Fix NVFP4 QAT convert path #3450
Conversation
**Summary:** The previous convert path through `QATConfig` did not swap `NVFP4FakeQuantizedLinear` back to `torch.nn.Linear`. The numerics tests still passed because this fake quantized linear happen to match the PTQ numerics exactly. **Test Plan:** ``` python test/quantization/test_qat.py -k test_qat_nvfp4 ```
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/3450
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ✅ No FailuresAs of commit 048568f with merge base aa21b80 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
andrewor14
added
the
topic: improvement
|
@andrewor14 has imported this pull request. If you are a Meta employee, you can view this in D88512245. |
| sqnr = compute_error(out, baseline_out).item() | ||
| self.assertGreaterEqual(sqnr, float("inf")) | ||
|
|
||
| # Compare converted values |
There was a problem hiding this comment.
oh we didn't compare convert results before?
There was a problem hiding this comment.
It's tested here:
ao/test/quantization/test_qat.py
Line 2085 in 048568f
We just never explicitly checked it's using tensor subclasses after convert (tests still passed because QAT prepare mimics PTQ exactly)
Summary: The previous convert path through
QATConfigdid not swapNVFP4FakeQuantizedLinearback totorch.nn.Linear. The numerics tests still passed because this fake quantized linear happen to match the PTQ numerics exactly.Test Plan: