[MoE] Decouple Mega MoE from DeepEP backend (#25406)

This commit is contained in:
Yuhao Yang
2026-05-16 00:18:43 -07:00
committed by GitHub
parent 162540e0a8
commit b2c6db0cc4
18 changed files with 41 additions and 29 deletions
-1
View File
@@ -100,7 +100,6 @@ class TestB200FlashCP(DSV4FlashAime25TestBase):
DEEPEP_LARGE_SMS_CONFIG,
]
EXTRA_ENV = {
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "1",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "1024",
}
-1
View File
@@ -116,7 +116,6 @@ class TestB200ProCP(DSV4ProAime25TestBase):
DEEPEP_LARGE_SMS_CONFIG,
]
EXTRA_ENV = {
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "1",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "256",
}
-1
View File
@@ -102,7 +102,6 @@ class TestB300FlashCP(DSV4FlashAime25TestBase):
DEEPEP_LARGE_SMS_CONFIG,
]
EXTRA_ENV = {
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "1",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "1024",
}
-1
View File
@@ -118,7 +118,6 @@ class TestB300ProCP(DSV4ProAime25TestBase):
DEEPEP_LARGE_SMS_CONFIG,
]
EXTRA_ENV = {
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "1",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "256",
}
@@ -24,9 +24,7 @@ DSV4_FLASH_MODEL_PATH = "sgl-project/DeepSeek-V4-Flash-FP8"
DSV4_FLASH_BASE_ENV = {
"SGLANG_ENABLE_SPEC_V2": "1",
"SGLANG_OPT_USE_TOPK_V2": "1",
"SGLANG_DSV4_FP4_EXPERTS": "0",
"SGLANG_JIT_DEEPGEMM_PRECOMPILE": "0",
}
DSV4_FLASH_SERVER_ARGS = [
-2
View File
@@ -41,9 +41,7 @@ HONGLOUMENG_PATH = os.environ.get(
DSV4_PRO_BASE_ENV = {
"SGLANG_ENABLE_SPEC_V2": "1",
"SGLANG_OPT_USE_TOPK_V2": "1",
"SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2": "1",
"SGLANG_JIT_DEEPGEMM_PRECOMPILE": "0",
}
DSV4_PRO_SERVER_ARGS = [
-1
View File
@@ -100,7 +100,6 @@ class TestGB300FlashCP(DSV4FlashAime25TestBase):
DEEPEP_LARGE_SMS_CONFIG,
]
EXTRA_ENV = {
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "1",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "1024",
}
-1
View File
@@ -118,7 +118,6 @@ class TestGB300ProCP(DSV4ProAime25TestBase):
DEEPEP_LARGE_SMS_CONFIG,
]
EXTRA_ENV = {
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "1",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "256",
}
-1
View File
@@ -112,7 +112,6 @@ class TestH200Fp8FlashCP(DSV4FlashAime25TestBase):
]
EXTRA_ENV = {
**H200_FP8_ENV,
"SGLANG_OPT_USE_JIT_INDEXER_METADATA": "1",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "1024",
}
@@ -28,20 +28,12 @@ SERVER_LAUNCH_TIMEOUT = 3600
_W4A8_MEGAMOE_ENV = {
"SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE": "1",
"SGLANG_OPT_FIX_MEGA_MOE_MEMORY": "1",
"SGLANG_OPT_FIX_NEXTN_MEGA_MOE": "1",
"SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK": "4096",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "0",
}
_W4A4_MEGAMOE_ENV = {
"SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE": "1",
"SGLANG_OPT_FIX_MEGA_MOE_MEMORY": "1",
"SGLANG_OPT_FIX_NEXTN_MEGA_MOE": "1",
"SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK": "4096",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "0",
"SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS": "1",
"SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND": "1",
}
@@ -81,7 +73,7 @@ class TestDSV4FlashFP4B200W4A8MegaMoE(ServerSanityMixin, CustomTestCase):
"4",
"--enable-dp-attention",
"--moe-a2a-backend",
"deepep",
"megamoe",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
@@ -122,7 +114,7 @@ class TestDSV4FlashFP4B200W4A4MegaMoE(ServerSanityMixin, CustomTestCase):
"4",
"--enable-dp-attention",
"--moe-a2a-backend",
"deepep",
"megamoe",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",