DokuWiki

Benutzer-Werkzeuge

Webseiten-Werkzeuge

Zuletzt angesehen:
amd-ryzen-ai-max-395-ai-lab-notizen

Unterschiede

Hier werden die Unterschiede zwischen zwei Versionen angezeigt.

Link zu dieser Vergleichsansicht

Beide Seiten der vorigen RevisionVorhergehende Überarbeitung
amd-ryzen-ai-max-395-ai-lab-notizen [2026/01/30 20:44] – [nightly2] bennyamd-ryzen-ai-max-395-ai-lab-notizen [2026/01/31 14:45] (aktuell) – benny
Zeile 1478: Zeile 1478:
 </code> </code>
  
 +===== UEFI/BIOS: Performance Mode =====
 +
 +<code>
 +benny@ailab:~$ source rocm-nightly2/bin/activate
 +(rocm-nightly2) benny@ailab:~$ cd rocm-nightly2/ComfyUI/
 +(rocm-nightly2) benny@ailab:~/rocm-nightly2/ComfyUI$ python main.py     --listen 192.168.11.190     --port 8189     --gpu-only     --disable-smart-memory     --disable-auto-launch     --enable-manager     --disable-api-nodes     --enable-cors 
 +Adding extra search path text_encoders /home/benny/comfy-models/text_encoders
 +Adding extra search path vae /home/benny/comfy-models/vae
 +Adding extra search path diffusion_models /home/benny/comfy-models/diffusion_models
 +Adding extra search path unet /home/benny/comfy-models/unet
 +Adding extra search path loras /home/benny/comfy-models/loras
 +Adding extra search path latent_upscale_models /home/benny/comfy-models/latent_upscale_models
 +Adding extra search path clip_vision /home/benny/comfy-models/clip_vision
 +[START] Security scan
 +[ComfyUI-Manager] Using uv as Python module for pip operations.
 +Using Python 3.13.5 environment at: /home/benny/rocm-nightly2
 +[DONE] Security scan
 +** ComfyUI startup time: 2026-01-31 14:49:55.074
 +** Platform: Linux
 +** Python version: 3.13.5 (main, Jun 25 2025, 18:55:22) [GCC 14.2.0]
 +** Python executable: /home/benny/rocm-nightly2/bin/python
 +** ComfyUI Path: /home/benny/rocm-nightly2/ComfyUI
 +** ComfyUI Base Folder Path: /home/benny/rocm-nightly2/ComfyUI
 +** User directory: /home/benny/rocm-nightly2/ComfyUI/user
 +** ComfyUI-Manager config path: /home/benny/rocm-nightly2/ComfyUI/user/__manager/config.ini
 +** Log path: /home/benny/rocm-nightly2/ComfyUI/user/comfyui.log
 +Using Python 3.13.5 environment at: /home/benny/rocm-nightly2
 +Using Python 3.13.5 environment at: /home/benny/rocm-nightly2
 +[PRE] ComfyUI-Manager
 +Checkpoint files will always be loaded safely.
 +Total VRAM 63219 MB, total RAM 126438 MB
 +pytorch version: 2.9.1+rocm7.11.0a20260118
 +Set: torch.backends.cudnn.enabled = False for better AMD performance.
 +AMD arch: gfx1151
 +ROCm version: (7, 2)
 +Set vram state to: HIGH_VRAM
 +Disabling smart memory management
 +Device: cuda:0 Radeon 8060S Graphics : native
 +Using async weight offloading with 2 streams
 +Enabled pinned memory 120115.0
 +Found comfy_kitchen backend triton: {'available': True, 'disabled': True, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8']}
 +Found comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'scaled_mm_nvfp4']}
 +Found comfy_kitchen backend cuda: {'available': True, 'disabled': True, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8']}
 +Using pytorch attention
 +Python version: 3.13.5 (main, Jun 25 2025, 18:55:22) [GCC 14.2.0]
 +ComfyUI version: 0.11.1
 +ComfyUI frontend version: 1.37.11
 +[Prompt Server] web root: /home/benny/rocm-nightly2/lib/python3.13/site-packages/comfyui_frontend_package/static
 +[START] ComfyUI-Manager
 +[ComfyUI-Manager] network_mode: public
 +[ComfyUI-Manager] The matrix sharing feature has been disabled because the `matrix-nio` dependency is not installed.
 + To use this feature, please run the following command:
 + /home/benny/rocm-nightly2/bin/python -m pip install matrix-nio
 +
 +
 +Import times for custom nodes:
 +   0.0 seconds: /home/benny/rocm-nightly2/ComfyUI/custom_nodes/websocket_image_save.py
 +
 +Context impl SQLiteImpl.
 +Will assume non-transactional DDL.
 +Assets scan(roots=['models']) completed in 0.016s (created=0, skipped_existing=32, total_seen=32)
 +Starting server
 +
 +To see the GUI go to: http://192.168.11.190:8189
 +got prompt
 +Using split attention in VAE
 +Using split attention in VAE
 +VAE load device: cuda:0, offload device: cuda:0, dtype: torch.bfloat16
 +Found quantization metadata version 1
 +Using MixedPrecisionOps for text encoder
 +Requested to load QwenImageTEModel_
 +loaded completely;  7910.28 MB loaded, full load: True
 +CLIP/text encoder model load device: cuda:0, offload device: cuda:0, current: cuda:0, dtype: torch.float16
 +model weight dtype torch.bfloat16, manual cast: None
 +model_type FLUX
 +FETCH ComfyRegistry Data [DONE]
 +[ComfyUI-Manager] default cache updated: https://api.comfy.org/nodes
 +FETCH DATA from: /home/benny/rocm-nightly2/ComfyUI/user/__manager/cache/1514988643_custom-node-list.json [DONE]
 +[ComfyUI-Manager] All startup tasks have been completed.
 +Requested to load QwenImage
 +loaded completely;  38967.90 MB loaded, full load: True
 +100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 50/50 [12:25<00:00, 14.92s/it]
 +Requested to load WanVAE
 +loaded completely;  242.03 MB loaded, full load: True
 +Prompt executed in 00:13:20
 +</code>
 +
 +===== BIOS/UEFI Performance Mode + GPU 96 GB VRAM fest =====
 +
 +<code>
 +benny@ailab:~$ source rocm-nightly2/bin/activate
 +(rocm-nightly2) benny@ailab:~$ cd rocm-nightly2/ComfyUI/
 +(rocm-nightly2) benny@ailab:~/rocm-nightly2/ComfyUI$ python main.py     --listen 192.168.11.190     --port 8189     --gpu-only     --disable-smart-memory     --disable-auto-launch     --enable-manager     --disable-api-nodes     --enable-cors 
 +Adding extra search path text_encoders /home/benny/comfy-models/text_encoders
 +Adding extra search path vae /home/benny/comfy-models/vae
 +Adding extra search path diffusion_models /home/benny/comfy-models/diffusion_models
 +Adding extra search path unet /home/benny/comfy-models/unet
 +Adding extra search path loras /home/benny/comfy-models/loras
 +Adding extra search path latent_upscale_models /home/benny/comfy-models/latent_upscale_models
 +Adding extra search path clip_vision /home/benny/comfy-models/clip_vision
 +[START] Security scan
 +[ComfyUI-Manager] Using uv as Python module for pip operations.
 +Using Python 3.13.5 environment at: /home/benny/rocm-nightly2
 +[DONE] Security scan
 +** ComfyUI startup time: 2026-01-31 15:11:30.580
 +** Platform: Linux
 +** Python version: 3.13.5 (main, Jun 25 2025, 18:55:22) [GCC 14.2.0]
 +** Python executable: /home/benny/rocm-nightly2/bin/python
 +** ComfyUI Path: /home/benny/rocm-nightly2/ComfyUI
 +** ComfyUI Base Folder Path: /home/benny/rocm-nightly2/ComfyUI
 +** User directory: /home/benny/rocm-nightly2/ComfyUI/user
 +** ComfyUI-Manager config path: /home/benny/rocm-nightly2/ComfyUI/user/__manager/config.ini
 +** Log path: /home/benny/rocm-nightly2/ComfyUI/user/comfyui.log
 +Using Python 3.13.5 environment at: /home/benny/rocm-nightly2
 +Using Python 3.13.5 environment at: /home/benny/rocm-nightly2
 +[PRE] ComfyUI-Manager
 +Checkpoint files will always be loaded safely.
 +Total VRAM 98304 MB, total RAM 31734 MB
 +pytorch version: 2.9.1+rocm7.11.0a20260118
 +Set: torch.backends.cudnn.enabled = False for better AMD performance.
 +AMD arch: gfx1151
 +ROCm version: (7, 2)
 +Set vram state to: HIGH_VRAM
 +Disabling smart memory management
 +Device: cuda:0 Radeon 8060S Graphics : native
 +Using async weight offloading with 2 streams
 +Enabled pinned memory 30147.0
 +Found comfy_kitchen backend triton: {'available': True, 'disabled': True, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8']}
 +Found comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'scaled_mm_nvfp4']}
 +Found comfy_kitchen backend cuda: {'available': True, 'disabled': True, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8']}
 +Using pytorch attention
 +Python version: 3.13.5 (main, Jun 25 2025, 18:55:22) [GCC 14.2.0]
 +ComfyUI version: 0.11.1
 +ComfyUI frontend version: 1.37.11
 +[Prompt Server] web root: /home/benny/rocm-nightly2/lib/python3.13/site-packages/comfyui_frontend_package/static
 +[START] ComfyUI-Manager
 +[ComfyUI-Manager] network_mode: public
 +[ComfyUI-Manager] The matrix sharing feature has been disabled because the `matrix-nio` dependency is not installed.
 + To use this feature, please run the following command:
 + /home/benny/rocm-nightly2/bin/python -m pip install matrix-nio
 +
 +
 +Import times for custom nodes:
 +   0.0 seconds: /home/benny/rocm-nightly2/ComfyUI/custom_nodes/websocket_image_save.py
 +
 +Context impl SQLiteImpl.
 +Will assume non-transactional DDL.
 +Assets scan(roots=['models']) completed in 0.011s (created=0, skipped_existing=32, total_seen=32)
 +Starting server
 +
 +To see the GUI go to: http://192.168.11.190:8189
 +got prompt
 +Using split attention in VAE
 +Using split attention in VAE
 +VAE load device: cuda:0, offload device: cuda:0, dtype: torch.bfloat16
 +Found quantization metadata version 1
 +Using MixedPrecisionOps for text encoder
 +Requested to load QwenImageTEModel_
 +loaded completely;  7910.28 MB loaded, full load: True
 +CLIP/text encoder model load device: cuda:0, offload device: cuda:0, current: cuda:0, dtype: torch.float16
 +model weight dtype torch.bfloat16, manual cast: None
 +model_type FLUX
 +Requested to load QwenImage
 +loaded completely;  38967.90 MB loaded, full load: True
 +  2%|██▉                                                                                                                                             | 1/50 [00:00<00:06,  7.29it/s]FETCH ComfyRegistry Data [DONE]
 +[ComfyUI-Manager] default cache updated: https://api.comfy.org/nodes
 +FETCH DATA from: /home/benny/rocm-nightly2/ComfyUI/user/__manager/cache/1514988643_custom-node-list.json [DONE]
 +[ComfyUI-Manager] All startup tasks have been completed.
 +100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 50/50 [10:58<00:00, 13.17s/it]
 +Requested to load WanVAE
 +loaded completely;  242.03 MB loaded, full load: True
 +Prompt executed in 00:11:45
 +
 +</code>
 +
 +===== Distrobox mit Performance (UEFI) und 96 GB VRAM BIOS =====
 +
 +<code>
 +benny@ailab:~/rocm-nightly2/ComfyUI$ distrobox enter strix-halo-comfyui
 +Starting container...                    [ OK ]
 +Installing basic packages...            [ OK ]
 +Setting up devpts mounts...              [ OK ]
 +Setting up read-only mounts...          [ OK ]
 +Setting up read-write mounts...          [ OK ]
 +Setting up host's sockets integration... [ OK ]
 +Integrating host's themes, icons, fonts... [ OK ]
 +Setting up distrobox profile...          [ OK ]
 +Setting up sudo...                      [ OK ]
 +Setting up user's group list...          [ OK ]
 +
 +Container Setup Complete!
 +
 +███████╗████████╗██████╗ ██╗██╗  ██╗      ██╗  ██╗ █████╗ ██╗      ██████╗ 
 +██╔════╝╚══██╔══╝██╔══██╗██║╚██╗██╔╝      ██║  ██║██╔══██╗██║     ██╔═══██╗
 +███████╗   ██║   ██████╔╝██║ ╚███╔╝       ███████║███████║██║     ██║   ██║
 +╚════██║   ██║   ██╔══██╗██║ ██╔██╗       ██╔══██║██╔══██║██║     ██║   ██║
 +███████║   ██║   ██║  ██║██║██╔╝ ██╗      ██║  ██║██║  ██║███████╗╚██████╔╝
 +╚══════╝   ╚═╝   ╚═╝  ╚═╝╚═╝╚═╝  ╚═╝      ╚═╝  ╚═╝╚═╝  ╚═╝╚══════╝ ╚═════╝ 
 +
 +                          C O M F Y   U I                        
 +
 +
 +AMD Ryzen AI Max “Strix Halo” — Image & Video Toolbox (gfx1151, ROCm via TheRock)
 +ROCm nightly: 7.11.0a20260115
 +
 +Machine: Micro Computer (HK) Tech Limited MS-S1 MAX
 +GPU    : AMD RYZEN AI MAX+ 395 w/ Radeon 8060S
 +
 +Repo   : https://github.com/kyuz0/amd-strix-halo-comfyui-toolboxes
 +Image  : docker.io/kyuz0/amd-strix-halo-comfyui:latest
 +
 +Included:
 +  - ComfyUI          → start_comfy_ui (http://localhost:8000)
 +
 +SSH tip: ssh -L 8000:localhost:8000 user@host
 +
 +benny@ailab:~/rocm-nightly2/ComfyUI$ start_comfy_ui 
 +Adding extra search path text_encoders /home/benny/comfy-models/text_encoders
 +Adding extra search path vae /home/benny/comfy-models/vae
 +Adding extra search path diffusion_models /home/benny/comfy-models/diffusion_models
 +Adding extra search path unet /home/benny/comfy-models/unet
 +Adding extra search path loras /home/benny/comfy-models/loras
 +Adding extra search path latent_upscale_models /home/benny/comfy-models/latent_upscale_models
 +Adding extra search path clip_vision /home/benny/comfy-models/clip_vision
 +Setting output directory to: /home/benny/comfy-outputs
 +Checkpoint files will always be loaded safely.
 +Total VRAM 98304 MB, total RAM 31734 MB
 +pytorch version: 2.11.0a0+rocm7.11.0a20260115
 +Set: torch.backends.cudnn.enabled = False for better AMD performance.
 +AMD arch: gfx1151
 +ROCm version: (7, 2)
 +Set vram state to: NORMAL_VRAM
 +Device: cuda:0 Radeon 8060S Graphics : native
 +Using async weight offloading with 2 streams
 +Enabled pinned memory 30147.0
 +Found comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'scaled_mm_nvfp4']}
 +Found comfy_kitchen backend cuda: {'available': True, 'disabled': True, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8']}
 +Found comfy_kitchen backend triton: {'available': True, 'disabled': True, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8']}
 +Using pytorch attention
 +Python version: 3.13.11 (main, Dec  5 2025, 00:00:00) [GCC 15.2.1 20251111 (Red Hat 15.2.1-4)]
 +ComfyUI version: 0.9.1
 +ComfyUI frontend version: 1.36.14
 +[Prompt Server] web root: /opt/venv/lib64/python3.13/site-packages/comfyui_frontend_package/static
 +Traceback (most recent call last):
 +  File "/opt/ComfyUI/nodes.py", line 2157, in load_custom_node
 +    module_spec.loader.exec_module(module)
 +    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^
 +  File "<frozen importlib._bootstrap_external>", line 1023, in exec_module
 +  File "<frozen importlib._bootstrap>", line 488, in _call_with_frames_removed
 +  File "/opt/ComfyUI/custom_nodes/ComfyUI-GGUF/__init__.py", line 7, in <module>
 +    from .nodes import NODE_CLASS_MAPPINGS
 +  File "/opt/ComfyUI/custom_nodes/ComfyUI-GGUF/nodes.py", line 16, in <module>
 +    from .ops import GGMLOps, move_patch_to_device
 +  File "/opt/ComfyUI/custom_nodes/ComfyUI-GGUF/ops.py", line 2, in <module>
 +    import gguf
 +ModuleNotFoundError: No module named 'gguf'
 +
 +Cannot import /opt/ComfyUI/custom_nodes/ComfyUI-GGUF module for custom nodes: No module named 'gguf'
 +AMD GPU Monitor thread started
 +AMD GPU Monitor: Web directory set to /opt/ComfyUI/custom_nodes/ComfyUI-AMDGPUMonitor/web
 +
 +Import times for custom nodes:
 +   0.0 seconds: /opt/ComfyUI/custom_nodes/websocket_image_save.py
 +   0.0 seconds: /opt/ComfyUI/custom_nodes/ComfyUI-AMDGPUMonitor
 +   0.0 seconds (IMPORT FAILED): /opt/ComfyUI/custom_nodes/ComfyUI-GGUF
 +   0.0 seconds: /opt/ComfyUI/custom_nodes/ComfyUI_essentials
 +
 +Using AMD SMI tool: /opt/venv/bin/rocm-smi
 +Context impl SQLiteImpl.
 +Will assume non-transactional DDL.
 +Assets scan(roots=['models']) completed in 0.021s (created=4, skipped_existing=28, total_seen=33)
 +Disabling intermediate node cache.
 +Starting server
 +
 +To see the GUI go to: http://127.0.0.1:8000
 +got prompt
 +Using split attention in VAE
 +Using split attention in VAE
 +VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
 +Found quantization metadata version 1
 +Using MixedPrecisionOps for text encoder
 +Requested to load QwenImageTEModel_
 +loaded completely;  7910.28 MB loaded, full load: True
 +CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cuda:0, dtype: torch.float16
 +model weight dtype torch.bfloat16, manual cast: None
 +model_type FLUX
 +Requested to load QwenImage
 +loaded completely; 96341.98 MB usable, 38967.90 MB loaded, full load: True
 +100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 50/50 [10:53<00:00, 13.06s/it]
 +Requested to load WanVAE
 +loaded completely; 89893.71 MB usable, 242.03 MB loaded, full load: True
 +Prompt executed in 00:12:49
 +^C
 +</code>
amd-ryzen-ai-max-395-ai-lab-notizen.1769805879.txt.gz · Zuletzt geändert: von benny
Donate Powered by PHP Valid HTML5 Valid CSS Driven by DokuWiki