Fix Params4bit.to() leaving CPU-packed weights packed

[22elix3r]

Params4bit.cuda() and .xpu() undo AVX512 CPU packing before the move. nn.Module.to() calls Parameter.to(), so model.to(device) skipped that step and kept the packed nibble layout.

This moves the same unpack into Params4bit.to() when the destination is not CPU.

Fixes #2078

[Copilot]

🟢 Approval recommended

The change is narrowly scoped, matches existing .cuda()/.xpu() behavior, and includes a targeted regression test covering the reported corruption scenario.

Pull request overview

This PR fixes a correctness bug in Params4bit.to() when moving a CPU-packed (AVX512) 4-bit weight to a non-CPU device via the common nn.Module.to(...) pathway. It ensures the CPU-specific packed nibble layout is inverted before transferring to accelerators, matching the behavior already present in .cuda()/.xpu() and preventing silent weight corruption.

Changes:

• Add a CPU-packing inverse step inside Params4bit.to() when the destination device is not CPU/meta.
• Add a regression test that packs a quantized CPU weight, moves it off CPU via .to(device), and checks dequantized values match the pre-pack reference.

File summaries

File	Description
bitsandbytes/nn/modules.py	Ensures Params4bit.to() undoes AVX512 CPU packing when moving to non-CPU devices, covering model.to(device) flows.
tests/test_linear4bit.py	Adds regression coverage for leaving CPU with packed 4-bit weights via .to(...).

Review details

• Files reviewed: 2/2 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[22elix3r]

@matthewdouglas @TimDettmers Params4bit.to() was skipping the CPU packing inverse that cuda()/xpu() already do, so model.to(device) could keep packed weights. Small patch plus a test. Review welcome.