Skip to content

[Bug] TAEHV + LoRA configuration change causes black image and subsequent crash #1849

New issue
New issue
Closed
Closed
[Bug] TAEHV + LoRA configuration change causes black image and subsequent crash#1849
Labels
bugSomething isn't working

Description

@cs8425
cs8425
opened on Aug 3, 2026

Git commit

$docker run --rm -it --entrypoint="/sd-server" ghcr.io/leejet/stable-diffusion.cpp:master-vulkan --version
stable-diffusion.cpp version unknown, commit db99efd

Operating System & Version

Ubuntu 24.04, Win10 22H2

GGML backends

Vulkan

Command-line arguments used

-v --diffusion-model /app/anima-base/anima-base-v1.0.safetensors --tae /app/vae/taew2_1.safetensors --llm /app/anima-base/qwen_3_06b_base.safetensors --cfg-scale 1 --steps 8 --sampling-method er_sde --fa

Steps to reproduce

Case 1

  1. Start the server with:

  2. Generate an image with a LoRA enabled.

  3. Change the LoRA configuration:

    • adjust LoRA weight, or
    • remove the LoRA, or
    • add another LoRA.

  4. Generate image again.

Result: the generated image is completely black.

  1. Generate one more image (with or without changing any parameters such as prompt, seed, steps, etc.).

Result: the server crashes, exit code 139.

Case 2

  1. Start the server with:

  2. Generate an image without any LoRA.

  3. Add a LoRA.

  4. Generate image again.

Result: the generated image is completely black.

  1. Generate one more image.

Result: the server crashes, exit code 139.

What you expected to happen

Changing LoRA parameters should only affect the generated image output.

The server should continue generating images normally without producing black images or crashing.

What actually happened

After server startup, the first generation always succeeds.

If the LoRA configuration is changed after the first generation, including:

  • enabling a LoRA
  • disabling a LoRA
  • changing LoRA weight

the next generated image becomes completely black.

Any subsequent generation attempt then crashes the server process (exit code 139).

The issue reproduces consistently on both:

  • Windows 10 22H2 + NVIDIA 4070S
  • Linux + AMD 9060XT

This suggests the issue is not vendor-specific and may be related to LoRA reload handling when TAEHV is used.

Logs / error messages / stack trace

server startup logs:

Details 

[DEBUG] main.cpp:82   - version: stable-diffusion.cpp version unknown, commit db99efd
WARNING: radv is not a conformant Vulkan implementation, testing use only.
ggml_vulkan: Found 2 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon Graphics (RADV GFX1200) (radv) | uma: 0 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 0 | matrix cores: KHR_coopmat
ggml_vulkan: 1 = AMD Radeon Graphics (RADV RENOIR) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 0 | matrix cores: none
load_backend: loaded Vulkan backend from /sd.cpp/bin/libggml-vulkan.so
load_backend: loaded CPU backend from /sd.cpp/bin/libggml-cpu-haswell.so
[DEBUG] main.cpp:83   - System Info: 
   SSE3 = 1 |    SSSE3 = 1 |    AVX = 1 |    AVX2 = 1 |    F16C = 1 |    FMA = 1 |    BMI2 = 1 |    OPENMP = 1 |    REPACK = 1 | 
[DEBUG] main.cpp:84   - SDSvrParams {
  listen_ip: 0.0.0.0,
  listen_port: "8081",
  serve_html_path: "",
}
[DEBUG] main.cpp:85   - SDContextParams {
  n_threads: 4,
  model_path: "",
  clip_l_path: "",
  clip_g_path: "",
  clip_vision_path: "",
  t5xxl_path: "",
  llm_path: "/app/anima-base/qwen_3_06b_base.safetensors",
  llm_vision_path: "",
  diffusion_model_path: "/app/anima-base/anima-base-v1.0.safetensors",
  high_noise_diffusion_model_path: "",
  uncond_diffusion_model_path: "",
  embeddings_connectors_path: "",
  vae_path: "",
  vae_format: "auto",
  audio_vae_path: "",
  taesd_path: "/app/vae/taew2_1.safetensors",
  esrgan_path: "",
  control_net_path: "",
  embedding_dir: "",
  embeddings: {
  }
  wtype: NONE,
  tensor_type_rules: "",
  lora_model_dir: ".",
  hires_upscalers_dir: "",
  photo_maker_path: "",
  rng_type: cuda,
  sampler_rng_type: NONE,
  offload_params_to_cpu: false,
  max_vram: "0",
  stream_layers: false,
  eager_load: false,
  backend: "",
  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: true,
  diffusion_flash_attn: false,
  diffusion_conv_direct: false,
  vae_conv_direct: false,
  prediction: NONE,
  lora_apply_mode: auto,
  force_sdxl_vae_conv_scale: false
}
[DEBUG] main.cpp:86   - SDGenerationParams {
  loras: "{
  }",
  high_noise_loras: "{
  }",
  prompt: "",
  negative_prompt: "",
  ad_model_path: "",
  ad_prompt: "",
  ad_negative_prompt: "",
  extra_ad_args: "",
  clip_skip: -1,
  width: -1,
  height: -1,
  batch_count: 1,
  qwen_image_layers: 3,
  init_image_path: "",
  end_image_path: "",
  mask_image_path: "",
  control_image_path: "",
  ref_image_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: NONE, sample_method: er_sde, sample_steps: 8, eta: inf, shifted_timestep: 0, flow_shift: inf, 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: 1,
  fps: 16,
  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, "" },
}
[DEBUG] model_loader.cpp:228  - using 4 threads for model loading
[INFO ] stable-diffusion.cpp:712  - loading diffusion model from '/app/anima-base/anima-base-v1.0.safetensors'
[INFO ] model_loader.cpp:242  - load /app/anima-base/anima-base-v1.0.safetensors using safetensors format
[DEBUG] model_loader.cpp:316  - init from '/app/anima-base/anima-base-v1.0.safetensors', prefix = 'model.diffusion_model.'
[INFO ] stable-diffusion.cpp:769  - loading llm from '/app/anima-base/qwen_3_06b_base.safetensors'
[INFO ] model_loader.cpp:242  - load /app/anima-base/qwen_3_06b_base.safetensors using safetensors format
[DEBUG] model_loader.cpp:316  - init from '/app/anima-base/qwen_3_06b_base.safetensors', prefix = 'text_encoders.llm.'
[INFO ] stable-diffusion.cpp:791  - loading tae from '/app/vae/taew2_1.safetensors'
[INFO ] model_loader.cpp:242  - load /app/vae/taew2_1.safetensors using safetensors format
[DEBUG] model_loader.cpp:316  - init from '/app/vae/taew2_1.safetensors', prefix = 'tae.'
[INFO ] stable-diffusion.cpp:900  - Version: Anima 
[DEBUG] ggml_extend_backend.cpp:545  - Found 3 backend devices:
[DEBUG] ggml_extend_backend.cpp:548  - #0: Vulkan0
[DEBUG] ggml_extend_backend.cpp:548  - #1: Vulkan1
[DEBUG] ggml_extend_backend.cpp:548  - #2: CPU
[DEBUG] ggml_extend_backend.cpp:395  - Initializing backend: Vulkan0
[INFO ] stable-diffusion.cpp:951  - Weight type stat:                      f16: 128  |    bf16: 995  
[INFO ] stable-diffusion.cpp:952  - Conditioner weight type stat:         bf16: 310  
[INFO ] stable-diffusion.cpp:953  - Diffusion model weight type stat:     bf16: 685  
[INFO ] stable-diffusion.cpp:954  - VAE weight type stat:             
[DEBUG] stable-diffusion.cpp:956  - ggml tensor size = 432 bytes
[DEBUG] qwen2_tokenizer.cpp:14   - merges size 151387
[DEBUG] qwen2_tokenizer.cpp:39   - vocab size: 151674
[DEBUG] llm.hpp:263  - llm: num_layers = 28, vocab_size = 151936, hidden_size = 1024, intermediate_size = 3072
[DEBUG] anima.hpp:49   - anima: num_layers = 28, hidden_size = 2048, num_heads = 16, head_dim = 128
[INFO ] stable-diffusion.cpp:1458 - using TAE for encoding / decoding
[INFO ] stable-diffusion.cpp:1572 - Using flash attention
[INFO ] stable-diffusion.cpp:1586 - Using flash attention in the diffusion model
[DEBUG] stable-diffusion.cpp:1594 - validating model metadata
[DEBUG] stable-diffusion.cpp:1651 - model metadata validated; weights will be prepared lazily
[INFO ] stable-diffusion.cpp:1692 - total params memory size = 5506.62MB (VRAM 5506.62MB, RAM 0.00MB): text_encoders 1433.75MB(VRAM), diffusion_model 4051.28MB(VRAM), vae 21.59MB(VRAM), controlnet 0.00MB(N/A), extensions 0.00MB(N/A)
[INFO ] stable-diffusion.cpp:1808 - running in FLOW mode
[INFO ] main.cpp:148  - listening on: http://0.0.0.0:8081

generate first image with lora:

Details 

[INFO ] stable-diffusion.cpp:5451 - generate_image 512x768
[DEBUG] stable-diffusion.cpp:2064 - lora ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors:1.00
[INFO ] stable-diffusion.cpp:2079 - apply_loras completed, taking 0.00s
[INFO ] stable-diffusion.cpp:3080 - Using 'cosmos_reference' preset for reference images
[INFO ] denoiser.hpp:1051 - get_sigmas with discrete scheduler
[INFO ] stable-diffusion.cpp:4280 - sampling using ER-SDE method
[DEBUG] conditioner.hpp:1718 - parse 'masterpiece, best quality, absurdres, ultra detailed, furina \(genshin impact\), blue eyes, heterochromia, white hair, long hair, ahoge, blue and white outfit, hat, gloves, 1girl, sitting atop a giant moon-shaped swing, gentle smile, star-filled sky, glowing clouds below, fantasy nightscape' to [['masterpiece, best quality, absurdres, ultra detailed, furina (genshin impact\), blue eyes, heterochromia, white hair, long hair, ahoge, blue and white outfit, hat, gloves, 1girl, sitting atop a giant moon-shaped swing, gentle smile, star-filled sky, glowing clouds below, fantasy nightscape', 1], ]
[DEBUG] bpe_tokenizer.cpp:208  - split prompt "masterpiece, best quality, absurdres, ultra detailed, furina (genshin impact\), blue eyes, heterochromia, white hair, long hair, ahoge, blue and white outfit, hat, gloves, 1girl, sitting atop a giant moon-shaped swing, gentle smile, star-filled sky, glowing clouds below, fantasy nightscape" to 72 tokens ["master", "piece", ",", "Ġbest", "Ġquality", ",", "Ġabsurd", "res", ",", "Ġultra", "Ġdetailed", ",", "Ġfur", "ina", "Ġ(", "gens", "hin", "Ġimpact", "\", "),", "Ġblue", "Ġeyes", ",", "Ġheter", "och", "rom", "ia", ",", "Ġwhite", "Ġhair", ",", "Ġlong", "Ġhair", ",", "Ġah", "oge", ",", "Ġblue", "Ġand", "Ġwhite", "Ġoutfit", ",", "Ġhat", ",", "Ġgloves", ",", "Ġ", "1", "girl", ",", "Ġsitting", "Ġatop", "Ġa", "Ġgiant", "Ġmoon", "-shaped", "Ġswing", ",", "Ġgentle", "Ġsmile", ",", "Ġstar", "-filled", "Ġsky", ",", "Ġglowing", "Ġclouds", "Ġbelow", ",", "Ġfantasy", "Ġnights", "cape", ]
[DEBUG] t5_unigram_tokenizer.cpp:336  - split prompt "masterpiece, best quality, absurdres, ultra detailed, furina (genshin impact\), blue eyes, heterochromia, white hair, long hair, ahoge, blue and white outfit, hat, gloves, 1girl, sitting atop a giant moon-shaped swing, gentle smile, star-filled sky, glowing clouds below, fantasy nightscape" to tokens ["▁masterpiece", ",", "▁best", "▁quality", ",", "▁absurd", "re", "s", ",", "▁ultra", "▁detailed", ",", "▁fur", "in", "a", "▁(", "gen", "s", "hin", "▁impact", "\", "),", "▁blue", "▁eyes", ",", "▁hetero", "chro", "m", "i", "a", ",", "▁white", "▁hair", ",", "▁long", "▁hair", ",", "▁", "a", "hog", "e", ",", "▁blue", "▁and", "▁white", "▁outfit", ",", "▁", "hat", ",", "▁gloves", ",", "▁1", "girl", ",", "▁sitting", "▁", "a", "top", "▁", "a", "▁giant", "▁moon", "-", "shaped", "▁swing", ",", "▁gentle", "▁smile", ",", "▁star", "-", "filled", "▁sky", ",", "▁glowing", "▁clouds", "▁below", ",", "▁fantasy", "▁night", "scape", ]
[DEBUG] model_loader.cpp:1038 - loading 310/310 tensors from /app/anima-base/qwen_3_06b_base.safetensors
  |##################################################| 310/310 - 1.16GB/s
[INFO ] model_loader.cpp:1300 - loading tensors completed, taking 0.96s (read: 0.08s, memcpy: 0.00s, convert: 0.02s, copy_to_backend: 0.20s)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (504.00 MB, 112 tensors, VRAM)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (929.75 MB, 198 tensors, VRAM)
[INFO ] model_loader.cpp:242  - load ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors using safetensors format
[DEBUG] model_loader.cpp:316  - init from './anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors', prefix = 'lora.'
[INFO ] lora.hpp:47   - loading LoRA from './anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors'
[DEBUG] model_loader.cpp:228  - using 4 threads for model loading
[DEBUG] model_loader.cpp:1038 - loading 1016/1016 tensors from ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors
  |##################################################| 1016/1016 - 0.00MB/s
[INFO ] model_loader.cpp:1300 - loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s)
[DEBUG] model_loader.cpp:1038 - loading 1016/1016 tensors from ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors
  |##################################################| 1016/1016 - 705.85MB/s
[INFO ] model_loader.cpp:1300 - loading tensors completed, taking 0.20s (read: 0.01s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.06s)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (141.88 MB, 1016 tensors, VRAM)
[DEBUG] lora.hpp:123  - finished loaded lora
[DEBUG] ggml_extend.hpp:2193 - lora compute buffer size: 0.00 MB(VRAM)
[INFO ] lora.hpp:997  - (0 / 1016) LoRA tensors have been applied, lora_file_path = ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors
[DEBUG] model_manager.cpp:981  - model manager releasing params backend buffer (141.88 MB, 1016 tensors, VRAM)
[DEBUG] ggml_extend.hpp:2193 - qwen3 compute buffer size: 2.83 MB(VRAM)
[DEBUG] conditioner.hpp:1776 - computing condition graph completed, taking 2339 ms
[INFO ] stable-diffusion.cpp:5131 - get_learned_condition completed, taking 2.34s
[INFO ] stable-diffusion.cpp:5502 - generating image: 1/1 - seed 43
[DEBUG] model_loader.cpp:1038 - loading 685/685 tensors from /app/anima-base/anima-base-v1.0.safetensors
  |##################################################| 685/685 - 2.96GB/s
[INFO ] model_loader.cpp:1300 - loading tensors completed, taking 1.32s (read: 0.35s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.68s)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (1020.23 MB, 224 tensors, VRAM)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (1012.02 MB, 156 tensors, VRAM)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (1020.02 MB, 158 tensors, VRAM)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (999.01 MB, 147 tensors, VRAM)
[INFO ] model_loader.cpp:242  - load ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors using safetensors format
[DEBUG] model_loader.cpp:316  - init from './anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors', prefix = 'lora.'
[INFO ] lora.hpp:47   - loading LoRA from './anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors'
[DEBUG] model_loader.cpp:228  - using 4 threads for model loading
[DEBUG] model_loader.cpp:1038 - loading 1016/1016 tensors from ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors
  |##################################################| 1016/1016 - 0.00MB/s
[INFO ] model_loader.cpp:1300 - loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s)
[DEBUG] model_loader.cpp:1038 - loading 1016/1016 tensors from ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors
  |##################################################| 1016/1016 - 705.85MB/s
[INFO ] model_loader.cpp:1300 - loading tensors completed, taking 0.20s (read: 0.01s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.04s)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (141.88 MB, 1016 tensors, VRAM)
[DEBUG] lora.hpp:123  - finished loaded lora
[DEBUG] ggml_extend.hpp:2193 - lora compute buffer size: 194.25 MB(VRAM)
[INFO ] lora.hpp:997  - (1016 / 1016) LoRA tensors have been applied, lora_file_path = ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors
[DEBUG] model_manager.cpp:981  - model manager releasing params backend buffer (141.88 MB, 1016 tensors, VRAM)
[DEBUG] ggml_extend.hpp:2193 - anima compute buffer size: 78.31 MB(VRAM)
  |==================================================| 8/8 - 1.06s/it
[INFO ] stable-diffusion.cpp:5534 - sampling completed, taking 15.98s
[INFO ] stable-diffusion.cpp:5546 - generating 1 latent images completed, taking 15.98s
[INFO ] stable-diffusion.cpp:5156 - decoding 1 latents
[DEBUG] model_loader.cpp:1038 - loading 64/128 tensors from /app/vae/taew2_1.safetensors
  |##################################################| 64/64 - 93.42MB/s
[INFO ] model_loader.cpp:1300 - loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer ( 18.79 MB, 64 tensors, VRAM)
[INFO ] model_loader.cpp:242  - load ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors using safetensors format
[DEBUG] model_loader.cpp:316  - init from './anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors', prefix = 'lora.'
[INFO ] lora.hpp:47   - loading LoRA from './anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors'
[DEBUG] model_loader.cpp:228  - using 4 threads for model loading
[DEBUG] model_loader.cpp:1038 - loading 1016/1016 tensors from ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors
  |##################################################| 1016/1016 - 0.00MB/s
[INFO ] model_loader.cpp:1300 - loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s)
[DEBUG] model_loader.cpp:1038 - loading 1016/1016 tensors from ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors
  |##################################################| 1016/1016 - 709.37MB/s
[INFO ] model_loader.cpp:1300 - loading tensors completed, taking 0.20s (read: 0.01s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.03s)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (141.88 MB, 1016 tensors, VRAM)
[DEBUG] lora.hpp:123  - finished loaded lora
[DEBUG] ggml_extend.hpp:2193 - lora compute buffer size: 0.00 MB(VRAM)
[INFO ] lora.hpp:997  - (0 / 1016) LoRA tensors have been applied, lora_file_path = ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors
[DEBUG] model_manager.cpp:981  - model manager releasing params backend buffer (141.88 MB, 1016 tensors, VRAM)
[DEBUG] ggml_extend.hpp:2193 - taehv compute buffer size: 2592.38 MB(VRAM)
[DEBUG] vae.hpp:219  - computing vae decode graph completed, taking 0.96s
[INFO ] stable-diffusion.cpp:5222 - latent 1 decoded, taking 0.96s
[INFO ] stable-diffusion.cpp:5226 - decode_first_stage completed, taking 0.96s
[INFO ] stable-diffusion.cpp:5684 - generate_image completed in 19.29s

change lora configuration, generate again, cause black image:

Details 

[INFO ] stable-diffusion.cpp:5451 - generate_image 512x768
[DEBUG] stable-diffusion.cpp:2064 - lora ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors:0.40
[DEBUG] model_manager.cpp:981  - model manager releasing params backend buffer (504.00 MB, 112 tensors, VRAM)
[DEBUG] model_manager.cpp:981  - model manager releasing params backend buffer (929.75 MB, 198 tensors, VRAM)
[DEBUG] model_manager.cpp:981  - model manager releasing params backend buffer (1020.23 MB, 224 tensors, VRAM)
[DEBUG] model_manager.cpp:981  - model manager releasing params backend buffer (1012.02 MB, 156 tensors, VRAM)
[DEBUG] model_manager.cpp:981  - model manager releasing params backend buffer (1020.02 MB, 158 tensors, VRAM)
[DEBUG] model_manager.cpp:981  - model manager releasing params backend buffer (999.01 MB, 147 tensors, VRAM)
[DEBUG] model_manager.cpp:981  - model manager releasing params backend buffer ( 18.79 MB, 64 tensors, VRAM)
[INFO ] stable-diffusion.cpp:2079 - apply_loras completed, taking 0.00s
[INFO ] stable-diffusion.cpp:3080 - Using 'cosmos_reference' preset for reference images
[INFO ] denoiser.hpp:1051 - get_sigmas with discrete scheduler
[INFO ] stable-diffusion.cpp:4280 - sampling using ER-SDE method
[DEBUG] conditioner.hpp:1718 - parse 'masterpiece, best quality, absurdres, ultra detailed, furina \(genshin impact\), blue eyes, heterochromia, white hair, long hair, ahoge, blue and white outfit, hat, gloves, 1girl, sitting atop a giant moon-shaped swing, gentle smile, star-filled sky, glowing clouds below, fantasy nightscape' to [['masterpiece, best quality, absurdres, ultra detailed, furina (genshin impact\), blue eyes, heterochromia, white hair, long hair, ahoge, blue and white outfit, hat, gloves, 1girl, sitting atop a giant moon-shaped swing, gentle smile, star-filled sky, glowing clouds below, fantasy nightscape', 1], ]
[DEBUG] bpe_tokenizer.cpp:208  - split prompt "masterpiece, best quality, absurdres, ultra detailed, furina (genshin impact\), blue eyes, heterochromia, white hair, long hair, ahoge, blue and white outfit, hat, gloves, 1girl, sitting atop a giant moon-shaped swing, gentle smile, star-filled sky, glowing clouds below, fantasy nightscape" to 72 tokens ["master", "piece", ",", "Ġbest", "Ġquality", ",", "Ġabsurd", "res", ",", "Ġultra", "Ġdetailed", ",", "Ġfur", "ina", "Ġ(", "gens", "hin", "Ġimpact", "\", "),", "Ġblue", "Ġeyes", ",", "Ġheter", "och", "rom", "ia", ",", "Ġwhite", "Ġhair", ",", "Ġlong", "Ġhair", ",", "Ġah", "oge", ",", "Ġblue", "Ġand", "Ġwhite", "Ġoutfit", ",", "Ġhat", ",", "Ġgloves", ",", "Ġ", "1", "girl", ",", "Ġsitting", "Ġatop", "Ġa", "Ġgiant", "Ġmoon", "-shaped", "Ġswing", ",", "Ġgentle", "Ġsmile", ",", "Ġstar", "-filled", "Ġsky", ",", "Ġglowing", "Ġclouds", "Ġbelow", ",", "Ġfantasy", "Ġnights", "cape", ]
[DEBUG] t5_unigram_tokenizer.cpp:336  - split prompt "masterpiece, best quality, absurdres, ultra detailed, furina (genshin impact\), blue eyes, heterochromia, white hair, long hair, ahoge, blue and white outfit, hat, gloves, 1girl, sitting atop a giant moon-shaped swing, gentle smile, star-filled sky, glowing clouds below, fantasy nightscape" to tokens ["▁masterpiece", ",", "▁best", "▁quality", ",", "▁absurd", "re", "s", ",", "▁ultra", "▁detailed", ",", "▁fur", "in", "a", "▁(", "gen", "s", "hin", "▁impact", "\", "),", "▁blue", "▁eyes", ",", "▁hetero", "chro", "m", "i", "a", ",", "▁white", "▁hair", ",", "▁long", "▁hair", ",", "▁", "a", "hog", "e", ",", "▁blue", "▁and", "▁white", "▁outfit", ",", "▁", "hat", ",", "▁gloves", ",", "▁1", "girl", ",", "▁sitting", "▁", "a", "top", "▁", "a", "▁giant", "▁moon", "-", "shaped", "▁swing", ",", "▁gentle", "▁smile", ",", "▁star", "-", "filled", "▁sky", ",", "▁glowing", "▁clouds", "▁below", ",", "▁fantasy", "▁night", "scape", ]
[DEBUG] model_loader.cpp:1038 - loading 310/310 tensors from /app/anima-base/qwen_3_06b_base.safetensors
  |##################################################| 310/310 - 1.31GB/s
[INFO ] model_loader.cpp:1300 - loading tensors completed, taking 0.85s (read: 0.09s, memcpy: 0.00s, convert: 0.02s, copy_to_backend: 0.16s)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (504.00 MB, 112 tensors, VRAM)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (929.75 MB, 198 tensors, VRAM)
[INFO ] model_loader.cpp:242  - load ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors using safetensors format
[DEBUG] model_loader.cpp:316  - init from './anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors', prefix = 'lora.'
[INFO ] lora.hpp:47   - loading LoRA from './anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors'
[DEBUG] model_loader.cpp:228  - using 4 threads for model loading
[DEBUG] model_loader.cpp:1038 - loading 1016/1016 tensors from ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors
  |##################################################| 1016/1016 - 0.00MB/s
[INFO ] model_loader.cpp:1300 - loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s)
[DEBUG] model_loader.cpp:1038 - loading 1016/1016 tensors from ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors
  |##################################################| 1016/1016 - 709.37MB/s
[INFO ] model_loader.cpp:1300 - loading tensors completed, taking 0.20s (read: 0.01s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.08s)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (141.88 MB, 1016 tensors, VRAM)
[DEBUG] lora.hpp:123  - finished loaded lora
[DEBUG] ggml_extend.hpp:2193 - lora compute buffer size: 0.00 MB(VRAM)
[INFO ] lora.hpp:997  - (0 / 1016) LoRA tensors have been applied, lora_file_path = ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors
[DEBUG] model_manager.cpp:981  - model manager releasing params backend buffer (141.88 MB, 1016 tensors, VRAM)
[DEBUG] ggml_extend.hpp:2193 - qwen3 compute buffer size: 2.83 MB(VRAM)
[DEBUG] conditioner.hpp:1776 - computing condition graph completed, taking 1414 ms
[INFO ] stable-diffusion.cpp:5131 - get_learned_condition completed, taking 1.41s
[INFO ] stable-diffusion.cpp:5502 - generating image: 1/1 - seed 43
[DEBUG] model_loader.cpp:1038 - loading 685/685 tensors from /app/anima-base/anima-base-v1.0.safetensors
  |##################################################| 685/685 - 2.90GB/s
[INFO ] model_loader.cpp:1300 - loading tensors completed, taking 1.35s (read: 0.36s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.70s)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (1020.23 MB, 224 tensors, VRAM)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (1012.02 MB, 156 tensors, VRAM)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (1020.02 MB, 158 tensors, VRAM)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (999.01 MB, 147 tensors, VRAM)
[INFO ] model_loader.cpp:242  - load ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors using safetensors format
[DEBUG] model_loader.cpp:316  - init from './anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors', prefix = 'lora.'
[INFO ] lora.hpp:47   - loading LoRA from './anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors'
[DEBUG] model_loader.cpp:228  - using 4 threads for model loading
[DEBUG] model_loader.cpp:1038 - loading 1016/1016 tensors from ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors
  |##################################################| 1016/1016 - 0.00MB/s
[INFO ] model_loader.cpp:1300 - loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s)
[DEBUG] model_loader.cpp:1038 - loading 1016/1016 tensors from ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors
  |##################################################| 1016/1016 - 709.37MB/s
[INFO ] model_loader.cpp:1300 - loading tensors completed, taking 0.20s (read: 0.01s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.08s)
[DEBUG] model_manager.cpp:406  - model manager prepared params backend buffer (141.88 MB, 1016 tensors, VRAM)
[DEBUG] lora.hpp:123  - finished loaded lora
[DEBUG] ggml_extend.hpp:2193 - lora compute buffer size: 194.25 MB(VRAM)
[INFO ] lora.hpp:997  - (1016 / 1016) LoRA tensors have been applied, lora_file_path = ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors
[DEBUG] model_manager.cpp:981  - model manager releasing params backend buffer (141.88 MB, 1016 tensors, VRAM)
[DEBUG] ggml_extend.hpp:2193 - anima compute buffer size: 78.31 MB(VRAM)
  |==================================================| 8/8 - 1.06s/it
[INFO ] stable-diffusion.cpp:5534 - sampling completed, taking 15.68s
[INFO ] stable-diffusion.cpp:5546 - generating 1 latent images completed, taking 15.68s
[INFO ] stable-diffusion.cpp:5156 - decoding 1 latents
[DEBUG] ggml_extend.hpp:2193 - taehv compute buffer size: 2611.16 MB(VRAM)
[DEBUG] vae.hpp:219  - computing vae decode graph completed, taking 0.13s
[INFO ] stable-diffusion.cpp:5222 - latent 1 decoded, taking 0.13s
[INFO ] stable-diffusion.cpp:5226 - decode_first_stage completed, taking 0.13s
[INFO ] stable-diffusion.cpp:5684 - generate_image completed in 17.24s

generate one more image cause crash after black image:

Details 

[INFO ] stable-diffusion.cpp:5451 - generate_image 512x768
[DEBUG] stable-diffusion.cpp:2064 - lora ./anima-base/lora-speed/anima-turbo-lora-v0.2.safetensors:0.40
[INFO ] stable-diffusion.cpp:2079 - apply_loras completed, taking 0.00s
[INFO ] stable-diffusion.cpp:3080 - Using 'cosmos_reference' preset for reference images
[INFO ] denoiser.hpp:1051 - get_sigmas with discrete scheduler
[INFO ] stable-diffusion.cpp:4280 - sampling using ER-SDE method
[DEBUG] conditioner.hpp:1718 - parse 'masterpiece, best quality, absurdres, ultra detailed, furina \(genshin impact\), blue eyes, heterochromia, white hair, long hair, ahoge, blue and white outfit, hat, gloves, 1girl, sitting atop a giant moon-shaped swing, gentle smile, star-filled sky, glowing clouds below, fantasy nightscape' to [['masterpiece, best quality, absurdres, ultra detailed, furina (genshin impact\), blue eyes, heterochromia, white hair, long hair, ahoge, blue and white outfit, hat, gloves, 1girl, sitting atop a giant moon-shaped swing, gentle smile, star-filled sky, glowing clouds below, fantasy nightscape', 1], ]
[DEBUG] bpe_tokenizer.cpp:208  - split prompt "masterpiece, best quality, absurdres, ultra detailed, furina (genshin impact\), blue eyes, heterochromia, white hair, long hair, ahoge, blue and white outfit, hat, gloves, 1girl, sitting atop a giant moon-shaped swing, gentle smile, star-filled sky, glowing clouds below, fantasy nightscape" to 72 tokens ["master", "piece", ",", "Ġbest", "Ġquality", ",", "Ġabsurd", "res", ",", "Ġultra", "Ġdetailed", ",", "Ġfur", "ina", "Ġ(", "gens", "hin", "Ġimpact", "\", "),", "Ġblue", "Ġeyes", ",", "Ġheter", "och", "rom", "ia", ",", "Ġwhite", "Ġhair", ",", "Ġlong", "Ġhair", ",", "Ġah", "oge", ",", "Ġblue", "Ġand", "Ġwhite", "Ġoutfit", ",", "Ġhat", ",", "Ġgloves", ",", "Ġ", "1", "girl", ",", "Ġsitting", "Ġatop", "Ġa", "Ġgiant", "Ġmoon", "-shaped", "Ġswing", ",", "Ġgentle", "Ġsmile", ",", "Ġstar", "-filled", "Ġsky", ",", "Ġglowing", "Ġclouds", "Ġbelow", ",", "Ġfantasy", "Ġnights", "cape", ]
[DEBUG] t5_unigram_tokenizer.cpp:336  - split prompt "masterpiece, best quality, absurdres, ultra detailed, furina (genshin impact\), blue eyes, heterochromia, white hair, long hair, ahoge, blue and white outfit, hat, gloves, 1girl, sitting atop a giant moon-shaped swing, gentle smile, star-filled sky, glowing clouds below, fantasy nightscape" to tokens ["▁masterpiece", ",", "▁best", "▁quality", ",", "▁absurd", "re", "s", ",", "▁ultra", "▁detailed", ",", "▁fur", "in", "a", "▁(", "gen", "s", "hin", "▁impact", "\", "),", "▁blue", "▁eyes", ",", "▁hetero", "chro", "m", "i", "a", ",", "▁white", "▁hair", ",", "▁long", "▁hair", ",", "▁", "a", "hog", "e", ",", "▁blue", "▁and", "▁white", "▁outfit", ",", "▁", "hat", ",", "▁gloves", ",", "▁1", "girl", ",", "▁sitting", "▁", "a", "top", "▁", "a", "▁giant", "▁moon", "-", "shaped", "▁swing", ",", "▁gentle", "▁smile", ",", "▁star", "-", "filled", "▁sky", ",", "▁glowing", "▁clouds", "▁below", ",", "▁fantasy", "▁night", "scape", ]
[DEBUG] ggml_extend.hpp:2193 - qwen3 compute buffer size: 2.83 MB(VRAM)
[DEBUG] conditioner.hpp:1776 - computing condition graph completed, taking 130 ms
[INFO ] stable-diffusion.cpp:5131 - get_learned_condition completed, taking 0.13s
[INFO ] stable-diffusion.cpp:5502 - generating image: 1/1 - seed 43
[DEBUG] ggml_extend.hpp:2193 - anima compute buffer size: 78.31 MB(VRAM)
  |==================================================| 8/8 - 1.06s/it
[INFO ] stable-diffusion.cpp:5534 - sampling completed, taking 8.81s
[INFO ] stable-diffusion.cpp:5546 - generating 1 latent images completed, taking 8.81s
[INFO ] stable-diffusion.cpp:5156 - decoding 1 latents
[DEBUG] ggml_extend.hpp:2193 - taehv compute buffer size: 2592.38 MB(VRAM)
(...just crash here, exit code 139)

Additional context / environment details

Did not check upscaler yet.

Flags won't affect result:

  • --fa
  • --vae-tiling
  • --rng cpu
  • --mmap
  • --eager-load

system 1:

  • OS: Ubuntu 24.04
  • GPU: Radeon RX 9060 XT 16GB
  • Backend: Vulkan
  • Mesa version: Mesa 25.2.8-0ubuntu0.24.04.2

run inside docker via docker compose:

services:
  llama-server:
    image: ghcr.io/leejet/stable-diffusion.cpp:master-vulkan
    container_name: sd-vulkan-server
    ports:
      - "8081:8081"
    devices:
      - /dev/dri:/dev/dri
      - /dev/kfd:/dev/kfd
    volumes:
      - ./anima-base:/app/anima-base
      - ./vae:/app/vae
      - ./upscale:/app/upscale
    user: 1000:1000
    group_add:
      - video
      - 110 #render # Required for GPU access
    working_dir: /app
    entrypoint: ["/sd-server"]
    command: >
      --listen-ip 0.0.0.0
      --listen-port 8081
      -v
      --diffusion-model /app/anima-base/anima-base-v1.0.safetensors
      --tae /app/vae/taew2_1.safetensors
      --llm /app/anima-base/qwen_3_06b_base.safetensors
      --cfg-scale 1
      --steps 8
      --sampling-method er_sde
      --fa

system 2:

  • OS: Windows 10 22H2
  • GPU: GeForce RTX 4070 SUPER
  • Backend: Vulkan
  • driver version: 560.94

Activity

  1. cs8425
    added
    bugSomething isn't working
    on Aug 3, 2026

  2. cjlworld commented on Oct 23, 2026

    @cjlworld
    cjlworld
    on Oct 23, 2026

    This issue appears to have been fixed by #1872.

    Confirmed on Arch Linux with a Radeon 680M (16 GB RAM), Mesa 26.2.5, and the Vulkan backend.

    Using the same Anima/Qwen/TAEHV models and LoRA, with 8 steps. I used 256×256 to fit the shared-memory constraint and sent LoRA weights sequentially as 1.0 → 0.4 → 0.4:

    A 1.0 → 1.0 → 1.0 control on db99efd completes all three requests with identical outputs.

    The regression window contains #1872, a ggml sync, and a CI commit. #1872 changes VAE cleanup from free_compute_buffer() to runner_done(), which releases runner-owned parameter references before a LoRA change invalidates model storage. This matches the black decode followed by a crash.

  3. cs8425 commented on Oct 25, 2026

    @cs8425
    cs8425
    on Oct 25, 2026
    Author

    Follow up:
    After I pulled the updated docker image which only know commit hash: 95ff762 .
    The server no longer crashes, and LoRA changes are applied correctly.
    Thanks for the fix! This issue can be closed.

  4. cs8425
    closed this as completedon Oct 25, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions