Allow prefill breakable CUDA graph for Qwen3.5 via multimodal opt-in allowlist (#30620)

python/sglang/srt/configs/model_config.py

@@ -452,6 +452,9 @@ class ModelConfig:
         self.is_multimodal_piecewise_cuda_graph_supported = enable_multimodal and (
             is_multimodal_piecewise_cuda_graph_supported(self.hf_config.architectures)
         )
+        self.is_multimodal_breakable_cuda_graph_supported = enable_multimodal and (
+            is_multimodal_breakable_cuda_graph_supported(self.hf_config.architectures)
+        )
         self.dtype = _get_and_verify_dtype(self.hf_text_config, dtype)
 
         # Derive context length and model shapes
@@ -1758,6 +1761,13 @@ multimodal_piecewise_cuda_graph_supported_model_archs = [
     "MiniMaxM3SparseForConditionalGeneration",
 ]
 
+# Multimodal archs whose LM prefill is validated under breakable CUDA graph;
+# embed-carrying batches are rejected at replay (can_run_graph) and run eager.
+multimodal_breakable_cuda_graph_supported_model_archs = [
+    "Qwen3_5ForConditionalGeneration",
+    "Qwen3_5MoeForConditionalGeneration",
+]
+
 if external_mm_model_arch := envs.SGLANG_EXTERNAL_MM_MODEL_ARCH.get():
     multimodal_model_archs.append(external_mm_model_arch)
 
@@ -1824,6 +1834,14 @@ def is_multimodal_piecewise_cuda_graph_supported(model_architectures: List[str])
     )
 
 
+def is_multimodal_breakable_cuda_graph_supported(model_architectures: List[str]):
+    """Whether a multimodal arch may keep prefill breakable CUDA graph enabled."""
+    return any(
+        arch in multimodal_breakable_cuda_graph_supported_model_archs
+        for arch in model_architectures
+    )
+
+
 # SequenceClassification models that use CrossEncodingPooler
 _cross_encoding_pooler_archs = [
     "BertForSequenceClassification",

python/sglang/srt/model_executor/runner/prefill_cuda_graph_runner.py

@@ -612,6 +612,12 @@ class PrefillCudaGraphRunner(BaseCudaGraphRunner):
             return False
         if forward_batch.replace_embeds is not None:
             return False
+        # The captured graph embeds from input_ids only; multimodal batches
+        # merge mm embeddings in the outer wrapper, which capture bypasses.
+        if forward_batch.mm_inputs is not None and any(
+            x is not None for x in forward_batch.mm_inputs
+        ):
+            return False
         # tc_piecewise captures with ForwardMode.EXTEND and spec_info=None.
         if forward_batch.forward_mode.is_target_verify():
             return False

python/sglang/srt/server_args.py

@@ -3645,8 +3645,12 @@ class ServerArgs:
             ),
             # DP-attn × BCG capture/replay not yet validated.
             ("DP attention", lambda: self._resolved().enable_dp_attention),
-            # Multimodal prefill replay faults under BCG.
+            # Multimodal prefill replay faults under BCG; allowlisted archs opt back in.
-            ("multimodal model", lambda: self.get_model_config().is_multimodal),
+            (
+                "multimodal model",
+                lambda: self.get_model_config().is_multimodal
+                and not self.get_model_config().is_multimodal_breakable_cuda_graph_supported,
+            ),
         ]
         for name, predicate in rules:
             if predicate():

test/registered/unit/server_args/test_server_args.py

@@ -1197,6 +1197,64 @@ class TestCudaGraphDisaggregationRoles(CustomTestCase):
         self.assertIn((Phase.DECODE, "backend"), args._cuda_graph_config_locked)
 
 
+class TestBreakableCudaGraphMultimodalAllowlist(CustomTestCase):
+    """The BCG "multimodal model" rule exempts archs on the BCG multimodal
+    opt-in allowlist (multimodal_breakable_cuda_graph_supported_model_archs)."""
+
+    def _handled_args(self, *, architectures, is_multimodal, allowlisted):
+        args = ServerArgs(model_path="dummy")
+        args.model_config = SimpleNamespace(
+            hf_config=SimpleNamespace(architectures=architectures),
+            is_piecewise_cuda_graph_disabled_model=False,
+            is_multimodal=is_multimodal,
+            is_multimodal_piecewise_cuda_graph_supported=False,
+            is_multimodal_breakable_cuda_graph_supported=allowlisted,
+        )
+        with (
+            patch("sglang.srt.utils.is_cuda", return_value=True),
+            patch.object(ServerArgs, "use_mla_backend", return_value=False),
+        ):
+            args._handle_cuda_graph_config()
+        return args
+
+    def test_multimodal_arch_disables_prefill_breakable(self):
+        args = self._handled_args(
+            architectures=["Qwen3VLForConditionalGeneration"],
+            is_multimodal=True,
+            allowlisted=False,
+        )
+        self.assertEqual(args.cuda_graph_config.prefill.backend, Backend.DISABLED)
+
+    def test_allowlisted_multimodal_arch_keeps_prefill_breakable(self):
+        args = self._handled_args(
+            architectures=["Qwen3_5MoeForConditionalGeneration"],
+            is_multimodal=True,
+            allowlisted=True,
+        )
+        self.assertEqual(args.cuda_graph_config.prefill.backend, Backend.BREAKABLE)
+
+    def test_allowlist_membership(self):
+        from sglang.srt.configs.model_config import (
+            is_multimodal_breakable_cuda_graph_supported,
+        )
+
+        self.assertTrue(
+            is_multimodal_breakable_cuda_graph_supported(
+                ["Qwen3_5MoeForConditionalGeneration"]
+            )
+        )
+        self.assertTrue(
+            is_multimodal_breakable_cuda_graph_supported(
+                ["Qwen3_5ForConditionalGeneration"]
+            )
+        )
+        self.assertFalse(
+            is_multimodal_breakable_cuda_graph_supported(
+                ["Qwen3VLForConditionalGeneration"]
+            )
+        )
+
+
 class TestCutedslMoeMaxNumTokens(CustomTestCase):
     """The shared CuteDSL MoE per-forward token bound. Fields are set directly
     to exercise the math independently of __post_init__ resolution.