[BCG][6/N] Allow prefill breakable CUDA graph for the Kimi archs (#34245)

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

python/sglang/srt/configs/model_config.py

@@ -1905,11 +1905,17 @@ multimodal_piecewise_cuda_graph_supported_model_archs = [
 
 # Multimodal archs whose LM prefill is validated under breakable CUDA graph;
 # embed-carrying batches are rejected at replay (can_run_graph) and run eager.
+# The Kimi archs are structurally multimodal -- their configs always carry a
+# vision_config, so is_multimodal is True even for text-only serving -- and the
+# generic multimodal rule disabled prefill CG for them despite the LM prefill
+# capturing cleanly.
 multimodal_breakable_cuda_graph_supported_model_archs = [
     "InternS2MobiusForConditionalGeneration",
     "Qwen3_5ForConditionalGeneration",
     "Qwen3_5MoeForConditionalGeneration",
     "MuseGlimmerForConditionalGeneration",
+    "KimiK3ForConditionalGeneration",
+    "KimiK25ForConditionalGeneration",
 ]
 
 if external_mm_model_arch := envs.SGLANG_EXTERNAL_MM_MODEL_ARCH.get():

python/sglang/srt/server_args.py

@@ -4520,13 +4520,14 @@ class ServerArgs:
         if (Phase.PREFILL, "backend") in self._cuda_graph_config_locked:
             return
 
-        # Breakable is the general CUDA default, but it is not compatible with
+        # Breakable is the CUDA default but not multimodal-compatible;
-        # multimodal prefill. Models on this allowlist have had their decoder
+        # piecewise-allowlisted archs run their validated decoder prefill
-        # prefill validated under tc_piecewise; the vision encoder remains
+        # there instead. Archs also on the breakable allowlist keep it --
-        # eager outside the captured LM forward.
+        # this runs first, so piecewise would otherwise silently win.
         if (
             self.cuda_graph_config.prefill.backend == Backend.BREAKABLE
             and self.get_model_config().is_multimodal_piecewise_cuda_graph_supported
+            and not self.get_model_config().is_multimodal_breakable_cuda_graph_supported
             # Keep trtllm_mla on the preferred breakable path, which now serves
             # MLA by falling back to the flashinfer MLA impl for extend.
             and self._resolved_attention_backends()[0] != "trtllm_mla"

test/registered/unit/configs/test_multimodal_piecewise_cuda_graph.py

@@ -72,7 +72,8 @@ class TestMultimodalPiecewiseCudaGraph(CustomTestCase):
     def test_supported_multimodal_model_upgrades_default_to_tc_piecewise(self):
         args = ServerArgs(model_path="dummy")
         args.model_config = SimpleNamespace(
-            is_multimodal_piecewise_cuda_graph_supported=True
+            is_multimodal_piecewise_cuda_graph_supported=True,
+            is_multimodal_breakable_cuda_graph_supported=False,
         )
         args.cuda_graph_config = CudaGraphConfig(
             prefill=PhaseConfig(backend=Backend.BREAKABLE)