ruro

icon indicating a website link ruro.live icon indicating an email [email protected]

icon company Tevian LLC icon location I am a CV/ML researcher at Tevian LLC. I teach CV courses at MSU, YSDA, MIPT and HSE. I am very competent, hardworking, smart and most importantly modest.

Results 104 comments of ruro

XDG Base Directory support, and/or data directory environment variable

AFAIK this is not stale @dosubot

Fix onnx Attention and torch SDPA quantization handling

Hm. iirc `onnx` added support for opset 23 in version `1.18.0`. So the new test is currently failing in CI due to ``` onnx==1.17.0; python_version < '3.13' onnx==1.18.0; python_version >=...

`MULTIHEAD_ATTENTION_OUTPUT` ignored patterns don't match "proper" SDPA / Attention

After spending some time trying to figure this out, I think that SDPA actually has its own custom handling logic in [`_get_scope_overrides` in `nncf/quantization/algorithms/min_max/algorithm.py`](https://github.com/openvinotoolkit/nncf/blob/09246b2bc7c244ae873171f08994be5ec8ac6623/src/nncf/quantization/algorithms/min_max/algorithm.py#L641), so a separate `MULTIHEAD_ATTENTION_OUTPUT` pattern is...

`MULTIHEAD_ATTENTION_OUTPUT` ignored patterns don't match "proper" SDPA / Attention

I think I figured out how to fix the `torch` issue too. While the `PTNNCFGraph` graph only has a single edge, the `InsertionPointGraph` and `QuantizerPropagationStateGraph` graphs that are generated from...