modeld: skip redundant cast, reshape, and flatten

[Quantizr]

Removing the redundant operations improves modeld.py CPU usage from 32% to 24% and makes execution time ~2% faster

For dmonitoringmodeld.py, it improves CPU usage from 17% to 12% and makes execution time ~4% faster.

---

In tinygrad, Tensor.numpy() is essentially defined as

self.cast(self.dtype.base).contiguous().to("CPU").realize().uop.base.buffer.numpy().reshape(self.shape)

in compile3.py, we already cast to float32

  run_onnx_jit = TinyJit(lambda **kwargs:
                         next(iter(run_onnx({k:v.to(Device.DEFAULT) for k,v in kwargs.items()}).values())).cast('float32'), prune=True)

and in modeld.py and dmonitoringmodeld.py, flatten() is called after .numpy(), making the reshape redundant.

There is also an extra .to("CPU") which I think currently causes tinygrad to do a useless copy from CPU to NPY

gotta afford the hey comma cpu usage somehow

[commaci-public]

ref for commit f7324f2: https://raw.githubusercontent.com/commaai/ci-artifacts/refs/heads/model_replay_models-go-brrr/8494c69d3c710e81|000001d4--2648a9a404_model_tici_f7324f263f6b24cd3889bf010d4fb66d97d310db.zst

All Model Replay Plots

[Copilot]

Pull Request Overview

This PR removes redundant tensor-to-NumPy operations (cast, reshape, flatten, and extra .to("CPU")) in model inference to lower CPU usage and slightly speed up execution.

• Replace .numpy().flatten() with direct buffer extraction via .contiguous().realize().uop.base.buffer.numpy()
• Apply the same change for both vision and policy outputs in modeld.py and for the monitoring model in dmonitoringmodeld.py

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
selfdrive/modeld/modeld.py	Drop .flatten() and use direct buffer NumPy extraction
selfdrive/modeld/dmonitoringmodeld.py	Remove redundant .flatten() in monitoring model output process

Comments suppressed due to low confidence (3)

selfdrive/modeld/modeld.py:166

• Removing .flatten() changes vision_output from a 1D array to its original multi-dimensional shape, which may break slice_outputs. To preserve existing behavior, reapply a flatten step (e.g., .reshape(-1) or .flatten()) after the NumPy conversion.

    self.vision_output = self.vision_run(**self.vision_inputs).contiguous().realize().uop.base.buffer.numpy()

selfdrive/modeld/modeld.py:173

• The .flatten() call was removed here as well, altering policy_output shape from a flat vector to multi-dimensional. Consider adding .reshape(-1) or .flatten() to maintain the expected 1D output.

    self.policy_output = self.policy_run(**self.policy_inputs).contiguous().realize().uop.base.buffer.numpy()

selfdrive/modeld/dmonitoringmodeld.py:96

• By removing .flatten(), output will retain its original multi-dimensional shape. If downstream logic expects a flat array, reapply a flatten (e.g., .reshape(-1) or .flatten()) after conversion.

    output = self.model_run(**self.tensor_inputs).contiguous().realize().uop.base.buffer.numpy()

[Copilot]

[nitpick] The long chain .contiguous().realize().uop.base.buffer.numpy() is repeated multiple times. Consider extracting this pattern into a helper function (e.g., to_cpu_array(tensor)) to reduce duplication and improve readability.