File size: 13,185 Bytes
e6b37e4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 | from dataclasses import dataclass
from typing import List, Optional
import torch
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
class OperationsStrategy:
operations: List[Operation]
deep_gemm_num_sms: Optional[int] = None
tbo_delta_stages: Optional[int] = None
@classmethod
def concat(cls, items: List["OperationsStrategy"]) -> "OperationsStrategy":
return OperationsStrategy(
operations=[x for item in items for x in item.operations],
deep_gemm_num_sms=_assert_all_same(
[item.deep_gemm_num_sms for item in items]
),
tbo_delta_stages=_assert_all_same(
[item.tbo_delta_stages for item in items]
),
)
@staticmethod
def init_new_tbo(
layers: torch.nn.ModuleList,
forward_mode: ForwardMode,
) -> "OperationsStrategy":
layer_name = layers[0].__class__.__name__
if layer_name == "DeepseekV2DecoderLayer":
return OperationsStrategy.concat(
[
_compute_moe_deepseek_layer_operations_strategy_tbo(
layer, forward_mode
)
for layer in layers
]
)
elif layer_name == "Qwen3MoeDecoderLayer":
return OperationsStrategy.concat(
[
_compute_moe_qwen3_layer_operations_strategy_tbo(
layer, forward_mode
)
for layer in layers
]
)
elif layer_name == "MiMoV2DecoderLayer":
return OperationsStrategy.concat(
[
_compute_moe_mimov2_layer_operations_strategy_tbo(
layer, forward_mode
)
for layer in layers
]
)
elif layer_name == "AgnesDecoderLayer":
return OperationsStrategy.concat(
[
_compute_moe_deepseek_v4_layer_operations_strategy_tbo(
layer, forward_mode
)
for layer in layers
]
)
else:
raise NotImplementedError
def _assert_all_same(items: List):
assert all(item == items[0] for item in items)
return items[0]
# -------------------------------- Strategy for DeepSeek ---------------------------------------
# TODO can refactor to make it more fancy if we have more complex strategies
def _compute_moe_deepseek_layer_operations_strategy_tbo(
layer: torch.nn.Module,
forward_mode: ForwardMode,
) -> OperationsStrategy:
assert layer.is_layer_sparse, "dense layer TBO not yet implemented"
if forward_mode == ForwardMode.EXTEND:
return _compute_moe_deepseek_blog_prefill(layer)
elif (
forward_mode == ForwardMode.DECODE or forward_mode == ForwardMode.TARGET_VERIFY
):
return _compute_moe_deepseek_blog_decode(layer)
else:
raise NotImplementedError(f"Unsupported {forward_mode=}")
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
return OperationsStrategy(
deep_gemm_num_sms=deep_gemm_num_sms,
tbo_delta_stages=0,
operations=[
layer.op_comm_prepare_attn,
layer.self_attn.op_prepare,
layer.self_attn.op_core,
layer.op_comm_prepare_mlp,
layer.mlp.op_gate,
layer.mlp.op_select_experts,
layer.mlp.op_dispatch_a,
operations.YieldOperation(),
layer.mlp.op_dispatch_b,
layer.mlp.op_experts,
layer.mlp.op_combine_a,
operations.YieldOperation(),
layer.mlp.op_shared_experts,
layer.mlp.op_combine_b,
layer.mlp.op_output,
layer.op_comm_postprocess_layer,
],
)
def _compute_moe_deepseek_blog_decode(layer):
return OperationsStrategy(
deep_gemm_num_sms=None,
tbo_delta_stages=2,
operations=[
layer.op_comm_prepare_attn,
layer.self_attn.op_prepare,
operations.YieldOperation(),
layer.self_attn.op_core,
layer.op_comm_prepare_mlp,
layer.mlp.op_gate,
layer.mlp.op_select_experts,
operations.YieldOperation(),
layer.mlp.op_dispatch_a,
layer.mlp.op_shared_experts,
operations.YieldOperation(),
layer.mlp.op_dispatch_b,
layer.mlp.op_experts,
layer.mlp.op_combine_a,
operations.YieldOperation(),
layer.mlp.op_combine_b,
operations.YieldOperation(),
layer.mlp.op_output,
layer.op_comm_postprocess_layer,
],
)
# -------------------------------- Strategy for DeepSeek V4 ---------------------------------------
# DSV4 prefill TBO (EP / mori path). Cross-layer mHC fusion is disabled under
# TBO, so each layer is self-contained: attn-side mHC pre+norm -> attn ->
# ffn-side mHC pre+norm -> MoE (a2a dispatch/combine overlapped) -> mHC post.
# The MoE ops are reused from self.mlp (DeepseekV2MoE) and decompose
# forward_deepep; the layer-level op_mhc_* wrap DSV4's hc_pre / hc_post.
def _compute_moe_deepseek_v4_layer_operations_strategy_tbo(
layer: torch.nn.Module,
forward_mode: ForwardMode,
) -> OperationsStrategy:
if forward_mode == ForwardMode.EXTEND:
return _compute_moe_deepseek_v4_prefill(layer)
else:
# Decode TBO for DSV4 is not implemented yet (ATOM data: decode TBO
# regresses; needs cuda-graph capture work). Prefill-only for now.
raise NotImplementedError(
f"DeepseekV4 TBO only supports prefill (EXTEND), got {forward_mode=}"
)
def _compute_moe_deepseek_v4_prefill(layer):
from sglang.srt.layers.moe import get_moe_a2a_backend
if get_moe_a2a_backend().is_none():
# Non-EP DP TP-MoE: overlap the DP all_gatherv (gather) + reduce_scatterv
# (combine) with the other ubatch's attn+MoE compute (ATOM's DSV4 path).
ops = [
layer.op_mhc_prepare_attn,
layer.self_attn.op_attn,
layer.op_mhc_post_attn_pre_mlp,
layer.op_gather_a,
operations.YieldOperation(),
layer.op_gather_b,
layer.op_moe,
layer.op_combine_a,
operations.YieldOperation(),
layer.op_combine_b,
layer.op_mhc_postprocess,
]
else:
# EP / mori a2a: reuse DeepseekV2MoE's deepep dispatch/combine ops.
ops = [
layer.op_mhc_prepare_attn,
layer.self_attn.op_attn,
layer.op_mhc_post_attn_pre_mlp,
layer.mlp.op_gate,
layer.mlp.op_select_experts,
layer.mlp.op_dispatch_a,
operations.YieldOperation(),
layer.mlp.op_dispatch_b,
layer.mlp.op_experts,
layer.mlp.op_combine_a,
operations.YieldOperation(),
layer.mlp.op_shared_experts,
layer.mlp.op_combine_b,
layer.mlp.op_output,
layer.op_mhc_postprocess,
]
return OperationsStrategy(
deep_gemm_num_sms=None,
tbo_delta_stages=0,
operations=ops,
)
# -------------------------------- Strategy for Qwen3 ---------------------------------------
# TODO: unstable, current strategy is almost the same as DeepSeek, keep redundant code here for
# convenience to adjust strategy
def _compute_moe_qwen3_layer_operations_strategy_tbo(
layer: torch.nn.Module,
forward_mode: ForwardMode,
) -> OperationsStrategy:
assert layer.is_layer_sparse, "qwen3 moe only support sparse layers"
if forward_mode == ForwardMode.EXTEND:
return _compute_moe_qwen3_prefill(layer)
elif (
forward_mode == ForwardMode.DECODE or forward_mode == ForwardMode.TARGET_VERIFY
):
return _compute_moe_qwen3_decode(layer)
else:
raise NotImplementedError(f"Unsupported {forward_mode=}")
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
return OperationsStrategy(
deep_gemm_num_sms=deep_gemm_num_sms,
tbo_delta_stages=0,
operations=[
layer.op_comm_prepare_attn,
layer.self_attn.op_prepare,
layer.self_attn.op_core,
layer.op_comm_prepare_mlp,
layer.mlp.op_gate,
layer.mlp.op_select_experts,
layer.mlp.op_dispatch_a,
operations.YieldOperation(),
layer.mlp.op_dispatch_b,
layer.mlp.op_experts,
layer.mlp.op_combine_a,
operations.YieldOperation(),
layer.mlp.op_combine_b,
layer.mlp.op_output,
layer.op_comm_postprocess_layer,
],
)
def _compute_moe_qwen3_decode(layer):
return OperationsStrategy(
deep_gemm_num_sms=None,
tbo_delta_stages=2,
operations=[
layer.op_comm_prepare_attn,
layer.self_attn.op_prepare,
operations.YieldOperation(),
layer.self_attn.op_core,
layer.op_comm_prepare_mlp,
layer.mlp.op_gate,
layer.mlp.op_select_experts,
operations.YieldOperation(),
layer.mlp.op_dispatch_a,
operations.YieldOperation(),
layer.mlp.op_dispatch_b,
layer.mlp.op_experts,
layer.mlp.op_combine_a,
operations.YieldOperation(),
layer.mlp.op_combine_b,
layer.mlp.op_output,
layer.op_comm_postprocess_layer,
operations.YieldOperation(),
],
)
# -------------------------------- Strategy for MiMoV2DecoderLayer ---------------------------------------
# TODO: unstable; current strategy matches DeepSeek for the common operations (MiMoV2 has no op_shared_experts),
# so we keep this redundant code here for convenience when adjusting the strategy
def _compute_moe_mimov2_layer_operations_strategy_tbo(
layer: torch.nn.Module,
forward_mode: ForwardMode,
) -> OperationsStrategy:
assert layer.is_layer_sparse, "MiMoV2DecoderLayer moe only support sparse layers"
if forward_mode == ForwardMode.EXTEND:
return _compute_moe_mimov2_prefill(layer)
elif (
forward_mode == ForwardMode.DECODE or forward_mode == ForwardMode.TARGET_VERIFY
):
return _compute_moe_mimov2_decode(layer)
else:
raise NotImplementedError(f"Unsupported {forward_mode=}")
def _compute_moe_mimov2_prefill(layer):
device_properties = torch.cuda.get_device_properties(device="cuda")
total_num_sms = device_properties.multi_processor_count
deep_gemm_num_sms = total_num_sms - DeepEPConfig.get_instance().num_sms
return OperationsStrategy(
deep_gemm_num_sms=deep_gemm_num_sms,
tbo_delta_stages=0,
operations=[
layer.op_comm_prepare_attn,
layer.self_attn.op_prepare,
layer.self_attn.op_core,
layer.op_comm_prepare_mlp,
layer.mlp.op_gate,
layer.mlp.op_select_experts,
layer.mlp.op_dispatch_a,
operations.YieldOperation(),
layer.mlp.op_dispatch_b,
layer.mlp.op_experts,
layer.mlp.op_combine_a,
operations.YieldOperation(),
layer.mlp.op_combine_b,
layer.mlp.op_output,
layer.op_comm_postprocess_layer,
],
)
def _compute_moe_mimov2_decode(layer):
return OperationsStrategy(
deep_gemm_num_sms=None,
tbo_delta_stages=2,
operations=[
layer.op_comm_prepare_attn,
layer.self_attn.op_prepare,
operations.YieldOperation(),
layer.self_attn.op_core,
layer.op_comm_prepare_mlp,
layer.mlp.op_gate,
layer.mlp.op_select_experts,
operations.YieldOperation(),
layer.mlp.op_dispatch_a,
operations.YieldOperation(),
layer.mlp.op_dispatch_b,
layer.mlp.op_experts,
layer.mlp.op_combine_a,
operations.YieldOperation(),
layer.mlp.op_combine_b,
layer.mlp.op_output,
layer.op_comm_postprocess_layer,
operations.YieldOperation(),
],
)
|