Device properties example #474
Conversation
leofang
added
enhancement
| from cuda.core.experimental import Device, system | ||
|
|
||
|
|
||
| def _bool_str(value: bool) -> str: |
There was a problem hiding this comment.
_yes_or_no (or just _yes_no works, too)
makes this self-explanatory (in the call sites).
There was a problem hiding this comment.
Yep, will change both filename and function to _yes_no
| return "YES" if value else "NO" | ||
|
|
||
|
|
||
| def _bytes_to_mbytes(value): |
There was a problem hiding this comment.
def _bytes_to_mbytes(value):
return f"{value / (1024 ** 2):.2f}MB"
def _bytes_to_gbytes(value):
return f"{value / (1024 ** 3):.2f}GB"
These were generated by ChatGPT: https://chatgpt.com/share/67bfde2c-7b20-8008-a099-302d0ed4dc60
There was a problem hiding this comment.
No, I am an old schooler. I used Windows calculator to calculate magic numbers
There was a problem hiding this comment.
I agree, it's more readable but my guts protest to use expensive power and division ops
There was a problem hiding this comment.
(Today I started experimenting with claude.ai)
https://claude.ai/share/5830e427-4b3b-44f5-8bef-c47d10ebe01f
I'm slightly modifying what it suggests:
BYTES_TO_GBYTES = 1 / (1024 ** 3)
def _bytes_to_gbytes(value):
return f"{value * BYTES_TO_GBYTES:.2f}GB"Making our examples readable seems valuable.
| print(f"- Constant memory available: {properties.total_constant_memory}B") | ||
| print(f"- Support for unified address space with host: {_bool_str(properties.unified_addressing)}") | ||
| print(f"- Support for virtual memory management: {_bool_str(properties.virtual_memory_management_supported)}") | ||
|
|
There was a problem hiding this comment.
Optional, but I'd lean towards adding a print() here, so that the output doesn't run directly into the command line prompt.
There was a problem hiding this comment.
I double checked log.txt. The last few lines in the output will be:
...
- Support for unified address space with host: YES
- Support for virtual memory management: YES
*****************************************************
Two blank lines are printed before command line appears. See line 169
| print(f"DEVICE {device.name} (id={device_id})") | ||
|
|
||
| device.set_current() | ||
| ctx = device.context |
There was a problem hiding this comment.
This seems to be unused. (Hm, not sure why ruff doesn't catch this.)
There was a problem hiding this comment.
Hm, not sure why ruff doesn't catch this.
Oh! Because it's currently a module-level variable (which could in theory be used from elsewhere). Another reason to move this code into a function. I think ruff will then flag this.
There was a problem hiding this comment.
I have initiated a separate discussion on set_current(). I wanted to add some meaningful example for it but it appears that context management is not yet fully implemented. Or am I wrong? I am happy to extend example with context stuff
There was a problem hiding this comment.
I upvoted your suggestion (yesterday).
it appears that context management is not yet fully implemented.
I had to look this up myself: #189
What do you think about adding a comment (very rough, just to give the idea):
# Extend example to show device context information after #189 is resolved.
|
@rwgk Your suggestions have been addressed |
|
/ok to test |
It works for all GitHub IDs linked with a valid NVIDIA email address, so Sergey you should be able to trigger it yourself! |
This comment has been minimized.
This comment has been minimized.
|
(I tried to look at the log of the one failing test but it doesn't load. I'll try again later.) |
|
It was a transient network issue, I reran the failed job and all is good now. Merging, thanks Sergey & Ralf! |
|
This example aims to get familiar with cuda.core device and device properties concepts. The script iterates across all devices in the system and prints their device properties.
log.txt