auto-round
fix(rrq): address PR #2308 review feedback
#2626
Open

fix(rrq): address PR #2308 review feedback #2626

luoyu-intel wants to merge 29 commits into main from fix/rrq-review-feedback
luoyu-intel
luoyu-intel fix(rrq): address PR #2308 review feedback
a695d84c
luoyu-intel luoyu-intel requested a review from copilot-pull-request-reviewer copilot-pull-request-reviewer 7 days ago
azure-pipelines
copilot-pull-request-reviewer
copilot-pull-request-reviewer commented on 2026-11-13
AutoRoundBot
azure-pipelines
luoyu-intel fix(rrq): load base in bf16 and avoid OOM on large models
c9c529c1
luoyu-intel docs(rrq): add Qwen3.8-27B 2+2+2+2 accuracy and cost report
448293e6
pre-commit-ci[bot] [pre-commit.ci] auto fixes from pre-commit.com hooks
7eaa1401
luoyu-intel fix(rrq): address review comments on fallback scope and export buffers
34708ed9
luoyu-intel test(rrq): cover incremental residual generation from a real 2-bit base
ba48d194
pre-commit-ci[bot] [pre-commit.ci] auto fixes from pre-commit.com hooks
ac5b7682
AutoRoundBot
luoyu-intel perf(rrq): add materialize_rrq and budget-aware weight caching
b61d3aec
luoyu-intel docs(rrq): document materialize_rrq and the inference speed trade-off
6dcd9847
AutoRoundBot
luoyu-intel perf(rrq): parallelize and GPU-accelerate incremental residual genera…
2aa8e496
luoyu-intel docs(rrq): document parallel multi-GPU residual generation
be6a7ddd
luoyu-intel docs(rrq): compare RTN vs AutoRound base plane quality
4b4b4b0a
luoyu-intel docs(rrq): add per-task deltas and mixed-precision cross-check
1a89c11a
luoyu-intel feat(rrq): add phase-dependent precision (6-bit prefill, 2-bit decode)
d19dd2fe
luoyu-intel docs(rrq): report GSM8K accuracy for the prefill/decode split
d215da7d
luoyu-intel perf(rrq): add fused INT2 GEMV kernel for single-token decode
228d88b7
luoyu-intel docs(rrq): document the fused decode kernel
234283f8
luoyu-intel docs(rrq): correct the fused kernel numerical characterisation
970a2ac8
luoyu-intel perf(rrq): batch the fused decode kernel
80089a18
luoyu-intel docs(rrq): report corrected GSM8K accuracy
ffb2ebca
luoyu-intel docs(rrq): align CN GSM8K delta with the measured value
87740a85
luoyu-intel docs(rrq): report prefill6/decode4 GSM8K accuracy and decode throughput
7c77dbd5
luoyu-intel docs(rrq): add the uniform 4-bit control and retract the p6d4 recomme…
8c01d4cf
luoyu-intel feat(rrq): let set_rrq_random_residual keep a uniform prefill
575f90d9
luoyu-intel feat(rrq): add only= to restrict mixed residual spending to one modul…
a83517e9
luoyu-intel docs(rrq): report MATH-500 accuracy and the bf16 lossless reference
93c896ea
luoyu-intel docs(rrq): add the commands that produce configurations A and B
4560d165
luoyu-intel docs(rrq): measure the tuned-base gain on reasoning benchmarks
140f4e2e
luoyu-intel fix(rrq): accept .bin-only raw models in generate_rrq_residual
b0844f95

Login to write a write a comment.

Login via GitHub

Assignees
No one assigned
Labels
Milestone