[Qubitium]

Mirrors dphnAI/nf12#3. Pure cleanup — no code, no behaviour change.

The problem

Every HIP header stated the same three facts twice, and three of them repeated the authorship sentence that the SPDX block already makes:

// SPDX-FileCopyrightText: 2026 ModelCloud.ai
// SPDX-FileCopyrightText: 2026 AlpinDale and the nf12 contributors   <- kept
// SPDX-License-Identifier: Apache-2.0                                 <- kept
// ---------------------------------------------------------------------------
// NF12-AD (NyoomFloat12) decode for AMD RDNA4 ...
//
// Algorithm: "NyoomFloat12: Accelerating LLM Inference via Lossless 12-bit
// Weight Compression", Liberman, Fang, Zhang, Dao, and Fu (Together AI, 2026).
// Together AI: https://www.together.ai/
//
// Original NF12 reference implementation and CUDA kernels:
//   https://github.com/dphnAI/nf12                                     <- dup
//   Copyright (c) 2026 AlpinDale and the nf12 contributors.            <- EXACT dup
//   Licensed under the Apache License, Version 2.0.                    <- EXACT dup
//
// This HIP port was written for GPT-QModel (ModelCloud.ai) and consumes the <- dup

The fix

One three-line block, each party named exactly once:

// SPDX-FileCopyrightText: 2026 ModelCloud.ai
// SPDX-FileCopyrightText: 2026 AlpinDale and the nf12 contributors
// SPDX-License-Identifier: Apache-2.0
// ---------------------------------------------------------------------------
// NF12-AD (NyoomFloat12) decode for AMD RDNA4 (gfx1200 / gfx1201 / gfx1250).
//
// Credits: NyoomFloat12 algorithm by Liberman, Fang, Zhang, Dao and Fu
// (Together AI, 2026), https://www.together.ai/. NF12 format, reference codec
// and CUDA kernels by AlpinDale, https://github.com/dphnAI/nf12 (Apache-2.0).
//
// It consumes the same NF12-AD byte stream as the CUDA kernels ...

Both required credits are preserved — Together AI (NyoomFloat12) and AlpinDale (dphnAI/nf12) — and each file now mentions each of them exactly once.

Also

• gfx1200/gfx1200/gfx1250 → gfx1200/gfx1201/gfx1250 in the format header.
• The encode section headed "Why the port is not a copy" is now "Port notes" — the paragraph directly above it already says the port is a transliteration producing byte-identical output, so the old heading read as a contradiction.

Net −28 lines across the five files.

Verification

pytest tests/test_nf12_ad_codec.py tests/test_backend_naming.py '\
       tests/test_quant_bits_validation.py tests/test_rocm_llc.py
# 235 passed, 1 skipped

No functional change, so no perf or accuracy numbers to report.