Runtimeerror: cuda out of memory.
This error message Runtimeerror: cuda out of memory is often encountered when the system is not able to allocate enough memory on the GPU to complete the requested operation. In …
Python RuntimeError Reference — 49+ PyTorch, CUDA, asyncio, Flask Fixes (2026)
itsourcecode.com hosts 49+ documented fixes for Python RuntimeError messages — the catch-all error class Python raises when something goes wrong at runtime that doesn’t fit a more specific category. In modern Python, most RuntimeErrors come from PyTorch/CUDA (GPU memory, scalar type, cudnn), asyncio (event loop already running), Flask/Django (working outside request context), or iteration mutation (set/dict changed size during iteration). Browse the fixes below.
What is a Python RuntimeError?
A RuntimeError is the base class for errors that don’t fit a more specific built-in exception. Libraries raise RuntimeError when something happens that the standard exception hierarchy doesn’t anticipate — PyTorch raises it for GPU/tensor problems, asyncio for event loop state issues, Flask for missing application context. Because it’s a catch-all, the fix depends heavily on which library raised it and what the message says.
How to debug any RuntimeError in 4 steps
- Read the full message. RuntimeError messages are library-specific and usually descriptive. “CUDA out of memory” means GPU memory; “event loop is already running” means asyncio context; “working outside of request context” means Flask scope.
- Identify the library. The traceback above the error tells you which library raised it. PyTorch → tensor/dtype/GPU. asyncio → event loop. Flask/Django → request context. PySpark → Java gateway. Each has a different fix pattern.
- Check resource state. For CUDA:
nvidia-smito see GPU memory. For asyncio: check if you’re inside an existing event loop. For dict/set iteration: never mutate while iterating — copy first or use a comprehension. - Search this hub by library category below. Most RuntimeErrors have well-documented patterns. Find your exact message.
Featured RuntimeError fixes by library
🔥 PyTorch — scalar type / dtype mismatches
PyTorch tensor operations require specific dtypes; mixing float16, float32, bfloat16, and double raises RuntimeError.
- Found dtype Long but expected Float
- Expected scalar type Half but found Float
- Expected scalar type BFloat16 but found Float
- Expected scalar type Float but found Double
- mat1 and mat2 shapes cannot be multiplied
🎮 PyTorch / CUDA — GPU errors
The most painful RuntimeErrors in ML — usually GPU memory, driver, or device assertion issues.
- CUDA out of memory
- No CUDA GPUs are available
- CUDA error: invalid device ordinal
- CUDA error: device-side assert triggered
- All CUDA-capable devices are busy or unavailable
- cuDNN error: CUDNN_STATUS_MAPPING_ERROR
- cuDNN error: CUDNN_STATUS_NOT_INITIALIZED
- cuDNN error: CUDNN_STATUS_EXECUTION_FAILED
- Couldn’t install torch
⏱️ asyncio — event loop state
🌐 Flask / Django — request context
- Working outside of request context (Flask)
- SQLALCHEMY_DATABASE_URI or SQLALCHEMY_BINDS must be set
- Cannot add middleware after an application has started
🔁 Mutation-during-iteration
☕ PySpark — Java gateway
📦 Build / Install / Config
- Poetry configuration is invalid
- Ninja is required to load C++ extensions
- Product license has not been initialized
- Python raise RuntimeError (manual raise)
Related error categories
RuntimeError is one of 10 hubs in our Python & JavaScript error reference cluster — 980+ documented fixes total:
- TypeError Reference — 220+ Python & JS TypeError fixes
- ModuleNotFoundError Reference — 198+ Python import errors
- AttributeError Reference — 173+ “object has no attribute X” fixes
- ValueError Reference — 100+ pandas/NumPy/sklearn fixes
- ImportError Reference — 67+ “cannot import name X from Y” fixes
- NameError Reference — 49+ “name X is not defined” fixes
- SyntaxError Reference — 48+ Python & JavaScript syntax errors
- ReferenceError Reference — 34+ JavaScript “is not defined” fixes
- HTTP Error Reference — 35+ HTTP status code fixes (4xx, 5xx)
- Python Tutorial — beginner-to-intermediate Python lessons
About this RuntimeError reference
Built since 2015 by PIES Information Technology Solutions, Binalbagan, Negros Occidental, Philippines. Every post comes from a real error encountered in production. Used by 12,000+ Python developers monthly.
Runtimeerror: this event loop is already running
The runtimeerror: this event loop is already running error is a common error that developers may encounter when working with asyncio in Python. This error message typically occurs when you …
Runtimeerror: expected scalar type float but found double
One of the common errors is runtimeerror: expected scalar type float but found double. This error occurs if a program expects a float but instead receives a double. In this …
Cannot add middleware after an application has started
One of the common errors you may encounter is runtimeerror: cannot add middleware after an application has started. This error occurs we attempt to add middleware to our application after …
Java gateway process exited before sending its port number
In this article, we will discuss the causes of the runtimeerror: java gateway process exited before sending its port number error and provide you with step-by-step instructions to fix it. …
Runtimeerror: dictionary changed size during iteration
The Runtimeerror: dictionary changed size during iteration typically occurs if we are attempting to change the size of the dictionary while iterating over it. This type of error is an …
runtimeerror: cudnn error: cudnn_status_not_initialized
The runtimeerror: cudnn error: cudnn_status_not_initialized error occurs if there is an issue with the CUDA Deep Neural Network (cuDNN) library. This error shows that the cuDNN library is unable to …
Runtimeerror: cuda error: invalid device ordinal
If you encounter this “Runtimeerror: Cuda Error: Invalid Device Ordinal” error. You already know how frustrating it should be. The most frustrating and confusing, as it often occurs without any …
Runtimeerror: couldnt install torch
Usually, we often run into errors like “runtimeerror: couldn’t install torch.”. It is one of the most common errors that developers may encounter during running their code. The “Runtimeerror: couldnt …
Runtimeerror: cuda error: device-side assert triggered
If you are encountering this error Runtimeerror: cuda error: device-side assert triggered. You will already know how frustrating it can be. The most confusing part for me have no clear, …
Frequently Asked Questions
What is the difference between RuntimeError and other exceptions?
RuntimeError is the catch-all base class for errors that do not fit a more specific exception (TypeError, ValueError, KeyError, etc.). Libraries reach for RuntimeError when no built-in exception matches their failure mode — common for GPU, async, framework-context issues.
How do I fix "CUDA out of memory" in PyTorch?
Five things to try in order: (1) Reduce batch size by half. (2) Add torch.cuda.empty_cache() between training epochs. (3) Use mixed-precision training (torch.cuda.amp). (4) Gradient checkpointing to trade compute for memory. (5) If using a Jupyter notebook, restart the kernel — old tensors may still be alive in GPU memory.
How do I fix "dictionary changed size during iteration"?
You are modifying a dict while iterating it. Fix by iterating over a copy: for key in list(d.keys()): or for key in dict(d):. Or build a new dict with a comprehension: d = {k: v for k, v in d.items() if condition}. Same applies to sets.
How do I fix "This event loop is already running"?
You are calling asyncio.run(...) from inside an already-running event loop (Jupyter, FastAPI, etc.). Use await directly if you are already in an async context. For Jupyter, use nest_asyncio.apply() at the top of the notebook — but only as a last resort.
How do I fix "Working outside of request context" in Flask?
You are trying to access request, session, or g from code that runs outside an HTTP request — typically background jobs, CLI commands, or unit tests. For backgrounded code, push an application context manually: with app.app_context():. For request-specific objects, you may need to pass the data explicitly instead.
How often is this RuntimeError reference updated?
New posts are added weekly. Existing posts are revised when major versions of PyTorch, CUDA, Flask, or asyncio ship breaking changes. Last refreshed: May 2026.