[Bug] MiniMax-H3 I2VA/FL2VA: ROCm VAE encode crashes with 0xC0000005 (ACCESS_VIOLATION) on Windows

[farawayso]

[Bug] MiniMax-H3 I2VA/FL2VA: ROCm VAE encode crashes with 0xC0000005 (ACCESS_VIOLATION) on Windows

sd-cli version: stable-diffusion.cpp version unknown, commit 97d2990 (pre-built Windows ROCm binary, ~2026-08-19)
OS: Windows 11
GPU: AMD Radeon RX 7900 XTX (gfx1100, 24GB VRAM / 23.7GB usable)
ROCm: ROCm 7.14.0 (bundled with the pre-built binary)

Description

MiniMax-H3 I2VA (image-to-video+audio) and FL2VA (first+last frame → video+audio) modes crash with exit code 3221225725 (0xC0000005, STATUS_ACCESS_VIOLATION) when using the ROCm HIP backend (--backend llm=cpu,vae=ROCm0,diffusion=ROCm0).

The crash is deterministic (100% reproduction rate) and occurs during VAE encode of the input image — the very first GPU operation after model loading. The diffusion sampling and VAE decode stages are never reached.

Reproduction

Command line

sd-cli.exe --mode vid_gen \
  -p "<lora:minimax_h3_turbo_v11_sdcpp:1.0> 美女说，树青，听说你去洱海了。" \
  --init-image input.png \
  --width 480 --height 864 \
  --video-frames 175 \
  --sample-steps 5 --flow-shift 6 --cfg-scale 1.0 \
  --seed 42 \
  --backend llm=cpu,vae=ROCm0,diffusion=ROCm0 \
  --offload-to-cpu \
  --eager-load \
  --lora-model-dir ./models/MiniMax-H3/loras \
  --lora-apply-mode at_runtime \
  --diffusion-fa \
  --max-vram ROCm0=22

The crash also occurs without --offload-to-cpu, with --max-vram values 14–22, and with or without --eager-load. The only variable that matters is the backend: vae=ROCm0 always crashes; vae=Vulkan1 always works.

Via menu

Select backend [5] ROCm+VAE乒乓, model pruned, mode I2VA, provide an input image. The auto-route (menu [4]) correctly avoids this by sending I2VA/FL2VA to Vulkan — the crash only happens when ROCm is explicitly selected.

Crash log

[INFO ] stable-diffusion.cpp:6123 - MiniMax-H3 I2VA
[DEBUG] vae.hpp:142  - VAE Tile size: 16x16
[DEBUG] ggml_extend.hpp:884  - num tiles : 2, 5
[DEBUG] ggml_extend.hpp:885  - optimal overlap : 0.125000, 0.406250 (targeting 0.250000)
[DEBUG] ggml_extend.hpp:886  - processing 10 tiles
[INFO ] ggml_graph_cut.cpp:973  - minimax_h3_video_vae build cached graph cut plan done (taking 1 ms)
[DEBUG] model_manager.cpp:506  - model manager staged compute params (344.04 MB, 118 tensors) to ROCm0, taking 0.11s
[DEBUG] ggml_extend.hpp:2244 - minimax_h3_video_vae compute buffer size: 627.13 MB(VRAM)

  [STDERR]   Device 0: AMD Radeon RX 7900 XTX, gfx1100 (0x1100), VMM: no, Wave Size: 32, VRAM: 24560 MiB
[FAILED] sd-cli.exe 退出码: 3221225725

The image preprocessing (crop + resize) completes successfully, and all four model files load and initialize normally. The crash happens at the instant the VAE encode graph is computed on ROCm0 — 118 tensors (344MB) are staged to VRAM, then the compute buffer is allocated (627MB), and immediately the process dies with an access violation. No error message from the HIP runtime, no user-mode exception handling.

Successful Vulkan comparison (same config, same hardware)

Running the identical config with --backend llm=cpu,vae=Vulkan1,diffusion=Vulkan1 completes successfully:

[DEBUG] vae.hpp:142  - VAE Tile size: 16x16
[DEBUG] ggml_extend.hpp:2244 - minimax_h3_video_vae compute buffer size: 627.13 MB(VRAM)
  |==================================================| 10/10 - 19.61it/s
[DEBUG] vae.hpp:166  - computing vae encode graph completed, taking 0.62s

Stage	ROCm (crash)	Vulkan (works)
VAE encode (image→latent)	0xC0000005 crash during compute	0.62s, 19.6it/s, clean
LLM vision conditioning	never reached	68.6s (normal for Qwen3-VL)
Diffusion sampling (5 steps)	never reached	122.8s@22.1s/it
VAE decode (4 passes × 10 tiles)	never reached	21.9s
Total generation	crash	231.1s / 4m17s

What has been ruled out

• Not the graph-cut bugs (MiniMax-H3: segmented offload (max_vram) deterministically produces an unrelated image (SD1.5 unaffected) #1865/fix: prevent gallocr hash overflow in tiny graph-cut segments #1880/fix: mark graph cuts with both a prefix and a suffix #1883): Those were in the diffusion sampling graph-cut path and were fixed in master-826 (16304cc). The VAE encode path is separate code — it uses its own graph-cut plan (minimax_h3_video_vae), not the diffusion model's graph. The encode path has never been patched.
• Not an OOM: VRAM during load is 0MB (all weights in RAM via --offload-to-cpu), and the encode only needs 344MB staged + 627MB compute buffer = ~1GB, well within 23.7GB available. The same config runs to completion on Vulkan with the same VRAM budget.
• Not model-specific: Both minimax_h3_fl2va_pruned-Q4_K_M (11GB) and minimax_h3_fl2va-Q4_K_M (18GB, full) models crash identically when I2VA/FL2VA is used on ROCm.
• Not input-image-specific: Tested with various input images (768×1368, 2048×2048, diverse content) — all crash at the same point.
• Not mitigatable: --offload-to-cpu, different --max-vram values (14–22), --eager-load, --vae-on-cpu (deprecated) — none prevent the crash. The only working VAE encode path on this hardware is Vulkan.

Hypothesis

The MiniMax-H3 video VAE (minimax_h3_video_vae_fp16.safetensors, 5.6GB, 562 tensors) is a custom architecture that differs from the standard SD VAE — it has a 24-channel latent space, temporal clip length 17, and uses tiled encoding. The ggml HIP backend for ROCm likely has a bug in one of the operations used by the VAE encoder graph (possibly a 2D convolution, a memory layout/view operation, or the tiled overlap handling) that causes an out-of-bounds memory access. The Vulkan backend handles the same computation graph without issue, suggesting the bug is in the HIP kernel implementation, not in the model itself or the higher-level VAE code.

Additional context: T2V/T2VA works fine on ROCm

Text-only modes (T2V/T2VA, no image input, no VAE encode needed) run flawlessly on the same ROCm binary — including VAE decode during the final video assembly. The VAE decode path uses the same model weights but different graph operations, and it completes successfully. This strongly points to a bug specific to the VAE encode graph's kernel set on the ROCm HIP backend.

Attachments

• Full crash log (from --verbose output): the log above is the complete relevant portion; the model loading section (before the crash) is unremarkable — all four models load without warnings.
• Successful Vulkan run log: available on request.

[stduhpf]

Did you manage to run any other model with the same ROCm build (are you sure it is specifically a MiniMax VAE issue)?

[farawayso]

Did you manage to run any other model with the same ROCm build (are you sure it is specifically a MiniMax VAE issue)?

@stduhpf Yes — I ran other models on the exact same ROCm build (commit 97d2990, the pre-built Windows ROCm binary), and the crash is specific to the MiniMax-H3 video VAE's encode graph, not a general problem with this ROCm build or with VAE encode on HIP:

Model	Mode	VAE encode (ROCm0, same build)
Z-Image Turbo (z-image ae.safetensors)	txt2img	not exercised (no encode needed) — but full pipeline + VAE decode works
Boogu Image Edit (img2img, z-image ae.safetensors)	img2img 512x512, untiled	✅ works — encode 3.20s, full run OK
Boogu Image Edit, forced --vae-tiling	img2img 1280x1280, 9 tiles	✅ works — encode 3.08s (tiled encode fine on HIP!)
MiniMax-H3 T2V/T2VA (minimax_h3_video_vae_fp16)	video, no encode	✅ works incl. VAE decode (same weights, entire pipeline OK)
MiniMax-H3 I2VA/FL2VA (same VAE)	video + input image, 10 tiles (2x5, overlap 0.125/0.40625)	❌ 0xC0000005, deterministic, at first graph compute

So: VAE encode in general (including tiled encode with the 2D AE) runs fine on this HIP build on the same GPU. The access violation only happens when computing the MiniMax-H3 video VAE encoder graph — a 3D/temporal VAE (24-channel latent, temporal clip 17, 562 tensors) whose tiled encode path differs from the standard 2D AE.

One more data point that suggests this is a HIP-kernel-set issue rather than a tiling logic issue: the identical tiled encode of the same MiniMax VAE completes cleanly on Vulkan (0.62s), and the MiniMax VAE decoder (different op set from the encoder, same weights) runs fine on ROCm. The crash happens the instant the encoder graph is computed (118 tensors staged → 627MB buffer allocated → process dies), with no HIP runtime error — consistent with an out-of-bounds access inside a HIP kernel used only by the encoder graph (possibly the 3D conv/layout ops).

If it helps narrow it down, I can also reproduce with a minimal input (single tile, no overlap) or grab a verbose graph dump — just tell me what would be most useful. For now I've verified the report is not a build-level issue: the same binary+backend handles VAE encode/tiled-encode fine for other models.

[stduhpf]

That's helpful thanks. Minimax VAE encode works fine on my local HIP builds so that's strange. I don't think the prebuilt releases are compatibe with my hardware so I can't test it myself.

[farawayso]

============================================================
MiniMax-H3 视频生成 7900XTX + ROCm/Vulkan
(全帧安全上限按官方对齐动态计算: 对齐 17k+5 / x32)

选择后端:
[1] ROCm (HIP) [T2V/T2VA 首选: 原生栈更快]
[2] Vulkan 分流 [I2V/FL2VA: 扩散+VAE 7900XTX, llm CPU]
[3] Vulkan 双卡 [full 备用/实验: 扩散7900XTX 流式]
[4] Auto 自动选择 [T2V/T2VA走ROCm, I2V/FL2VA走Vulkan]
[5] ROCm+VAE乒乓 [VAE→RAM驻留, 解码时自动搬回独显, 推荐, 长视频用这个]
选择 [4]: 5
模型类型:
[1] FL2VA (文本/图像生成视频) [默认]
[2] Ref2VA (参考图/视频/音频)
选择 [1]:
模型规模:
[1] pruned (11GB) [默认, 快速, 8步]
[2] full (18GB) [高质量慢速, 20步, 需 stream-layers]
选择 [1]:
生成模式:
[1] T2VA 文本+音频
[2] T2V 仅文本
[3] I2VA 文本+首帧图片
[4] FL2VA 文本+首帧+末帧图片
[5] Ref2VA 文本+参考图/视频/音频
[6] 长视频 格式拼接(同提示词)
[7] 对话链 每段独自台词+音频
[8] 参考图对话链 每段独自台词+角色参考图
选择 [1]: 3
===== 分辨率选择 (官方对齐: 9:16竖屏 / 32的倍数, 全帧安全上限在此范围内自动计算) =====
[1] 480x864 标准 9:16档 [默认, 速度最快]
[2] 576x1024 标准 精确9:16 [推荐, 帧数较多]
[3] 736x1280 标准 抖音720P
[4] 768x1344 标准 9:16官方短边768标准
[5] 自定义
选择 [1]: 3
步数:
[1] 12 步 [快速, 仍欠采样, 不推荐]
[2] 20 步 [标准推荐]
[3] 30 步 [高质量慢速]
[4] 自定义
[5] Turbo 4步 [LoRA加速, 5步, flow-shift 6, 快速, 默认]
[6] Turbo 8步 [LoRA加速, 8步, flow-shift 12, 质量较好]
[7] Spectrum跳步 20步 [1.54x, 质量持平, 不用LoRA]
[8] Turbo 4步 v1.1 [LoRA新版, 768p训练, 快速]
选择 [5]: 8
[INFO] VAE 不在主卡 (乒乓): 安全上限提高到 73 帧
===== 帧数 (当前安全上限 73 帧, 官方对齐 17k+5) =====
[1] 56 帧 [保守]
[2] 73 帧 [安全上限,推荐]
[3] 自定义帧数
[4] 按秒数
选择 [2]:
音频:
[1] 开启音频 [默认]
[2] 关闭音频
选择 [1]:

============================================================
配置检查:
后端: roc_pingpong [VAE→RAM, 解码回独显]
模型: fl2va-pruned 模式: i2va
分辨率: 736x1280 (实际 736x1280) 帧数: 73 音频: 1 步数: 5 CFG: 1.0
LoRA: turbo4v11

按回车开始生成，按 Ctrl+C 取消...

提示词: 敢问，有没有人请朕吃饭？
负面提示词 (可选，回车跳过):
首帧图片路径 (输入 q 取消首帧): C:\Users\Faraway-So\Documents\Text-to-image-video\outputs\boogu_turbo_20260813_212414.png

============================================================
MiniMax-H3 i2va / fl2va-pruned 736x1280 73帧 5步 CFG=1.0
后端: ROC_PINGPONG
提示词: 敢问，有没有人请朕吃饭？
首帧: C:\Users\Faraway-So\Documents\Text-to-image-video\outputs\boogu_turbo_20260813_212414.png
音频: 是
输出: 20260822_230643_i2va_pruned_736x1280_73f.mp4
VRAM: 峰值≈22.8GB / 23.7GB [PINGPONG: VAE→RAM]

[Running] sd-cli.exe ...

[INFO] roc_pingpong: VAE 权重放 RAM (--offload-to-cpu), 解码时自动搬回独显 (扩散省 ~5.5GB, 不吃 890M)
[VRAM·PINGPONG] tokens≈20,740 (t=22) | 激活缓冲≈9.92GB + 驻留权重 12.47GB = 峰值≈22.84GB / 可用 23.7GB
[DEBUG] main.cpp:689 - version: stable-diffusion.cpp version unknown, commit 97d2990
[DEBUG] main.cpp:690 - System Info:
SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | REPACK = 1 |
[DEBUG] main.cpp:691 - SDCliParams {
mode: vid_gen,
output_path: "C:\Users\Faraway-So\Documents\Text-to-image-video\outputs\20260822_230643_i2va_pruned_736x1280_73f.mp4",
image_path: "",
metadata_format: "text",
verbose: true,
color: false,
canny_preprocess: false,
convert_name: false,
preview_method: none,
preview_interval: 1,
preview_path: "preview.avi",
preview_fps: 24,
taesd_preview: false,
preview_noisy: false,
imatrix_out: "",
metadata_raw: false,
metadata_brief: false,
metadata_all: false
}
[DEBUG] main.cpp:692 - SDContextParams {
n_threads: 8,
model_path: "",
clip_l_path: "",
clip_g_path: "",
clip_vision_path: "",
t5xxl_path: "",
llm_path: "C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\qwen3vl_32b_minimax_h3-Q4_K_M.gguf",
llm_vision_path: "",
diffusion_model_path: "C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\minimax_h3_fl2va_pruned-Q4_K_M.gguf",
high_noise_diffusion_model_path: "",
uncond_diffusion_model_path: "",
embeddings_connectors_path: "",
vae_path: "C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\minimax_h3_video_vae_fp16.safetensors",
vae_format: "auto",
audio_vae_path: "C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\minimax_h3_audio_vae_fp32.safetensors",
taesd_path: "",
esrgan_path: "",
control_net_path: "",
embedding_dir: "",
embeddings: {
}
wtype: NONE,
tensor_type_rules: "",
lora_model_dir: "C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\loras",
hires_upscalers_dir: "",
photo_maker_path: "",
rng_type: cpu,
sampler_rng_type: NONE,
offload_params_to_cpu: true,
max_vram: "ROCm0=22",
stream_layers: false,
eager_load: true,
backend: "llm=cpu,vae=ROCm0,diffusion=ROCm0",
params_backend: "",
split_mode: "",
model_args: "",
auto_fit: false,
enable_mmap: false,
control_net_cpu: false,
clip_on_cpu: false,
vae_on_cpu: false,
flash_attn: false,
diffusion_flash_attn: true,
diffusion_conv_direct: false,
vae_conv_direct: false,
prediction: NONE,
lora_apply_mode: at_runtime,
force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:693 - SDGenerationParams {
loras: "{
"C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\loras\minimax_h3_turbo_v11_sdcpp.safetensors": "1"
}",
high_noise_loras: "{
}",
prompt: " 敢问，有没有人请朕吃饭？",
negative_prompt: "",
ad_model_path: "",
ad_prompt: "",
ad_negative_prompt: "",
extra_ad_args: "",
clip_skip: -1,
width: 736,
height: 1280,
batch_count: 1,
qwen_image_layers: 3,
init_image_path: "C:\Users\Faraway-So\Documents\Text-to-image-video\outputs\boogu_turbo_20260813_212414.png",
end_image_path: "",
mask_image_path: "",
control_image_path: "",
ref_image_paths: [],
ref_video_paths: [],
ref_video_audio_paths: [],
ref_audio_paths: [],
control_video_path: "",
auto_resize_ref_image: true,
increase_ref_index: false,
pm_id_images_dir: "",
pm_id_embed_path: "",
pm_style_strength: 20,
skip_layers: [7, 8, 9],
sample_params: (txt_cfg: 1.00, img_cfg: 1.00, distilled_guidance: 3.50, slg.layer_count: 0, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: simple, sample_method: euler, sample_steps: 5, eta: inf, shifted_timestep: 0, flow_shift: 6.00, extra_sample_args: ),
high_noise_skip_layers: [7, 8, 9],
high_noise_sample_params: (txt_cfg: 7.00, img_cfg: 7.00, distilled_guidance: 3.50, slg.layer_count: 0, slg.layer_start: 0.01, slg.layer_end: 0.20, slg.scale: 0.00, scheduler: NONE, sample_method: NONE, sample_steps: 20, eta: inf, shifted_timestep: 0, flow_shift: inf, extra_sample_args: ),
custom_sigmas: [],
cache_mode: "",
cache_option: "",
cache: disabled (threshold=inf, start=0.15, end=0.95),
moe_boundary: 0.875,
video_frames: 73,
fps: 24,
vace_strength: 1,
strength: 0.75,
control_strength: 0.9,
seed: 42,
upscale_repeats: 1,
upscale_tile_size: 128,
hires: { enabled: false, upscaler: "Latent", model_path: "", scale: 2, target_width: 0, target_height: 0, steps: 0, denoising_strength: 0.7, custom_sigmas: [], upscale_tile_size: 128 },
vae_tiling_params: { 0, 0, 0, 0, 0.5, 0, 0, "" },
}
[INFO ] media_io.cpp:575 - crop input image from 1024x1024 to 588x1024, image_path = C:\Users\Faraway-So\Documents\Text-to-image-video\outputs\boogu_turbo_20260813_212414.png
[INFO ] media_io.cpp:592 - resize input image from 588x1024 to 736x1280
[DEBUG] model_loader.cpp:228 - using 8 threads for model loading
[INFO ] stable-diffusion.cpp:720 - loading diffusion model from 'C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\minimax_h3_fl2va_pruned-Q4_K_M.gguf'
[INFO ] model_loader.cpp:236 - load C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\minimax_h3_fl2va_pruned-Q4_K_M.gguf using gguf format
[DEBUG] model_loader.cpp:288 - init from 'C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\minimax_h3_fl2va_pruned-Q4_K_M.gguf'
[INFO ] stable-diffusion.cpp:777 - loading llm from 'C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\qwen3vl_32b_minimax_h3-Q4_K_M.gguf'
[INFO ] model_loader.cpp:236 - load C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\qwen3vl_32b_minimax_h3-Q4_K_M.gguf using gguf format
[DEBUG] model_loader.cpp:288 - init from 'C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\qwen3vl_32b_minimax_h3-Q4_K_M.gguf'
[INFO ] stable-diffusion.cpp:791 - loading vae from 'C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\minimax_h3_video_vae_fp16.safetensors'
[INFO ] model_loader.cpp:242 - load C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\minimax_h3_video_vae_fp16.safetensors using safetensors format
[DEBUG] model_loader.cpp:316 - init from 'C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\minimax_h3_video_vae_fp16.safetensors', prefix = 'vae.'
[INFO ] stable-diffusion.cpp:815 - loading audio VAE from 'C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\minimax_h3_audio_vae_fp32.safetensors'
[INFO ] model_loader.cpp:242 - load C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\minimax_h3_audio_vae_fp32.safetensors using safetensors format
[DEBUG] model_loader.cpp:316 - init from 'C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\minimax_h3_audio_vae_fp32.safetensors', prefix = ''
[INFO ] stable-diffusion.cpp:908 - Version: MiniMax-H3
[DEBUG] ggml_extend_backend.cpp:395 - Initializing backend: ROCm0
[DEBUG] ggml_extend_backend.cpp:395 - Initializing backend: CPU
[INFO ] stable-diffusion.cpp:959 - Weight type stat: f32: 927 | f16: 664 | q4_K: 458 | q6_K: 100 | bf16: 764
[INFO ] stable-diffusion.cpp:960 - Conditioner weight type stat: q4_K: 250 | q6_K: 100 | bf16: 552
[INFO ] stable-diffusion.cpp:961 - Diffusion model weight type stat: f32: 10 | f16: 102 | q4_K: 208 | bf16: 212
[INFO ] stable-diffusion.cpp:962 - VAE weight type stat: f16: 562
[DEBUG] stable-diffusion.cpp:964 - ggml tensor size = 432 bytes
[DEBUG] qwen2_tokenizer.cpp:14 - merges size 151387
[DEBUG] qwen2_tokenizer.cpp:39 - vocab size: 151674
[DEBUG] llm.hpp:290 - llm: num_layers = 50, vocab_size = 151936, hidden_size = 5120, intermediate_size = 25600
[DEBUG] llm.hpp:1768 - enable llm vision
[DEBUG] minimax_h3.hpp:116 - minimax_h3: layers=50, hidden=5376, heads=56, head_dim=128, ffn=14336, adaln_curve=1025
[INFO ] stable-diffusion.cpp:1512 - using VAE for encoding / decoding
[INFO ] stable-diffusion.cpp:1637 - Using flash attention in the diffusion model
[DEBUG] stable-diffusion.cpp:1645 - validating model metadata
[DEBUG] model_loader.cpp:1045 - loading 532/532 tensors from C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\minimax_h3_fl2va_pruned-Q4_K_M.gguf
|########## | 532/2910 - 1.86GB/s
[DEBUG] model_loader.cpp:1045 - loading 902/902 tensors from C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\qwen3vl_32b_minimax_h3-Q4_K_M.gguf
|######################### | 1434/2910 - 1.66GB/s
[DEBUG] model_loader.cpp:1045 - loading 560/562 tensors from C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\minimax_h3_video_vae_fp16.safetensors
|################################### | 1994/2910 - 1.69GB/s
[DEBUG] model_loader.cpp:1045 - loading 916/917 tensors from C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\minimax_h3_audio_vae_fp32.safetensors
|##################################################| 2910/2910 - 1.69GB/s
[INFO ] model_loader.cpp:1309 - loading tensors completed, taking 19.52s (read: 18.63s, memcpy: 0.00s, convert: 0.52s, copy_to_backend: 0.00s)
[DEBUG] model_manager.cpp:409 - model manager prepared params backend buffer (18864.74 MB, 902 tensors, RAM)
[DEBUG] model_manager.cpp:409 - model manager prepared params backend buffer (16534.02 MB, 2008 tensors, RAM)
[DEBUG] stable-diffusion.cpp:1702 - model metadata validated; weights pre-loaded to params backend
[INFO ] stable-diffusion.cpp:1760 - total params memory size = 35398.76MB (VRAM 0.00MB, RAM 35398.76MB): text_encoders 18864.74MB(RAM), diffusion_model 10975.93MB(RAM), vae 5558.09MB(RAM), controlnet 0.00MB(N/A), extensions 0.00MB(N/A)
[INFO ] stable-diffusion.cpp:1864 - running in FLOW mode
[DEBUG] stable-diffusion.cpp:2120 - lora C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\loras\minimax_h3_turbo_v11_sdcpp.safetensors:1.00
[INFO ] stable-diffusion.cpp:2037 - apply lora at runtime
[INFO ] model_loader.cpp:242 - load C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\loras\minimax_h3_turbo_v11_sdcpp.safetensors using safetensors format
[DEBUG] model_loader.cpp:316 - init from 'C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\loras\minimax_h3_turbo_v11_sdcpp.safetensors', prefix = 'lora.'
[INFO ] lora.hpp:47 - loading LoRA from 'C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\loras\minimax_h3_turbo_v11_sdcpp.safetensors'
[DEBUG] model_loader.cpp:228 - using 8 threads for model loading
[DEBUG] model_loader.cpp:1045 - loading 416/416 tensors from C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\loras\minimax_h3_turbo_v11_sdcpp.safetensors
|##################################################| 416/416 - 0.00MB/s
[INFO ] model_loader.cpp:1309 - loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s)
[INFO ] model_loader.cpp:242 - load C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\loras\minimax_h3_turbo_v11_sdcpp.safetensors using safetensors format
[DEBUG] model_loader.cpp:316 - init from 'C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\loras\minimax_h3_turbo_v11_sdcpp.safetensors', prefix = 'lora.'
[INFO ] lora.hpp:47 - loading LoRA from 'C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\loras\minimax_h3_turbo_v11_sdcpp.safetensors'
[DEBUG] model_loader.cpp:228 - using 8 threads for model loading
[DEBUG] model_loader.cpp:1045 - loading 416/416 tensors from C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\loras\minimax_h3_turbo_v11_sdcpp.safetensors
|##################################################| 416/416 - 0.00MB/s
[INFO ] model_loader.cpp:1309 - loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s)
[DEBUG] model_loader.cpp:1045 - loading 400/416 tensors from C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\loras\minimax_h3_turbo_v11_sdcpp.safetensors
|##################################################| 400/400 - 1.43GB/s
[INFO ] model_loader.cpp:1309 - loading tensors completed, taking 1.23s (read: 0.95s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.16s)
[DEBUG] model_manager.cpp:409 - model manager prepared params backend buffer (1793.75 MB, 400 tensors, VRAM)
[DEBUG] lora.hpp:123 - finished loaded lora
[INFO ] model_loader.cpp:242 - load C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\loras\minimax_h3_turbo_v11_sdcpp.safetensors using safetensors format
[DEBUG] model_loader.cpp:316 - init from 'C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\loras\minimax_h3_turbo_v11_sdcpp.safetensors', prefix = 'lora.'
[INFO ] lora.hpp:47 - loading LoRA from 'C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\loras\minimax_h3_turbo_v11_sdcpp.safetensors'
[DEBUG] model_loader.cpp:228 - using 8 threads for model loading
[DEBUG] model_loader.cpp:1045 - loading 416/416 tensors from C:\Users\Faraway-So\Documents\Text-to-image-video\models\MiniMax-H3\loras\minimax_h3_turbo_v11_sdcpp.safetensors
|##################################################| 416/416 - 0.00MB/s
[INFO ] model_loader.cpp:1309 - loading tensors completed, taking 0.21s (read: 0.00s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s)
[INFO ] stable-diffusion.cpp:2135 - apply_loras completed, taking 3.24s
[INFO ] denoiser.hpp:1079 - get_sigmas with Simple scheduler
[INFO ] stable-diffusion.cpp:4380 - sampling using Euler method
[INFO ] stable-diffusion.cpp:6123 - MiniMax-H3 I2VA
[DEBUG] vae.hpp:142 - VAE Tile size: 16x16
[DEBUG] ggml_extend.hpp:884 - num tiles : 4, 7
[DEBUG] ggml_extend.hpp:885 - optimal overlap : 0.375000, 0.333333 (targeting 0.250000)
[DEBUG] ggml_extend.hpp:886 - processing 28 tiles
[INFO ] ggml_graph_cut.cpp:973 - minimax_h3_video_vae build cached graph cut plan done (taking 1 ms)
[DEBUG] model_manager.cpp:506 - model manager staged compute params (344.04 MB, 118 tensors) to ROCm0, taking 0.10s
[DEBUG] ggml_extend.hpp:2244 - minimax_h3_video_vae compute buffer size: 627.13 MB(VRAM)
|==============> | 8/28 - 7.19it/s[DEBUG] ggml_extend.hpp:63 - ggml_backend_cuda_graph_compute: CUDA graph warmup complete

[STDERR] Device 0: AMD Radeon RX 7900 XTX, gfx1100 (0x1100), VMM: no, Wave Size: 32, VRAM: 24560 MiB
[FAILED] sd-cli.exe 退出码: 3221225725
[FAILED] 退出码: -1
请按任意键继续. . .