[Bug Report] Severe VRAM Allocation Instability in PyTorch after llama-cpp-python is Imported
Maintainers usually reply within 1 day
Nobody has claimed this yet.
Assessment
- Difficulty
- 5/5
- Estimated time
- Over a week
- Newbie friendliness
- 30/100
- Issue type
- Bug
- Clarity
- Mostly clear
- Activity status
- Stale
- Domain
- ai-infra-agents
Research direction
The issue describes a low-level CUDA context conflict between llama-cpp-python and PyTorch. Start by examining the CUDA initialization code in llama-cpp-python's bindings, particularly around import time. Look for any CUDA context creation or memory allocator settings. Reproduce the issue in a minimal environment that imports llama-cpp-python and then loads a large PyTorch model, monitoring VRAM allocation patterns. 'Done' would be identifying the specific call or configuration causing the conflict and proposing a fix.
Written by the indexing model from the issue text.
Description
Prerequisites
- I am running the latest code.
- I carefully followed the README.md.
- I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
- I reviewed the Discussions, and have a new bug to share.
Expected Behavior
When loading a large PyTorch-based model (e.g., an SDXL checkpoint >10GB) within a host application (ComfyUI), I expect the GPU's dedicated memory (VRAM) usage to increase smoothly and stabilize at the required level for the model. This process should be fast and efficient.
Current Behavior
The VRAM allocation is extremely erratic and inefficient. The memory usage graph shows a severe sawtooth/spiky pattern—it spikes up, drops slightly, spikes again, and repeats this cycle until the model is finally loaded. This process is significantly slower than normal and suggests a fundamental conflict in memory management at the CUDA level.
This behavior only occurs when a component that imports llama-cpp-python is loaded by the host application at startup. The issue manifests before any llama-cpp-python functions are actively called; the mere act of importing the library is enough to trigger this conflict with subsequent PyTorch operations.
Environment and Context
-
Hardware:
- GPU: NVIDIA RTX PRO 6000 Blackwell Workstation Edition (94.5 GB VRAM)
-
Operating System:
- Windows 11
-
SDK and Software Versions:
- Python Version:
3.12.11 - Compiler: MSVC / Visual Studio
2022used to compile for CUDA - CUDA Toolkit: 12.8 (locally compiled)
- NVIDIA Driver Version:
580.97 - Host Application: ComfyUI (Release
2025-08-27or newer)
- Python Version:
-
Key Python Packages:
- llama-cpp-python:
0.3.16 - torch:
2.8.0+cu128 - diskcache: 5.6.3
- llama-cpp-python:
Failure Information (for bugs)
This report details a deep-seated conflict between llama-cpp-python and torch regarding CUDA context and/or VRAM management.
Investigation Summary & Key Findings
We conducted an extensive, multi-day investigation to isolate this issue, and the results were conclusive:
-
The problem is triggered by the
importofllama-cpp-python. Any host application that loads this library at startup will cause subsequent, unrelatedtorchVRAM allocations to become unstable. Disabling the components that importllama-cpp-pythoncompletely resolves the issue. -
The
diskcachedependency is a Red Herring. We initially suspected thediskcachelibrary was the cause. To test this, we performed a definitive experiment: we hid the realdiskcachelibrary and replaced it with a completely inert, non-functional fake Python package that only contained an emptyCacheclass.- Result: The VRAM instability problem persisted even with the fake library.
- Log Confirmation: The application log confirmed our fake library was being used:
--- 警告:正在使用一個偽造的 (FAKE) diskcache 函式庫 --- Set up system temp directory: C:\Users\Win\AppData\Local\Temp\latentsync_16880d62 - Conclusion: This proves that
diskcache's code is not the active cause. The problem isllama-cpp-pythonitself. The initial correlation was due todiskcachebeing a mandatory dependency; uninstalling it simply preventedllama-cpp-pythonfrom being imported successfully.
Final Hypothesis
The evidence overwhelmingly suggests a low-level conflict between how llama-cpp-python (especially a version compiled against a modern CUDA 12.8 toolkit) initializes its CUDA context and how PyTorch expects to manage the VRAM. The initialization of llama-cpp-python appears to "pollute" or "fragment" the VRAM state, making subsequent large-scale allocations by PyTorch highly inefficient.
Steps to Reproduce
- Set up a ComfyUI environment with PyTorch and the system components listed above.
- Compile and install
llama-cpp-pythonwith CUDA 12.8 support. - Install any ComfyUI custom node that imports
llama-cpp-pythonat startup (e.g.,ComfyUI-JoyCaptionorComfyUI-Searge-LLM-Node). - Launch ComfyUI. The server will start, and
llama-cpp-pythonwill be imported into the process. - Load a standard workflow designed to load a large PyTorch-based checkpoint model (e.g., an SDXL checkpoint, >10GB).
- Observe the "Dedicated GPU Memory" graph in Windows Task Manager. You will see the erratic, spiky allocation pattern instead of a smooth ramp-up.
Note on llama.cpp comparison: This issue is not about the performance or functionality of llama-cpp-python's inference, but rather a resource conflict it creates with another major library (torch). Therefore, comparing behavior against llama.cpp's ./main executable is not applicable, as the issue requires both runtimes to be active in the same Python process.
- Dominant language
- Python
- Stars
- 10.7k
- Forks
- 1.5k
- Avg merge
- 1h 40m
- Merged PRs (30d)
- 5
Getting set up
- No Dockerfile or Docker Compose file
- No pull request template
- Read the contributing guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
More from abetlen/llama-cpp-python
-
Difficulty 2/5 1-3 hours Newbie friendliness 78/100
abetlen/llama-cpp-python#2424 ·
Maintainers usually reply within 1 day
-
Seven llama_sampler_init_* bindings admit keyword arguments that the ctypes function object silently dropsPossibly taken @Belal0066 claimed this 59 days ago. Open
Difficulty 2/5 1-3 hours Newbie friendliness 88/100
abetlen/llama-cpp-python#2371 ·
Maintainers usually reply within 1 day
-
uv add llama-cpp-python wheels fails for versions above 0.3.30Possibly taken A pull request linked to this issue is open or already merged. Open
Difficulty 2/5 1-3 hours Newbie friendliness 65/100
abetlen/llama-cpp-python#2352 · 1 comment · 3 reactions ·
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 75/100
abetlen/llama-cpp-python#2314 ·
Maintainers usually reply within 1 day
-
Llama.embed() calls LlamaBatch.add_sequence with old 3-arg signature; missing logits_arrayMay be free again @lxcxjxhx claimed this 136 days ago, and no pull request is open. Open
Difficulty 2/5 1-3 hours Newbie friendliness 65/100
abetlen/llama-cpp-python#2211 · 2 comments ·
Maintainers usually reply within 1 day
All issues in abetlen/llama-cpp-python
Similar issues
-
Bug IronPython Python 3
Difficulty 2/5 1-3 hours Newbie friendliness 72/100
pyrevitlabs/pyRevit#4184 ·
Maintainers usually reply within 1 day
-
Difficulty 1/5 Under an hour Newbie friendliness 74/100
jqueryscript/awesome-claude-code#1009 ·
-
tests
Difficulty 2/5 1-3 hours Newbie friendliness 66/100
Maintainers usually reply within 3 days
-
Difficulty 2/5 1-3 hours Newbie friendliness 72/100
Maintainers usually reply within 1 day
-
[Plugin] 群聊记忆 · 人格养成Open
Difficulty 1/5 1-3 hours Newbie friendliness 74/100