feat(hygon): add Qwen3-235B-A3B BF16/W8A8 inference support - #1474

Open

qinyiqun wants to merge 16 commits into

InfiniTensor:mainfrom

qinyiqun:bw1100_qw235

Commits on Aug 5, 2026

Add Hygon MoE operator support
qinyiqun
committed
50a3ca1
feat(hygon): add graph-safe MoE inference ops
qinyiqun
committed
4265f84
fix: honor singleton strides in contiguity checks
qinyiqun
committed
e6a89ae
Optimize-Hygon-TP2-graph-allreduce
qinyiqun
committed
ff0dbd6
feat(hygon): optimize Qwen3 MoE graph inference on TP8
qinyiqun
committed
dcd3c0e
fix-hygon-w16a16-marlin-graph-safe
qinyiqun
committed
60cc8c5
perf-hygon-tp8-bf16-graph-allreduce
qinyiqun
committed
6267fb8
feat(hygon): auto-detect LightOP device config
qinyiqun
committed
8ebab1b
feat(hygon): support fp32 cross-entropy output
qinyiqun
committed
f81dec8
perf(hygon): fuse MoE align padding fill
qinyiqun
committed
e3aeb11
perf(hygon): add LightOP paged KV attention path
qinyiqun
committed
26648e4
fix(hygon): stabilize threaded graph inference
qinyiqun
committed
40b11db
feat(hygon): expose optimized inference operators
qinyiqun
committed
a06372a
fix(core): resolve RoPE cache getter rebase conflict
qinyiqun
committed
c5ad69d
fix(graph): remove stale rebase preprocessor guard
qinyiqun
committed
d40aed5
fix-hygon-align-allreduce-graph-capture-hook
qinyiqun
committed
4638ed2