[Feat] Intranode Dispatch&Combine Kernel #522

Merged

coderfeli merged 26 commits into

mainfrom

yanbo/dispatch_combine

Jun 10, 2026

Commits on Jun 2, 2026

refactor(dispatch_combine): minimize FlyDSL core deltas, inline helpers

yanboshao
authored and
yashao@amd.com
committed
style: satisfy pre-checks (black + ruff)

yanboshao
authored and
yashao@amd.com
committed
ci(dispatch_combine): restrict tests to 8-GPU runners

yanboshao
authored and
yashao@amd.com
committed
fix(ci): restore arith.py style-compliant import layout

yanboshao
authored and
yashao@amd.com
committed
feat(dispatch_combine): in-kernel grid barrier + dynamic arch/LDS check

yanboshao
authored and
yashao@amd.com
committed
feat(dispatch_combine): verify_self with mori-parity byte verify

yanboshao
authored and
yashao@amd.com
committed
chore(dispatch_combine): reserve combine_no_stage1 and refresh CI sweep note

yanboshao
authored and
yashao@amd.com
committed
refactor(dispatch_combine): align intranode op with mori parity surface

yanboshao
authored and
yashao@amd.com
committed
style: satisfy black + ruff pre-checks

yanboshao
authored and
yashao@amd.com
committed
refactor(dispatch_combine): drop frozen launch-geometry params + slim op-wrapper comments

yanboshao
authored and
yashao@amd.com
committed
fix(dispatch_combine): wpb=16 wide-path no-op + verify_self desync hang + f32 bitcast JIT

yanboshao
authored and
yashao@amd.com
committed
style: satisfy black on multi-line skip-print

yanboshao
authored and
yashao@amd.com
committed
rm dead code

yanboshao
authored and
yashao@amd.com
committed
fix(dispatch_combine): cache-key combine_launch dtype via str(...) instead of editing flydsl JIT core

yanboshao
authored and
yashao@amd.com
committed
feat(dispatch_combine): mori-parity P2P-read with launch-time dtype JIT cache

yanboshao
authored and
yashao@amd.com
committed
perf(dispatch_combine): 396-case mixed-dtype sweep report

yanboshao
authored and
yashao@amd.com
committed
perf(dispatch_combine): rerun 396-case sweep with mori head-to-head

yanboshao
authored and
yashao@amd.com
committed
refactor(dispatch_combine): CI sweep cleanup + test-harness hardening

yanboshao
authored and
yashao@amd.com
committed
docs(ci): align flydsl.yaml comment with zero-copy combine rename

yanboshao
authored and
yashao@amd.com
committed
fix(dispatch_combine): plumb combine_dtype_key through mori-parity profilers + black squeeze

yanboshao
authored and
yashao@amd.com
committed
feat(dispatch_combine): add bs={4,32,8K} fp8->bf16 mixed CI cases + multi-gpu CI rename

yanboshao
authored and
yashao@amd.com
committed
chore(dispatch_combine): drop perf-sweep working artifacts from PR (keep local)

yanboshao
authored and
yashao@amd.com
committed
fix mi325 CI issues

authored and
yashao@amd.com
committed
refactor(dispatch_combine): parameterise Stage 3 unroll + slim comments

yanboshao
authored and
yashao@amd.com
committed
fix(dispatch_combine): scatter all scale i32 slots + restore out_scales caller dtype

yanboshao
authored and
yashao@amd.com
committed

Commits on Jun 8, 2026

refactor(dispatch_combine): address review cleanup (dead code + early checks)
yanboshao
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Feat] Intranode Dispatch&Combine Kernel #522

[Feat] Intranode Dispatch&Combine Kernel #522
coderfeli merged 26 commits into
mainfrom
yanbo/dispatch_combine

Commits on Jun 2, 2026

refactor(dispatch_combine): minimize FlyDSL core deltas, inline helpers

style: satisfy pre-checks (black + ruff)

ci(dispatch_combine): restrict tests to 8-GPU runners

fix(ci): restore arith.py style-compliant import layout

feat(dispatch_combine): in-kernel grid barrier + dynamic arch/LDS check

feat(dispatch_combine): verify_self with mori-parity byte verify

chore(dispatch_combine): reserve combine_no_stage1 and refresh CI sweep note

refactor(dispatch_combine): align intranode op with mori parity surface

style: satisfy black + ruff pre-checks

refactor(dispatch_combine): drop frozen launch-geometry params + slim op-wrapper comments

fix(dispatch_combine): wpb=16 wide-path no-op + verify_self desync hang + f32 bitcast JIT

style: satisfy black on multi-line skip-print

rm dead code

fix(dispatch_combine): cache-key combine_launch dtype via str(...) instead of editing flydsl JIT core

feat(dispatch_combine): mori-parity P2P-read with launch-time dtype JIT cache

perf(dispatch_combine): 396-case mixed-dtype sweep report

perf(dispatch_combine): rerun 396-case sweep with mori head-to-head

refactor(dispatch_combine): CI sweep cleanup + test-harness hardening

docs(ci): align flydsl.yaml comment with zero-copy combine rename

fix(dispatch_combine): plumb combine_dtype_key through mori-parity profilers + black squeeze

feat(dispatch_combine): add bs={4,32,8K} fp8->bf16 mixed CI cases + multi-gpu CI rename

chore(dispatch_combine): drop perf-sweep working artifacts from PR (keep local)

fix mi325 CI issues

refactor(dispatch_combine): parameterise Stage 3 unroll + slim comments

fix(dispatch_combine): scatter all scale i32 slots + restore out_scales caller dtype

Commits on Jun 8, 2026

refactor(dispatch_combine): address review cleanup (dead code + early checks)