Fix oom issues with fp8 for llama (#1454)

test/srt/test_bench_serving.py

@@ -1,6 +1,7 @@
 import unittest
 
 from sglang.test.test_utils import (
+    DEFAULT_FP8_MODEL_NAME_FOR_TEST,
     DEFAULT_MODEL_NAME_FOR_TEST,
     DEFAULT_MOE_MODEL_NAME_FOR_TEST,
     is_in_ci,
@@ -59,6 +60,17 @@ class TestBenchServing(unittest.TestCase):
         if is_in_ci():
             assert res["output_throughput"] > 2600
 
+    def test_offline_throughput_default_fp8(self):
+        res = run_bench_serving(
+            model=DEFAULT_FP8_MODEL_NAME_FOR_TEST,
+            num_prompts=500,
+            request_rate=float("inf"),
+            other_server_args=[],
+        )
+
+        if is_in_ci():
+            assert res["output_throughput"] > 3100
+
     def test_online_latency_default(self):
         res = run_bench_serving(
             model=DEFAULT_MODEL_NAME_FOR_TEST,

test/srt/test_chunked_prefill.py

@@ -12,8 +12,10 @@ from sglang.test.test_utils import (
 
 
 class TestChunkedPrefill(unittest.TestCase):
-    def run_mmlu(self, disable_radix_cache, enable_mixed_chunk):
-        other_args = ["--chunked-prefill-size", "32"]
+    def run_mmlu(
+        self, disable_radix_cache, enable_mixed_chunk, chunked_prefill_size=32
+    ):
+        other_args = ["--chunked-prefill-size", str(chunked_prefill_size)]
         if disable_radix_cache:
             other_args += ["--disable-radix-cache"]
 
@@ -55,6 +57,11 @@ class TestChunkedPrefill(unittest.TestCase):
     def test_mixed_chunked_prefill_without_radix_cache(self):
         self.run_mmlu(disable_radix_cache=True, enable_mixed_chunk=True)
 
+    def test_no_chunked_prefill(self):
+        self.run_mmlu(
+            disable_radix_cache=False, enable_mixed_chunk=False, chunked_prefill_size=-1
+        )
+
 
 if __name__ == "__main__":
     unittest.main()