minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[CP V1 Deprecation 3/5] Remove generic prefill CP v1 runtime (#36228)

Browse Source
This commit is contained in:
Baizhou Zhang
2026-09-06 21:53:53 -07:00
committed by GitHub
parent aaf9a95763
commit b6c31b155c
34 changed files with 400 additions and 741 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/unit/server_args/test_server_args.py
+34 -50
View File
@@ -1075,28 +1075,22 @@ class TestContextParallelServerArgs(CustomTestCase):
with self.assertRaisesRegex(ValueError, "DeepSeek V3.2.*interleave"):
handle_context_parallelism(server_args)
@override_platform(is_hip=False, is_npu=False)
def test_generic_canonical_cp_mirrors_to_transitional_runtime_fields(self):
@override_platform(is_hip=False, is_npu=False, is_musa=False)
def test_generic_canonical_cp_does_not_enable_platform_runtime_fields(self):
cases = (
(
"zigzag_mla_or_gqa",
"zigzag",
"fa3",
True,
False,
"in-seq-split",
),
(
"interleave_dsa",
"interleave",
"dsa",
False,
True,
"round-robin-split",
),
)
for name, strategy, backend, expect_generic, expect_dsa, mode in cases:
for name, strategy, backend in cases:
with self.subTest(name=name):
server_args = self._new_cp_args(
enable_prefill_cp=True,
@@ -1105,6 +1099,7 @@ class TestContextParallelServerArgs(CustomTestCase):
)
handle_platform_cp_compatibility(server_args)
handle_legacy_cp_runtime_compatibility(server_args)
self.assertFalse(
resolution_result(server_args, "enable_prefill_context_parallel")
@@ -1115,32 +1110,13 @@ class TestContextParallelServerArgs(CustomTestCase):
)
)
handle_legacy_cp_runtime_compatibility(server_args)
self.assertEqual(
resolution_result(server_args, "enable_prefill_context_parallel"),
expect_generic,
)
self.assertEqual(
resolution_result(
server_args, "enable_dsa_prefill_context_parallel"
),
expect_dsa,
)
self.assertEqual(
resolution_result(server_args, "dsa_prefill_cp_mode"), mode
)
self.assertEqual(
resolution_result(server_args, "prefill_cp_mode"), mode
)
@override_platform(is_hip=False, is_npu=False)
@override_platform(is_hip=False, is_npu=False, is_musa=False)
def test_non_platform_legacy_prefill_cp_is_rejected(self):
server_args = ServerArgs(
model_path="instance://127.0.0.1:8000/dummy",
enable_prefill_context_parallel=True,
)
with self.assertRaisesRegex(ValueError, "HIP or Ascend NPU"):
with self.assertRaisesRegex(ValueError, "protected HIP, Ascend NPU, or MUSA"):
handle_platform_cp_compatibility(server_args)
def test_generic_v1_cp_options_are_not_public_cli(self):
@@ -1172,28 +1148,36 @@ class TestContextParallelServerArgs(CustomTestCase):
resolution_result(args, "dsa_prefill_cp_mode"), "round-robin-split"
)
def test_canonical_interleave_cp_mirrors_to_dsa_runtime_aliases(self):
server_args = self._new_cp_args(
enable_prefill_cp=True,
cp_strategy="interleave",
attention_backend="dsa",
)
def test_platform_interleave_cp_mirrors_to_dsa_runtime_aliases(self):
for platform in ("is_hip", "is_npu", "is_musa"):
facts = dict(is_hip=False, is_npu=False, is_musa=False)
facts[platform] = True
with self.subTest(platform=platform), override_platform(**facts):
server_args = self._new_cp_args(
enable_prefill_cp=True,
cp_strategy="interleave",
attention_backend="dsa",
)
handle_legacy_cp_runtime_compatibility(server_args)
handle_context_parallelism(server_args)
handle_legacy_cp_runtime_compatibility(server_args)
handle_context_parallelism(server_args)
self.assertTrue(
resolution_result(server_args, "enable_dsa_prefill_context_parallel")
)
self.assertFalse(
resolution_result(server_args, "enable_prefill_context_parallel")
)
self.assertEqual(
resolution_result(server_args, "dsa_prefill_cp_mode"), "round-robin-split"
)
self.assertEqual(
resolution_result(server_args, "prefill_cp_mode"), "round-robin-split"
)
self.assertTrue(
resolution_result(
server_args, "enable_dsa_prefill_context_parallel"
)
)
self.assertFalse(
resolution_result(server_args, "enable_prefill_context_parallel")
)
self.assertEqual(
resolution_result(server_args, "dsa_prefill_cp_mode"),
"round-robin-split",
)
self.assertEqual(
resolution_result(server_args, "prefill_cp_mode"),
"round-robin-split",
)
def test_context_parallel_handler_initializes_cp_strategy(self):
server_args = self._new_cp_args(