[bug]: quantized qwen image edit 2511 memory management

[keturn]

Is there an existing issue for this problem?

• I have searched the existing issues

Install method

Docker

Operating system

Linux

GPU vendor

Nvidia (CUDA)

GPU model

RTX 3060

GPU VRAM

12 GB

Version number

6.13.0rc2 (docker sha-25bbf32-cuda)

Browser

Firefox 150.0.2

System Information

[About…]

{ "version": "6.12.0.post1", "dependencies": { "absl-py" : "2.3.1", "accelerate" : "1.8.1", "annotated-types" : "0.7.0", "anyio" : "4.9.0", "attrs" : "25.3.0", "bcrypt" : "3.2.2", "bidict" : "0.23.1", "bitsandbytes" : "0.46.1", "blake3" : "1.0.5", "certifi" : "2025.6.15", "cffi" : "1.17.1", "charset-normalizer" : "3.4.2", "cityhash" : "0.4.7", "click" : "8.2.1", "coloredlogs" : "15.0.1", "compel" : "2.1.1", "contourpy" : "1.3.2", "cryptography" : "45.0.5", "CUDA" : "12.8", "cycler" : "0.12.1", "Deprecated" : "1.2.18", "diffusers" : "0.37.0", "dnspython" : "2.7.0", "dynamicprompts" : "0.31.0", "ecdsa" : "0.19.1", "einops" : "0.8.1", "email-validator" : "2.3.0", "fastapi" : "0.118.3", "fastapi-events" : "0.12.2", "filelock" : "3.18.0", "flatbuffers" : "25.2.10", "fonttools" : "4.58.5", "fsspec" : "2025.5.1", "gguf" : "0.17.1", "h11" : "0.16.0", "hf-xet" : "1.1.5", "httpcore" : "1.0.9", "httptools" : "0.6.4", "httpx" : "0.28.1", "huggingface-hub" : "0.34.4", "humanfriendly" : "10.0", "idna" : "3.10", "importlib_metadata" : "8.7.0", "InvokeAI" : "6.12.0.post1", "jax" : "0.6.2", "jaxlib" : "0.6.2", "Jinja2" : "3.1.6", "kiwisolver" : "1.4.8", "kornia" : "0.8.1", "kornia_rs" : "0.1.9", "MarkupSafe" : "3.0.1", "matplotlib" : "3.10.3", "mediapipe" : "0.10.14", "ml_dtypes" : "0.5.1", "mpmath" : "1.3.0", "networkx" : "3.5", "numpy" : "1.26.4", "nvidia-cublas-cu12" : "12.8.3.14", "nvidia-cuda-cupti-cu12" : "12.8.57", "nvidia-cuda-nvrtc-cu12" : "12.8.61", "nvidia-cuda-runtime-cu12": "12.8.57", "nvidia-cudnn-cu12" : "9.7.1.26", "nvidia-cufft-cu12" : "11.3.3.41", "nvidia-cufile-cu12" : "1.13.0.11", "nvidia-curand-cu12" : "10.3.9.55", "nvidia-cusolver-cu12" : "11.7.2.55", "nvidia-cusparse-cu12" : "12.5.7.53", "nvidia-cusparselt-cu12" : "0.6.3", "nvidia-nccl-cu12" : "2.26.2", "nvidia-nvjitlink-cu12" : "12.8.61", "nvidia-nvtx-cu12" : "12.8.55", "onnx" : "1.16.1", "onnxruntime" : "1.19.2", "opencv-contrib-python" : "4.11.0.86", "opt_einsum" : "3.4.0", "packaging" : "25.0", "passlib" : "1.7.4", "picklescan" : "0.0.26", "pillow" : "11.3.0", "prompt_toolkit" : "3.0.51", "protobuf" : "4.25.8", "psutil" : "7.0.0", "pyasn1" : "0.6.1", "pycparser" : "2.22", "pydantic" : "2.11.7", "pydantic-settings" : "2.10.1", "pydantic_core" : "2.33.2", "pymorton" : "1.0.5", "pyparsing" : "3.2.3", "PyPatchMatch" : "1.0.2", "python-dateutil" : "2.9.0.post0", "python-dotenv" : "1.1.1", "python-engineio" : "4.12.2", "python-jose" : "3.5.0", "python-multipart" : "0.0.20", "python-socketio" : "5.13.0", "PyWavelets" : "1.8.0", "PyYAML" : "6.0.2", "regex" : "2024.11.6", "requests" : "2.32.4", "rsa" : "4.9.1", "safetensors" : "0.5.3", "scipy" : "1.16.0", "semver" : "3.0.4", "sentencepiece" : "0.2.0", "setuptools" : "80.9.0", "simple-websocket" : "1.1.0", "six" : "1.17.0", "sniffio" : "1.3.1", "sounddevice" : "0.5.2", "spandrel" : "0.4.2", "starlette" : "0.46.2", "sympy" : "1.14.0", "tokenizers" : "0.22.0", "torch" : "2.7.1+cu128", "torchsde" : "0.2.6", "torchvision" : "0.22.1+cu128", "tqdm" : "4.67.1", "trampoline" : "0.1.2", "transformers" : "4.56.0", "triton" : "3.3.1", "typing-inspection" : "0.4.1", "typing_extensions" : "4.14.0", "urllib3" : "2.5.0", "uvicorn" : "0.35.0", "uvloop" : "0.21.0", "watchfiles" : "1.1.0", "wcwidth" : "0.2.13", "websockets" : "15.0.1", "wrapt" : "1.17.2", "wsproto" : "1.2.0", "zipp" : "3.23.0" }, "config": { "schema_version": "4.0.3", "legacy_models_yaml_path": null, "host": "", "port": 9090, "allow_origins": [], "allow_credentials": true, "allow_methods": ["*"], "allow_headers": ["*"], "log_tokenization": true, "patchmatch": true, "models_dir": "models", "convert_cache_dir": "models/.convert_cache", "download_cache_dir": "models/.download_cache", "legacy_conf_dir": "configs", "db_dir": "databases", "outputs_dir": "outputs", "image_subfolder_strategy": "flat", "custom_nodes_dir": "nodes", "style_presets_dir": "style_presets", "workflow_thumbnails_dir": "workflow_thumbnails", "log_handlers": ["console"], "log_format": "color", "log_level": "info", "log_sql": false, "log_level_network": "warning", "use_memory_db": false, "dev_reload": false, "profile_graphs": false, "profile_prefix": null, "profiles_dir": "profiles", "max_cache_ram_gb": null, "max_cache_vram_gb": null, "log_memory_usage": false, "model_cache_keep_alive_min": 0, "device_working_mem_gb": 3, "enable_partial_loading": true, "keep_ram_copy_of_weights": true, "ram": null, "vram": null, "lazy_offload": true, "pytorch_cuda_alloc_conf": "backend:cudaMallocAsync", "device": "cuda", "precision": "float16", "sequential_guidance": false, "attention_type": "auto", "attention_slice_size": "auto", "force_tiled_decode": false, "pil_compress_level": 1, "max_queue_size": 10000, "clear_queue_on_startup": false, "max_queue_history": null, "allow_nodes": null, "deny_nodes": null, "node_cache_size": 512, "hashing_algorithm": "blake3_single", "remote_api_tokens": [], "scan_models_on_startup": false, "unsafe_disable_picklescan": false, "allow_unknown_models": true, "multiuser": true, "strict_password_checking": false, "external_alibabacloud_api_key": null, "external_alibabacloud_base_url": null, "external_gemini_api_key": null, "external_openai_api_key": null, "external_gemini_base_url": null, "external_openai_base_url": null, "external_seedream_api_key": null, "external_seedream_base_url": null }, "set_config_fields": [ "pytorch_cuda_alloc_conf", "precision", "device", "port", "log_tokenization", "ssl_certfile", "legacy_models_yaml_path", "host", "ssl_keyfile", "multiuser", "remote_api_tokens" ] }

What happened

Initially I ran Quen Image Edit 2511 Q4.
Then installed the Q2 quant of the same model (from the Starter Models list), and switched to that.

Re-ran the same seed/prompt/etc on the Q2 model and got assorted problems:

• torch.OOM during decoding step on two out of a batch of six generations (the other four completed)
• WARNING --> Loading 0.0 MB into VRAM, but only -1447.625 MB were requested.
• Widely varying reported values for QwenImageTransformer2DModel model size in VRAM, from 99.5% to 60%

invoke-q2.log

What you expected to happen

• 7 GB Q2 QwenImageTransformer2DMode should probably fit entirely in VRAM (or is it truly 0.5% too big?)
• the 242 MB autoencoder definitely should

[keturn]

After restarting Invoke:

• QwenImageTransformer2DMode did load 100% in VRAM for the first generation
• torch.OOM on the decode (l2i) step
• subsequent generations in the batch had different %VRAM for the transformer and autoencoder, but did not OOM on decode

[keturn]

I turned on model cache logging:

Loaded model 'e50334c8-857a-48bd-98a6-3ac13aeceac2:vae' (AutoencoderKLQwenImage) onto cuda device in 0.05s. Total model size: 242.03MB, VRAM: 242.03MB (100.0%)
Loaded model onto execution device: model_bytes_loaded=242.03MB, 
After loading: model_total=242 MB, model_vram=242 MB (100.0% %), vram_available=235 MB, 
Finished locking model e50334c8-857a-48bd-98a6-3ac13aeceac2:vae (Type: AutoencoderKLQwenImage)
 Model cache state:
  Storage Device (cpu)           Limit:  8830.8 MB, Used:  8173.9 MB (92.6%), Available:   656.9 MB ( 7.4%)
  Compute Device (cuda)          Limit:  8421.6 MB, Used:  8186.2 MB (97.2%), Available:   235.4 MB ( 2.8%)
  CUDA Memory Allocated:         8186.2 MB
  Total models:                  3
  Models:
    806a0920-dec1-4166-a53d-b27a2b6e80b5:transformer (QwenImageTransformer2DModel):  total= 7121.9 MB, vram= 7121.9 MB (100.0%), ram=    0.0 MB ( 0.0%), locked=False
    54afef59-a373-4f04-9478-60de706c7459 (ModelPatchRaw):                            total=  810.0 MB, vram=    0.0 MB ( 0.0%), ram=  810.0 MB (100.0%), locked=False
    e50334c8-857a-48bd-98a6-3ac13aeceac2:vae (AutoencoderKLQwenImage):               total=  242.0 MB, vram=  242.0 MB (100.0%), ram=    0.0 MB ( 0.0%), locked=True

Emitted private invocation event invocation_progress to user room user:52398277-8246-41f8-99b6-622308cdaac8 and admin room
Unlocked model e50334c8-857a-48bd-98a6-3ac13aeceac2:vae (Type: AutoencoderKLQwenImage)
On node error: queue item 49680, session d64a9ae8-6b9a-4e21-8cf8-cd8495c06dbf, node d0c34e69-ae50-4e96-94bc-948f5df70ccc (qwen_image_l2i)
Error while invoking session d64a9ae8-6b9a-4e21-8cf8-cd8495c06dbf, invocation d0c34e69-ae50-4e96-94bc-948f5df70ccc (qwen_image_l2i): Allocation on device 0 would exceed allowed memory. (out of memory)
Currently allocated     : 10.10 GiB
Requested               : 591.12 MiB
Device limit            : 11.62 GiB
Free (according to CUDA): 33.44 MiB

So I guess device_working_mem_gb is too low for l2i, and I just found out here because the q2 quant does a very good job of mostly filling up the amount allocated for cache?

Is that "working mem" figure really static? It seems like it could be quite different depending on the specific workload.

[keturn]

I re-submitted the job, and I'm having trouble making sense of the log. The output of the denoise node was cached from the last run, so it tried to pick up again with l2i:

[InvokeAI]::DEBUG --> Invocation cache hit for type "qwen_image_denoise": cf98eaa9-8c19-4eea-8df5-d07271e6ae1d
[InvokeAI]::DEBUG --> On after run node: queue item 49681, session f55274fc-d1fe-4853-b2a1-61e040f0e3d1, node cf98eaa9-8c19-4eea-8df5-d07271e6ae1d (qwen_image_denoise)
[InvokeAI]::DEBUG --> On before run node: queue item 49681, session f55274fc-d1fe-4853-b2a1-61e040f0e3d1, node 09dce3e1-b895-4097-a210-01f8c2645891 (qwen_image_l2i)
[InvokeAI]::DEBUG --> Skipping invocation cache for "qwen_image_l2i": 09dce3e1-b895-4097-a210-01f8c2645891
[ModelManagerService]::DEBUG --> [MODEL CACHE] Cache hit: e50334c8-857a-48bd-98a6-3ac13aeceac2:vae (Type: AutoencoderKLQwenImage)
[ModelManagerService]::DEBUG --> [MODEL CACHE] Locking model e50334c8-857a-48bd-98a6-3ac13aeceac2:vae (Type: AutoencoderKLQwenImage)
[ModelManagerService]::DEBUG --> [MODEL CACHE] Before unloading: model_total=242 MB, model_vram=242 MB (100.0% %), vram_available=-3038 MB, 
[ModelManagerService]::DEBUG --> [MODEL CACHE] Offloading unlocked models with goal of making room for 0.00MB of VRAM.
[ModelManagerService]::DEBUG --> [MODEL CACHE] Unloaded 806a0920-dec1-4166-a53d-b27a2b6e80b5:transformer from VRAM to free 3041 MB.
[ModelManagerService]::DEBUG --> [MODEL CACHE] Unloaded models (if necessary): vram_bytes_freed=3041.28MB
[ModelManagerService]::DEBUG --> [MODEL CACHE] After unloading: model_total=242 MB, model_vram=242 MB (100.0% %), vram_available=-1342 MB, 
[ModelManagerService]::DEBUG --> [MODEL CACHE] Unloaded 8.43MB from the model being locked (e50334c8-857a-48bd-98a6-3ac13aeceac2:vae).
[InvokeAI]::WARNING --> Loading 0.0 MB into VRAM, but only -1341.25 MB were requested. This is the minimum set of weights in VRAM required to run the model.
[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model 'e50334c8-857a-48bd-98a6-3ac13aeceac2:vae' (AutoencoderKLQwenImage) onto cuda device in 0.18s. Total model size: 242.03MB, VRAM: 233.60MB (96.5%)
[ModelManagerService]::DEBUG --> [MODEL CACHE] Loaded model onto execution device: model_bytes_loaded=0.00MB, 
[ModelManagerService]::DEBUG --> [MODEL CACHE] After loading: model_total=242 MB, model_vram=234 MB (96.5% %), vram_available=-1342 MB, 
[ModelManagerService]::DEBUG --> [MODEL CACHE] Finished locking model e50334c8-857a-48bd-98a6-3ac13aeceac2:vae (Type: AutoencoderKLQwenImage)
[ModelManagerService]::DEBUG --> [MODEL CACHE] Model cache state:
  Storage Device (cpu)           Limit:  8830.8 MB, Used:  8173.9 MB (92.6%), Available:   656.9 MB ( 7.4%)
  Compute Device (cuda)          Limit:  5168.6 MB, Used:  6510.8 MB (126.0%), Available: -1342.2 MB (-26.0%)
  CUDA Memory Allocated:         6510.8 MB
  Total models:                  3
  Models:
    806a0920-dec1-4166-a53d-b27a2b6e80b5:transformer (QwenImageTransformer2DModel):  total= 7121.9 MB, vram= 4080.6 MB (57.3%), ram= 3041.3 MB (42.7%), locked=False
    54afef59-a373-4f04-9478-60de706c7459 (ModelPatchRaw):                            total=  810.0 MB, vram=    0.0 MB ( 0.0%), ram=  810.0 MB (100.0%), locked=False
    e50334c8-857a-48bd-98a6-3ac13aeceac2:vae (AutoencoderKLQwenImage):               total=  242.0 MB, vram=  233.6 MB (96.5%), ram=    8.4 MB ( 3.5%), locked=True

[2026-06-01 02:28:25,339]::[InvokeAI]::DEBUG --> Emitted private invocation event invocation_progress to user room user:52398277-8246-41f8-99b6-622308cdaac8 and admin room
[2026-06-01 02:28:25,871]::[ModelManagerService]::DEBUG --> [MODEL CACHE] Unlocked model e50334c8-857a-48bd-98a6-3ac13aeceac2:vae (Type: AutoencoderKLQwenImage)
[2026-06-01 02:28:25,875]::[InvokeAI]::DEBUG --> On node error: queue item 49681, session f55274fc-d1fe-4853-b2a1-61e040f0e3d1, node 09dce3e1-b895-4097-a210-01f8c2645891 (qwen_image_l2i)
[2026-06-01 02:28:25,875]::[InvokeAI]::ERROR --> Error while invoking session f55274fc-d1fe-4853-b2a1-61e040f0e3d1, invocation 09dce3e1-b895-4097-a210-01f8c2645891 (qwen_image_l2i): Allocation on device 0 would exceed allowed memory. (out of memory)
Currently allocated     : 7.89 GiB
Requested               : 196.27 MiB
Device limit            : 11.62 GiB
Free (according to CUDA): 33.75 MiB

So many questions. Why is "Compute Device (cuda) Limit" so much lower this time? Maybe something to do with this? https://github.com/invoke-ai/InvokeAI/blob/5db87729311543d8dc79014923dcbd5cd593e5a1/invokeai/backend/model_manager/load/model_cache/model_cache.py#L606-L610

Why is the transformer still taking up 4 GB VRAM? I guess it must have done a partial_unload_from_vram? Why would it not unload more if it's reporting negative VRAM available?

By what logic does the autoencoder end up 96.5% loaded after that?

[keturn]

Then trying to re-run with a new prompt:

[InvokeAI]::ERROR --> Error while invoking session c2f306d8-…, invocation ad66d1db-… (qwen_image_text_encoder): Some modules are dispatched on the CPU or the disk. Make sure you have enough GPU RAM to fit the quantized model. If you want to dispatch the model on the CPU or the disk while keeping these modules in 32-bit, you need to set llm_int8_enable_fp32_cpu_offload=True and pass a custom device_map to from_pretrained. Check https://huggingface.co/docs/transformers/main/en/main_classes/quantization#offload-between-cpu-and-gpu for more details.

But I don't have any messages from the model cache logger about qwen_image_text_encoder at all. Does that part not use the model manager?

"Encoder quantization" was set to int8.

Setting it to None (bf16) seems to make it use the model cache. (It kicked everything else out to try to fit the 16 GB in.)

[keturn]

Ah, qwen latents-to-image is more sensitive to the manually-specified working memory size because it doesn't have a estimate_vae_working_memory like SD and Flux do.

[github-actions]

There has been no activity in this issue for 30 days. If this issue is still being experienced, please reply with an updated confirmation that the issue is still being experienced with the latest release.

[keturn]

Figuring out the working memory for the l2i step is the subject of #9266, I think.

The lack of a working memory estimate for the VAE was addressed in #9305.

[lstein]

@keturn Have you tried the recent PR (9266) that addresses the qwen memory estimation issues, and do they fix the problem?