From 43bdee703ebbac6b9a1b458098cde03aeb262ea4 Mon Sep 17 00:00:00 2001 From: Shu Wang Date: Wed, 4 Mar 2026 18:28:10 -0600 Subject: [PATCH] Fix Fp8 MTP layer a2a backend without EP. (#18515) Co-authored-by: Baizhou Zhang --- python/sglang/srt/server_args.py | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py index ea35de411..9a2208a4c 100644 --- a/python/sglang/srt/server_args.py +++ b/python/sglang/srt/server_args.py @@ -1432,6 +1432,15 @@ class ServerArgs: logger.info( "Use deep_gemm moe runner and deepep a2a backend for bf16 nextn layer in deepseek fp4 checkpoint." ) + # Validate usage of ep + if self.ep_size == 1: + raise ValueError( + "Invalid configuration: 'deep_gemm' speculative MoE runner backend with " + "'deepep' a2a backend requires expert parallelism (ep_size > 1). " + f"Current ep_size is {self.ep_size}. " + "Please set --ep-size > 1 (e.g., --ep-size 8) to use this configuration, " + "or change --speculative-moe-a2a-backend to 'none' if expert parallelism is not available." + ) else: self.speculative_moe_runner_backend = "triton" self.speculative_moe_a2a_backend = "none"