From 9d2f1584fa3ce8a03908f4808447d2706544cf5e Mon Sep 17 00:00:00 2001
From: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>
Date: Fri, 14 Aug 2026 22:07:50 +0800
Subject: [PATCH] Fix MiniMax-H3 Cache-DiT BCG warning (#34848)

---
 .../model_specific_stages/minimax_h3/stages/denoising.py     | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/stages/denoising.py b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/stages/denoising.py
index fdcc37ca2..638bf813c 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/stages/denoising.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/stages/denoising.py
@@ -413,6 +413,11 @@ class MiniMaxH3DenoisingStage(DenoisingStage):
         current_mode = getattr(self, "_minimax_h3_cache_mode", None)
         self._minimax_h3_quality = quality
 
+        if self.server_args.enable_breakable_cuda_graph:
+            if desired_mode is not None:
+                super()._maybe_enable_cache_dit(num_inference_steps, batch)
+            return
+
         # H3 is monolithic-only, and the scheduler executes one worker batch at
         # a time. Combined with `quality` in the dynamic-batch signature, this
         # makes the process-wide hook transition safe at this batch boundary.