Metal backend: use HQQ in Parakeet quantization (pytorch#17258)

manuelcandales · web-flow · commit f48a600ea713 · 2026-02-05T21:32:57.000-05:00
Bump ao pin to get pytorch/ao#3829, and use HQQ in Parakeet quantization
diff --git a/examples/models/parakeet/quantize.py b/examples/models/parakeet/quantize.py
@@ -37,6 +37,7 @@ def quantize_model_(  # noqa: C901
         config = UIntxWeightOnlyConfig(
             group_size=qlinear_group_size,
             bitwidth=4,
+            uintx_choose_qparams_algorithm="hqq",
         )
 
         def linear_filter(m, fqn):
diff --git a/third-party/ao b/third-party/ao
@@ -1 +1 @@
-Subproject commit 1b4b6d998bf988f059e97a10181cbc4aec269b69
+Subproject commit 026b76d12758af2687de28837b5c6de8ef8ad577

Original file line number	Diff line number	Diff line change
`@@ -37,6 +37,7 @@ def quantize_model_( # noqa: C901`
`37`	`37`	`config = UIntxWeightOnlyConfig(`
`38`	`38`	`group_size=qlinear_group_size,`
`39`	`39`	`bitwidth=4,`
	`40`	`+ uintx_choose_qparams_algorithm="hqq",`
`40`	`41`	`)`
`41`	`42`
`42`	`43`	`def linear_filter(m, fqn):`