chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Revert "[AMD] support two batch overlapping for mori ep #17953" (#19161)

Browse Source
This commit is contained in:
Baizhou Zhang
2026-02-23 01:19:23 +08:00
committed by GitHub
parent 45095bac70
commit 43f83525c0
11 changed files with 157 additions and 725 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
python/sglang/srt/batch_overlap/operations_strategy.py
View File

@@ -7,9 +7,6 @@ from sglang.srt.batch_overlap import operations
from sglang.srt.batch_overlap.operations import Operation
from sglang.srt.layers.moe.token_dispatcher import DeepEPConfig
from sglang.srt.model_executor.forward_batch_info import ForwardMode
from sglang.srt.utils import is_hip
_is_hip = is_hip()
@dataclass
@@ -94,9 +91,7 @@ def _compute_moe_deepseek_layer_operations_strategy_tbo(
def _compute_moe_deepseek_blog_prefill(layer):
device_properties = torch.cuda.get_device_properties(device="cuda")
total_num_sms = device_properties.multi_processor_count
deep_gemm_num_sms = None
if not _is_hip:
deep_gemm_num_sms = total_num_sms - DeepEPConfig.get_instance().num_sms
deep_gemm_num_sms = total_num_sms - DeepEPConfig.get_instance().num_sms
return OperationsStrategy(
deep_gemm_num_sms=deep_gemm_num_sms,
@@ -173,9 +168,7 @@ def _compute_moe_qwen3_layer_operations_strategy_tbo(
def _compute_moe_qwen3_prefill(layer):
device_properties = torch.cuda.get_device_properties(device="cuda")
total_num_sms = device_properties.multi_processor_count
deep_gemm_num_sms = None
if not _is_hip:
deep_gemm_num_sms = total_num_sms - DeepEPConfig.get_instance().num_sms
deep_gemm_num_sms = total_num_sms - DeepEPConfig.get_instance().num_sms
return OperationsStrategy(
deep_gemm_num_sms=deep_gemm_num_sms,

5
python/sglang/srt/batch_overlap/two_batch_overlap.py
View File

@@ -30,7 +30,6 @@ from sglang.srt.layers.moe import (
from sglang.srt.layers.moe.token_dispatcher import (
DeepEPDispatcher,
MooncakeEPDispatcher,
MoriEPDispatcher,
)
from sglang.srt.layers.moe.token_dispatcher.base import BaseDispatcher
from sglang.srt.managers.schedule_batch import ScheduleBatch
@@ -1028,10 +1027,6 @@ class MaybeTboDeepEPDispatcher(BaseDispatcher):
self._inners = [
MooncakeEPDispatcher(**kwargs) for _ in range(num_inner_dispatchers)
]
elif get_moe_a2a_backend().is_mori():
self._inners = [
MoriEPDispatcher(**kwargs) for _ in range(num_inner_dispatchers)
]
def _execute(self, name, tbo_subbatch_index: Optional[int] = None, **kwargs):
return getattr(self._inners[tbo_subbatch_index or 0], name)(**kwargs)