feat(hygon): add Qwen3-235B-A3B BF16/W8A8 inference support - #532

Open

qinyiqun wants to merge 14 commits into

InfiniTensor:mainfrom

qinyiqun:bw1100_qw235

Commits on Aug 5, 2026

feat(hygon): wire Marlin MoE inference path
qinyiqun
committed
de044c2
docs-qwen3-moe-w8a8-dispatch-findings
qinyiqun
committed
96814d1
Share Hygon Marlin MoE workspace across layers
qinyiqun
committed
b4a70ea
feat(hygon): enable optimized Qwen3 MoE W8A8 graph inference
qinyiqun
committed
b687546
fix-hygon-select-moe-path-from-model
qinyiqun
committed
7777954
feat(hygon): select LightOP configs from device
qinyiqun
committed
c32739e
fix(hygon): slice large W16A16 MoE prefills
qinyiqun
committed
eb7122c
feat(eval): add token-level PPL scoring CLI
qinyiqun
committed
0a72228
feat(bench): support paired multi-length cases
qinyiqun
committed
55abe99
perf(hygon): narrow reusable MoE routing workspace
qinyiqun
committed
4118f3b
perf(hygon): use paged KV cache in FlashAttention graphs
qinyiqun
committed
2d233c9
fix(hygon): select safe paged attention layouts
qinyiqun
committed
0ea0209
refactor(hygon): route optimized inference through InfiniCore
qinyiqun
committed
91a45c8
perf-hygon-shard-lm-head-across-tp-ranks
qinyiqun
committed
6a486f2