diff --git a/python/sglang/srt/arg_groups/model_overrides/deepseek_v2.py b/python/sglang/srt/arg_groups/model_overrides/deepseek_v2.py index 54e2cd2a2e..23568e15d1 100644 --- a/python/sglang/srt/arg_groups/model_overrides/deepseek_v2.py +++ b/python/sglang/srt/arg_groups/model_overrides/deepseek_v2.py @@ -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( diff --git a/python/sglang/srt/layers/communicator_dsa_cp.py b/python/sglang/srt/layers/communicator/adapters/context_parallel.py similarity index 100% rename from python/sglang/srt/layers/communicator_dsa_cp.py rename to python/sglang/srt/layers/communicator/adapters/context_parallel.py diff --git a/python/sglang/srt/layers/communicator/fusions/__init__.py b/python/sglang/srt/layers/communicator/fusions/__init__.py new file mode 100644 index 0000000000..ba64f66fd9 --- /dev/null +++ b/python/sglang/srt/layers/communicator/fusions/__init__.py @@ -0,0 +1,4 @@ +"""Backend implementations of stage-boundary fusion. + +The boundary selects these candidates; compute kernels stay with their backend. +""" diff --git a/python/sglang/srt/layers/moe/cutedsl_ar_fusion.py b/python/sglang/srt/layers/communicator/fusions/cutedsl.py similarity index 100% rename from python/sglang/srt/layers/moe/cutedsl_ar_fusion.py rename to python/sglang/srt/layers/communicator/fusions/cutedsl.py diff --git a/python/sglang/srt/layers/communicator/ops.py b/python/sglang/srt/layers/communicator/ops.py index d122af316a..438197d287 100644 --- a/python/sglang/srt/layers/communicator/ops.py +++ b/python/sglang/srt/layers/communicator/ops.py @@ -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, diff --git a/python/sglang/srt/model_executor/runner/base_runner.py b/python/sglang/srt/model_executor/runner/base_runner.py index 7f69e36de5..fcef357fd7 100644 --- a/python/sglang/srt/model_executor/runner/base_runner.py +++ b/python/sglang/srt/model_executor/runner/base_runner.py @@ -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, ) diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py index 4f23371e53..487ee74c6e 100644 --- a/python/sglang/srt/models/deepseek_v2.py +++ b/python/sglang/srt/models/deepseek_v2.py @@ -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. diff --git a/python/sglang/srt/models/deepseek_v4.py b/python/sglang/srt/models/deepseek_v4.py index cd7e928648..4751bd2a84 100644 --- a/python/sglang/srt/models/deepseek_v4.py +++ b/python/sglang/srt/models/deepseek_v4.py @@ -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, ) diff --git a/python/sglang/srt/models/qwen2_moe.py b/python/sglang/srt/models/qwen2_moe.py index 91a9d6e2f8..9b33dfd043 100644 --- a/python/sglang/srt/models/qwen2_moe.py +++ b/python/sglang/srt/models/qwen2_moe.py @@ -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, diff --git a/python/sglang/srt/models/qwen3_5.py b/python/sglang/srt/models/qwen3_5.py index 67d347cbff..9609348326 100644 --- a/python/sglang/srt/models/qwen3_5.py +++ b/python/sglang/srt/models/qwen3_5.py @@ -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) diff --git a/test/registered/unit/layers/moe/test_cutedsl_ar_fusion.py b/test/registered/unit/layers/moe/test_cutedsl_ar_fusion.py index 5a62ace744..a9ab72fb50 100644 --- a/test/registered/unit/layers/moe/test_cutedsl_ar_fusion.py +++ b/test/registered/unit/layers/moe/test_cutedsl_ar_fusion.py @@ -16,6 +16,12 @@ from sglang.srt.layers.communicator import ( UnreducedOutput, reduce_output, ) +from sglang.srt.layers.communicator.fusions.cutedsl import ( + CuteDSLFusion, + MoeFinalizeHandoff, + install_cutedsl_fusion, + prepare_cutedsl_fusion, +) from sglang.srt.layers.communicator.ops import _consumer_step, _read_input from sglang.srt.layers.flashinfer_mnnvl_cutedsl import ( FlashInferMNNVLCuteDSLARFusion, @@ -23,12 +29,6 @@ from sglang.srt.layers.flashinfer_mnnvl_cutedsl import ( _with_early_finalize_shared_load, ) from sglang.srt.layers.layernorm import RMSNorm -from sglang.srt.layers.moe.cutedsl_ar_fusion import ( - CuteDSLFusion, - MoeFinalizeHandoff, - install_cutedsl_fusion, - prepare_cutedsl_fusion, -) from sglang.srt.model_executor.cuda_graph_config import CudaGraphConfig, PhaseConfig from sglang.srt.model_executor.forward_batch_info import ForwardMode from sglang.srt.runtime_context import get_forward, publish, reset_context @@ -38,7 +38,7 @@ from sglang.test.communicator_patch import patch_communicator register_cpu_ci(est_time=9, suite="base-a-test-cpu") -_MODULE = "sglang.srt.layers.moe.cutedsl_ar_fusion" +_MODULE = "sglang.srt.layers.communicator.fusions.cutedsl" _DECODE = SimpleNamespace(forward_mode=ForwardMode.DECODE, input_ids=torch.zeros(8)) diff --git a/test/registered/unit/layers/test_declared_attention_cp.py b/test/registered/unit/layers/test_declared_attention_cp.py index cf245d6e08..a009c0f6a2 100644 --- a/test/registered/unit/layers/test_declared_attention_cp.py +++ b/test/registered/unit/layers/test_declared_attention_cp.py @@ -16,8 +16,8 @@ from unittest.mock import patch import torch from sglang.srt.layers import communicator as comm -from sglang.srt.layers import communicator_dsa_cp as dsa_cp from sglang.srt.layers import layernorm_sp +from sglang.srt.layers.communicator.adapters import context_parallel as dsa_cp from sglang.test.ci.ci_register import register_cpu_ci from sglang.test.communicator_patch import patch_communicator from sglang.test.test_utils import CustomTestCase