auto-round
Recurrent Residual Quantization (RRQ) for LLMs
#2308
Open

Recurrent Residual Quantization (RRQ) for LLMs #2308

luoyu-intel wants to merge 44 commits into main from feat/rrq-phase1
luoyu-intel
luoyu-intel
luoyu-intel luoyu-intel requested a review from wenhuach21 wenhuach21 32 days ago
luoyu-intel
AutoRoundBot
azure-pipelines
hshen14 hshen14 requested a review from xin3he xin3he 31 days ago
hshen14 hshen14 requested a review from lvliang-intel lvliang-intel 31 days ago
hshen14 hshen14 requested a review from yiliu30 yiliu30 31 days ago
hshen14 hshen14 requested a review from lkk12014402 lkk12014402 31 days ago
hshen14 hshen14 requested a review from copilot-pull-request-reviewer copilot-pull-request-reviewer 31 days ago
copilot-pull-request-reviewer
copilot-pull-request-reviewer commented on 2026-09-09
hshen14
hshen14 commented on 2026-09-09
hshen14 hshen14 requested a review from Zhenzhong1 Zhenzhong1 31 days ago
lvliang-intel
lvliang-intel commented on 2026-09-09
lvliang-intel
lvliang-intel commented on 2026-09-09
a32543254
a32543254 approved these changes on 2026-09-09
a32543254
a32543254 commented on 2026-09-09
AutoRoundBot
wenhuach21
wenhuach21 commented on 2026-09-09
wenhuach21
wenhuach21 commented on 2026-09-09
wenhuach21
wenhuach21 commented on 2026-09-09
AutoRoundBot
luoyu-intel
luoyu-intel feat(rrq): Phase 1 - Recurrent Residual Quantization (packed INT2 2+2…
9a1b044a
luoyu-intel feat(rrq): Phase 2 - generate_rrq_residual (incremental from existing…
f3af3cad
luoyu-intel feat(rrq): Phase 3 - add per-plane sign-SGD tuning
7aec193a
luoyu-intel fix(rrq): accumulate optimized residual prefix
5a24a5c0
pre-commit-ci[bot] [pre-commit.ci] auto fixes from pre-commit.com hooks
b40ab6a4
luoyu-intel feat(rrq): match AutoRound per-plane quality + configurable mixed-pre…
9bf7d8d7
luoyu-intel docs(rrq): add Recurrent Residual Quantization entry to README
0957d120
pre-commit-ci[bot] [pre-commit.ci] auto fixes from pre-commit.com hooks
fe5fc97b
luoyu-intel fix(cli): match shared argparse options before falling back to dest i…
f7d228ea
luoyu-intel perf(rrq): replace 4-plane matmul accumulation with single dequant+ma…
fa9a55bb
pre-commit-ci[bot] [pre-commit.ci] auto fixes from pre-commit.com hooks
0a3f5d31
luoyu-intel fix(rrq): declare tunable fields as named __init__ params
503204f7
luoyu-intel fix(export): include unquantized vision layers in AWQ modules_to_not_…
a8f5fc13
luoyu-intel refactor(rrq): minimize common-code changes to preserve main behavior
6b3dc45f
pre-commit-ci[bot] [pre-commit.ci] auto fixes from pre-commit.com hooks
263e841a
luoyu-intel revert(export): restore AWQ _collect_modules_to_not_convert to main l…
d822fb96
luoyu-intel Apply batched suggestions from code review
1af201b1
luoyu-intel revert log info
1aa09505
luoyu-intel fix(rrq): raise ValueError for explicit model_free=True with RRQConfig
cac3743d
luoyu-intel fix(rrq): enforce disable_opt_rtn=True in RRQConfig.check_config()
e0c29eef
luoyu-intel Refactor RRQ exports to sub-package entry points and add eval scripts
87516ac3
luoyu-intel test(rrq): add lm-eval accuracy test for RRQ W2A16 quantization
32699b23
pre-commit-ci[bot] [pre-commit.ci] auto fixes from pre-commit.com hooks
f409ccf4
luoyu-intel luoyu-intel force pushed from 7006c7a7 to f4aae121 25 days ago
luoyu-intel docs(README): note that RRQ does not support production-level deploym…
d44a67c0
luoyu-intel feat(rrq): hide RRQ from user-facing listings, auto-select via format
f673c4e5
pre-commit-ci[bot] [pre-commit.ci] auto fixes from pre-commit.com hooks
30381ba5
luoyu-intel fix(rrq): auto-select W2A16 scheme and force disable_opt_rtn for RRQ
0189dc8b
pre-commit-ci[bot] [pre-commit.ci] auto fixes from pre-commit.com hooks
fb2634c2
luoyu-intel refactor(rrq): remove RRQ from _BUILTIN_ALGORITHM_ORDER
52de3565
luoyu-intel refactor(rrq): use quant_method='auto-round' + packing_format='auto_r…
f4aae121
pre-commit-ci[bot] [pre-commit.ci] auto fixes from pre-commit.com hooks
5883f7d5
luoyu-intel fix(test): add proper category prefix to skip_ci reason
d95c78a2
AutoRoundBot
azure-pipelines
luoyu-intel luoyu-intel requested a review from copilot-pull-request-reviewer copilot-pull-request-reviewer 25 days ago
copilot-pull-request-reviewer
copilot-pull-request-reviewer commented on 2026-09-15
wenhuach21
wenhuach21 commented on 2026-09-15
wenhuach21
wenhuach21 commented on 2026-09-15
wenhuach21
wenhuach21 commented on 2026-09-15
wenhuach21
wenhuach21 commented on 2026-09-15
wenhuach21
wenhuach21 commented on 2026-09-15
wenhuach21
wenhuach21 approved these changes on 2026-09-15
luoyu-intel fix: normalize Conv1D weight orientation in RRQ quantizer
e5eab0c4
luoyu-intel fix: allow save_quantized(format=...) to override constructor-set format
f7fa8c7f
pre-commit-ci[bot] [pre-commit.ci] auto fixes from pre-commit.com hooks
70b09f46
luoyu-intel fix(rrq): implement RTN quantization for layers outside transformer b…
fede3a77
pre-commit-ci[bot] [pre-commit.ci] auto fixes from pre-commit.com hooks
39c2e514
yiliu30
yiliu30 approved these changes on 2026-09-19
luoyu-intel fix(rrq): ensure base+residual are both exported via a single save_rr…
7816b9ec
luoyu-intel Potential fix for pull request finding
4528e9e3
luoyu-intel fix(rrq): default active_bits to max supported by checkpoint
c1007208
pre-commit-ci[bot] [pre-commit.ci] auto fixes from pre-commit.com hooks
34fee7c7
luoyu-intel Potential fix for pull request finding
25b8e055
luoyu-intel fix(rrq): address PR #2308 review comments
03e92452
pre-commit-ci[bot] [pre-commit.ci] auto fixes from pre-commit.com hooks
027407b7

Login to write a write a comment.

Login via GitHub

Assignees
No one assigned
Labels
Milestone