[LinusJungemann]

Supersedes #262: the same change, ported onto the #257 port (refactor/custom_ops_v2) and current dev. Depends on the custom-ops PR; review that one first. Until it is merged, this PR's diff also contains its two commits.

Problem

#262 finishes the cleanup that #257 started for the custom ops:

• all of src/ passes ruff and pyright;
• assert is no longer used for error handling;
• ruff replaces flake8 in pre-commit.

It was written against the same 819-commit-old dev as #257. Since then dev has changed 15 of the files #262 touches. It has also added a lot of code #262 never saw: QoR estimation, resource models, torch_hw_modules, PWPolyF/Pad1D/Where/SelectToken/Requant conversion, multi-FPGA, MLO DDR weight offsets. With ruff now gating every file under src/ in CI, all of that code has to be clean too.

Changes

• Refactor Part 2: Fix all remaining ruff warnings, type errors and asserts in src #262 re-applied: Refactor Part 2: Fix all remaining ruff warnings, type errors and asserts in src #262's own changes are re-applied file by file, keeping every functional change dev made. Where Refactor Part 2: Fix all remaining ruff warnings, type errors and asserts in src #262 had cleaned up code that dev has since rewritten (e.g. FINNLoop aximm handling in create_stitched_ip, the routed-netlist power flow, the data_packing fast paths, robust_rmtree), dev's version is kept and cleaned up the same way.
• Code dev added since Refactor Part 2: Fix all remaining ruff warnings, type errors and asserts in src #262: cleaned up to the same standard. That means annotations, cast() at untyped qonnx boundaries, pathlib, f-strings and docstrings, and FINNUserError/FINNInternalError instead of assert.
• Tooling:
  • ruff (ruff-check v0.15.10 on ^src/) replaces flake8; .flake8 is removed.
  • The docstring-check action is removed, since ruff's D rules cover it.
• Behaviour: unchanged apart from exception types for invalid input, except:
  • AuroraPartitioner: the connections_per_device constraint now sums over the right nodes (Refactor Part 2: Fix all remaining ruff warnings, type errors and asserts in src #262's fix). The old generator used the leftover node from the previous loop, normally the last node, so ports_per_device had no effect. Partitions that were feasible before can now be infeasible.
  • fold_quant_weights: Refactor Part 2: Fix all remaining ruff warnings, type errors and asserts in src #262 added an "unknown output shape" check for every target op. It is now only applied to Conv/ConvTranspose, where the shape is used; applied to all targets it broke MatMul weight folding (test_brevitas_pwpolyf).
  • build_dataflow_steps: Refactor Part 2: Fix all remaining ruff warnings, type errors and asserts in src #262 moved the Vitis platform resolution into the multi-FPGA branch. It still runs first for single-FPGA Alveo builds, so an unresolvable platform keeps failing before synthesis.
• Merge fixes:
  • streamline_plus imported Refactor Part 2: Fix all remaining ruff warnings, type errors and asserts in src #262's AbsorbSignBiasIntoMultiThreshold but called dev's renamed AbsorbScalarBiasIntoMultiThreshold, which would have raised a NameError.
  • multi_dnn imported DNNContainer from the old abstract/ path.
• Tests: test_fifo_transaction_counts expects FINNUserError instead of AssertionError.

Testing

Lint:

• pre-commit run --all-files is clean, including ruff on all of src/.
• pyright src reports 0 errors.
• There are 0 assert statements in src/.

Follow-ups

• The changelog entries cite Refactor Part 2: Fix all remaining ruff warnings, type errors and asserts in src #262 and should be switched to this PR's number.

🤖 Generated with Claude Code