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

[AMD] deepseek-v4 clean env vars (#28920)

Browse Source
This commit is contained in:
Thomas Wang
2026-06-22 07:32:43 -07:00
committed by GitHub
parent dd39ef6cec
commit 04d952ea10
10 changed files with 29 additions and 108 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/amd/test_deepseek_v4_flash_fp4.py
+2 -16
View File
@@ -36,26 +36,12 @@ DEEPSEEK_V4_FP4_MODEL_PATH = os.environ.get(
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 = {
"SGLANG_DEFAULT_THINKING": "1",
"SGLANG_DSV4_REASONING_EFFORT": "max",
"SGLANG_OPT_DEEPGEMM_HC_PRENORM": "false",
"SGLANG_USE_AITER": "1",
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_USE_ROCM700A": "0",
"SGLANG_DP_USE_GATHERV": "1",
"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",
"SGLANG_OPT_USE_AITER_INDEXER": "true",
"SGLANG_OPT_USE_TILELANG_INDEXER": "false",
"SGLANG_OPT_USE_TILELANG_MHC_PRE": "false",
"SGLANG_OPT_USE_TILELANG_MHC_POST": "false",
"SGLANG_FP8_PAGED_MQA_LOGITS_TORCH": "1",
"SGLANG_OPT_USE_MULTI_STREAM_OVERLAP": "false",
"SGLANG_ROCM_USE_MULTI_STREAM": "false",
"AITER_BF16_FP8_MOE_BOUND": "0",
}
test/registered/amd/test_deepseek_v4_flash_fp8.py
+2 -16
View File
@@ -36,26 +36,12 @@ DEEPSEEK_V4_FP8_MODEL_PATH = os.environ.get(
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 = {
"SGLANG_DEFAULT_THINKING": "1",
"SGLANG_DSV4_REASONING_EFFORT": "max",
"SGLANG_OPT_DEEPGEMM_HC_PRENORM": "false",
"SGLANG_USE_AITER": "1",
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_USE_ROCM700A": "0",
"SGLANG_DP_USE_GATHERV": "1",
"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",
"SGLANG_OPT_USE_AITER_INDEXER": "true",
"SGLANG_OPT_USE_TILELANG_INDEXER": "false",
"SGLANG_OPT_USE_TILELANG_MHC_PRE": "false",
"SGLANG_OPT_USE_TILELANG_MHC_POST": "false",
"SGLANG_FP8_PAGED_MQA_LOGITS_TORCH": "1",
"SGLANG_OPT_USE_MULTI_STREAM_OVERLAP": "false",
"SGLANG_ROCM_USE_MULTI_STREAM": "false",
"AITER_BF16_FP8_MOE_BOUND": "0",
}
test/registered/amd/test_deepseek_v4_pro_fp4.py
+2 -16
View File
@@ -38,26 +38,12 @@ DEEPSEEK_V4_PRO_FP4_MODEL_PATH = os.environ.get(
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 = {
"SGLANG_DEFAULT_THINKING": "1",
"SGLANG_DSV4_REASONING_EFFORT": "max",
"SGLANG_OPT_DEEPGEMM_HC_PRENORM": "false",
"SGLANG_USE_AITER": "1",
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_USE_ROCM700A": "0",
"SGLANG_DP_USE_GATHERV": "1",
"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",
"SGLANG_OPT_USE_AITER_INDEXER": "true",
"SGLANG_OPT_USE_TILELANG_INDEXER": "false",
"SGLANG_OPT_USE_TILELANG_MHC_PRE": "false",
"SGLANG_OPT_USE_TILELANG_MHC_POST": "false",
"SGLANG_FP8_PAGED_MQA_LOGITS_TORCH": "1",
"SGLANG_OPT_USE_MULTI_STREAM_OVERLAP": "false",
"SGLANG_ROCM_USE_MULTI_STREAM": "false",
"AITER_BF16_FP8_MOE_BOUND": "0",
}
test/registered/amd/test_deepseek_v4_pro_fp4_cp.py
+2 -16
View File
@@ -37,24 +37,10 @@ SERVER_LAUNCH_TIMEOUT = 5400
COMMON_ENV_VARS = {
"SGLANG_DEFAULT_THINKING": "1",
"SGLANG_DSV4_REASONING_EFFORT": "max",
"SGLANG_OPT_DEEPGEMM_HC_PRENORM": "false",
"SGLANG_USE_AITER": "1",
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_USE_ROCM700A": "0",
"SGLANG_DP_USE_GATHERV": "1",
"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",
"SGLANG_OPT_USE_AITER_INDEXER": "true",
"SGLANG_OPT_USE_TILELANG_INDEXER": "false",
"SGLANG_OPT_USE_TILELANG_MHC_PRE": "false",
"SGLANG_OPT_USE_TILELANG_MHC_POST": "false",
"SGLANG_FP8_PAGED_MQA_LOGITS_TORCH": "1",
"SGLANG_OPT_USE_MULTI_STREAM_OVERLAP": "false",
"SGLANG_ROCM_USE_MULTI_STREAM": "false",
"AITER_BF16_FP8_MOE_BOUND": "0",
"SGLANG_EAGER_INPUT_NO_COPY": "false",
}
# FP4 variant (matches test_deepseek_v4_pro_fp4.py; V4-Pro also auto-detects it).
test/registered/amd/test_deepseek_v4_pro_fp4_mtp.py
+2 -17
View File
@@ -38,28 +38,13 @@ FLASHMLA_BACKEND = os.environ.get("SGLANG_HACK_FLASHMLA_BACKEND", "unified_kv_tr
GSM8K_ACCURACY_THRESHOLD = 0.92
AVG_SPEC_ACCEPT_LENGTH_THRESHOLD = 2.8
# Common DeepSeek-V4 env vars (AMD ROCm 7.2 path: AITER indexer + triton attn + ROCm700A).
COMMON_ENV_VARS = {
"SGLANG_DEFAULT_THINKING": "1",
"SGLANG_DSV4_REASONING_EFFORT": "max",
"SGLANG_OPT_DEEPGEMM_HC_PRENORM": "false",
"SGLANG_USE_AITER": "1",
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_USE_ROCM700A": "0",
"SGLANG_DP_USE_GATHERV": "1",
"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",
"SGLANG_OPT_USE_AITER_INDEXER": "true",
"SGLANG_OPT_USE_TILELANG_INDEXER": "false",
"SGLANG_OPT_USE_TILELANG_MHC_PRE": "false",
"SGLANG_OPT_USE_TILELANG_MHC_POST": "false",
"SGLANG_FP8_PAGED_MQA_LOGITS_TORCH": "1",
"SGLANG_OPT_USE_MULTI_STREAM_OVERLAP": "false",
"SGLANG_ROCM_USE_MULTI_STREAM": "false",
"AITER_BF16_FP8_MOE_BOUND": "0",
"SGLANG_EAGER_INPUT_NO_COPY": "1",
}
FP4_ENV_VARS = {
test/registered/amd/test_deepseek_v4_pro_fp8.py
+2 -16
View File
@@ -38,26 +38,12 @@ DEEPSEEK_V4_PRO_FP8_MODEL_PATH = os.environ.get(
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 = {
"SGLANG_DEFAULT_THINKING": "1",
"SGLANG_DSV4_REASONING_EFFORT": "max",
"SGLANG_OPT_DEEPGEMM_HC_PRENORM": "false",
"SGLANG_USE_AITER": "1",
"SGLANG_USE_ROCM700A": "1",
"SGLANG_OPT_USE_FUSED_COMPRESS": "true",
"SGLANG_OPT_USE_FUSED_COMPRESS_TRITON": "true",
"SGLANG_USE_ROCM700A": "0",
"SGLANG_DP_USE_GATHERV": "1",
"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",
"SGLANG_OPT_USE_AITER_INDEXER": "true",
"SGLANG_OPT_USE_TILELANG_INDEXER": "false",
"SGLANG_OPT_USE_TILELANG_MHC_PRE": "false",
"SGLANG_OPT_USE_TILELANG_MHC_POST": "false",
"SGLANG_FP8_PAGED_MQA_LOGITS_TORCH": "1",
"SGLANG_OPT_USE_MULTI_STREAM_OVERLAP": "false",
"SGLANG_ROCM_USE_MULTI_STREAM": "false",
"AITER_BF16_FP8_MOE_BOUND": "0",
}