[bug]: model cache logs negative VRAM requested

[keturn]

Is there an existing issue for this problem?

• [x] I have searched the existing issues

Operating system

Linux

GPU vendor

Nvidia (CUDA)

GPU model

RTX 3060

GPU VRAM

12 GB

Version number

5.9

What happened

When generating Flux images, I frequently see messages like this in the log:

[21:15:07,670]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model '924bda16-8c73-4aed-a06c-6216705962ea:transformer' (Flux) onto cuda device in 1.20s. Total model size: 8573.12MB, VRAM: 6444.62MB (75.2%)        
[21:15:38,492]::[InvokeAI]::WARNING --> Loading 0.0 MB into VRAM, but only -283.4375 MB were requested. This is the minimum set of weights in VRAM required to run the model.                                                                          
[21:15:38,494]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model 'c75a604c-a146-44be-a294-36a1842c3f7e:vae' (AutoEncoder) onto cuda device in 0.14s. Total model size: 159.87MB, VRAM: 0.00MB (0.0%)   

Two things about that message are weird:

• reports a negative number of megabytes were requested?
• loaded zero megabytes?

What you expected to happen

not sure what to expect from the model cache's logging

[firesign]

InvokeAI version 5.10.1, GTX 3060 12gb I'm seeing a similar negative VRAM request, but with different SDXL models: this happens an unknown but seemingly consistent number of gens between restarts; if I were to guess, around 50 gens. It's been an issue for me since late 2024:

[2025-05-09 10:53:48,195]::[InvokeAI]::INFO --> Executing queue item 31946, session 1350f061-30a5-4249-97f1-6f2a7245d4e4
[2025-05-09 10:53:49,384]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model 'cb09f201-e122-4083-a7ee-2ea26e0985ce:unet' (UNet2DConditionModel) onto cuda device in 1.16s. Total model size: 4897.05MB, VRAM: 4897.05MB (100.0%)
[2025-05-09 10:53:50,578]::[ModelManagerService]::INFO --> [MODEL CACHE] Loaded model 'cb09f201-e122-4083-a7ee-2ea26e0985ce:scheduler' (EDMDPMSolverMultistepScheduler) onto cuda device in 0.00s. Total model size: 0.00MB, VRAM: 0.00MB (0.0%)

  0%|          | 0/34 [00:00<?, ?it/s]
  3%|2         | 1/34 [00:00<00:23,  1.39it/s]
  6%|5         | 2/34 [00:01<00:23,  1.36it/s]
  9%|8         | 3/34 [00:02<00:22,  1.36it/s]
 12%|#1        | 4/34 [00:02<00:21,  1.37it/s]
 15%|#4        | 5/34 [00:03<00:21,  1.37it/s]
 18%|#7        | 6/34 [00:04<00:20,  1.37it/s]
 21%|##        | 7/34 [00:05<00:19,  1.36it/s]
 24%|##3       | 8/34 [00:05<00:19,  1.37it/s]
 26%|##6       | 9/34 [00:06<00:18,  1.36it/s]
 29%|##9       | 10/34 [00:07<00:17,  1.35it/s]
 32%|###2      | 11/34 [00:08<00:16,  1.36it/s]
 35%|###5      | 12/34 [00:08<00:16,  1.36it/s]
 38%|###8      | 13/34 [00:09<00:15,  1.36it/s]
 41%|####1     | 14/34 [00:10<00:14,  1.36it/s]
 44%|####4     | 15/34 [00:11<00:14,  1.36it/s]
 47%|####7     | 16/34 [00:11<00:13,  1.36it/s]
 50%|#####     | 17/34 [00:12<00:12,  1.35it/s]
 53%|#####2    | 18/34 [00:13<00:11,  1.36it/s]
 56%|#####5    | 19/34 [00:13<00:11,  1.36it/s]
 59%|#####8    | 20/34 [00:14<00:10,  1.35it/s]
 62%|######1   | 21/34 [00:15<00:09,  1.36it/s]
 65%|######4   | 22/34 [00:16<00:08,  1.35it/s]
 68%|######7   | 23/34 [00:16<00:08,  1.36it/s]
 71%|#######   | 24/34 [00:17<00:07,  1.36it/s]
 74%|#######3  | 25/34 [00:18<00:06,  1.36it/s]
 76%|#######6  | 26/34 [00:19<00:05,  1.35it/s]
 79%|#######9  | 27/34 [00:19<00:05,  1.36it/s]
 82%|########2 | 28/34 [00:20<00:04,  1.35it/s]
 85%|########5 | 29/34 [00:21<00:03,  1.36it/s]
 88%|########8 | 30/34 [00:22<00:02,  1.36it/s]
 91%|#########1| 31/34 [00:22<00:02,  1.36it/s]
 94%|#########4| 32/34 [00:23<00:01,  1.35it/s]
 97%|#########7| 33/34 [00:24<00:00,  1.36it/s]
100%|##########| 34/34 [00:25<00:00,  1.35it/s]
100%|##########| 34/34 [00:25<00:00,  1.36it/s]
C:\Users\nnn\invokeai\.venv\Lib\site-packages\invokeai\app\invocations\baseinvocation.py:183: PydanticDeprecatedSince211: Accessing the 'model_fields' attribute on the instance is deprecated. Instead, you should access this attribute from the model class. Deprecated in Pydantic V2.11 to be removed in V3.0.
  for field_name, field in self.model_fields.items():
[2025-05-09 10:54:16,629]::[InvokeAI]::WARNING --> Loading 0.0 MB into VRAM, but only -1205.203125 MB were requested. This is the minimum set of weights in VRAM required to run the model.
Process exited with code: 0

[psychedelicious]

From Ryan in discord: https://discord.com/channels/1020123559063990373/1149506274971631688/1326650031310114816

This warning is usually technically accurate when it pops up, but I can definitely see why it would be confusing. It means that we would ideally like to use 600MB less VRAM to be safe, but couldn't find a way to offload those weights safely. How much VRAM do you have?