File "~/ComfyUI/nodes.py", line 988, in load_unet
    model = comfy.sd.load_diffusion_model(unet_path, model_options=model_options)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "~/ComfyUI/comfy/sd.py", line 2055, in load_diffusion_model
    model = load_diffusion_model_state_dict(sd, model_options=model_options, metadata=metadata, disable_dynamic=disable_dynamic)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "~/ComfyUI/comfy/sd.py", line 2047, in load_diffusion_model_state_dict
    model.load_model_weights(new_sd, "", assign=model_patcher.is_dynamic())
  File "~/ComfyUI/comfy/model_base.py", line 352, in load_model_weights
    m, u = self.diffusion_model.load_state_dict(to_load, strict=False, assign=assign)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "~/comfyui-venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2609, in load_state_dict
    load(self, state_dict)
  File "~/comfyui-venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2597, in load
    load(child, child_state_dict, child_prefix)  # noqa: F821
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "~/comfyui-venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2597, in load
    load(child, child_state_dict, child_prefix)  # noqa: F821
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "~/comfyui-venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2597, in load
    load(child, child_state_dict, child_prefix)  # noqa: F821
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  [Previous line repeated 1 more time]
  File "~/comfyui-venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 2580, in load
    module._load_from_state_dict(
  File "~/ComfyUI/comfy/ops.py", line 1194, in _load_from_state_dict
    _load_quantized_module(self, super()._load_from_state_dict, *args, load_extra_params=True)
  File "~/ComfyUI/comfy/ops.py", line 1072, in _load_quantized_module
    raise ValueError(f"Unknown quantization format for layer {layer_name}")
ValueError: Unknown quantization format for layer blocks.0.attn.wq

[2026-07-05 12:04:37.629] Prompt executed in 0.64 seconds

Other

My Setup:
CPU: AMD Ryzen 9 7950X
GPU: AMD Radeon RX 7900 XTX (RDNA3, gfx1100)
OS: Linux Mint 22.3 (kernel 6.17.0-35-generic)
ROCm: 7.2.4
ComfyUI: v0.27.0-15-g6c62ca0b (2026-07-04)
Python : 3.12.3
Python torch package : 2.9.1+rocm7.2.4.lw.git39497456

INT8 model that crashes:
Krea2-Turbo-int8-ConvRot.safetensors from https://huggingface.co/lilcheaty/Krea2-INT8-ConvRot/tree/main
INT8 model that works (INT8_v4 in timings table below):
krea2turbo_INT8_comfyfixed.safetensors from https://huggingface.co/tsolful/Krea2_Turbo_Raw_INT8/tree/main

Timing Results:

quad      = --use-quad-cross-attention
quad-dyn  = --use-quad-cross-attention --enable-dynamic-vram
torch     = --use-pytorch-cross-attention
torch-dyn = --use-pytorch-cross-attention --enable-dynamic-vram
flash     = --use-flash-attention
flash-dyn = --use-flash-attention --enable-dynamic-vram

             quad            quad-dyn
         Run 1   Run 2     Run 1   Run 2
FP8      37.68   18.88     25.55   18.86
Q8       31.81   19.52     28.86   19.55 
Int8_v2  60.66   55.56     46.34   43.10 
Int8_v3  70.61   54.33     45.82   42.22 
Int8_v4  60.03   54.17     45.22   42.85 

             torch           torch-dyn
         Run 1   Run 2     Run 1   Run 2
FP8      32.95   19.52     24.49   19.61
Q8       26.28   19.94     29.36   20.06 
Int8_v2  63.73   56.85     47.00   44.26 
Int8_v3  69.41   56.35     47.36   43.29 
Int8_v4  64.09   55.40     45.77   43.47 

             flash           flash-dyn
         Run 1   Run 2     Run 1   Run 2
FP8      30.80   16.86     22.92   16.90
Q8       28.05   17.42     26.28   17.42 
Int8_v2  64.74   53.88     45.15   42.04 
Int8_v3  66.92   53.06     46.06   41.37 
Int8_v4  63.12   51.96     45.83   41.28 

[AlfredSartan572]

There is a INT8-Fast-ROCM custom node: https://github.com/patientx/ComfyUI-INT8-Fast-ROCM
On that page, I saw : "INT8 is now officially supported in ComfyUI"
I understood this as: this custom node is no longer needed.
But that does not appear to be the case.
After installing this custom node, and using their "Load Diffusion Model INT8 (W8A8)" node:

• the timings are much better. INT8 is now about 19% faster than FP8
• the crash with that Krea2-INT8 model no longer happens
  => something in the standard ComfyUI Diffusion Model loader does not like this model and causes it to crash.

Timings including the Fast-ROCM diffusion model loader:

             quad            quad-dyn      quad-dyn-fast-rocm
         Run 1   Run 2     Run 1   Run 2     Run 1   Run 2
FP8      37.68   18.88     25.55   18.86       -      -
Q8       31.81   19.52     28.86   19.55       -      - 
Int8_v1  crash   crash     crash   crash     23.83   16.45
Int8_v2  60.66   55.56     46.34   43.10     33.24   16.26 
Int8_v3  70.61   54.33     45.82   42.22     35.14   15.59 
Int8_v4  60.03   54.17     45.22   42.85     28.44   16.39 

             torch           torch-dyn     torch-dyn-fast-rocm
         Run 1   Run 2     Run 1   Run 2     Run 1   Run 2
FP8      32.95   19.52     24.49   19.61       -      -
Q8       26.28   19.94     29.36   20.06       -      - 
Int8_v1  crash   crash     crash   crash     24.36   16.92   
Int8_v2  63.73   56.85     47.00   44.26     33.86   16.99
Int8_v3  69.41   56.35     47.36   43.29     33.17   16.38 
Int8_v4  64.09   55.40     45.77   43.47     28.76   17.01 

             flash           flash-dyn     flash-dyn-fast-rocm
         Run 1   Run 2     Run 1   Run 2     Run 1   Run 2
FP8      30.80   16.86     22.92   16.90       -      -
Q8       28.05   17.42     26.28   17.42       -      - 
Int8_v1  crash   crash     crash   crash     24.48   14.36 
Int8_v2  64.74   53.88     45.15   42.04     31.14   14.26
Int8_v3  66.92   53.06     46.06   41.37     30.45   13.68 
Int8_v4  63.12   51.96     45.83   41.28     26.15   14.34

[0xDELUXA]

INT8 ConvRot works and delivers better performance than FP8 on RDNA2-4. On AMD, this node is both faster and more stable than the native implementation. As noted:

Long story short, this is still relevant for AMD users until the ComfyUI-Kitchen team fixes the native performance issues.

[0xDELUXA]

@AlfredSartan572 The model you're using isn't in the native ComfyUI INT8 format. Use the custom node. You can convert it if you prefer, but I wouldn't recommend it.

[AlfredSartan572]

Clear.
Thank you for the feedback.