fix(quantization): disable the KV cache for layerwise calibration - #2248

Open

Fridah-nv wants to merge 13 commits into

mainfrom

fridah/layerwise-kv-cache-replay-fix

Commits on Aug 25, 2026

fix(quantization): stop a layerwise replay attending over its own KV cache

Fridah-nv
and
claude
committed
7baf1d0
docs: compact the KV-cache fix docstring and changelog entry

Fridah-nv
and
claude
committed
20df3d3
fix(quantization): clear the KV cache on both replay paths, widen the changelog

Fridah-nv
and
claude
committed
6013c55
test: drop the inert calib_mutates_weights override from the KV-cache test

Fridah-nv
and
claude
committed
0f12053
fix(quantization): clear the KV cache at capture, generalize the changelog

Fridah-nv
and
claude
committed
3383c7a
refactor(quantization): clear the captured KV cache once, at capture

Fridah-nv
and
claude
committed
e03c53b
fix(quantization): match the KV cache by shape, and clear it on the resume path

Fridah-nv
and
claude
committed
3310a48
docs: compact the KV-cache comments and docstrings

Fridah-nv
and
claude
committed
519a789
test: keep only the KV-cache tests that can fail

Fridah-nv
and
claude
committed
555d6bc
test: fold the two structural KV-cache tests into one parametrized case

Fridah-nv
and
claude
committed
1e721a0
fix(quantization): disable the KV cache for layerwise calibration

Fridah-nv
and
claude
committed
c3300d8

Commits on Aug 26, 2026

fix(quantization): refuse a KV cache at capture, and scope the changelog

Fridah-nv
and
claude
committed
dd3e40c
docs(changelog): drop the layerwise KV-cache entry

Fridah-nv
and
claude
committed
0e35120