[NPU] Add test for --dllm-fdfo (#33634)

2026-08-25 10:58:45 +08:00
parent 1fa32d50e1
commit a0f52eca98
1 changed files with 84 additions and 0 deletions
@@ -0,0 +1,84 @@
+import unittest
+
+from sglang.test.ascend.test_ascend_utils import (
+    LLaDA2_0_MINI_WEIGHTS_PATH,
+    run_bench_serving,
+)
+from sglang.test.ci.ci_register import register_npu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
+
+
+class TestLLaDA2MiniFDFO(CustomTestCase):
+    """The test used the LLaDA2.0-mini model, with --dllm-fdfo, and throughputs improved.
+
+    [Test Category] Diffusion LLM
+    [Test Target] --dllm-fdfo
+    """
+
+    def test_dlla_fdfo_vs_no_fdfo_performance(self):
+        TTFTS = []
+        throughputs = []
+        p99_ttfts = []
+        model = LLaDA2_0_MINI_WEIGHTS_PATH
+        common_args = [
+            [
+                "--trust-remote-code",
+                "--tp-size",
+                2,
+                "--mem-fraction-static",
+                0.9,
+                "--disable-radix-cache",
+                "--max-running-requests",
+                16,
+                "--dllm-algorithm",
+                "LowConfidence",
+            ],
+            [
+                "--trust-remote-code",
+                "--tp-size",
+                2,
+                "--disable-radix-cache",
+                "--mem-fraction-static",
+                0.9,
+                "--max-running-requests",
+                16,
+                "--dllm-algorithm",
+                "LowConfidence",
+                "--no-dllm-fdfo",
+            ],
+        ]
+        for common_arg in common_args:
+            other_args = common_arg + (
+                ["--attention-backend", "ascend", "--disable-cuda-graph"]
+            )
+            res = run_bench_serving(
+                model=model,
+                num_prompts=128,
+                random_input_len=3584,
+                random_output_len=1024,
+                request_rate=float("inf"),
+                max_concurrency=16,
+                other_server_args=other_args,
+            )
+            TTFT = res["mean_ttft_ms"]
+            TTFTS.append(TTFT)
+            throughput = res["total_throughput"]
+            throughputs.append(throughput)
+            p99_ttft = res["p99_ttft_ms"]
+            p99_ttfts.append(p99_ttft)
+
+        # FDFO (First-Done-First-Out) scheduling improves performance by prioritizing
+        # requests that complete their diffusion steps earlier, reducing waiting time.
+        # With FDFO enabled:
+        # - TTFT is lower because earlier-completing requests get priority and start generating tokens sooner
+        # - Throughput is higher due to better hardware utilization from reduced blocking
+        # - P99 TTFT is lower because FDFO prevents a single slow request from blocking others
+        assert float(TTFTS[0]) < float(TTFTS[1])
+        assert float(throughputs[0]) > float(throughputs[1])
+        assert float(p99_ttfts[0]) < float(p99_ttfts[1])
+
+
+if __name__ == "__main__":
+    unittest.main()