fix(rotation): persist rotation config and apply online Hadamard for inference and calibration

[Qubitium]

Summary

Cherry-pick/adapt the rotation persistence + online-Hadamard work from GPT-QModel-Ultra#66, plus a calibration fix. The original inference-only fix caused a gsm8k_platinum regression because mlp.down_proj weights contained the fused Hadamard H while GPTQ was still calibrating on the raw down_proj input m. This PR makes HookedLinear apply the same online Hadamard during calibration/replay so the GPTQ Hessian matches the inference activation distribution.

What Changed

• QuantizeConfig emits/round-trips rotation in quantize_config.json.
• rotate_mlp_output fuses a full exact Hadamard into mlp.down_proj weights and now stores online_full_had/had_K/K on the nn.Linear so the state survives the HookedLinear conversion used during GPTQ.
• Added apply_online_hadamard(...) in gptqmodel/quantization/rotation/hadamard_utils.py; BaseQuantLinear._apply_rotation_to_input and HookedLinear.forward both call it.
• HookedLinear.from_linear copies rotation state from the source nn.Linear.
• TorchLinear._forward and TritonV2Linear.forward call _apply_rotation_to_input before the quantized matmul.
• BaseGPTQModel.quantize()/from_quantized() attach online-Hadamard state to mlp.down_proj and reject rotation for unsupported backends.
• BaseQuantizeConfig disables offload_to_disk when rotation is set (rotation needs materialized weights).
• tests/test_rotation_persistence.py verifies config round-trip, online-state setup, forward routing, Hadamard cancellation, and HookedLinear calibration-path cancellation.
• Removed an unreachable trailing return x flagged by code-quality review.

Tests

• I added a new simple/fast unit test for this change.
• I ran the new targeted test locally.
• I ran other directly relevant local tests.

cd /root/repos/GPTQModel
PYTHON_GIL=0 pytest -q tests/test_rotation_persistence.py
# 9 passed, 15 warnings in 7.49s

python -m ruff check gptqmodel/quantization/rotation/hadamard_utils.py \
  gptqmodel/nn_modules/hooked_linear.py gptqmodel/nn_modules/qlinear/__init__.py \
  gptqmodel/quantization/rotation/rotation.py tests/test_rotation_persistence.py
# All checks passed!

Full-dataset Evalution on Llama-3.2-1B-Instruct (4-bit GPTQ, rotation='hadamard') against a non-rotated 4-bit baseline and a dense BF16 reference:

checkpoint	gsm8k_platinum acc,num	arc_challenge acc,exam
dense BF16	0.3821	0.3460
baseline GPTQ (no rotation)	0.3160	0.3528
rotated GPTQ (this fix)	0.3565	0.3514

The rotated model recovers the gsm8k_platinum regression and scores between the dense reference and the non-rotated baseline; arc_challenge is unchanged within noise.

Review Requirements

• I personally reviewed every file in this diff.
• I checked that the code matches existing project structure, APIs, and conventions.
• I avoided unnecessary monkeypatching and used the project's normal extension points where possible.

Notes

The JIT fast_hadamard_transform extension is vendored from Dao-AILab/fast-hadamard-transform (Tri Dao, BSD-3-Clause) under gptqmodel_ext/hadamard/ and only used by the gptq_torch/gptq_triton backends.

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring