MLP 竟然解鎖了EGPU 9060X 16GB的overflow 1344*768*12s 直出
-
MLP 竟然解鎖了EGPU 9060X 16GB的overflow 134476812s 直出
九頭龍閃真的有九頭龍,不是黑心牛糞餅裡沒有餅
https://youtu.be/3QXNio64nCk
[INFO] got prompt
[INFO] Model storage policy: fast_disk=True paths=['D:\Comfy-Desktop\models\vae\minimax_h3_audio_vae_fp32.safetensors']
[INFO] VAE load device: cuda:0, offload device: cpu, dtype: torch.float32
[INFO] Model storage policy: fast_disk=True paths=['D:\Comfy-Desktop\models\vae\minimax_h3_sol_vae_int8_convrot.safetensors']
[INFO] Found quantization metadata version 1
[INFO] VAE load device: cuda:0, offload device: cpu, dtype: torch.float16
[INFO] Found quantization metadata version 1
[INFO] Using MixedPrecisionOps for text encoder
[INFO] Model storage policy: fast_disk=True paths=['D:\Comfy-Desktop\models\text_encoders\qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors']
[INFO] CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
[INFO] Requested to load MiniMaxH3VideoVAE
[INFO] Model MiniMaxH3VideoVAE prepared for dynamic VRAM loading. 1895MB Staged. 0 patches attached. Force pre-loaded 104 weights: 251 KB.
[INFO] Model MiniMaxH3VideoVAE prepared for dynamic VRAM loading. 1895MB Staged. 0 patches attached. Force pre-loaded 104 weights: 251 KB.
[INFO] Model MiniMaxH3VideoVAE prepared for dynamic VRAM loading. 1895MB Staged. 0 patches attached. Force pre-loaded 104 weights: 251 KB.
[INFO] Model MiniMaxH3VideoVAE prepared for dynamic VRAM loading. 1895MB Staged. 0 patches attached. Force pre-loaded 104 weights: 251 KB.
[INFO] Requested to load MiniMaxH3TEModel_
[INFO] Model MiniMaxH3TEModel_ prepared for dynamic VRAM loading. 14956MB Staged. 0 patches attached. Force pre-loaded 410 weights: 4572 KB.
[INFO] model weight dtype torch.float8_e4m3fn, manual cast: torch.bfloat16
[INFO] model_type FLOW_AV
[INFO] Model storage policy: fast_disk=True paths=['D:\Comfy-Desktop\models\diffusion_models\minimax_h3_ref2va_pruned_bf16.safetensors']
[INFO] Requested to load MiniMaxH3
[INFO] 0 models unloaded.
[INFO] Model MiniMaxH3 prepared for dynamic VRAM loading. 19308MB Staged. 208 patches attached. Force pre-loaded 210 weights: 1175 KB.
0%| | 0/8 [00:00<?, ?it/s]
0%| | 0/8 [00:00<?, ?it/s, Model Initializing ... ]
0%| | 0/8 [04:13<?, ?it/s, Model Initialization complete! ]
12%|█▎ | 1/8 [04:13<29:36, 253.83s/it, Model Initialization complete! ]
12%|█▎ | 1/8 [08:27<29:36, 253.83s/it]
25%|██▌ | 2/8 [08:27<25:22, 253.69s/it]
38%|███▊ | 3/8 [12:40<21:03, 252.66s/it]
50%|█████ | 4/8 [16:52<16:50, 252.64s/it]
62%|██████▎ | 5/8 [21:07<12:45, 255.19s/it]
75%|███████▌ | 6/8 [25:37<08:59, 269.76s/it]
88%|████████▊ | 7/8 [29:51<04:13, 253.79s/it]
100%|██████████| 8/8 [34:09<00:00, 257.75s/it]
100%|██████████| 8/8 [34:09<00:00, 256.15s/it]
[INFO] Comfy model compiler graph breaks: 0, rogues: 0
[INFO] Requested to load MiniMaxH3AudioVAE
[INFO] Model MiniMaxH3AudioVAE prepared for dynamic VRAM loading. 576MB Staged. 0 patches attached. Force pre-loaded 401 weights: 539 KB.
[INFO] Model MiniMaxH3VideoVAE prepared for dynamic VRAM loading. 1895MB Staged. 0 patches attached. Force pre-loaded 104 weights: 251 KB.
[INFO] Prompt executed in 00:37:12MLP 切塊是一個自訂節點:
D:\comfyui-rocm\custom_nodes\comfyui-minimax-mlp-chunk\__init__.py。ComfyUI 啟動時會自動套用,不用在工作流程裡接任何節點。可以調的只有一個環境變數:
MINIMAX_MLP_CHUNK,意思是每一塊放幾個 token。數值怎麼選
設定 效果 不設(預設 8192) 每塊約佔 0.5 GB 記憶體。1344×768、12 秒實測:共享記憶體從 4.3 GB 降到 0.27 GB,每步從 445 秒降到 225 秒 調小,例如 4096尖峰再降約一半。如果還是有溢出到共享記憶體,就往小調;切得越細,速度會稍微變慢 調大,例如 16384少切幾次,略快一點,但尖峰會變大。只有在記憶體很寬裕時才值得試 0關掉切塊,等於原版行為 怎麼改
我查過你的
comfyui-user.bat,目前沒有設定這個變數,所以用的是預設值 8192。要改的話,在comfyui-user.bat第 92 行set MIOPEN_FIND_ENFORCE=1的下一行加:set MINIMAX_MLP_CHUNK=4096改完要重啟 ComfyUI 才會生效。
怎麼確認有生效
啟動後看
D:\comfyui-rocm\user\comfyui.log,會出現這一行:[minimax-mlp-chunk] MiniMax H3 MLP runs in chunks of 4096 tokens如果設成
0,會顯示installed but disabled。完全移除
把資料夾改名成
comfyui-minimax-mlp-chunk.disabled,再重啟 ComfyUI。注意: 現在 12 秒版正在跑,用的是 8192。要改的話,請等它跑完再重啟,不然這一輪會中斷。