[Kimi] Support kimi-k3 (#32541)

Co-authored-by: DarkSharpness <76582120+DarkSharpness@users.noreply.github.com>
Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>
Co-authored-by: Mick <mickjagger19@icloud.com>
Co-authored-by: Yuhao Yang <47235274+yhyang201@users.noreply.github.com>
Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com>
Co-authored-by: Ke Bao <ispobaoke@gmail.com>
Co-authored-by: Baizhou Zhang <sobereddiezhang@gmail.com>
Co-authored-by: Chunan Zeng <zcnrex@gmail.com>
Co-authored-by: Khoa Pham <khoa.pham@radixark.ai>
Co-authored-by: Ziyi Xu <ziyi.xu@radixark.ai>
Co-authored-by: Zijie Xia <37504505+zijiexia@users.noreply.github.com>
Co-authored-by: Yuwei An <ayw.sirius19@gmail.com>
Co-authored-by: zhangxiaohao <1024393531@qq.com>
Co-authored-by: Yangmin Li <yangminl@nvidia.com>
Co-authored-by: Julien Lin <jullin@nvidia.com>
Co-authored-by: Hao Phan <htphan@nvidia.com>
Co-authored-by: Thomas Wang <1am9trash@gmail.com>
Co-authored-by: RolaoDenthu <xinyisong0111@gmail.com>
Co-authored-by: pigeonsoup <32922982+pigeonsoup@users.noreply.github.com>
Co-authored-by: HaiShaw <hixiao@gmail.com>
Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>
Co-authored-by: Pranjal Shankhdhar <pranjal.ssh@gmail.com>
Co-authored-by: Lee Nau <lee.nau@gmail.com>
Co-authored-by: HMING <126185151+Hearum@users.noreply.github.com>
Co-authored-by: elvischenv <219235043+elvischenv@users.noreply.github.com>
Co-authored-by: Byron Hsu <byronhsu1230@gmail.com>
Co-authored-by: Byron Hsu <byron+per@periodiclabs.ai>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: Thomas Wang <thomawan@amd.com>
Co-authored-by: Xinyi Song <86638975+RolaoDenthu@users.noreply.github.com>
Co-authored-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>
Co-authored-by: Cheng Wan <cheng.wan@radixark.ai>
Co-authored-by: BBuf <xiaoyu.zhang@radixark.ai>
Co-authored-by: Hanming Lu <hanminglu@meta.com>
Co-authored-by: Xinyi Song <xinyis10@illinois.edu>

test/registered/dcp/test_kimi_linear_dcp4.py

@@ -1,7 +1,7 @@
 """Four-Blackwell acceptance coverage for Kimi Linear TokenSpeed MLA DCP.
 
 The captured-shape and eager-shape requests deliberately straddle
-``--cuda-graph-max-bs-decode=64``.  This guards both the regular CUDA graph
+``--cuda-graph-max-bs-decode``.  This guards both the regular CUDA graph
 decode path and the full-capacity eager DCP LSE scratch-buffer path.
 """
 
@@ -20,9 +20,10 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=900, stage="base-c", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=240, stage="extra-b", runner_config="4-gpu-b200")
 
 KIMI_LINEAR_MODEL = "moonshotai/Kimi-Linear-48B-A3B-Instruct"
+CUDA_GRAPH_MAX_BS_DECODE = 256
 
 
 def _has_four_blackwell_gpus() -> bool:
@@ -41,12 +42,11 @@ def _has_four_blackwell_gpus() -> bool:
 class TestKimiLinearDCP4(GSM8KMixin, CustomTestCase):
     model = KIMI_LINEAR_MODEL
     base_url = DEFAULT_URL_FOR_TEST
-    gsm8k_score_threshold = 0.90
+    gsm8k_score_threshold = 0.88
     gsm8k_num_examples = 200
     # Keep accuracy evaluation within the captured decode batch sizes so its
-    # score is batch-invariant. The separate smoke test still exercises the
-    # eager path with batch size 65.
-    gsm8k_num_threads = 4
+    # score is batch-invariant.
+    gsm8k_num_threads = 128
     gsm8k_num_shots = 5
 
     @classmethod
@@ -74,7 +74,7 @@ class TestKimiLinearDCP4(GSM8KMixin, CustomTestCase):
                 "--dtype",
                 "bfloat16",
                 "--cuda-graph-max-bs-decode",
-                "64",
+                str(CUDA_GRAPH_MAX_BS_DECODE),
                 "--cuda-graph-backend-prefill",
                 "disabled",
                 "--mem-fraction-static",
@@ -112,12 +112,24 @@ class TestKimiLinearDCP4(GSM8KMixin, CustomTestCase):
             self.assertTrue(output["text"].strip())
             self.assertGreater(output["meta_info"]["completion_tokens"], 0)
 
+    def _effective_max_running_requests(self) -> int:
+        response = requests.get(self.base_url + "/server_info", timeout=30)
+        response.raise_for_status()
+        return min(
+            state["effective_max_running_requests_per_dp"]
+            for state in response.json()["internal_states"]
+        )
+
     def test_decode_cuda_graph_and_eager_batch(self):
-        # Batch two replays a captured shape; batch 65 is above the configured
-        # regular CUDA graph maximum and therefore exercises eager decode.
         self._assert_batch_completes(2)
         self._assert_batch_completes(2)
-        self._assert_batch_completes(65)
+        self.assertGreater(
+            self._effective_max_running_requests(),
+            CUDA_GRAPH_MAX_BS_DECODE,
+            "eager DCP decode is unreachable: concurrency was capped at or "
+            "below the CUDA graph capture ceiling",
+        )
+        self._assert_batch_completes(CUDA_GRAPH_MAX_BS_DECODE + 1)
 
     def test_physical_capacity_sanity(self):
         response = requests.get(self.base_url + "/server_info", timeout=30)

test/registered/dcp/test_kimi_linear_dcp_dspark4.py

@@ -1,3 +1,5 @@
+"""Four-Blackwell Kimi Linear TokenSpeed DCP + DSpark static acceptance test."""
+
 import json
 import socket
 import tempfile
@@ -39,6 +41,7 @@ def _has_four_blackwell_gpus() -> bool:
 
 
 def _write_dummy_qwen3_dspark_draft(root: Path) -> str:
+    """Write a dummy Qwen3 DSpark config with Kimi Linear dimensions."""
     draft_dir = root / "qwen3-dspark-kimi-proxy"
     draft_dir.mkdir()
     config = {