Add the 8-gpu Inkling consistency test (#34009)

test/registered/8-gpu-models/test_inkling_nvfp4_nightly.py

@@ -2,6 +2,7 @@ import unittest
 
 from sglang.srt.environ import envs
 from sglang.test.accuracy_test_runner import AccuracyTestParams
+from sglang.test.cache_consistency_jitter import get_jitter_engine, run_jitter_test
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.performance_test_runner import PerformanceTestParams
 from sglang.test.run_combined_tests import run_combined_tests
@@ -12,6 +13,7 @@ from sglang.test.test_utils import ModelLaunchSettings, is_blackwell_system
 register_cuda_ci(est_time=3600, suite="nightly-8-gpu-common", nightly=True)
 
 INKLING_NVFP4_MODEL = "thinkingmachines/Inkling-NVFP4"
+INKLING_SMALL_NVFP4_MODEL = "thinkingmachines/Inkling-Small-NVFP4"
 
 # Verified Blackwell NVFP4 recipe; see the SGLang cookbook Inkling page.
 NVFP4_ARGS = [
@@ -75,5 +77,35 @@ class TestInklingNVFP4Nightly(unittest.TestCase):
             )
 
 
+class TestInklingSmallCacheConsistencyNightly(unittest.TestCase):
+    """Bitwise version of the per-commit check in
+    ``test/registered/models_e2e/test_inkling.py``, on the real checkpoint and
+    at a batch shape the tiny checkpoint never reaches."""
+
+    @unittest.skipIf(not is_blackwell_system(), "NVFP4 requires Blackwell")
+    def test_scored_contexts_are_bitwise_identical(self):
+        with envs.SGLANG_ENABLE_UNIFIED_RADIX_TREE.override(1):
+            with get_jitter_engine(
+                model_path=INKLING_SMALL_NVFP4_MODEL,
+                tp_size=8,
+                trust_remote_code=True,
+                quantization="modelopt_fp4",
+                attention_backend="fa4",
+                page_size=128,
+                fp4_gemm_runner_backend="flashinfer_trtllm",
+                moe_runner_backend="flashinfer_trtllm_routed",
+                enable_torch_symm_mem=True,
+                mamba_radix_cache_strategy="extra_buffer",
+                swa_full_tokens_ratio=0.1,
+                mamba_full_memory_ratio=0.1,
+                mem_fraction_static=0.6,
+                enable_deterministic_inference=True,
+                # The harness pins a tight pool; this checkpoint needs room for
+                # the 2048-4096 token prefixes the default workload draws.
+                max_total_tokens=131_072,
+            ) as engine:
+                run_jitter_test(engine)
+
+
 if __name__ == "__main__":
     unittest.main()