feat(ascend): op-norm-rope group (re-PR after #72 revert) — Swiglu, SiluAndMul, CausalSoftmax, RmsNorm, AddRmsNorm, ApplyRotaryPosEmb, RotaryEmbedding#73

Draft

zhangyue207 wants to merge 26 commits intomasterfrom

feat/ascend-op-norm-rope-v2

Commits on Apr 25, 2026

feat(ascend): op-norm-rope group — Swiglu, SiluAndMul, CausalSoftmax, RmsNorm, AddRmsNorm, ApplyRotaryPosEmb, RotaryEmbedding

zhangyue
authored and
zhangyue
committed
fix(ascend): norm/swiglu destructors + missing add_rms_norm custom kernel registration

zhangyue
authored and
zhangyue
committed
style(ascend): rename `AddRmsNorm` parameters to PyTorch-aligned names

zhangyue
authored and
zhangyue
committed
style(ascend): comment + assert message audit for norm/swiglu/softmax kernels

zhangyue
authored and
zhangyue
committed
test(silu_and_mul): add `implementation_index` parametrize and strided coverage

zhangyue
authored and
zhangyue
committed
refactor(ascend/rotary_embedding): unify RotaryEmbedding and ApplyRotaryPosEmb base ops

zhangyue
authored and
zhangyue
committed
feat(scripts/generate_wrappers): emit `apply_rotary_pos_emb` Python shim

zhangyue
authored and
zhangyue
committed
test(rotary_embedding): merge apply_rotary_pos_emb cases + cover MLA/3D/partial

zhangyue
authored and
zhangyue
committed
fix(generate_wrappers): propagate scalar param defaults to pybind signature

zhangyue
authored and
zhangyue
committed
fix(ascend/rotary_embedding): correct pre-gathered layout + revert sincos executor destroy

zhangyue
authored and
zhangyue
committed
test(rotary_embedding): fix GPT-J reference for partial rotary

zhangyue
authored and
zhangyue
committed
refactor(pr66-simplify): correct `rstd_out` semantic name + clarity fixes

zhangyue
authored and
zhangyue
committed
style(tests): ruff format `test_add_rms_norm.py` after `residual_out` rename

zhangyue
authored and
zhangyue
committed
build(ascend-custom): drive `build.sh` from `pip install` with proper dep tracking

zhangyue
authored and
zhangyue
committed
refactor(data_type): pin `DataType` enum values explicitly

zhangyue
authored and
zhangyue
committed
feat(ascend-custom): add bf16 support + Google-style identifier renames

zhangyue
authored and
zhangyue
committed
refactor(base): align Linear/SiluAndMul/AddRmsNorm/RotaryEmbedding with vLLM

zhangyue
authored and
zhangyue
committed
refactor(base): trim narrative comments and collapse CPU Linear ctors

zhangyue
authored and
zhangyue
committed
fix(pr66-review): address review findings 1-3

zhangyue
authored and
zhangyue
committed
refactor(pr66): drop `apply_rotary_pos_emb` wrapper + tests

zhangyue
authored and
zhangyue
committed
test(rotary_embedding): add `pre_gathered=True` coverage

zhangyue
authored and
zhangyue
committed
chore(pr66): drop unused headers

zhangyue
authored and
zhangyue
committed
style(pr66): sweep assert-message periods + comment backticks

zhangyue
authored and
zhangyue
committed
refactor(pr66): rename AscendC custom kernels to PascalCase + C2 param order

zhangyue
authored and
zhangyue
committed
refactor(pr66): trim commit-narration comments

zhangyue
authored and
zhangyue
committed
fix(build): gate ops `--whole-archive` link on `WITH_ASCEND`

zhangyue
authored and
zhangyue
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(ascend): op-norm-rope group (re-PR after #72 revert) — Swiglu, SiluAndMul, CausalSoftmax, RmsNorm, AddRmsNorm, ApplyRotaryPosEmb, RotaryEmbedding#73

feat(ascend): op-norm-rope group (re-PR after #72 revert) — Swiglu, SiluAndMul, CausalSoftmax, RmsNorm, AddRmsNorm, ApplyRotaryPosEmb, RotaryEmbedding#73
zhangyue207 wants to merge 26 commits intomasterfrom
feat/ascend-op-norm-rope-v2

Commits on Apr 25, 2026

feat(ascend): op-norm-rope group — Swiglu, SiluAndMul, CausalSoftmax, RmsNorm, AddRmsNorm, ApplyRotaryPosEmb, RotaryEmbedding

fix(ascend): norm/swiglu destructors + missing add_rms_norm custom kernel registration

style(ascend): rename `AddRmsNorm` parameters to PyTorch-aligned names

style(ascend): comment + assert message audit for norm/swiglu/softmax kernels

test(silu_and_mul): add `implementation_index` parametrize and strided coverage

refactor(ascend/rotary_embedding): unify RotaryEmbedding and ApplyRotaryPosEmb base ops

feat(scripts/generate_wrappers): emit `apply_rotary_pos_emb` Python shim

test(rotary_embedding): merge apply_rotary_pos_emb cases + cover MLA/3D/partial

fix(generate_wrappers): propagate scalar param defaults to pybind signature

fix(ascend/rotary_embedding): correct pre-gathered layout + revert sincos executor destroy

test(rotary_embedding): fix GPT-J reference for partial rotary

refactor(pr66-simplify): correct `rstd_out` semantic name + clarity fixes

style(tests): ruff format `test_add_rms_norm.py` after `residual_out` rename

build(ascend-custom): drive `build.sh` from `pip install` with proper dep tracking

refactor(data_type): pin `DataType` enum values explicitly

feat(ascend-custom): add bf16 support + Google-style identifier renames

refactor(base): align Linear/SiluAndMul/AddRmsNorm/RotaryEmbedding with vLLM

refactor(base): trim narrative comments and collapse CPU Linear ctors

fix(pr66-review): address review findings 1-3

refactor(pr66): drop `apply_rotary_pos_emb` wrapper + tests

test(rotary_embedding): add `pre_gathered=True` coverage

chore(pr66): drop unused headers

style(pr66): sweep assert-message periods + comment backticks

refactor(pr66): rename AscendC custom kernels to PascalCase + C2 param order

refactor(pr66): trim commit-narration comments

fix(build): gate ops `--whole-archive` link on `WITH_ASCEND`