[Qubitium]

Summary

Replaces the GPTQMODEL_MSE_GRID_BUDGET_MB / GPTQMODEL_CLIP_GRID_BUDGET_MB
environment variables from #3435 with a proper config object on
QuantizeConfig, folding the legacy float mse exponent into it: the field
is now mse, taking an MSEConfig.

QuantizeConfig(
    mse=MSEConfig(norm=2.4, grid_budget="64M", clip_grid_budget="1G"),
)

What Changed

• quantization/config.py: MSEConfig now carries norm (the MSE exponent)
  plus grid_budget / clip_grid_budget. Budgets are byte counts (ints) or
  postfixed strings ("256M", "1G", optional B); None keeps the
  device-derived default (16M CPU / 256M GPU) and 0 forces single-candidate
  chunking (and one-token AWQ product tiles). The mse field lives on
  BaseQuantizeConfig, so GPTQConfig, QQQConfig, and AWQConfig all
  accept it.
• Back-compat normalization: mse=2.4 → MSEConfig(norm=2.4); dicts merge
  field-wise, and per-layer dynamic overrides replace norm while keeping
  configured budgets.
• quantizer.py: find_params/_mse_shrink_grid read the exponent and grid
  budget from qcfg.mse.
• qqq.py: Quantizer now accepts qcfg in its constructor like the GPTQ
  Quantizer and reads the grid budget from self.qcfg.mse; the
  configure(grid_budget_mb=...) shim is gone. Internal Quantizer() call
  sites pass self.qcfg.
• awq_processor.py / qqq_processor.py / gptq_processor.py: read budgets
  and norm from qcfg.mse; dynamic mse overrides go through
  _normalize_mse.
• models/writer.py + config.py meta payload: checkpoint meta still
  serializes the float exponent (meta["mse"] = mse.norm) so saved configs
  round-trip identically. Budgets are calibration-scratch tuning knobs, not
  part of quantization identity.
• README updated for the new API.

Tests

• Unit tests have been executed and passed.
• Existing MSE/clip grid tests were updated to the new config surface
  (mse=MSEConfig / mse={...} dicts); the tiny-budget forced-chunking
  paths still hold.

cd tests && python -m pytest test_find_params_mse_grid.py \
    test_qqq_quantizer.py test_awq_clip_consistency.py -x -q
# 134 passed, 72 skipped (CUDA unavailable on this VM)

python -m pytest test_serialization.py -k meta -q
# 1 passed (meta["mse"] == cfg.mse.norm round-trip)

Review Requirements

• Reviewed this PR.
• Checked that the code matches existing project structure, APIs, and conventions.
• Avoided unnecessary monkeypatching and used the project's normal extension points where possible.
• Minimized and compacted the impacted code surface.
• Considered and eliminated potential regressions.

Kernel Accuracy Requirements

• No kernel numerics changed in this PR; the config plumbing only feeds
  chunk-size targets into the vectorized grid evaluators from perf(quantizer): vectorize MSE shrink grid and AWQ clip grid evaluation #3435, whose
  bitwise-equality vs the sequential oracle was established there.

Notes

• norm semantics are unchanged: 0 disables the MSE search; AWQ ignores it
  and only uses clip_grid_budget.
• No GPU on this VM: CUDA-parametrized test cases skip locally.
• 3 unrelated test_serialization.py cases fail with HF-download OSError
  in this environment (no network access to checkpoints).

Link to Devin session: https://app.devin.ai/sessions/2eb9e33e49b54734a60b2552d71b0aca
Open in Devin Desktop: https://app.devin.ai/desktop/session/2eb9e33e49b54734a60b2552d71b0aca?variant=devin
Requested by: @Qubitium

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access that start with 'Devin'.

• Disable automatic comment, CI, and merge conflict monitoring