[WS1][Ascend] Qwen3-8B Dense train-inference closeout for BF16 execution vs FP32 reference - #406

Merged

Flink-ddd merged 10 commits into

RL-Align:testfrom

zhangj1an:feat/ws1-ascend-closeout

Sep 12, 2026

Commits on Sep 12, 2026

[WS1][Ascend] #266 closeout on NPU: ascend_bf16 as a third required profile

zhangj1an
and
claude
committed
a82a52d
fix(ascend): close the WS1 C2/C4/C8 gaps for ascend_bf16

zhangj1an
and
claude
committed
b1a437a
fix(ascend): force the fused-logp kernel path for non-contiguous logits

zhangj1an
and
claude
committed
04d7bb8
fix(ascend): make deterministic attention padding-invariant (keyBegin tiling)

zhangj1an
and
claude
committed
6551777
fix(ascend): shape-invariant RMSNorm rstd via a fixed-order reduction

zhangj1an
and
claude
committed
8ed1693
fix(ascend): fixed FP32 pairwise tree for the RMSNorm backward reduction

zhangj1an
and
claude
committed
d452e95
fix(ascend): layout-invariant attention backward via valid-token compaction

zhangj1an
and
claude
committed
2993cff
fix(ascend): dedicated attention backward kernel for bitwise layout invariance

zhangj1an
and
claude
committed
9039fee
test(ascend): bf16-reference local deviation for the C10 gate on 64 GB HBM

zhangj1an
and
claude
committed
17b18f1
test(ascend): offloaded FP32 reference for the C10 gate on 64 GB HBM

zhangj1an
and
claude
committed
1fe03e7