From 2ac189edc894412cee4dfa50a50ca69685d7752d Mon Sep 17 00:00:00 2001
From: HandH1998 <1335248067@qq.com>
Date: Tue, 11 Mar 2025 01:12:09 +0800
Subject: [PATCH] Amd test fp8 (#4261)

---
 .github/workflows/pr-test-amd.yml             |  1 +
 python/sglang/srt/configs/model_config.py     |  1 +
 .../srt/layers/quantization/fp8_utils.py      |  4 +
 python/sglang/test/test_utils.py              |  4 +
 test/srt/run_suite.py                         |  1 +
 test/srt/test_eval_fp8_accuracy.py            | 73 +++++++++++++++++++
 6 files changed, 84 insertions(+)
 create mode 100644 test/srt/test_eval_fp8_accuracy.py

diff --git a/.github/workflows/pr-test-amd.yml b/.github/workflows/pr-test-amd.yml
index a4aa39ad2..f21eaba0b 100644
--- a/.github/workflows/pr-test-amd.yml
+++ b/.github/workflows/pr-test-amd.yml
@@ -55,6 +55,7 @@ jobs:
         timeout-minutes: 20
         run: |
           docker exec -w /sglang-checkout/test/srt ci_sglang python3 test_eval_accuracy_large.py
+          docker exec -w /sglang-checkout/test/srt ci_sglang python3 test_eval_fp8_accuracy.py
           docker exec -w /sglang-checkout/test/srt ci_sglang python3 models/test_qwen_models.py
 
   mla-test-1-gpu-amd:
diff --git a/python/sglang/srt/configs/model_config.py b/python/sglang/srt/configs/model_config.py
index 489cc6d4b..028a4519a 100644
--- a/python/sglang/srt/configs/model_config.py
+++ b/python/sglang/srt/configs/model_config.py
@@ -237,6 +237,7 @@ class ModelConfig:
             "compressed_tensors",
             "compressed-tensors",
             "fbgemm_fp8",
+            "w8a8_fp8",
         ]
         optimized_quantization_methods = [
             "fp8",
diff --git a/python/sglang/srt/layers/quantization/fp8_utils.py b/python/sglang/srt/layers/quantization/fp8_utils.py
index e53b971be..b7b2f2b89 100644
--- a/python/sglang/srt/layers/quantization/fp8_utils.py
+++ b/python/sglang/srt/layers/quantization/fp8_utils.py
@@ -32,6 +32,10 @@ if _is_cuda:
     else:
         from sgl_kernel import fp8_scaled_mm
 
+# Input scaling factors are no longer optional in _scaled_mm starting
+# from pytorch 2.5. Allocating a dummy tensor to pass as input_scale
+TORCH_DEVICE_IDENTITY = torch.ones(1, dtype=torch.float32)
+
 
 def cutlass_fp8_supported():
     if not _is_cuda:
diff --git a/python/sglang/test/test_utils.py b/python/sglang/test/test_utils.py
index d9964cd50..3fe5d2a0f 100644
--- a/python/sglang/test/test_utils.py
+++ b/python/sglang/test/test_utils.py
@@ -28,6 +28,10 @@ from sglang.test.run_eval import run_eval
 from sglang.utils import get_exception_traceback
 
 DEFAULT_FP8_MODEL_NAME_FOR_TEST = "neuralmagic/Meta-Llama-3.1-8B-FP8"
+DEFAULT_FP8_MODEL_NAME_FOR_ACCURACY_TEST = "neuralmagic/Meta-Llama-3-8B-Instruct-FP8"
+DEFAULT_FP8_MODEL_NAME_FOR_DYNAMIC_QUANT_ACCURACY_TEST = (
+    "neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8-dynamic"
+)
 DEFAULT_MODEL_NAME_FOR_TEST = "meta-llama/Llama-3.1-8B-Instruct"
 DEFAULT_SMALL_MODEL_NAME_FOR_TEST = "meta-llama/Llama-3.2-1B-Instruct"
 DEFAULT_MOE_MODEL_NAME_FOR_TEST = "mistralai/Mixtral-8x7B-Instruct-v0.1"
diff --git a/test/srt/run_suite.py b/test/srt/run_suite.py
index 413cef32c..d387fd710 100644
--- a/test/srt/run_suite.py
+++ b/test/srt/run_suite.py
@@ -69,6 +69,7 @@ suites = {
         TestFile("test_vision_llm.py", 18.4),
         TestFile("test_vision_openai_server.py", 344),
         TestFile("test_w8a8_quantization.py", 46),
+        TestFile("test_eval_fp8_accuracy.py", 172),
     ],
     "nightly": [
         TestFile("test_nightly_gsm8k_eval.py"),
diff --git a/test/srt/test_eval_fp8_accuracy.py b/test/srt/test_eval_fp8_accuracy.py
new file mode 100644
index 000000000..8d3c5c00c
--- /dev/null
+++ b/test/srt/test_eval_fp8_accuracy.py
@@ -0,0 +1,73 @@
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.run_eval import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_FP8_MODEL_NAME_FOR_ACCURACY_TEST,
+    DEFAULT_FP8_MODEL_NAME_FOR_DYNAMIC_QUANT_ACCURACY_TEST,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
+
+
+class TestEvalFP8Accuracy(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEFAULT_FP8_MODEL_NAME_FOR_ACCURACY_TEST
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model, cls.base_url, timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_mmlu(self):
+        args = SimpleNamespace(
+            base_url=self.base_url,
+            model=self.model,
+            eval_name="mmlu",
+            num_examples=64,
+            num_threads=32,
+            temperature=0.1,
+        )
+
+        metrics = run_eval(args)
+        self.assertGreaterEqual(metrics["score"], 0.62)
+
+
+class TestEvalFP8DynamicQuantAccuracy(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEFAULT_FP8_MODEL_NAME_FOR_DYNAMIC_QUANT_ACCURACY_TEST
+        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=["--quantization", "w8a8_fp8"],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_mmlu(self):
+        args = SimpleNamespace(
+            base_url=self.base_url,
+            model=self.model,
+            eval_name="mmlu",
+            num_examples=64,
+            num_threads=32,
+            temperature=0.1,
+        )
+
+        metrics = run_eval(args)
+        self.assertGreaterEqual(metrics["score"], 0.70)
+
+
+if __name__ == "__main__":
+    unittest.main()