Giuseppe/functional quant reorder - #1580

Open

Giuseppe5 wants to merge 41 commits into

giuseppe/quant_tensor_subclassfrom

giuseppe/functional_quant_reorder

Commits on Sep 9, 2026

Feat (quant_tensor): convert from NamedTuple to Tensor subclass
Giuseppe5
authored
8437c0c
fix notebook
Giuseppe5
authored
4509802
more fixes
Giuseppe5
authored
b1f180e
added tests
Giuseppe5
authored
85a0e3f
remove .tensor alias
Giuseppe5
authored
c6e37ca
fix tests
Giuseppe5
authored
d966c54
Feat (quant_tensor): support for slicing of metadata
Giuseppe5
authored
9878562
Fix (quant_tensor): normalize negative group dimensions when slicing
Giuseppe5
authored
d8af572
pre-commit
Giuseppe5
authored
f8d2c80
Fix: disable cached QuantTensor reconstruction during Dynamo export
Giuseppe5
authored
ffbed46
fix tests
Giuseppe5
authored
fba4387
fix test 2
Giuseppe5
authored
4ee25f2
Simplify interface
Giuseppe5
authored
206344e
coherent-ish names
Giuseppe5
authored
f25dc15
fix tests
Giuseppe5
authored
dfee8a8
Fix notebooks
Giuseppe5
authored
ae094d6
simplify interface
Giuseppe5
authored
1eb67b3

Commits on Sep 17, 2026

Review
Giuseppe5
committed
c87c2e7
Feat: add prepared functional activation quantization
Giuseppe5
committed
1987644
Feat: support functional quantization specifications
Giuseppe5
committed
b50a948
Feat: add functional weight quantization lifecycle
Giuseppe5
committed
1c54b59
Feat: support functional gradient checkpointing
Giuseppe5
committed
fc5e4f8
Feat: export functional quantization API
Giuseppe5
committed
65be074
Feat: migrate LLM SDPA to functional quantization
Giuseppe5
committed
29780ba
Test: cover functional quantization contracts
Giuseppe5
committed
51f7ff9
precommit
Giuseppe5
committed
2c60e31
Support for input/weight/all functional quantization
Giuseppe5
committed
9c9698a
Fix: harden functional parameter provenance
Giuseppe5
committed
4a948d8
Refactor: discover functional parameter owners before parametrization
Giuseppe5
committed
0646db1
Feat: allow custom functional quantization specifications
Giuseppe5
committed
51c75de
Fix: pass through disabled unprepared functional calls
Giuseppe5
committed
7d7e0f9
Refactor: remove functional QuantLinear special cases
Giuseppe5
committed
10d3f25
Feat: support grouped_mm
Giuseppe5
committed
6ac1c83
Fix: correct reset per-module rather than per-model
Giuseppe5
committed
247aa92
Fix: do not call runtime quant for parameters
Giuseppe5
committed
68ff518
fix tests
Giuseppe5
committed
fc61f3e
Precommit
Giuseppe5
committed
db4dc16
Support for better provenance tracking, individual and global expert quantization
Giuseppe5
committed
42b2e4c
cleanup
Giuseppe5
committed
5349301
more clarity
Giuseppe5
committed
e4238cf

Commits on Sep 30, 2026

fix
Giuseppe5
committed
b11ea05