[Kernel] Add gpt-oss Router GEMM kernel (#37205)

Signed-off-by: Xin Yang <xyangx@amazon.com>
This commit is contained in:
Xin Yang
2026-03-18 08:15:56 -07:00
committed by GitHub
parent 17808394bc
commit b1169d7be8
13 changed files with 875 additions and 13 deletions
+3 -7
View File
@@ -20,12 +20,11 @@ from vllm.distributed import (
tensor_model_parallel_all_gather,
)
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.fused_moe import FusedMoE, GateLinear
from vllm.model_executor.layers.fused_moe.config import FusedMoEParallelConfig
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
QKVParallelLinear,
ReplicatedLinear,
RowParallelLinear,
)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
@@ -175,13 +174,11 @@ class MLPBlock(torch.nn.Module):
self.hidden_size = config.hidden_size
self.experts_per_token = config.num_experts_per_tok
self.world_size = dist.get_world_size() if dist.is_initialized() else 1
self.router = ReplicatedLinear(
self.router = GateLinear(
config.hidden_size,
config.num_local_experts,
bias=True,
quant_config=None,
prefix=f"{prefix}.router",
return_bias=False,
)
assert config.intermediate_size % self.world_size == 0
self.experts = FusedMoE(
@@ -209,7 +206,7 @@ class MLPBlock(torch.nn.Module):
self, x[:, : self.hidden_size], self.router.weight, self.router.bias
)
else:
g = self.router(x)
g, _ = self.router(x)
x = self.experts(hidden_states=x, router_logits=g)[:, : self.hidden_size]
if self.is_sequence_parallel:
@@ -273,7 +270,6 @@ class GptOssModel(nn.Module, EagleModelMixin):
self.config = vllm_config.model_config.hf_config
self.quant_config = vllm_config.quant_config
self.parallel_config = vllm_config.parallel_config
self.config.hidden_size = self.config.hidden_size
self.embedding = VocabParallelEmbedding(
self.config.vocab_size,
self.config.hidden_size,