Add VLA v2 execution notebook and v2_experiments benchmark harness

[Vishal-sys-code]

Motivation

• Provide a reproducible execution harness for VLA v2 that targets the current fast stack (VLASequential, VLAParallel, and VLATriton when available) and saves paper-ready artifacts.
• Consolidate benchmarking, small-model ablations, and a tiny LM proxy evaluation into a single script to match VLAv3-era references and make comparisons consistent.
• Ensure results are portable and discoverable by emitting CSV/JSON/PNG/PDF artifacts plus a manifest describing the run.

Description

• Add notebooks/08_VLA_v2_Execution_Notebook.ipynb which demonstrates creating a V2Config, running run_full_v2_suite, and inspecting generated artifacts.
• Add scripts/v2_experiments.py that defines V2Config, a GatedNormalizedSM ablation module, helpers (set_seed, make_vla_backend), benchmarks (benchmark_streaming, run_mqar_capacity_proxy), evaluation (run_tiny_lm_eval, run_sm_ablation), plotting (save_plot), and the orchestration function run_full_v2_suite.
• run_full_v2_suite conditionally includes the triton backend if HAS_TRITON is true and writes mqar_capacity_multiseed.csv, streaming_benchmarks.csv, lm_eval_tiny.json, sm_ablation_metrics.json, plot files (.png/.pdf), and manifest.json to the configured out_dir.

Testing

• No CI tests were added as part of this change.
• Performed local automated smoke runs by importing scripts.v2_experiments, instantiating V2Config, and executing run_sm_ablation and run_tiny_lm_eval which completed successfully on the test environment.
• Plot generation is guarded by a matplotlib import check so plotting is skipped cleanly when the plotting backend is not available.

---

Codex Task

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
variational-linear-attention	Ready	Preview, Comment	May 20, 2026 6:34am