[Refactor] Group communicator fusion and CP adapters (#41547)

This commit is contained in:
Cheng Wan
2026-09-28 20:10:50 -07:00
committed by GitHub
parent e49d7fb170
commit 0a8ca85565
12 changed files with 36 additions and 24 deletions
@@ -103,7 +103,7 @@ def _deepseek_family_overrides(server_args: Any, hf_config: Any) -> dict:
)
# Note(kpham-sgl): Keep attn_tp_size == 1 under DSA CP.
# The DSA / MLA CP gather and reduce-scatter
# (communicator_dsa_cp) assume it.
# (the dsa_cp_* helpers in adapters/context_parallel.py) assume it.
attn_cp_size = cfg.tp_size // cfg.dp_size
overrides["attn_cp_size"] = attn_cp_size
logger.warning(
@@ -168,7 +168,7 @@ def _deepseek_family_overrides(server_args: Any, hf_config: Any) -> dict:
)
# FIXME(kpham-sgl): Keep attn_tp_size == 1 under MLA CP.
# The DSA / MLA CP gather and reduce-scatter
# (communicator_dsa_cp) assume it.
# (the dsa_cp_* helpers in adapters/context_parallel.py) assume it.
attn_cp_size = cfg.tp_size // cfg.dp_size
overrides["attn_cp_size"] = attn_cp_size
logger.warning(
@@ -0,0 +1,4 @@
"""Backend implementations of stage-boundary fusion.
The boundary selects these candidates; compute kernels stay with their backend.
"""
+4 -4
View File
@@ -33,6 +33,10 @@ from sglang.srt.layers.communicator.adapters.attention import (
_redistribute_to_attn_tp_shards,
get_attn_tp_context,
)
from sglang.srt.layers.communicator.adapters.context_parallel import (
dsa_cp_gather_hidden_states,
dsa_cp_reduce_scatter_hidden_states,
)
from sglang.srt.layers.communicator.layout import (
CommunicateContext,
Layout,
@@ -48,10 +52,6 @@ from sglang.srt.layers.communicator.output import (
)
from sglang.srt.layers.communicator.residual import StageRead, StageUpdate
from sglang.srt.layers.communicator.residual.add_norm import ADD, NORM_READ
from sglang.srt.layers.communicator_dsa_cp import (
dsa_cp_gather_hidden_states,
dsa_cp_reduce_scatter_hidden_states,
)
from sglang.srt.layers.dp_attention import (
attn_tp_all_gather_into_tensor,
attn_tp_reduce_scatter_tensor,
@@ -303,7 +303,7 @@ class BaseRunner(ABC):
resolve_flashinfer_allreduce_fusion_backend()
if not mr.is_draft_worker:
# A draft installs no fusion communicator.
from sglang.srt.layers.moe.cutedsl_ar_fusion import (
from sglang.srt.layers.communicator.fusions.cutedsl import (
prepare_cutedsl_fusion,
)
+8 -4
View File
@@ -74,7 +74,7 @@ from sglang.srt.layers.communicator import (
get_attn_tp_context,
layer_input_buffer,
)
from sglang.srt.layers.communicator_dsa_cp import (
from sglang.srt.layers.communicator.adapters.context_parallel import (
maybe_prefetch_next_full_attention_kv,
)
from sglang.srt.layers.cp.cp_decode_attn_tp import get_cp_decode_attn_tp_ctx
@@ -1107,7 +1107,9 @@ class DeepseekV2MoE(nn.Module):
if deferred_finalize and get_forward().defer_moe_finalize:
# deferred_finalize excludes _shared_expert_tp1, so the shared add folds in.
assert shared_output is not None
from sglang.srt.layers.moe.cutedsl_ar_fusion import MoeFinalizeHandoff
from sglang.srt.layers.communicator.fusions.cutedsl import (
MoeFinalizeHandoff,
)
return MoeFinalizeHandoff.from_flashinfer(
final_hidden_states,
@@ -2639,7 +2641,7 @@ class DeepseekV2DecoderLayer(nn.Module):
and _use_mnnvl_cutedsl_fusion()
):
# Dense layers too: selecting cutedsl turns the legacy fusion off.
from sglang.srt.layers.moe.cutedsl_ar_fusion import CuteDSLFusion
from sglang.srt.layers.communicator.fusions.cutedsl import CuteDSLFusion
fusions = CuteDSLFusion()
return LayerCommunicator(
@@ -2971,7 +2973,9 @@ class DeepseekV2Model(nn.Module):
# llama_4_scaling: for supporting Mistral-Large-3 model
self.llama_4_scaling_config = getattr(config, "llama_4_scaling", None)
from sglang.srt.layers.moe.cutedsl_ar_fusion import install_cutedsl_fusion
from sglang.srt.layers.communicator.fusions.cutedsl import (
install_cutedsl_fusion,
)
install_cutedsl_fusion(
# PP pads self.layers with PPMissingLayer, which has no communicator.
+1 -1
View File
@@ -71,7 +71,7 @@ from sglang.srt.layers.attention.dsv4.dsv41_sparse import (
)
from sglang.srt.layers.attention.dsv4.indexer import C4Indexer
from sglang.srt.layers.communicator import get_attn_tp_context
from sglang.srt.layers.communicator_dsa_cp import (
from sglang.srt.layers.communicator.adapters.context_parallel import (
dsa_cp_gather_hidden_states,
dsa_cp_reduce_scatter_hidden_states,
)
+3 -1
View File
@@ -859,7 +859,9 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
if defer_finalize:
if shared_output is None:
raise RuntimeError("Qwen deferred finalize requires shared output")
from sglang.srt.layers.moe.cutedsl_ar_fusion import MoeFinalizeHandoff
from sglang.srt.layers.communicator.fusions.cutedsl import (
MoeFinalizeHandoff,
)
return MoeFinalizeHandoff.from_flashinfer(
final_hidden_states,
+5 -3
View File
@@ -217,7 +217,7 @@ def _use_mnnvl_cutedsl_fusion(config: Qwen3_5TextConfig, is_nextn: bool) -> bool
def _layer_fusions(config: Qwen3_5TextConfig, is_nextn: bool):
"""The CuTe DSL kernels a layer tries before its own, when they are on."""
if _use_mnnvl_cutedsl_fusion(config, is_nextn):
from sglang.srt.layers.moe.cutedsl_ar_fusion import CuteDSLFusion
from sglang.srt.layers.communicator.fusions.cutedsl import CuteDSLFusion
return CuteDSLFusion()
return None
@@ -1730,7 +1730,7 @@ class Qwen3_5ForCausalLM(nn.Module):
"layers: "
f"{unsupported_layers}"
)
from sglang.srt.layers.moe.cutedsl_ar_fusion import (
from sglang.srt.layers.communicator.fusions.cutedsl import (
install_cutedsl_fusion,
)
@@ -1877,7 +1877,9 @@ class Qwen3_5ForCausalLM(nn.Module):
# The final layer has no successor to consume its deferred MoE tail.
is_deferred_finalize = False
if self.flashinfer_mnnvl_cutedsl_fusion is not None:
from sglang.srt.layers.moe.cutedsl_ar_fusion import MoeFinalizeHandoff
from sglang.srt.layers.communicator.fusions.cutedsl import (
MoeFinalizeHandoff,
)
is_deferred_finalize = isinstance(hidden_states, MoeFinalizeHandoff)