Import flash_mla from sgl-kernel (#12135)

This commit is contained in:
Baizhou Zhang
2025-10-29 23:54:21 -07:00
committed by GitHub
parent fb52d35f63
commit 621dfb8886
8 changed files with 9 additions and 63 deletions
+1 -1
View File
@@ -82,7 +82,7 @@ suites = {
TestFile("test_ebnf_constrained.py", 108),
TestFile("test_eval_fp8_accuracy.py", 303),
TestFile("test_fa3.py", 376),
# TestFile("test_flashmla.py", 352),
TestFile("test_flashmla.py", 352),
TestFile("rotary_embedding/test_mrope.py", 300),
TestFile("test_function_call_parser.py", 10),
TestFile("test_fused_moe.py", 30),
+2 -20
View File
@@ -11,6 +11,7 @@ import torch
from sglang.srt.utils import kill_process_tree
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.send_one import BenchArgs, send_one_prompt
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST_MLA,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -18,7 +19,7 @@ from sglang.test.test_utils import (
CustomTestCase,
is_in_ci,
popen_launch_server,
run_bench_one_batch,
write_github_step_summary,
)
@@ -31,7 +32,6 @@ class TestFlashMLAAttnBackend(unittest.TestCase):
if torch.cuda.is_available() and torch.version.cuda:
other_args.extend(
[
"--enable-torch-compile",
"--cuda-graph-max-bs",
"2",
"--attention-backend",
@@ -65,24 +65,6 @@ class TestFlashMLAAttnBackend(unittest.TestCase):
self.assertGreater(metrics["accuracy"], 0.60)
class TestFlashMLAAttnLatency(unittest.TestCase):
def test_latency(self):
_, output_throughput, _ = run_bench_one_batch(
DEFAULT_MODEL_NAME_FOR_TEST_MLA,
[
"--attention-backend",
"flashmla",
"--enable-torch-compile",
"--cuda-graph-max-bs",
"16",
"--trust-remote-code",
],
)
if is_in_ci():
self.assertGreater(output_throughput, 100)
class TestFlashMLAMTP(CustomTestCase):
@classmethod
def setUpClass(cls):