Skip to content

Qat - #5

Merged
ilfpns merged 3 commits into
featurefrom
QAT
Sep 17, 2026
Merged

Qat#5
ilfpns merged 3 commits into
featurefrom
QAT

Conversation

@ilfpns

@ilfpns ilfpns commented Sep 17, 2026

Copy link
Copy Markdown
Owner

No description provided.

ilfpns and others added 3 commits September 17, 2026 00:04
- ptq/quantization.py: qat_quantize() prepares the model with
  prepare_qat (fake-quant inserted), trains with a small Adam loop,
  then converts to a quantized model. build_self_distillation_data()
  builds (input, target) pairs using the FP32 model as teacher, since
  there is no labeled dataset yet.
- main.py: run QAT alongside FP32/Dynamic PTQ/Static PTQ and include
  it in the size/speed/accuracy reports.
- README.md: document the Static PTQ and QAT roadmap steps as done,
  including the observed (counter-intuitive) result that QAT's
  MaxDiff came out worse than Static PTQ's in this toy self-
  distillation setup, and why.

Co-Authored-By: Claude Sonnet 5 <[email protected]>
@ilfpns
ilfpns merged commit 2710204 into feature Sep 17, 2026
@ilfpns
ilfpns deleted the QAT branch September 17, 2026 11:49
ilfpns added a commit that referenced this pull request Sep 17, 2026
Merge pull request #5 from ilfpns/QAT
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant