[DSV4] Add silu clamp limit to shared expert (#40950)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
This commit is contained in:
Yongye Zhu
2026-04-27 00:37:43 -07:00
committed by GitHub
parent 22631f80a0
commit 706a04d34b
7 changed files with 269 additions and 29 deletions
+40
View File
@@ -151,6 +151,46 @@ class SiluAndMul(CustomOp):
return self.forward_cuda(x)
@CustomOp.register("silu_and_mul_with_clamp")
class SiluAndMulWithClamp(CustomOp):
"""SwiGLU activation with input clamping (used by some MoE shared experts).
Computes:
gate = clamp(x[..., :d], max=swiglu_limit)
up = clamp(x[..., d:], min=-swiglu_limit, max=swiglu_limit)
out = silu(gate) * up
where d = x.shape[-1] // 2.
Shapes:
x: (num_tokens, 2 * d) or (batch_size, seq_len, 2 * d)
return: (num_tokens, d) or (batch_size, seq_len, d)
"""
def __init__(self, swiglu_limit: float, *, compile_native: bool = True):
super().__init__(compile_native=compile_native)
self.swiglu_limit = float(swiglu_limit)
if current_platform.is_cuda_alike() or current_platform.is_xpu():
self.op = torch.ops._C.silu_and_mul_with_clamp
elif current_platform.is_cpu():
self._forward_method = self.forward_native
def forward_native(self, x: torch.Tensor) -> torch.Tensor:
d = x.shape[-1] // 2
gate = torch.clamp(x[..., :d], max=self.swiglu_limit)
up = torch.clamp(x[..., d:], min=-self.swiglu_limit, max=self.swiglu_limit)
return F.silu(gate) * up
def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
d = x.shape[-1] // 2
output_shape = x.shape[:-1] + (d,)
out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
self.op(out, x, self.swiglu_limit)
return out
def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
return self.forward_cuda(x)
# --8<-- [start:mul_and_silu]
@CustomOp.register("mul_and_silu")
class MulAndSilu(CustomOp):
@@ -45,7 +45,7 @@ def _gelu_and_mul(
# Uses static methods or standalone functions to avoid instantiating CustomOp
# classes, which would call get_current_vllm_config() before config is set.
_CPU_MOE_ACT_FN: dict[MoEActivation, Callable[[torch.Tensor], torch.Tensor]] = {
MoEActivation.SILU: SiluAndMul.forward_native,
MoEActivation.SILU: lambda x: SiluAndMul(compile_native=False).forward_native(x),
MoEActivation.SWIGLUOAI: _swigluoai_forward_native,
MoEActivation.GELU: _gelu_and_mul,
}
+58 -3
View File
@@ -17,6 +17,7 @@ from vllm.distributed import (
get_tensor_model_parallel_world_size,
)
from vllm.forward_context import get_forward_context
from vllm.model_executor.layers.activation import SiluAndMul, SiluAndMulWithClamp
from vllm.model_executor.layers.deepseek_v4_attention import (
DeepseekV4Indexer,
DeepseekV4MLAModules,
@@ -34,7 +35,10 @@ from vllm.model_executor.layers.linear import (
RowParallelLinear,
)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationMethods
from vllm.model_executor.layers.quantization import (
QuantizationConfig,
QuantizationMethods,
)
from vllm.model_executor.layers.quantization.fp8 import Fp8Config
from vllm.model_executor.layers.quantization.mxfp4 import Mxfp4MoEMethod
from vllm.model_executor.layers.quantization.utils.quant_utils import (
@@ -46,7 +50,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
VocabParallelEmbedding,
)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.models.deepseek_v2 import DeepseekV2MLP
from vllm.model_executor.utils import set_weight_attrs
from vllm.platforms import current_platform
from vllm.sequence import IntermediateTensors
@@ -63,6 +66,57 @@ from .utils import (
)
class DeepseekV4MLP(nn.Module):
def __init__(
self,
hidden_size: int,
intermediate_size: int,
hidden_act: str,
swiglu_limit: float | None = None,
quant_config: QuantizationConfig | None = None,
reduce_results: bool = True,
is_sequence_parallel: bool = False,
prefix: str = "",
) -> None:
super().__init__()
# If is_sequence_parallel, the input and output tensors are sharded
# across the ranks within the tp_group. In this case the weights are
# replicated and no collective ops are needed.
# Otherwise we use standard TP with an allreduce at the end.
self.gate_up_proj = MergedColumnParallelLinear(
hidden_size,
[intermediate_size] * 2,
bias=False,
quant_config=quant_config,
disable_tp=is_sequence_parallel,
prefix=f"{prefix}.gate_up_proj",
)
self.down_proj = RowParallelLinear(
intermediate_size,
hidden_size,
bias=False,
quant_config=quant_config,
reduce_results=reduce_results,
disable_tp=is_sequence_parallel,
prefix=f"{prefix}.down_proj",
)
if hidden_act != "silu":
raise ValueError(
f"Unsupported activation: {hidden_act}. Only silu is supported for now."
)
if swiglu_limit is not None:
self.act_fn = SiluAndMulWithClamp(swiglu_limit)
else:
self.act_fn = SiluAndMul()
def forward(self, x):
gate_up, _ = self.gate_up_proj(x)
x = self.act_fn(gate_up)
x, _ = self.down_proj(x)
return x
class DeepseekV4FP8Config(Fp8Config):
"""FP8 config that routes MoE layers to MXFP4 quantization.
@@ -672,10 +726,11 @@ class DeepseekV4MoE(nn.Module):
else:
intermediate_size = config.moe_intermediate_size * config.n_shared_experts
self.shared_experts = DeepseekV2MLP(
self.shared_experts = DeepseekV4MLP(
hidden_size=config.hidden_size,
intermediate_size=intermediate_size,
hidden_act=config.hidden_act,
swiglu_limit=self.swiglu_limit,
quant_config=quant_config,
reduce_results=self.use_mega_moe,
prefix=f"{prefix}.shared_experts",