minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[AMD] Test DeepSeek V4 FlashMLA backend variants nightly (#28290)

Browse Source
This commit is contained in:
Bingxu Chen
2026-06-16 01:03:26 -07:00
committed by GitHub
parent c5b9106c1a
commit 0fc2bc4a8b
5 changed files with 48 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/amd/test_deepseek_v4_flash_fp4.py
+4 -3
View File
@@ -34,6 +34,7 @@ DEEPSEEK_V4_FP4_MODEL_PATH = os.environ.get(
"DEEPSEEK_V4_FP4_MODEL_PATH", "deepseek-ai/DeepSeek-V4-Flash"
)
SERVER_LAUNCH_TIMEOUT = 3600
FLASHMLA_BACKEND = os.environ.get("SGLANG_HACK_FLASHMLA_BACKEND", "unified_kv_triton")
# Common DeepSeek-V4 env vars (AMD ROCm 7.2 path: AITER indexer + triton attn + ROCm700A).
COMMON_ENV_VARS = {
@@ -44,7 +45,7 @@ COMMON_ENV_VARS = {
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_HACK_FLASHMLA_BACKEND": "unified_kv_triton",
"SGLANG_HACK_FLASHMLA_BACKEND": FLASHMLA_BACKEND,
"SGLANG_OPT_FP8_WO_A_GEMM": "false",
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "false",
"SGLANG_OPT_USE_TOPK_V2": "false",
@@ -126,7 +127,7 @@ class TestDeepseekV4Fp4(CustomTestCase):
if is_in_ci():
write_github_step_summary(
f"### test_gsm8k (deepseek-v4-flash-fp4)\n"
f"### test_gsm8k (deepseek-v4-flash-fp4, {FLASHMLA_BACKEND})\n"
f'{metrics["accuracy"]=:.3f}\n'
)
self.assertGreater(metrics["accuracy"], 0.91)
@@ -185,7 +186,7 @@ class TestDeepseekV4Fp4(CustomTestCase):
report_results = results_data
summary_lines = [
"### test_perf_8k_1k (deepseek-v4-flash-fp4)",
f"### test_perf_8k_1k (deepseek-v4-flash-fp4, {FLASHMLA_BACKEND})",
"input_len=8192 output_len=1024",
"",
"| batch size | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |",
test/registered/amd/test_deepseek_v4_flash_fp8.py
+4 -3
View File
@@ -34,6 +34,7 @@ DEEPSEEK_V4_FP8_MODEL_PATH = os.environ.get(
"DEEPSEEK_V4_FP8_MODEL_PATH", "sgl-project/DeepSeek-V4-Flash-FP8"
)
SERVER_LAUNCH_TIMEOUT = 3600
FLASHMLA_BACKEND = os.environ.get("SGLANG_HACK_FLASHMLA_BACKEND", "unified_kv_triton")
# Common DeepSeek-V4 env vars (AMD ROCm 7.2 path: AITER indexer + triton attn + ROCm700A).
COMMON_ENV_VARS = {
@@ -44,7 +45,7 @@ COMMON_ENV_VARS = {
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_HACK_FLASHMLA_BACKEND": "unified_kv_triton",
"SGLANG_HACK_FLASHMLA_BACKEND": FLASHMLA_BACKEND,
"SGLANG_OPT_FP8_WO_A_GEMM": "false",
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "false",
"SGLANG_OPT_USE_TOPK_V2": "false",
@@ -126,7 +127,7 @@ class TestDeepseekV4Fp8(CustomTestCase):
if is_in_ci():
write_github_step_summary(
f"### test_gsm8k (deepseek-v4-flash-fp8)\n"
f"### test_gsm8k (deepseek-v4-flash-fp8, {FLASHMLA_BACKEND})\n"
f'{metrics["accuracy"]=:.3f}\n'
)
self.assertGreater(metrics["accuracy"], 0.91)
@@ -185,7 +186,7 @@ class TestDeepseekV4Fp8(CustomTestCase):
report_results = results_data
summary_lines = [
"### test_perf_8k_1k (deepseek-v4-flash-fp8)",
f"### test_perf_8k_1k (deepseek-v4-flash-fp8, {FLASHMLA_BACKEND})",
"input_len=8192 output_len=1024",
"",
"| batch size | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |",
test/registered/amd/test_deepseek_v4_pro_fp4.py
+4 -3
View File
@@ -36,6 +36,7 @@ DEEPSEEK_V4_PRO_FP4_MODEL_PATH = os.environ.get(
)
# Pro is 1.6T; weight load + warmup is much longer than Flash 285B.
SERVER_LAUNCH_TIMEOUT = 5400
FLASHMLA_BACKEND = os.environ.get("SGLANG_HACK_FLASHMLA_BACKEND", "unified_kv_triton")
# Common DeepSeek-V4 env vars (AMD ROCm 7.2 path: AITER indexer + triton attn + ROCm700A).
COMMON_ENV_VARS = {
@@ -46,7 +47,7 @@ COMMON_ENV_VARS = {
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_HACK_FLASHMLA_BACKEND": "unified_kv_triton",
"SGLANG_HACK_FLASHMLA_BACKEND": FLASHMLA_BACKEND,
"SGLANG_OPT_FP8_WO_A_GEMM": "false",
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "false",
"SGLANG_OPT_USE_TOPK_V2": "false",
@@ -128,7 +129,7 @@ class TestDeepseekV4ProFp4(CustomTestCase):
if is_in_ci():
write_github_step_summary(
f"### test_gsm8k (deepseek-v4-pro-fp4)\n"
f"### test_gsm8k (deepseek-v4-pro-fp4, {FLASHMLA_BACKEND})\n"
f'{metrics["accuracy"]=:.3f}\n'
)
self.assertGreater(metrics["accuracy"], 0.92)
@@ -187,7 +188,7 @@ class TestDeepseekV4ProFp4(CustomTestCase):
report_results = results_data
summary_lines = [
"### test_perf_8k_1k (deepseek-v4-pro-fp4)",
f"### test_perf_8k_1k (deepseek-v4-pro-fp4, {FLASHMLA_BACKEND})",
"input_len=8192 output_len=1024",
"",
"| batch size | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |",
test/registered/amd/test_deepseek_v4_pro_fp8.py
+4 -3
View File
@@ -36,6 +36,7 @@ DEEPSEEK_V4_PRO_FP8_MODEL_PATH = os.environ.get(
)
# Pro is 1.6T; weight load + warmup is much longer than Flash 285B.
SERVER_LAUNCH_TIMEOUT = 5400
FLASHMLA_BACKEND = os.environ.get("SGLANG_HACK_FLASHMLA_BACKEND", "unified_kv_triton")
# Common DeepSeek-V4 env vars (AMD ROCm 7.2 path: AITER indexer + triton attn + ROCm700A).
COMMON_ENV_VARS = {
@@ -46,7 +47,7 @@ COMMON_ENV_VARS = {
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_HACK_FLASHMLA_BACKEND": "unified_kv_triton",
"SGLANG_HACK_FLASHMLA_BACKEND": FLASHMLA_BACKEND,
"SGLANG_OPT_FP8_WO_A_GEMM": "false",
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "false",
"SGLANG_OPT_USE_TOPK_V2": "false",
@@ -128,7 +129,7 @@ class TestDeepseekV4ProFp8(CustomTestCase):
if is_in_ci():
write_github_step_summary(
f"### test_gsm8k (deepseek-v4-pro-fp8)\n"
f"### test_gsm8k (deepseek-v4-pro-fp8, {FLASHMLA_BACKEND})\n"
f'{metrics["accuracy"]=:.3f}\n'
)
self.assertGreater(metrics["accuracy"], 0.91)
@@ -187,7 +188,7 @@ class TestDeepseekV4ProFp8(CustomTestCase):
report_results = results_data
summary_lines = [
"### test_perf_8k_1k (deepseek-v4-pro-fp8)",
f"### test_perf_8k_1k (deepseek-v4-pro-fp8, {FLASHMLA_BACKEND})",
"input_len=8192 output_len=1024",
"",
"| batch size | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |",