feat(s-series): device Philox RNG (rng_uniform/normal + dropout) on x86 + ROCm (P6)

[gstoner]

Summary

Counter-based Philox-4x32-10 RNG as a real device lane on both native targets. A device RNG can't bit-match numpy's Generator (numpy uses Philox-4x64 + its own Lemire/ziggurat transforms), so Tessera's device lane uses the standard Philox-4x32-10 (Salmon et al. 2011 — the JAX/cuRAND algorithm): embarrassingly parallel (element i depends only on key + counter=i), reproducible, validated BIT-EXACTLY against a numpy reference of the identical algorithm. A separate deterministic stream from tessera.rng (the host numpy-Generator path).

What lands

• avx512_philox_f32.cpp — scalar Philox-4x32-10 C-ABI uniform kernel.
• GenerateROCMPhiloxKernel.cpp + ROCM_PhiloxKernelOp + pass registration — the gfx1151 MLIR Philox kernel (10 unrolled rounds, i64 umulhi); lowers to a valid hsaco, HIP-launched.
• rng_device.py — the numpy Philox-4x32-10 reference + uniform/normal/dropout transforms (the bit-exact target).
• runtime.py — _rng_compute + _execute_{x86,rocm}_compiled_rng: device uniform bits + host transform (uniform-scale / Box-Muller normal / dropout mask).
• backend_manifest / execution_matrix: fused x86 + compiled rocm entries, rows, descriptions, fixture map.

Proof

test_{x86,rocm}_rng_compiled.py — uniform/normal bit-exact vs the reference + statistics (mean≈0.5/var≈1/12; normal mean≈0/std≈1) + dropout rate/scale/eval + determinism. Both executed on hardware (AVX-512 + gfx1151): 9 passed; uniform/normal bit-identical across x86, gfx1151, and the numpy reference. Manifest/matrix/coverage/doc-drift gates green; ruff + mypy clean.

Note

The device stream is a standalone deterministic Philox-4x32-10, distinct from tessera.rng's host numpy-Generator stream — RNG streams are implementation-defined per backend; the contract is correctness + reproducibility, proven by bit-exact agreement with the documented reference. Dashboard: rng_uniform/rng_normal/dropout backend_kernel reference → partial.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a0f427f78c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[gstoner]

Note: Tessera's device Philox-4x32-10 is the same standard algorithm rocRAND ships (rocRAND PRNG list includes Philox 4x32-10; Salmon et al. 2011), so this device stream is consistent with rocRAND's Philox by construction — validated bit-exact against the in-repo numpy reference of the identical rounds. Per Decision #23 it's a reimplementation in Tessera codegen, not a rocRAND/hipRAND wrap. Future RNG generators (ThreeFry, Sobol QRNG, philox_4x64) can mirror rocRAND's set.