tinygrad

tinygrad copied to clipboard

WIP: `tensor.squeeze` function + exception testing

1

In an effort to stick closely to the Pytorch API, I added functionality for testing exceptions being thrown. The squeeze function pops the targeted dimension if its size is 1....

dosier

llama.py has an option to run the profiler on one of the tokens that gets generated, but it seems the profiler wants to run only on the third token. If...

carsonRadtke

Fix Metal tests on CI

1

I could reproduce the failing Metal tests on the CI on a local Intel Mac, so I guess the failures are specific to this platform. I have narrowed it down...

akretz

Use `-n=auto` to speed up tests in CI. | | Before | After | Speed up | |-----------------|:-------:|:-------:|:------------:| | Optimizer Test | 1m21s | 45s | 44% | | GPU...

jla524

diverse test value in test_dtype DATA based on dtype

2

added negative int for int case TODO: test this on real GPU. test passed on tinybox

chenyuxyz

fp16 resnet

2

this used to at least produce non-nan losses. testing it...

chaosagent

search: change to use "spawn" and limit the number of tasks per child

1

also clean up some examples to use __main__ and not initialize resources outside of main This limits the amount of memory leaked in the child processes. For very large kernels,...

flammit

don't include hip common

5

You only need to install "hsa-rocr comgr" (on a real system, remu needs more? @Qazalin) We can make HIP CI a lot faster. We also need to replace * hsa-rocr...

geohot

MLperf BERT training

12

## Usage Instructions ### Step 1: Downloading, Verifying, and Uncompressing Datasets Ensure the dataset is properly downloaded, verified against an MD5 hash, and uncompressed to your specified base directory. **Command:**...

Eliulm

CUDA float16 conversion not working for coder.py

6

python3 tinygrad/examples/coder.py [44/175] create model: 240.74 ms https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B/resolve/main/pytorch_model-00001-of-00002.bin?download=true: 100%|███████████████████████████████████████████████████| 9.94G/9.94G [03:14

vish-pr