YannByron opened a new pull request, #10627: URL: https://github.com/apache/paimon/pull/10627
### Purpose CWD deep-copies every requested cell before applying transforms. With long action windows, this recursively copies the immutable scalars in each numeric vector, including vectors shared by overlapping windows. Copy flat lists of exact builtin immutable values directly; retain `deepcopy` for nested values, subclasses and custom objects so transforms and padding remain isolated. This is a two-file follow-up to #10567. It adds no sampler or cache configuration. The existing mutable-cell test now checks repeated offsets, overlapping/duplicate windows, nested lists and mutable constant padding against explicit values. Local real-image loader measurements (five alternating pairs after a discarded warmup run): | Workload | master | This change | Ratio of medians | | --- | ---: | ---: | ---: | | 128-step action window, 0 workers | 142.04 samples/s | 150.71 samples/s | +6.1% | | 32-step action window, 0 workers | 170.39 samples/s | 170.13 samples/s | -0.2% | These use four complete RoboMIND episodes (2,299 frames), three original 640x480 JPEG views, unchanged sample order and float32 tensors, batch 16, four warmup batches and 64 measured batches. Hardware: M2 Pro / 32 GiB, Python 3.13, Torch 2.10, PyArrow 19; Python reader, local files, uncontrolled OS cache. The 128-step paired gains range from +2.9% to +9.4%; the 32-step workload has no stable gain. This does not establish GPU or 30% end-to-end training gains. ### Tests - [x] Extended mutable-cell isolation regression passes; no-copy and nested shallow-copy mutations fail on output assertions. - [x] Focused flake8 and `git diff --check`. - [x] Real HDF5 / Paimon / official LeRobot tensor and sample-ID replay agrees for the measured single-worker workload. - [ ] Complete affected Python regression suite and remaining worker/ACT comparisons (draft while these run). -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
