[BugFix][Qwen3.8] Support Qwen3.8-MXFP4 DCP by registing Qwen3_5 text-only archs in mamba radix cache whitelists (#35297)

This commit is contained in:
Wang, FangYuan
2026-08-24 22:13:59 -07:00
committed by GitHub
parent 9eee990ce1
commit 01db8a7829
@@ -1680,6 +1680,10 @@ _MAMBA_RADIX_CACHE_ARCHS = frozenset(
"InternS2PreviewForConditionalGeneration",
"InternS2MobiusForConditionalGeneration",
"Qwen3_5ForConditionalGeneration",
# Text-only entries of the same hybrid stack (models/qwen3_5_text.py);
# Qwen3.8-2.4T-A95B ships as Qwen3_5MoeForCausalLM.
"Qwen3_5MoeForCausalLM",
"Qwen3_5ForCausalLM",
"MiniCPMV4_6ForConditionalGeneration",
"NemotronHForCausalLM",
"NemotronHPuzzleForCausalLM",
@@ -1700,6 +1704,10 @@ _MAMBA_EXTRA_BUFFER_ARCHS = frozenset(
"KimiLinearForCausalLM",
"Qwen3_5ForConditionalGeneration",
"Qwen3_5MoeForConditionalGeneration",
# Text-only entries of the same hybrid stack (models/qwen3_5_text.py);
# Qwen3.8-2.4T-A95B ships as Qwen3_5MoeForCausalLM.
"Qwen3_5MoeForCausalLM",
"Qwen3_5ForCausalLM",
"Qwen3NextForCausalLM",
"InternS2PreviewForConditionalGeneration",
"MiniCPMV4_6ForConditionalGeneration",