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

[AMD] [CI] Add dsv4 accuracy PR gate to pr-test-amd-rocm720 (#27149)

Browse Source
This commit is contained in:
Bingxu Chen
2026-06-11 23:42:16 -07:00
committed by GitHub
parent 66ab5c9c7a
commit 1cd5cb1220
5 changed files with 165 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/amd/test_deepseek_v4_flash_fp4.py
+5 -1
View File
@@ -44,7 +44,7 @@ COMMON_ENV_VARS = {
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_HACK_FLASHMLA_BACKEND": "triton",
"SGLANG_HACK_FLASHMLA_BACKEND": "unified_kv_triton",
"SGLANG_OPT_FP8_WO_A_GEMM": "false",
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "false",
"SGLANG_OPT_USE_TOPK_V2": "false",
@@ -131,6 +131,10 @@ class TestDeepseekV4Fp4(CustomTestCase):
)
self.assertGreater(metrics["accuracy"], 0.91)
@unittest.skipIf(
os.environ.get("SGLANG_DSV4_ACCURACY_ONLY") == "1",
"SGLANG_DSV4_ACCURACY_ONLY=1: accuracy-only run (skipping perf)",
)
def test_b_perf_8k_1k(self):
json_output = "/tmp/deepseek_v4_flash_fp4_perf.json"
if os.path.exists(json_output):
test/registered/amd/test_deepseek_v4_flash_fp8.py
+5 -1
View File
@@ -44,7 +44,7 @@ COMMON_ENV_VARS = {
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_HACK_FLASHMLA_BACKEND": "triton",
"SGLANG_HACK_FLASHMLA_BACKEND": "unified_kv_triton",
"SGLANG_OPT_FP8_WO_A_GEMM": "false",
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "false",
"SGLANG_OPT_USE_TOPK_V2": "false",
@@ -131,6 +131,10 @@ class TestDeepseekV4Fp8(CustomTestCase):
)
self.assertGreater(metrics["accuracy"], 0.91)
@unittest.skipIf(
os.environ.get("SGLANG_DSV4_ACCURACY_ONLY") == "1",
"SGLANG_DSV4_ACCURACY_ONLY=1: accuracy-only run (skipping perf)",
)
def test_b_perf_8k_1k(self):
json_output = "/tmp/deepseek_v4_flash_fp8_perf.json"
if os.path.exists(json_output):
test/registered/amd/test_deepseek_v4_pro_fp4.py
+5 -1
View File
@@ -46,7 +46,7 @@ COMMON_ENV_VARS = {
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_HACK_FLASHMLA_BACKEND": "triton",
"SGLANG_HACK_FLASHMLA_BACKEND": "unified_kv_triton",
"SGLANG_OPT_FP8_WO_A_GEMM": "false",
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "false",
"SGLANG_OPT_USE_TOPK_V2": "false",
@@ -133,6 +133,10 @@ class TestDeepseekV4ProFp4(CustomTestCase):
)
self.assertGreater(metrics["accuracy"], 0.92)
@unittest.skipIf(
os.environ.get("SGLANG_DSV4_ACCURACY_ONLY") == "1",
"SGLANG_DSV4_ACCURACY_ONLY=1: accuracy-only run (skipping perf)",
)
def test_b_perf_8k_1k(self):
json_output = "/tmp/deepseek_v4_pro_fp4_perf.json"
if os.path.exists(json_output):
test/registered/amd/test_deepseek_v4_pro_fp8.py
+5 -1
View File
@@ -46,7 +46,7 @@ COMMON_ENV_VARS = {
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_HACK_FLASHMLA_BACKEND": "triton",
"SGLANG_HACK_FLASHMLA_BACKEND": "unified_kv_triton",
"SGLANG_OPT_FP8_WO_A_GEMM": "false",
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "false",
"SGLANG_OPT_USE_TOPK_V2": "false",
@@ -133,6 +133,10 @@ class TestDeepseekV4ProFp8(CustomTestCase):
)
self.assertGreater(metrics["accuracy"], 0.91)
@unittest.skipIf(
os.environ.get("SGLANG_DSV4_ACCURACY_ONLY") == "1",
"SGLANG_DSV4_ACCURACY_ONLY=1: accuracy-only run (skipping perf)",
)
def test_b_perf_8k_1k(self):
json_output = "/tmp/deepseek_v4_pro_fp8_perf.json"
if os.path.exists(json_output):