[DeepSeek] Update tests and document for DeepSeek V3.2 NVFP4 checkpoint (#17657)

docs/basic_usage/deepseek_v32.md

@@ -125,6 +125,14 @@ python3 -m sglang.launch_server \
   --reasoning-parser deepseek-v3
 ```
 
+## NVFP4 Checkpoint
+
+To launch deepseek v3.2 [NVFP4 checkpoint](https://huggingface.co/nvidia/DeepSeek-V3.2-NVFP4) on Blackwell devices, the user needs to specify the quantization method as `modelopt_fp4`, and moe runner backend as one of `flashinfer_trtllm`(recommended), `flashinfer_cutlass` and `flashinfer_cutedsl`. Any other usage (parallelism, reasoning parser, ...) is the same as FP8 checkpoint.
+
+An example launching command can be:
+```bash
+python -m sglang.launch_server --model nvidia/DeepSeek-V3.2-NVFP4 --tp 4 --quantization modelopt_fp4 --moe-runner-backend flashinfer_trtllm --tool-call-parser deepseekv32  --reasoning-parser deepseek-v3
+```
 
 ## PD Disaggregation
 

test/srt/run_suite.py

@@ -36,6 +36,7 @@ suites = {
         TestFile("test_fp8_blockwise_gemm.py", 280),
         TestFile("test_gpt_oss_4gpu.py", 300),
         TestFile("test_nvfp4_gemm.py", 360),
+        TestFile("test_deepseek_v32_fp4_4gpu.py", 600),
     ],
     # "per-commit-8-gpu-b200": [
     #     TestFile("test_mistral_large3_basic.py", 275),  # Moved to nightly - large model

test/srt/test_deepseek_v32_fp4_4gpu.py

@@ -0,0 +1,79 @@
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.test_utils import (
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    is_in_ci,
+    popen_launch_server,
+    write_github_step_summary,
+)
+
+FULL_DEEPSEEK_V3_FP4_MODEL_PATH = "nvidia/DeepSeek-V3.2-NVFP4"
+SERVER_LAUNCH_TIMEOUT = 1200
+
+
+class TestDeepseekV32FP4(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = FULL_DEEPSEEK_V3_FP4_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--tp",
+            "4",
+            "--dp",
+            "4",
+            "--enable-dp-attention",
+            "--attention-backend",
+            "nsa",
+            "--moe-runner-backend",
+            "flashinfer_trtllm",
+            "--quantization",
+            "modelopt_fp4",
+            "--kv-cache-dtype",
+            "fp8_e4m3",
+            "--tool-call-parser",
+            "deepseekv32",
+            "--reasoning-parser",
+            "deepseek-v3",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true,"num_threads": 64}',
+        ]
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=SERVER_LAUNCH_TIMEOUT,
+            other_args=other_args,
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_a_gsm8k(
+        self,
+    ):  # Append an "a" to make this test run first (alphabetically) to warm up the server
+        args = SimpleNamespace(
+            num_shots=20,
+            data_path=None,
+            num_questions=1319,
+            parallel=1319,
+            max_new_tokens=512,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"{metrics=}")
+
+        if is_in_ci():
+            write_github_step_summary(
+                f"### test_gsm8k (deepseek-v3-fp4)\n" f'{metrics["accuracy"]=:.3f}\n'
+            )
+
+        self.assertGreater(metrics["accuracy"], 0.935)
+
+
+if __name__ == "__main__":
+    unittest.main()