Switch inkling per-commit test to nvfp4 (#32402)

2026-08-09 16:49:49 +08:00
parent 78cd60b4e3
commit c500674124
1 changed files with 103 additions and 0 deletions
@@ -0,0 +1,103 @@
+"""Per-commit accuracy + logprob-consistency test for Inkling-Small-NVFP4.
+
+``test_inkling.py`` boots a shrunken checkpoint, so it can only guard that the
+code paths run -- an undertrained model has no answer quality to gate on. This
+one serves the real NVFP4 checkpoint, which is what catches a weight-load or
+FP4-kernel regression that keeps the server healthy while the outputs go wrong.
+
+gsm8k here is few-shot completion, so it never renders a chat turn and never
+reaches the reasoning path -- it gates the FP4 numerics, not answer quality
+under thinking. tp=4 to match the runner; the checkpoint does not fit on fewer
+cards.
+"""
+
+import os
+import unittest
+from types import SimpleNamespace
+from urllib.parse import urlparse
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+)
+
+register_cuda_ci(est_time=600, stage="extra-b", runner_config="4-gpu-b200")
+
+_MODEL_PATH = os.environ.get(
+    "INKLING_SMALL_TEST_MODEL_PATH", "thinkingmachines/Inkling-Small-NVFP4"
+)
+
+# Measured 0.900 (10-shot, 200 questions, tp=4, invalid=0.000) -- completion,
+# so no thinking. The floor sits ~2.5 sigma of the 200-question sampling noise
+# below that: a real accuracy collapse trips it, the sample spread does not.
+GSM8K_THRESHOLD = 0.85
+
+
+class TestInklingSmallNvfp4(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = _MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                "--tp",
+                "4",
+                "--trust-remote-code",
+                "--quantization",
+                "modelopt_fp4",
+                "--attention-backend",
+                "fa4",
+                "--page-size",
+                "128",
+                "--fp4-gemm-backend",
+                "flashinfer_trtllm",
+                "--moe-runner-backend",
+                "flashinfer_trtllm_routed",
+                "--mamba-radix-cache-strategy",
+                "extra_buffer",
+                "--swa-full-tokens-ratio",
+                "0.1",
+                "--mamba-full-memory-ratio",
+                "0.1",
+                "--mem-fraction-static",
+                "0.85",
+            ],
+            env={**os.environ, "SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1"},
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        if getattr(cls, "process", None) is not None:
+            kill_process_tree(cls.process.pid)
+
+    def test_gsm8k(self):
+        """Answer quality on the real checkpoint: guards the modelopt_fp4 weight
+        load and the FP4 GEMM/MoE kernels against changes that keep the server
+        healthy but corrupt the numerics."""
+        from sglang.test.few_shot_gsm8k import run_eval as run_few_shot_gsm8k
+
+        url = urlparse(self.base_url)
+        metrics = run_few_shot_gsm8k(
+            SimpleNamespace(
+                num_shots=10,
+                data_path=None,
+                num_questions=200,
+                max_new_tokens=16000,
+                parallel=128,
+                host=f"http://{url.hostname}",
+                port=int(url.port),
+            )
+        )
+        print(f"[{self.__class__.__name__}] gsm8k: {metrics['accuracy']:.3f}")
+        self.assertGreaterEqual(metrics["accuracy"], GSM8K_THRESHOLD)
+
+
+if __name__ == "__main__":
+    unittest.main()