[DeepSeek V4] Default FP4 checkpoints to FlashInfer MXFP4 MoE (#35919)

2026-08-21 16:48:58 -07:00
parent 7d7ab4b5c6
commit 60ff1e33a5
4 changed files with 106 additions and 74 deletions
@@ -5540,15 +5540,6 @@ class ServerArgs:
            validate_deepseek_v4_cp(self)
            validate_deepseek_v4_mega_moe_token_budget(self)

-            # The SM120 marlin fallback moved to the resolution pipeline
-            # (arg_groups/overrides.py: _deepseek_v4_sm120_moe), invoked here
-            # at its legacy slot.
-            from sglang.srt.arg_groups.overrides import (
-                _deepseek_v4_sm120_moe,
-                run_post_process_pass,
-            )
-
-            run_post_process_pass(self, _deepseek_v4_sm120_moe)
            if is_sm120_supported():
                # SM120 lacks tcgen05/TMEM: disable features that depend on
                # DeepGEMM or require >99KB SMEM (topk_v2).