[DeepSeekV3.2] Enable pure TP & Partial DP Attention (#13646)

This commit is contained in:
YAMY
2025-11-30 15:59:23 -08:00
committed by GitHub
parent c72f0756d2
commit decb48965d
7 changed files with 286 additions and 24 deletions
@@ -25,6 +25,9 @@ class TestNightlyDeepseekV32Performance(unittest.TestCase):
"--trust-remote-code",
"--tp",
"8",
"--dp",
"8",
"--enable-dp-attention",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
],
@@ -35,6 +38,9 @@ class TestNightlyDeepseekV32Performance(unittest.TestCase):
"--trust-remote-code",
"--tp",
"8",
"--dp",
"8",
"--enable-dp-attention",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
@@ -51,6 +57,25 @@ class TestNightlyDeepseekV32Performance(unittest.TestCase):
},
{
"name": "nsa",
"other_args": [
"--trust-remote-code",
"--tp",
"8",
"--dp",
"8",
"--enable-dp-attention",
"--attention-backend",
"nsa",
"--nsa-prefill-backend",
"flashmla_sparse",
"--nsa-decode-backend",
"flashmla_kv",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
],
},
{
"name": "pure_tp",
"other_args": [
"--trust-remote-code",
"--tp",
@@ -197,6 +197,63 @@ class TestDeepseekV32NasBackend_fp8kvcache(CustomTestCase):
self.assertGreater(metrics["accuracy"], 0.935)
class TestDeepseekV32NasBackend_pure_tp(CustomTestCase):
"""Test DeepSeek V3.2 with pure TP mode (no DP attention)."""
@classmethod
def setUpClass(cls):
cls.model = DEEPSEEK_V32_MODEL_PATH
cls.base_url = DEFAULT_URL_FOR_TEST
# Pure TP configuration without --dp and --enable-dp-attention
other_args = [
"--trust-remote-code",
"--attention-backend",
"nsa",
"--nsa-prefill-backend",
"flashmla_sparse",
"--nsa-decode-backend",
"flashmla_kv",
"--tp",
"8",
]
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_a_gsm8k(self):
"""Test GSM8K accuracy with pure TP mode."""
args = SimpleNamespace(
num_shots=20,
data_path=None,
num_questions=1400,
parallel=1400,
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():
TEST_RESULTS.append(
{
"variant": "pure_tp",
"prefill_backend": "flashmla_sparse",
"decode_backend": "flashmla_kv",
"kv_cache": "fp16",
"accuracy": metrics["accuracy"],
}
)
self.assertGreater(metrics["accuracy"], 0.935)
def _write_summary_table():
"""Write a markdown table with all test results."""
if not TEST_RESULTS:
+25
View File
@@ -25,6 +25,9 @@ class TestNightlyDeepseekV32Performance(unittest.TestCase):
"--trust-remote-code",
"--tp",
"8",
"--dp",
"8",
"--enable-dp-attention",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
],
@@ -35,6 +38,9 @@ class TestNightlyDeepseekV32Performance(unittest.TestCase):
"--trust-remote-code",
"--tp",
"8",
"--dp",
"8",
"--enable-dp-attention",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
@@ -51,6 +57,25 @@ class TestNightlyDeepseekV32Performance(unittest.TestCase):
},
{
"name": "nsa",
"other_args": [
"--trust-remote-code",
"--tp",
"8",
"--dp",
"8",
"--enable-dp-attention",
"--attention-backend",
"nsa",
"--nsa-prefill-backend",
"flashmla_sparse",
"--nsa-decode-backend",
"flashmla_kv",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
],
},
{
"name": "pure_tp",
"other_args": [
"--trust-remote-code",
"--tp",