biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Kernels] Overlap shared experts with send/recv (#23273)

Browse Source 
Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2025-09-03 12:35:18 -04:00
committed by GitHub
parent fa4311d85f
commit e9b92dcd89
32 changed files with 885 additions and 227 deletions
Download Patch File Download Diff File Expand all files Collapse all files

103
vllm/model_executor/models/deepseek_v2.py
View File

@@ -49,6 +49,7 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.shared_fused_moe import SharedFusedMoE
from vllm.model_executor.layers.vocab_parallel_embedding import (
ParallelLMHead, VocabParallelEmbedding)
from vllm.model_executor.model_loader.weight_utils import (
@@ -147,63 +148,85 @@ class DeepseekV2MoE(nn.Module):
self.physical_expert_end = (self.physical_expert_start +
self.n_local_physical_experts)
self.experts = FusedMoE(
num_experts=config.n_routed_experts,
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
reduce_results=False,
renormalize=config.norm_topk_prob,
quant_config=quant_config,
use_grouped_topk=True,
num_expert_group=config.n_group,
topk_group=config.topk_group,
prefix=f"{prefix}.experts",
scoring_func=config.scoring_func,
# we do scaling outside, set factor to 1.0 to avoid double mul
routed_scaling_factor=1.0,
e_score_correction_bias=self.gate.e_score_correction_bias,
enable_eplb=self.enable_eplb,
num_redundant_experts=self.n_redundant_experts)
if config.n_shared_experts is not None:
if config.n_shared_experts is None:
self.experts = FusedMoE(
num_experts=config.n_routed_experts,
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
reduce_results=False,
renormalize=config.norm_topk_prob,
quant_config=quant_config,
use_grouped_topk=True,
num_expert_group=config.n_group,
topk_group=config.topk_group,
prefix=f"{prefix}.experts",
scoring_func=config.scoring_func,
# we do scaling outside, set factor to 1.0 to avoid double mul
routed_scaling_factor=1.0,
e_score_correction_bias=self.gate.e_score_correction_bias,
enable_eplb=self.enable_eplb,
num_redundant_experts=self.n_redundant_experts)
self.shared_experts = None
else:
intermediate_size = (config.moe_intermediate_size *
config.n_shared_experts)
self.shared_experts = DeepseekV2MLP(
hidden_size=config.hidden_size,
intermediate_size=intermediate_size,
hidden_act=config.hidden_act,
quant_config=quant_config,
reduce_results=self.experts.must_reduce_shared_expert_outputs(
),
reduce_results=False,
prefix=f"{prefix}.shared_experts",
)
self.experts = SharedFusedMoE(
shared_experts=self.shared_experts,
num_experts=config.n_routed_experts,
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
reduce_results=False,
renormalize=config.norm_topk_prob,
quant_config=quant_config,
use_grouped_topk=True,
num_expert_group=config.n_group,
topk_group=config.topk_group,
prefix=f"{prefix}.experts",
scoring_func=config.scoring_func,
# we do scaling outside, set factor to 1.0 to avoid double mul
routed_scaling_factor=1.0,
e_score_correction_bias=self.gate.e_score_correction_bias,
enable_eplb=self.enable_eplb,
num_redundant_experts=self.n_redundant_experts)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
num_tokens, hidden_dim = hidden_states.shape
hidden_states = hidden_states.view(-1, hidden_dim)
if self.n_shared_experts is not None:
shared_output = self.shared_experts(hidden_states)
# router_logits: (num_tokens, n_experts)
router_logits, _ = self.gate(hidden_states)
if hidden_states.dtype != torch.float16:
final_hidden_states = self.experts(
hidden_states=hidden_states,
router_logits=router_logits) * self.routed_scaling_factor
fused_moe_out = self.experts(hidden_states=hidden_states,
router_logits=router_logits)
if self.shared_experts is not None:
shared_output, final_hidden_states = fused_moe_out
else:
# Fix FP16 overflow
# See DeepseekV2DecoderLayer for more details.
final_hidden_states = self.experts(hidden_states=hidden_states,
router_logits=router_logits)
if shared_output is not None:
if hidden_states.dtype != torch.float16:
final_hidden_states = final_hidden_states + shared_output
else:
# Fix FP16 overflow
# See DeepseekV2DecoderLayer for more details.
final_hidden_states = final_hidden_states + shared_output \
* (1. / self.routed_scaling_factor)
shared_output = None
final_hidden_states = fused_moe_out
# Fix FP16 overflow
# See DeepseekV2DecoderLayer for more details.
if hidden_states.dtype != torch.float16:
final_hidden_states *= self.routed_scaling_factor
elif self.shared_experts is not None:
assert shared_output is not None
shared_output *= (1. / self.routed_scaling_factor)
if self.shared_experts is not None:
assert shared_output is not None
final_hidden_states += shared_output
if self.tp_size > 1:
final_hidden_states = (