fix(qqq): accumulate Hessian in-place per device to reduce memory

[Qubitium]

Summary

Reduce QQQ quantization peak memory by avoiding a separate columns x columns xtx output tensor for every calibration batch and by freeing per-device Hessian partials as they are merged.

Changes

QQQ.add_batch:

• Keep the existing per-device Hessian partial design (required for calibration_data_device="balanced" / multi-GPU MoE).
• Allocate the partial buffer once per device as float32.
• Accumulate inp @ inp.t() directly into that buffer with addmm_ instead of inp.matmul(inp.t()) + add_.
  • This removes the transient xtx tensor that is the same size as the Hessian for each batch.

QQQ.materialize_hessian:

• Merge per-device partials into the final Hessian on the target device.
• popitem() and del partial each partial as it is added, so the peak during materialization is result + one partial instead of result + all partials.

QQQ.quantize:

• Fixed the diag index tensor to live on H.device so the damping step is safe when H has been materialized on a different device from self.dev.

Validation

• ruff check gptqmodel/quantization/qqq.py tests/test_qqq.py passes.
• New TestQQQHessian unit tests in tests/test_qqq.py verify that add_batch + materialize_hessian match the closed-form 2/N * X^T X reference for:
  • single batch,
  • multiple batches,
  • multi-device (cpu + cuda:0) accumulation,
  • empty batches.
• pytest tests/test_qqq.py::TestGroupSize::test_load_group_128 -v passed on GPU 4.

Closes memory-reduction follow-up for QQQ after #2115 / PR #2979.

[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