AMD ROCm / DynamicVRAM: performance degrades after first generation; /free restores full speed

[PennywiseDev]

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

Repeated executions of the same workflow should have approximately stable sampling performance.

Actual Behavior

Without explicitly freeing/unloading models, the first run is fast but subsequent runs become roughly 3–4x slower per sampling step.

Calling /free between generations restores normal performance and keeps subsequent runs stable.

Steps to Reproduce

1. Start ComfyUI with a clean process.
2. Run the same 1024x1024 / 8-step Krea2 workflow repeatedly.
3. Do not change prompt, model, resolution, sampler settings, or LoRAs.
4. Compare sampling speed between the first and following generations.

Debug Logs

ComfyUI AMD/ROCm DynamicVRAM slowdown - relevant logs
========================================================================

[Environment / startup]
[INFO] Total VRAM 16304 MB, total RAM 65461 MB
[INFO] pytorch version: 2.12.0+rocm7.14.0
[INFO] Set: torch.backends.cudnn.enabled = False for better AMD performance.
[INFO] AMD arch: gfx1200
[INFO] ROCm version: (7, 14)
[INFO] Set vram state to: NORMAL_VRAM
[INFO] Device: cuda:0 AMD Radeon RX 9060 XT : native
[INFO] Using async weight offloading with 2 streams
[INFO] Enabled pinned memory 26184.0
[INFO] Using pytorch attention
[INFO] aimdo: src-win/cuda-detour.c:38:INFO:aimdo_setup_hooks: installing 6 hooks
[INFO] aimdo: src-win/shmem-detect.c:80:INFO:comfy-aimdo WDDM adapter match: AMD Radeon RX 9060 XT runtime_luid=00000000:0000f336 dxgi_luid=00000000:0000f336
[INFO] aimdo: src/control.c:284:INFO:comfy-aimdo inited for GPU: AMD Radeon RX 9060 XT (VRAM: 16304 MB)
[INFO] DynamicVRAM support detected and enabled
[INFO] Python version: 3.13.12 (main, Feb 12 2026, 00:38:53) [MSC v.1944 64 bit (AMD64)]
[INFO] ComfyUI version: 0.37.0
[INFO] comfy-aimdo version: 0.5.5

[Bad state: first run fast]
[INFO] got prompt
[INFO] Model storage policy: fast_disk=True paths=['C:\\Users\\Michael\\AppData\\Local\\Comfy-Desktop\\ComfyUI-Shared\\models\\vae\\qwen_image_vae.safetensors']
[INFO] Using split attention in VAE
[INFO] Using split attention in VAE
[INFO] VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
[INFO] Found quantization metadata version 1
[INFO] Using MixedPrecisionOps for text encoder
[INFO] Model storage policy: fast_disk=True paths=['C:\\Users\\Michael\\AppData\\Local\\Comfy-Desktop\\ComfyUI-Shared\\models\\text_encoders\\qwen3vl_4b_fp8_scaled.safetensors']
[INFO] CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
[INFO] Requested to load Krea2TEModel_
[INFO] Model Krea2TEModel_ prepared for dynamic VRAM loading. 4999MB Staged. 0 patches attached. Force pre-loaded 249 weights: 627 KB.
[INFO] Found quantization metadata version 1
[INFO] Detected mixed precision quantization
[INFO] Using mixed precision operations
[INFO] Native ops: asym_w4a8_int8, convrot_w4a4, float8_e5m2, int8_tensorwise, float8_e4m3fn , emulated ops: nvfp4, mxfp8
[INFO] model weight dtype torch.bfloat16, manual cast: torch.bfloat16
[INFO] model_type FLUX
[INFO] Model storage policy: fast_disk=True paths=['C:\\Users\\Michael\\AppData\\Local\\Comfy-Desktop\\ComfyUI-Shared\\models\\diffusion_models\\krea2_turbo_fp8_scaled.safetensors']
[INFO] Requested to load Krea2
[INFO] Model Krea2 prepared for dynamic VRAM loading. 12530MB Staged. 0 patches attached. Force pre-loaded 160 weights: 2824 KB.

  0%|          | 0/8 [00:00<?, ?it/s]
  0%|          | 0/8 [00:00<?, ?it/s,   Model Initializing ...  ]
  0%|          | 0/8 [00:04<?, ?it/s,  Model Initialization complete!  ]
 12%|█▎        | 1/8 [00:04<00:33,  4.74s/it,  Model Initialization complete!  ]
 12%|█▎        | 1/8 [00:00<00:33,  4.74s/it]
 25%|██▌       | 2/8 [00:00<00:00,  7.59it/s]
 38%|███▊      | 3/8 [00:02<00:11,  2.27s/it]
 50%|█████     | 4/8 [00:04<00:09,  2.27s/it]
 62%|██████▎   | 5/8 [00:07<00:06,  2.26s/it]
 75%|███████▌  | 6/8 [00:09<00:04,  2.27s/it]
 88%|████████▊ | 7/8 [00:11<00:02,  2.27s/it]
100%|██████████| 8/8 [00:13<00:00,  2.27s/it]
100%|██████████| 8/8 [00:13<00:00,  1.73s/it]
[INFO] Requested to load WanVAE
[INFO] 0 models unloaded.
[INFO] Model WanVAE prepared for dynamic VRAM loading. 241MB Staged. 0 patches attached. Force pre-loaded 60 weights: 61 KB.
[INFO] Prompt executed in 33.52 seconds

[Bad state: second run slow]
[INFO] got prompt
[INFO] Model Krea2TEModel_ prepared for dynamic VRAM loading. 4999MB Staged. 0 patches attached. Force pre-loaded 249 weights: 627 KB.
[INFO] Model Krea2 prepared for dynamic VRAM loading. 12530MB Staged. 0 patches attached. Force pre-loaded 160 weights: 2824 KB.

  0%|          | 0/8 [00:00<?, ?it/s]
  0%|          | 0/8 [00:00<?, ?it/s,   Model Initializing ...  ]
  0%|          | 0/8 [00:08<?, ?it/s,  Model Initialization complete!  ]
 12%|█▎        | 1/8 [00:08<01:00,  8.65s/it,  Model Initialization complete!  ]
 12%|█▎        | 1/8 [00:00<01:00,  8.65s/it]
 25%|██▌       | 2/8 [00:00<00:02,  2.63it/s]
 38%|███▊      | 3/8 [00:08<00:40,  8.16s/it]
 50%|█████     | 4/8 [00:17<00:32,  8.16s/it]
 62%|██████▎   | 5/8 [00:25<00:24,  8.15s/it]
 75%|███████▌  | 6/8 [00:33<00:16,  8.14s/it]
 88%|████████▊ | 7/8 [00:41<00:08,  8.15s/it]
100%|██████████| 8/8 [00:49<00:00,  8.13s/it]
100%|██████████| 8/8 [00:49<00:00,  6.21s/it]
[INFO] 0 models unloaded.
[INFO] Model WanVAE prepared for dynamic VRAM loading. 241MB Staged. 0 patches attached. Force pre-loaded 60 weights: 61 KB.
[INFO] Prompt executed in 70.31 seconds

[Bad state: additional slow runs]
[INFO] got prompt
[INFO] Model Krea2TEModel_ prepared for dynamic VRAM loading. 4999MB Staged. 0 patches attached. Force pre-loaded 249 weights: 627 KB.
[INFO] Model Krea2 prepared for dynamic VRAM loading. 12530MB Staged. 0 patches attached. Force pre-loaded 160 weights: 2824 KB.

  0%|          | 0/8 [00:00<?, ?it/s]
  0%|          | 0/8 [00:00<?, ?it/s,   Model Initializing ...  ]
  0%|          | 0/8 [00:08<?, ?it/s,  Model Initialization complete!  ]
 12%|█▎        | 1/8 [00:08<00:57,  8.27s/it,  Model Initialization complete!  ]
 12%|█▎        | 1/8 [00:00<00:57,  8.27s/it]
 25%|██▌       | 2/8 [00:00<00:02,  2.62it/s][INFO] FETCH ComfyRegistry Data [DONE]
[INFO] [ComfyUI-Manager] default cache updated: https://api.comfy.org/nodes
FETCH DATA from: C:\Users\Michael\AppData\Local\Comfy-Desktop\ComfyUI-Installs\ComfyUI\ComfyUI\user\__manager\cache\1514988643_custom-node-list.json [DONE]
[INFO] [ComfyUI-Manager] All startup tasks have been completed.

 38%|███▊      | 3/8 [00:08<00:39,  7.94s/it]
 50%|█████     | 4/8 [00:16<00:32,  8.04s/it]
 62%|██████▎   | 5/8 [00:24<00:24,  8.02s/it]
 75%|███████▌  | 6/8 [00:32<00:15,  7.96s/it]
 88%|████████▊ | 7/8 [00:40<00:07,  7.97s/it]
100%|██████████| 8/8 [00:48<00:00,  7.98s/it]
100%|██████████| 8/8 [00:48<00:00,  6.08s/it]
[INFO] 0 models unloaded.
[INFO] Model WanVAE prepared for dynamic VRAM loading. 241MB Staged. 0 patches attached. Force pre-loaded 60 weights: 61 KB.
[INFO] Prompt executed in 69.28 seconds
[INFO] got prompt
[INFO] Model Krea2TEModel_ prepared for dynamic VRAM loading. 4999MB Staged. 0 patches attached. Force pre-loaded 249 weights: 627 KB.
[INFO] Model Krea2 prepared for dynamic VRAM loading. 12530MB Staged. 0 patches attached. Force pre-loaded 160 weights: 2824 KB.

  0%|          | 0/8 [00:00<?, ?it/s]
  0%|          | 0/8 [00:00<?, ?it/s,   Model Initializing ...  ]
  0%|          | 0/8 [00:08<?, ?it/s,  Model Initialization complete!  ]
 12%|█▎        | 1/8 [00:08<00:58,  8.32s/it,  Model Initialization complete!  ]
 12%|█▎        | 1/8 [00:00<00:58,  8.32s/it]
 25%|██▌       | 2/8 [00:00<00:02,  2.68it/s]
 38%|███▊      | 3/8 [00:08<00:39,  7.98s/it]
 50%|█████     | 4/8 [00:16<00:31,  7.95s/it]
 62%|██████▎   | 5/8 [00:24<00:23,  7.99s/it]
 75%|███████▌  | 6/8 [00:32<00:15,  7.98s/it]
 88%|████████▊ | 7/8 [00:40<00:08,  8.01s/it]
100%|██████████| 8/8 [00:48<00:00,  7.94s/it]
100%|██████████| 8/8 [00:48<00:00,  6.07s/it]
[INFO] 0 models unloaded.
[INFO] Model WanVAE prepared for dynamic VRAM loading. 241MB Staged. 0 patches attached. Force pre-loaded 60 weights: 61 KB.
[INFO] Prompt executed in 69.26 seconds

[After /free between jobs: stable fast run, no LoRA]
[INFO] got prompt
[INFO] Model storage policy: fast_disk=True paths=['C:\\Users\\Michael\\AppData\\Local\\Comfy-Desktop\\ComfyUI-Shared\\models\\vae\\qwen_image_vae.safetensors']
[INFO] Using split attention in VAE
[INFO] Using split attention in VAE
[INFO] VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
[INFO] Found quantization metadata version 1
[INFO] Using MixedPrecisionOps for text encoder
[INFO] Model storage policy: fast_disk=True paths=['C:\\Users\\Michael\\AppData\\Local\\Comfy-Desktop\\ComfyUI-Shared\\models\\text_encoders\\qwen3vl_4b_fp8_scaled.safetensors']
[INFO] CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
[INFO] Requested to load Krea2TEModel_
[INFO] Model Krea2TEModel_ prepared for dynamic VRAM loading. 4999MB Staged. 0 patches attached. Force pre-loaded 249 weights: 627 KB.
[INFO] Found quantization metadata version 1
[INFO] Detected mixed precision quantization
[INFO] Using mixed precision operations
[INFO] Native ops: float8_e4m3fn, convrot_w4a4, float8_e5m2, int8_tensorwise, asym_w4a8_int8 , emulated ops: mxfp8, nvfp4
[INFO] model weight dtype torch.bfloat16, manual cast: torch.bfloat16
[INFO] model_type FLUX
[INFO] Model storage policy: fast_disk=True paths=['C:\\Users\\Michael\\AppData\\Local\\Comfy-Desktop\\ComfyUI-Shared\\models\\diffusion_models\\krea2_turbo_fp8_scaled.safetensors']
[INFO] Requested to load Krea2
[INFO] Model Krea2 prepared for dynamic VRAM loading. 12530MB Staged. 0 patches attached. Force pre-loaded 160 weights: 2824 KB.

  0%|          | 0/8 [00:00<?, ?it/s]
  0%|          | 0/8 [00:00<?, ?it/s,   Model Initializing ...  ]
  0%|          | 0/8 [00:03<?, ?it/s,  Model Initialization complete!  ]
 12%|█▎        | 1/8 [00:03<00:24,  3.53s/it,  Model Initialization complete!  ]
 12%|█▎        | 1/8 [00:00<00:24,  3.53s/it]
 25%|██▌       | 2/8 [00:00<00:00,  6.76it/s]
 38%|███▊      | 3/8 [00:02<00:12,  2.50s/it]
 50%|█████     | 4/8 [00:05<00:09,  2.49s/it]
 62%|██████▎   | 5/8 [00:07<00:07,  2.49s/it]
 75%|███████▌  | 6/8 [00:10<00:04,  2.49s/it]
 88%|████████▊ | 7/8 [00:12<00:02,  2.53s/it]
100%|██████████| 8/8 [00:15<00:00,  2.51s/it]
100%|██████████| 8/8 [00:15<00:00,  1.91s/it]
[INFO] Requested to load WanVAE
[INFO] 0 models unloaded.
[INFO] Model WanVAE prepared for dynamic VRAM loading. 241MB Staged. 0 patches attached. Force pre-loaded 60 weights: 61 KB.
[INFO] Prompt executed in 27.52 seconds

[After /free between jobs: stable fast run, LoRA active (256 patches)]
[INFO] got prompt
[INFO] Model storage policy: fast_disk=True paths=['C:\\Users\\Michael\\AppData\\Local\\Comfy-Desktop\\ComfyUI-Shared\\models\\vae\\qwen_image_vae.safetensors']
[INFO] Using split attention in VAE
[INFO] Using split attention in VAE
[INFO] VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
[INFO] Found quantization metadata version 1
[INFO] Using MixedPrecisionOps for text encoder
[INFO] Model storage policy: fast_disk=True paths=['C:\\Users\\Michael\\AppData\\Local\\Comfy-Desktop\\ComfyUI-Shared\\models\\text_encoders\\qwen3vl_4b_fp8_scaled.safetensors']
[INFO] CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
[INFO] Requested to load Krea2TEModel_
[INFO] Model Krea2TEModel_ prepared for dynamic VRAM loading. 4999MB Staged. 0 patches attached. Force pre-loaded 249 weights: 627 KB.
[INFO] Found quantization metadata version 1
[INFO] Detected mixed precision quantization
[INFO] Using mixed precision operations
[INFO] Native ops: float8_e4m3fn, convrot_w4a4, float8_e5m2, int8_tensorwise, asym_w4a8_int8 , emulated ops: mxfp8, nvfp4
[INFO] model weight dtype torch.bfloat16, manual cast: torch.bfloat16
[INFO] model_type FLUX
[INFO] Model storage policy: fast_disk=True paths=['C:\\Users\\Michael\\AppData\\Local\\Comfy-Desktop\\ComfyUI-Shared\\models\\diffusion_models\\krea2_turbo_fp8_scaled.safetensors']
[INFO] Requested to load Krea2
[INFO] Model Krea2 prepared for dynamic VRAM loading. 12530MB Staged. 256 patches attached. Force pre-loaded 160 weights: 2824 KB.

  0%|          | 0/8 [00:00<?, ?it/s]
  0%|          | 0/8 [00:00<?, ?it/s,   Model Initializing ...  ]
  0%|          | 0/8 [00:05<?, ?it/s,  Model Initialization complete!  ]
 12%|█▎        | 1/8 [00:05<00:38,  5.44s/it,  Model Initialization complete!  ]
 12%|█▎        | 1/8 [00:00<00:38,  5.44s/it]
 25%|██▌       | 2/8 [00:00<00:02,  2.11it/s]
 38%|███▊      | 3/8 [00:03<00:12,  2.45s/it]
 50%|█████     | 4/8 [00:05<00:10,  2.60s/it]
 62%|██████▎   | 5/8 [00:08<00:07,  2.60s/it]
 75%|███████▌  | 6/8 [00:11<00:05,  2.61s/it]
 88%|████████▊ | 7/8 [00:13<00:02,  2.61s/it]
100%|██████████| 8/8 [00:16<00:00,  2.61s/it]
100%|██████████| 8/8 [00:16<00:00,  2.05s/it]
[INFO] Requested to load WanVAE
[INFO] 0 models unloaded.
[INFO] Model WanVAE prepared for dynamic VRAM loading. 241MB Staged. 0 patches attached. Force pre-loaded 60 weights: 61 KB.
[INFO] Prompt executed in 29.54 seconds

[Additional stable LoRA runs]
[INFO] got prompt
[INFO] Model storage policy: fast_disk=True paths=['C:\\Users\\Michael\\AppData\\Local\\Comfy-Desktop\\ComfyUI-Shared\\models\\vae\\qwen_image_vae.safetensors']
[INFO] Using split attention in VAE
[INFO] Using split attention in VAE
[INFO] VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
[INFO] Found quantization metadata version 1
[INFO] Using MixedPrecisionOps for text encoder
[INFO] Model storage policy: fast_disk=True paths=['C:\\Users\\Michael\\AppData\\Local\\Comfy-Desktop\\ComfyUI-Shared\\models\\text_encoders\\qwen3vl_4b_fp8_scaled.safetensors']
[INFO] CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
[INFO] Requested to load Krea2TEModel_
[INFO] Model Krea2TEModel_ prepared for dynamic VRAM loading. 4999MB Staged. 0 patches attached. Force pre-loaded 249 weights: 627 KB.
[INFO] Found quantization metadata version 1
[INFO] Detected mixed precision quantization
[INFO] Using mixed precision operations
[INFO] Native ops: float8_e4m3fn, convrot_w4a4, float8_e5m2, int8_tensorwise, asym_w4a8_int8 , emulated ops: mxfp8, nvfp4
[INFO] model weight dtype torch.bfloat16, manual cast: torch.bfloat16
[INFO] model_type FLUX
[INFO] Model storage policy: fast_disk=True paths=['C:\\Users\\Michael\\AppData\\Local\\Comfy-Desktop\\ComfyUI-Shared\\models\\diffusion_models\\krea2_turbo_fp8_scaled.safetensors']
[INFO] Requested to load Krea2
[INFO] Model Krea2 prepared for dynamic VRAM loading. 12530MB Staged. 256 patches attached. Force pre-loaded 160 weights: 2824 KB.

  0%|          | 0/8 [00:00<?, ?it/s]
  0%|          | 0/8 [00:00<?, ?it/s,   Model Initializing ...  ]
  0%|          | 0/8 [00:05<?, ?it/s,  Model Initialization complete!  ]
 12%|█▎        | 1/8 [00:05<00:38,  5.47s/it,  Model Initialization complete!  ]
 12%|█▎        | 1/8 [00:00<00:38,  5.47s/it]
 25%|██▌       | 2/8 [00:00<00:02,  2.15it/s]
 38%|███▊      | 3/8 [00:03<00:12,  2.43s/it]
 50%|█████     | 4/8 [00:05<00:10,  2.60s/it]
 62%|██████▎   | 5/8 [00:08<00:07,  2.59s/it]
 75%|███████▌  | 6/8 [00:11<00:05,  2.59s/it]
 88%|████████▊ | 7/8 [00:13<00:02,  2.60s/it]
100%|██████████| 8/8 [00:16<00:00,  2.60s/it]
100%|██████████| 8/8 [00:16<00:00,  2.04s/it]
[INFO] Requested to load WanVAE
[INFO] 0 models unloaded.
[INFO] Model WanVAE prepared for dynamic VRAM loading. 241MB Staged. 0 patches attached. Force pre-loaded 60 weights: 61 KB.
[INFO] Prompt executed in 29.14 seconds
[INFO] Using RAM pressure cache.
[INFO] got prompt
[INFO] Model storage policy: fast_disk=True paths=['C:\\Users\\Michael\\AppData\\Local\\Comfy-Desktop\\ComfyUI-Shared\\models\\vae\\qwen_image_vae.safetensors']
[INFO] Using split attention in VAE
[INFO] Using split attention in VAE
[INFO] VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
[INFO] Found quantization metadata version 1
[INFO] Using MixedPrecisionOps for text encoder
[INFO] Model storage policy: fast_disk=True paths=['C:\\Users\\Michael\\AppData\\Local\\Comfy-Desktop\\ComfyUI-Shared\\models\\text_encoders\\qwen3vl_4b_fp8_scaled.safetensors']
[INFO] CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
[INFO] Requested to load Krea2TEModel_
[INFO] Model Krea2TEModel_ prepared for dynamic VRAM loading. 4999MB Staged. 0 patches attached. Force pre-loaded 249 weights: 627 KB.
[INFO] Found quantization metadata version 1
[INFO] Detected mixed precision quantization
[INFO] Using mixed precision operations
[INFO] Native ops: float8_e4m3fn, convrot_w4a4, float8_e5m2, int8_tensorwise, asym_w4a8_int8 , emulated ops: mxfp8, nvfp4
[INFO] model weight dtype torch.bfloat16, manual cast: torch.bfloat16
[INFO] model_type FLUX
[INFO] Model storage policy: fast_disk=True paths=['C:\\Users\\Michael\\AppData\\Local\\Comfy-Desktop\\ComfyUI-Shared\\models\\diffusion_models\\krea2_turbo_fp8_scaled.safetensors']
[INFO] Requested to load Krea2
[INFO] Model Krea2 prepared for dynamic VRAM loading. 12530MB Staged. 256 patches attached. Force pre-loaded 160 weights: 2824 KB.

  0%|          | 0/8 [00:00<?, ?it/s]
  0%|          | 0/8 [00:00<?, ?it/s,   Model Initializing ...  ]
  0%|          | 0/8 [00:05<?, ?it/s,  Model Initialization complete!  ]
 12%|█▎        | 1/8 [00:05<00:38,  5.45s/it,  Model Initialization complete!  ]
 12%|█▎        | 1/8 [00:00<00:38,  5.45s/it]
 25%|██▌       | 2/8 [00:00<00:02,  2.14it/s]
 38%|███▊      | 3/8 [00:03<00:12,  2.43s/it]
 50%|█████     | 4/8 [00:05<00:10,  2.60s/it]
 62%|██████▎   | 5/8 [00:08<00:07,  2.60s/it]
 75%|███████▌  | 6/8 [00:11<00:05,  2.60s/it]
 88%|████████▊ | 7/8 [00:13<00:02,  2.60s/it]
100%|██████████| 8/8 [00:16<00:00,  2.60s/it]
100%|██████████| 8/8 [00:16<00:00,  2.05s/it]
[INFO] Requested to load WanVAE
[INFO] 0 models unloaded.
[INFO] Model WanVAE prepared for dynamic VRAM loading. 241MB Staged. 0 patches attached. Force pre-loaded 60 weights: 61 KB.
[INFO] Prompt executed in 29.40 seconds

Other

Description

On AMD ROCm, image generation performance degrades heavily after the first generation in the same ComfyUI process.

The first generation runs normally, but subsequent generations become roughly 3–4x slower per sampling step, even with the exact same workflow and with no LoRAs applied.

Calling the /free endpoint with both unload_models and free_memory before the next generation restores normal performance.

Interestingly, when /free is used between every generation, performance remains stable and is even slightly faster overall than before the issue appeared.

This strongly suggests a problem related to memory residency / offloading / DynamicVRAM state rather than the sampler, prompt, or LoRA processing itself.

Environment

• GPU: AMD Radeon RX 9060 XT, 16 GB VRAM
• Windows
• ROCm 7.14
• PyTorch 2.12.0+rocm7.14.0
• ComfyUI 0.37.0
• comfy-aimdo 0.5.5
• DynamicVRAM enabled
• NORMAL_VRAM
• async weight offloading with 2 streams
• fast_disk enabled
• Model: Krea2 Turbo FP8
• Resolution: 1024x1024
• Sampler: 8 steps

ComfyUI reports DynamicVRAM as enabled and uses async weight offloading. :contentReference[oaicite:0]{index=0}

Reproduction

1. Start ComfyUI with a clean process.
2. Run the same 1024x1024 / 8-step Krea2 workflow repeatedly.
3. Do not change prompt, model, resolution, sampler settings, or LoRAs.
4. Compare sampling speed between the first and following generations.

Without /free

First generation:

• ~2.27 s/it
• 33.52 seconds total

Second generation:

• ~8.13–8.16 s/it
• 70.31 seconds total

Further generations remain around ~8 s/it and ~69 seconds total.

The model reports 0 patches attached, so LoRA patching is not involved in these slow runs. :contentReference[oaicite:1]{index=1}

Additional runs remain slow at approximately 7.9–8.0 s/it and ~69 seconds total. :contentReference[oaicite:2]{index=2}

Workaround

Calling:

POST /free

with:

{
"unload_models": true,
"free_memory": true
}

between every generation prevents the slowdown.

With this workaround enabled, repeated runs are stable:

• 27.52 seconds
• 29.54 seconds
• 29.54 seconds
• 29.14 seconds
• 29.40 seconds
• 29.70 seconds

Sampling remains around ~2.5–2.6 s/it. :contentReference[oaicite:3]{index=3} :contentReference[oaicite:4]{index=4} :contentReference[oaicite:5]{index=5}

LoRA comparison

The issue does not appear to be caused by LoRAs.

Without LoRA patches:

• Krea2 reports 0 patches attached
• sampling is ~2.5–2.6 s/it
• total time is ~27–29 seconds

With LoRA enabled:

• Krea2 reports 256 patches attached
• sampling remains ~2.6 s/it
• total time remains ~29–30 seconds

Example with 256 patches:

• initialization: ~5.44 s
• sampling: ~2.60–2.61 s/it
• total: 29.54 seconds

:contentReference[oaicite:6]{index=6}

Further LoRA runs show the same behavior:

• 29.14 seconds
• 29.40 seconds
• 29.70 seconds

:contentReference[oaicite:7]{index=7} :contentReference[oaicite:8]{index=8} :contentReference[oaicite:9]{index=9}

Observation

The important difference appears to be the runtime state between generations.

The workflow loads/stages approximately:

• Krea2TEModel_: 4999 MB
• Krea2: 12530 MB
• WanVAE: 241 MB

and all are reported as being prepared for dynamic VRAM loading. :contentReference[oaicite:10]{index=10}

Because /free consistently restores the fast state, my current suspicion is that some memory residency, offloading, allocator, or DynamicVRAM state persists after a generation and causes later generations to use a much slower execution path.

I cannot determine from the logs whether the underlying problem is in ComfyUI DynamicVRAM, comfy-aimdo, PyTorch/ROCm memory management, or WDDM interaction.

Expected behavior

Repeated executions of the same workflow should have approximately stable sampling performance.

Actual behavior

Without explicitly freeing/unloading models, the first run is fast but subsequent runs become roughly 3–4x slower per sampling step.

Calling /free between generations restores normal performance and keeps subsequent runs stable.

[kellylinopenclaw-design]

Same degradation on a large LLM-style text encoder + int8 UNet (RDNA4, 16 GB) — with s/it numbers

Confirming this on 16 GB RDNA4 (gfx1201) with a much heavier model set than the Krea2 example, so it does not look specific to small models or to the workflow:

Environment: Windows 11, AMD Radeon RX 9070 XT (gfx1201, 16 GB VRAM), 96 GB RAM, torch 2.13 + ROCm 10.1, ComfyUI 0.37.0, comfy-aimdo 0.5.5, comfy-kitchen 0.2.35, launched with --use-ck-attention --disable-fast-disk.

Models: MiniMax-H3 — 32B Qwen3-VL text encoder (nvfp4, ~15 GB), 21 GB int8 diffusion UNet, 2 VAEs. Single workflow, single fresh process.

Observed on a fresh process (first generation):

• pre-sampling stage (ref-image VAE encode + text encode): ~26 min before the first sampler step
• low-res phase (6 steps): 45.25 / 57.65 / 65.42 s/it
• high-res phase (2 steps): 444.64 s/it
• VRAM during the run: ~5.7 GB free of 15.9 GB — the encoder and the UNet are both fighting for the card

Notes that may help narrow it down:

1. --disable-fast-disk (new in 0.37.0) does take effect — the log prints Model storage policy: fast_disk=False — but it did not restore performance in this configuration.
2. On the same machine, running the text encoder in its own process (card holds only the encoder, UNet loaded by a second instance that samples) removes the problem completely: the same encode stage takes 117-176 s. So /free restoring speed is consistent with "low residency is fast, mixed residency degrades", not with a per-generation leak alone.
3. We previously measured the same family of behaviour on 0.36 with dynamic VRAM on and no dedicated encode process: s/it climbing 51 -> 104 -> 119 -> 126 -> 158, with the sampler thread repeatedly in model_vbar.py:88 fault -> vbar_fault -> cast_modules_with_vbar -> cast_bias_weight.
4. Baseline sensitivity: a stale python process holding 2.76 GB of VRAM was still alive during part of a run and pushed the tested process' free VRAM from ~8.5 GB down to ~5.7 GB. Worth checking no other GPU process is alive while reproducing.

Happy to run any specific probe you want on this box (e.g. --disable-dynamic-vram, pinned vs unpinned, per-generation /free timing series) and paste the logs.

[Klix555]

I can reproduce a similar slowdown on a Radeon RX 7800 XT with ComfyUI Desktop.

Environment:

• Windows 10
• AMD Radeon RX 7800 XT, gfx1101, 16 GB VRAM
• ComfyUI Desktop 0.37.2
• PyTorch 2.12.0+rocm7.14.0
• comfy-aimdo 0.5.5
• 64 GB system RAM

Workflow:

• Krea-2 Turbo FP8 (krea2_turbo_fp8_scaled.safetensors)
• 8 sampling steps
• The same workflow and settings were used for consecutive generations.
• The workflow includes ComfyUI-Krea2T-Enhancer; I have not yet isolated whether the custom node affects reproduction.

Results from the attached log:

• First generation: 29 seconds for sampling; 48.73 seconds total.
• Second generation: 2 minutes 18 seconds for sampling; 185.02 seconds total.
• The Krea model reports a full load of 12532.86 MB and lowvram patches: 0 in both runs.
• No out-of-memory error is reported.

I tested --disable-dynamic-vram, --disable-pinned-memory, and --disable-async-offload in separate runs, but the slowdown still occurs. I also tried POSTing to /free between generations; it did not restore normal speed in my case.

The slowdown has also occurred at different image sizes. I can provide the full log if useful. Could this be related to ROCm/PyTorch or memory handling on Windows, and is there a useful minimal test I should try?

Comfy LOG 9.txt