[V1] EP/TP MoE + DP Attention (#13931)
This commit is contained in:
@@ -80,7 +80,8 @@ class OlmoeMoE(nn.Module):
|
||||
reduce_results=True,
|
||||
renormalize=False,
|
||||
quant_config=quant_config,
|
||||
tp_size=tp_size)
|
||||
tp_size=tp_size,
|
||||
prefix=f"{prefix}.experts")
|
||||
|
||||
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
||||
# NOTE: hidden_states can have either 1D or 2D shape.
|
||||
@@ -212,6 +213,7 @@ class OlmoeDecoderLayer(nn.Module):
|
||||
hidden_size=config.hidden_size,
|
||||
intermediate_size=config.intermediate_size,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.mlp",
|
||||
)
|
||||
self.input_layernorm = RMSNorm(config.hidden_size, eps=1e-5)
|
||||
self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=1e-5)
|
||||
|
||||
Reference in New Issue
Block a user