202611112058_H3_image文生图+TE4b优化_omnixpu加速_B580_16G_冷启
2026-11-11 20:58:10 · sess-20261111-205736 · 工作流 minimax_h3_image_t2i
success冷启omnixpu加速B58016G
总耗时
173.93s
采样耗时
23s
采样步数
6 / 6
每步采样
3.94s/it
模型载入
14.20GB
单次峰值
8.59GB
生成物

模型加载模型文件 / 加载节点来自工作流结构(API 图);载入 MB / 耗时 / MB/s / 路径来源来自执行日志
| 角色 | 模型文件 | 加载节点 | 载入 MB | 耗时 s | MB/s | 路径来源 |
|---|---|---|---|---|---|---|
| 底模 base | MiniMax-H3-Image-Pruned-Q3_K_M.gguf ext/H3_diff/MiniMax-H3-Image-Pruned-Q3_K_M.gguf | #37 UnetLoaderGGUFAdvanced MiniMaxH3Image | 8791.8 | 67.092 | 131 | 未识别 |
| 文本编码器 text_encoder | mmh3-4b-ClipProj-v3.1.safetensors | #25 ClipProjApply | — | — | — | — |
| 文本编码器 text_encoder | qwen3vl_4b_int8_convrot.safetensors ext/qwen3vl_4b_int8_convrot.safetensors | #29 CLIPLoader Qwen3VLTEModel_ | 4633.3 | 0.104 | 44551 | policy 相邻 |
| VAE vae | minimax_h3_image_vae_fp16.safetensors H3_vae/minimax_h3_image_vae_fp16.safetensors | #21 VAELoaderKJ MiniMaxH3ImageVAE | 1114.2 | 3.749 | 297 | 按角色匹配 |
共 4 行:底模 1 · 文本编码器 2 · VAE 1。模型文件 / 加载节点来自工作流结构(API 图)。载入 MB / 耗时 / MB/s 来自工作流执行日志(`Requested to load` → `loaded completely`);同一文件一次执行里加载多次时合并为一行、数值是累计值,相加仍等于 KPI 的「模型载入」。
MB/s 用于区分「内存回迁」(GB/s 级)与「磁盘重读」(数百 MB/s)。text_encoder 在某些栈上 device_to=cpu,它的加载耗时接近 0,不代表省了显存搬运。
MB/s 用于区分「内存回迁」(GB/s 级)与「磁盘重读」(数百 MB/s)。text_encoder 在某些栈上 device_to=cpu,它的加载耗时接近 0,不代表省了显存搬运。
工作流结构UI 图含 bypass 节点,API 图是真正执行的子集工作流导出
| 画布节点数(UI 图) | 14 |
|---|---|
| 实际执行节点数(API 图) | 14 |
| 节点状态分布(UI 图 mode) | active 14 · muted 0 · bypassed 0 |
| 被 bypass 的节点 | 无 |
| 被 mute 的节点 | 无 |
节点详情 14 个节点 · 6 个插件
| 节点名 | 插件名 | 插件版本 |
|---|---|---|
CLIPLoader 加载CLIP | comfy-core | 0.42.5 |
ConditioningZeroOut 条件零化 | comfy-core | 0.45.2 |
KSampler K采样器 | comfy-core | 0.45.2 |
MiniMaxH3ImageConditioning | comfy-core | 0.45.2 |
PrimitiveStringMultiline Text (user_prompt) | comfy-core | 0.45.2 |
ResolutionSelector 分辨率选择器 | comfy-core | 0.45.2 |
SaveImage 保存图像 | comfy-core | 0.45.2 |
VAEDecodeTiled VAE解码(分块) | comfy-core | 0.11.1 |
ClipProjApply | ComfyUI-ClipProj nicolab28 | c01ba8fb8f41… |
LogCollection | ComfyUI-FullLog | — |
UnetLoaderGGUFAdvanced | ComfyUI-GGUF-XPU analytics-zoo | 39671fe73117… |
VAELoaderKJ | comfyui-kjnodes | 50a0837f9aea… |
VRAM_Debug | comfyui-kjnodes | 9b565354e32a… |
VRAMCleanup | comfyui_memory_cleanup | 1.1.3 |
运行环境
| 系统 | Windows-11-10.0.22000-SP0 · AMD64 · build 10.0.22000 |
|---|---|
| 物理内存 | 15.9 GB (16G) |
| GPU | Intel(R) Arc(TM) B580 Graphics (0xE20B, 11946.2 MB, driver 1.17.39894+4) |
| 平台 | Intel(R) oneAPI Unified Runtime over Level-Zero V2 |
| 计算单元 | 160 |
| Python | 3.13.11 |
| torch | 2.15.0+xpu (xpu available: True, devices: 1) |
| ComfyUI | 0.45.2 preview=Latent2RGB |
| 插件数 | 68 |
| 加速栈 | |
| 加速方案 | omnixpu加速 (工作流执行日志中 omni_xpu_kernel 0.2.0b3+torch215.bmg 生效,且出现 [OmniXPU] 插件标记) |
| OmniXPU 内核 | omni_xpu_kernel 0.2.0b3+torch215.bmg 已生效:sdp, norm, rotary, linear_fp8, int8, layout |
| OmniXPU 插件 | ComfyUI-OmniXPU 16 个 adapter applied、4 个 skipped |
| aimdo | comfy-aimdo 0.6.0 未启用 (disabled by env) |
| kitchen XPU | comfy-kitchen 0.2.41 xpu 后端可用 (adaln, apply_rope, apply_rope1, apply_rope1_… 共 50 项) |
加速方案 = 内核生效行 + 插件
[OmniXPU] 标记,两处都在工作流执行日志里才判「omnixpu加速」;aimdo / kitchen 是可选栈,不影响方案判定,也不进文件名。运行时环境变量 18 项
| 变量名 | 值 |
|---|---|
OCL_ICD_FILENAMES | ;F:\Intel-oneAPI\compiler\latest\bin\intelocl64.dll |
OMNIXPU_DEBUG | 0 |
OMNIXPU_ENABLE | 1 |
OMNIXPU_QWEN_IMAGE21_QKV | 1 |
ONEAPI_CLANG_VERSION | 23 |
ONEAPI_DEVICE_SELECTOR | level_zero:0 |
ONEAPI_PATH | F:\Intel-oneAPI |
ONEAPI_ROOT | F:\Intel-oneAPI |
SYCL_CACHE_MAX_SIZE | 8192 |
SYCL_CACHE_THRESHOLD | 7 |
SYCL_DEVICE_FILTER | level_zero |
SYCL_PI_LEVEL_ZERO_DEVICE_SCOPE_EVENTS | 0 |
SYCL_PI_LEVEL_ZERO_TRACK_INDIRECT_ACCESS_MEMORY | 1 |
SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS | 1 |
UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS | 1 |
UR_L0_USE_IMMEDIATE_COMMANDLISTS | 1 |
ZE_DEVICE_SLEEP | 0 |
ZE_ENABLE_PCI_ID_DEVICE_ORDER | 1 |
时间轴
| 时刻 | 相对起点 s | 增量 s | 事件 |
|---|---|---|---|
| 20:58:10.640 | 0.27 | 0.27 | 下发存储策略 · minimax_h3_image_vae_fp16.safetensors |
| 20:58:17.209 | 6.84 | 6.57 | 下发存储策略 · qwen3vl_4b_int8_convrot.safetensors |
| 20:58:17.331 | 6.96 | 0.12 | 请求加载 Qwen3VLTEModel_ |
| 20:58:17.435 | 7.07 | 0.11 | qwen3vl_4b_int8_convrot.safetensors 就绪(4633.34 MB) |
| 20:58:54.133 | 43.77 | 36.70 | 请求加载 MiniMaxH3Image |
| 21:00:01.225 | 110.86 | 67.09 | MiniMaxH3Image 就绪(8791.75 MB) |
| 21:00:50.719 | 160.35 | 49.49 | 请求加载 MiniMaxH3ImageVAE |
| 21:00:54.468 | 164.10 | 3.75 | minimax_h3_image_vae_fp16.safetensors 就绪(1114.23 MB) |
| 21:01:04.218 | 173.85 | 9.75 | 执行完成 |
原始日志内嵌,不依赖 user/comfyui.log
来源
user/comfyui.log · 35 行 · 日志自报耗时 173.93 s原始日志全文
[2026-11-11 20:58:10.368] gguf qtypes: F32 (320), Q3_K (208)[2026-11-11 20:58:10.399] model weight dtype torch.bfloat16, manual cast: None[2026-11-11 20:58:10.409] model_type FLOW[2026-11-11 20:58:10.424] Model storage policy: fast_disk=False paths=[][2026-11-11 20:58:10.640] Model storage policy: fast_disk=False paths=['E:\\model\\H3_vae\\minimax_h3_image_vae_fp16.safetensors'][2026-11-11 20:58:16.382] VAE load device: xpu:0, offload device: cpu, dtype: torch.float16[2026-11-11 20:58:16.841] Found quantization metadata version 1[2026-11-11 20:58:16.842] Using MixedPrecisionOps for text encoder[2026-11-11 20:58:17.209] Model storage policy: fast_disk=False paths=['E:\\model\\qwen3vl_4b_int8_convrot.safetensors'][2026-11-11 20:58:17.331] Requested to load Qwen3VLTEModel_[2026-11-11 20:58:17.435] loaded completely; 4633.34 MB loaded, full load: True[2026-11-11 20:58:17.458] CLIP/text encoder model load device: cpu, offload device: cpu, current: cpu, dtype: torch.float16[2026-11-11 20:58:17.536] [ClipProj] mmh3-4b-ClipProj-v3.1.safetensors | tap 24 | 2560 -> 5120 | cos_test 0.6874[2026-11-11 20:58:54.133] Requested to load MiniMaxH3Image[2026-11-11 21:00:01.225] loaded completely; 10103.00 MB usable, 8791.75 MB loaded, full load: True[2026-11-11 21:00:02.129] [OmniXPU] norm first use: RMSNorm shape=torch.Size([452, 5376])[2026-11-11 21:00:02.793] [OmniXPU] experimental capability-driven BMG D128 CUTE route enabled; set OMNI_ATTN_BACKEND=torch to disable it if this workflow shows an issue[2026-11-11 21:00:02.794] [OmniXPU] attention CUTE_BHLD_D128 #1: heads=56 q=452 kv=452 dtype=torch.bfloat16[2026-11-11 21:00:02.899] [OmniXPU] attention CUTE_BHLD_D128 #2: heads=56 q=452 kv=452 dtype=torch.bfloat16[2026-11-11 21:00:02.962] [OmniXPU] attention CUTE_BHLD_D128 #3: heads=56 q=452 kv=452 dtype=torch.bfloat16[2026-11-11 21:00:26.760] 100%|████████████████████████████████████████████████████████████████████████████████████| 6/6 [00:23<00:00, 3.77s/it]100%|████████████████████████████████████████████████████████████████████████████████████| 6/6 [00:23<00:00, 3.94s/it][2026-11-11 21:00:50.091] VRAM清理完成 [卸载模型: True, 清空缓存: True][2026-11-11 21:00:50.719] Requested to load MiniMaxH3ImageVAE[2026-11-11 21:00:54.468] loaded completely; 10103.00 MB usable, 1114.23 MB loaded, full load: True[2026-11-11 21:00:55.092] [OmniXPU] attention fallback: batch=4 (seq=261)[2026-11-11 21:00:55.802] [OmniXPU] attention fallback: batch=4 (seq=261)[2026-11-11 21:00:55.804] [OmniXPU] attention fallback: batch=4 (seq=261)[2026-11-11 21:00:55.806] [OmniXPU] attention fallback: batch=4 (seq=261)[2026-11-11 21:00:55.807] [OmniXPU] attention fallback: batch=4 (seq=261)[2026-11-11 21:00:56.390] VRAMdebug: free memory before: 11,346,875,904[2026-11-11 21:00:57.181] VRAMdebug: free memory after: 12,526,496,768[2026-11-11 21:00:57.181] VRAMdebug: freed memory: 1,179,620,864[2026-11-11 21:01:04.218] Prompt executed in 173.93 seconds