Add sliding-window attention tuning support

[umangyadav]

Summary

• Include sliding_window_size in Rock and Python attention tuning identities, reports, and quick-tuning grouping.
• Propagate current_seq_len through runtime rocmlir-gen commands while deliberately excluding it from tuning keys.
• Exercise sliding-window configurations in attention sweeps and add rocmlir-gen, MIGraphX, legacy tuning-DB, and mixed-TSV regression coverage.

Bugs fixed and error signatures

Sliding-window tuning-key collisions

The Rock tuning key previously omitted slidingWindowSize. A windowed problem therefore emitted the same sequence around -split_kv 1 -num_heads_q ... as a non-windowed problem, and different window sizes also shared one identity. Tuning databases could consequently return a stale or suboptimal perf configuration for a materially different attention kernel.

The key now conditionally emits -sliding_window_size <N> after -split_kv. A disabled window remains omitted, so existing non-windowed tuning keys are unchanged.

Python tuning-key and report mismatch

Once the C++ key contains the window, the old Python parser fails with:

ValueError: Unknown Attention config argument -sliding_window_size -> <N>

Python reports and quick-tuning grouping also lacked a window column, which merged distinct window sizes into one problem identity. Older TSVs would either raise KeyError: 'SlidingWindowSize', or—when concatenated with newer TSVs—receive NaN and be silently dropped by pandas groupby.

AttentionConfiguration, report columns, and quick-tuning columns now carry SlidingWindowSize; missing and NaN legacy values normalize to the disabled value 0.

Lost runtime sequence positions

current_seq_len is a runtime KV-cache position, not a tuning dimension. Previously it was appended only inside parameterSweeps.test_config; commands and repro strings produced directly by AttentionConfiguration omitted it. A sliding-window command reconstructed through that path could fail Rock verification with:

slidingWindowSize requires currentSeqLen to be set

The configuration now owns runtime argument emission, so current_seq_len appears exactly once in generated driver commands. It remains absent from to_command_line() and C++ tuning keys, preventing per-position tuning-key fragmentation.

Missing sweep and fusion coverage

Attention parameter sweeps never generated sliding-window cases, and the MIGraphX problem-key path did not verify that a folded sliding-window mask contributes its window—but not its runtime sequence position—to the key. The sweep now covers both plain and sliding-window KV-cache problems, and focused regression tests pin both key behavior and legacy compatibility.

Python lint gate

Changing perfRunner.py activates the repository's full-file flake8 check. Two pre-existing inline dictionary expressions triggered E201/E231; they are normalized without changing behavior so the Python lint gate remains clean.

Compatibility

• Non-sliding attention keys remain byte-for-byte unchanged.
• Legacy tuning TSVs without SlidingWindowSize continue to load as window size 0.
• current_seq_len remains runtime-only and never enters tuning DB identity or reports.

Test plan

• ninja -C build rocmlir-gen rocmlir-driver
• Focused lit tests: rocmlir-gen problem key, MIGraphX sliding-window problem key, and attention tuning-DB compatibility
• python3 -m pytest -q mlir/utils/performance/tests (126 passed)
• C++ premerge clang-format and clang-tidy
• Python flake8, YAPF, and syntax compilation
• git diff --check origin/develop

Made with Cursor

[Copilot]

Pull request overview

Adds first-class sliding-window attention support to rocMLIR tuning identities and the Python performance tooling, ensuring sliding-window kernels are keyed distinctly while keeping current_seq_len runtime-only.

Changes:

• Extend Rock (C++) and Python attention tuning identities/reporting to include sliding_window_size, preserving legacy non-window keys.
• Plumb current_seq_len into runtime rocmlir-gen command generation while excluding it from tuning keys/DB identities.
• Add/extend regression coverage across rocmlir-gen tuning keys, MIGraphX fusion keys, tuning-DB compatibility, and quick-tuning TSV handling.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
mlir/utils/performance/reportUtils.py	Adds SlidingWindowSize to attention report/test parameter columns.
mlir/utils/performance/perfRunner.py	Extends AttentionConfiguration with sliding_window_size + current_seqlen; emits them in runtime driver args (but not tuning keys).
mlir/utils/performance/parameterSweeps.py	Removes manual current_seq_len append since runtime arg emission moves into configs.
mlir/utils/performance/attentionSweeps.py	Generates and propagates sliding-window samples into attention sweep configurations.
mlir/utils/performance/analysis/quickTuningGen.py	Adds SlidingWindowSize to grouping columns and defaults legacy/mixed TSVs to 0.
mlir/test/rocmlir-gen/problem-key.mlir	Pins tuning-key behavior: window participates in key; current_seq_len does not.
mlir/test/perf-scripts/attention-tuning-db-compat.py	Expands compatibility tests for sliding-window keys and runtime-only current_seq_len.
mlir/test/fusion/mixr-attention-sliding-window-problem-key.mlir	Adds MIGraphX fusion tuning-key regression for sliding-window attention.
mlir/lib/Dialect/Rock/Tuning/RockTuningImpl.cpp	Adds conditional -sliding_window_size <N> emission to Rock attention tuning keys.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[codecov]

Codecov Report

❌ Patch coverage is 66.66667% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
mlir/lib/Dialect/Rock/Tuning/RockTuningImpl.cpp	66.67%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #2439      +/-   ##
===========================================
+ Coverage    82.57%   83.69%   +1.13%     
===========================================
  Files          120      121       +1     
  Lines        42852    43183     +331     
  Branches      7110     7182      +72     
===========================================
+ Hits         35381    36142     +761     
+ Misses        4815     4490     -325     
+ Partials      2656     2551     -105     

Flag	Coverage Δ
gfx120x	83.43% <66.67%> (+0.91%)	⬆️
gfx950	83.51% <66.67%> (+1.16%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
mlir/lib/Dialect/Rock/Tuning/RockTuningImpl.cpp	60.88% <66.67%> (+1.56%)	⬆️

... and 38 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[bogdan-petkovic]

LGTM

[justinrosner]

Nit: I think that we can drop this comment

[umangyadav]

Removed.

[justinrosner]

Why rename this function? Is it being used in non-legacy cases as well?

[umangyadav]

Yes It was renamed because it now handles both legacy and current keys.