EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Remove monkey_patch_vllm_dummy_weight_loader (#2064)

Browse Source
This commit is contained in:
Lianmin Zheng
2024-11-17 15:48:12 -08:00
committed by GitHub
parent c1f401fc58
commit 38625e2139
6 changed files with 17 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
test/srt/test_bench_latency.py
View File

@@ -13,7 +13,7 @@ class TestBenchLatency(unittest.TestCase):
output_throughput = run_bench_latency(DEFAULT_MODEL_NAME_FOR_TEST, [])
if is_in_ci():
assert output_throughput > 130, f"{output_throughput=}"
self.assertGreater(output_throughput, 135)
def test_moe_default(self):
output_throughput = run_bench_latency(
@@ -21,7 +21,7 @@ class TestBenchLatency(unittest.TestCase):
)
if is_in_ci():
assert output_throughput > 125, f"{output_throughput=}"
self.assertGreater(output_throughput, 125)
if __name__ == "__main__":

22
test/srt/test_bench_serving.py
View File

@@ -20,7 +20,7 @@ class TestBenchServing(unittest.TestCase):
)
if is_in_ci():
assert res["output_throughput"] > 2830
self.assertGreater(res["output_throughput"], 2850)
def test_offline_throughput_non_stream_small_batch_size(self):
res = run_bench_serving(
@@ -35,7 +35,7 @@ class TestBenchServing(unittest.TestCase):
)
if is_in_ci():
assert res["output_throughput"] > 1000
self.assertGreater(res["output_throughput"], 950)
def test_offline_throughput_without_radix_cache(self):
res = run_bench_serving(
@@ -46,7 +46,7 @@ class TestBenchServing(unittest.TestCase):
)
if is_in_ci():
assert res["output_throughput"] > 2880
self.assertGreater(res["output_throughput"], 2900)
def test_offline_throughput_without_chunked_prefill(self):
res = run_bench_serving(
@@ -57,7 +57,7 @@ class TestBenchServing(unittest.TestCase):
)
if is_in_ci():
assert res["output_throughput"] > 2600
self.assertGreater(res["output_throughput"], 2600)
def test_offline_throughput_with_triton_attention_backend(self):
res = run_bench_serving(
@@ -73,7 +73,7 @@ class TestBenchServing(unittest.TestCase):
)
if is_in_ci():
assert res["output_throughput"] > 2930
self.assertGreater(res["output_throughput"], 2950)
def test_offline_throughput_default_fp8(self):
res = run_bench_serving(
@@ -84,7 +84,7 @@ class TestBenchServing(unittest.TestCase):
)
if is_in_ci():
assert res["output_throughput"] > 3100
self.assertGreater(res["output_throughput"], 3200)
def test_online_latency_default(self):
res = run_bench_serving(
@@ -95,9 +95,9 @@ class TestBenchServing(unittest.TestCase):
)
if is_in_ci():
assert res["median_e2e_latency_ms"] < 12000
assert res["median_ttft_ms"] < 80
assert res["median_itl_ms"] < 12
self.assertLess(res["median_e2e_latency_ms"], 12000)
self.assertLess(res["median_ttft_ms"], 80)
self.assertLess(res["median_itl_ms"], 11)
def test_moe_offline_throughput_default(self):
res = run_bench_serving(
@@ -108,7 +108,7 @@ class TestBenchServing(unittest.TestCase):
)
if is_in_ci():
assert res["output_throughput"] > 1850
self.assertGreater(res["output_throughput"], 1900)
def test_moe_offline_throughput_without_radix_cache(self):
res = run_bench_serving(
@@ -119,7 +119,7 @@ class TestBenchServing(unittest.TestCase):
)
if is_in_ci():
assert res["output_throughput"] > 1950
self.assertGreater(res["output_throughput"], 1950)
if __name__ == "__main__":