[Hybrid][torch.compile] Refactor mamba2 forward to avoid obscuring linear projections under custom op (#28587)
Signed-off-by: Tomer Asida <57313761+tomeras91@users.noreply.github.com>
This commit is contained in:
@@ -138,8 +138,7 @@ class BambaMixerDecoderLayer(nn.Module):
|
||||
else:
|
||||
hidden_states, residual = self.input_layernorm(hidden_states, residual)
|
||||
|
||||
output = torch.empty_like(hidden_states)
|
||||
self.mamba(hidden_states, output)
|
||||
output = self.mamba(hidden_states)
|
||||
# Fully Connected
|
||||
hidden_states, residual = self.pre_ff_layernorm(output, residual)
|
||||
hidden_states = self.feed_forward(hidden_states)
|
||||
|
||||
@@ -198,10 +198,8 @@ class FalconH1SSMDecoderLayer(nn.Module):
|
||||
residual: torch.Tensor | None,
|
||||
**kwargs,
|
||||
):
|
||||
output = torch.empty_like(hidden_states)
|
||||
self.mamba(
|
||||
output = self.mamba(
|
||||
hidden_states,
|
||||
output,
|
||||
mup_vector=self.mup_vector,
|
||||
)
|
||||
return output, residual
|
||||
|
||||
@@ -115,8 +115,7 @@ class GraniteMoeHybridMambaDecoderLayer(nn.Module):
|
||||
):
|
||||
residual = hidden_states
|
||||
hidden_states = self.input_layernorm(hidden_states)
|
||||
output = torch.empty_like(hidden_states)
|
||||
self.mamba(hidden_states, output)
|
||||
output = self.mamba(hidden_states)
|
||||
hidden_states = residual + output * self.residual_multiplier
|
||||
|
||||
residual = hidden_states
|
||||
|
||||
@@ -87,8 +87,7 @@ class Mamba2DecoderLayer(nn.Module):
|
||||
else:
|
||||
hidden_states, residual = self.norm(hidden_states, residual)
|
||||
|
||||
output = torch.empty_like(hidden_states)
|
||||
self.mixer(hidden_states, output)
|
||||
output = self.mixer(hidden_states)
|
||||
return output, residual
|
||||
|
||||
|
||||
|
||||
@@ -376,8 +376,7 @@ class NemotronHMambaDecoderLayer(nn.Module):
|
||||
else:
|
||||
hidden_states, residual = self.norm(hidden_states, residual)
|
||||
|
||||
output = torch.empty_like(hidden_states)
|
||||
self.mixer(hidden_states, output)
|
||||
output = self.mixer(hidden_states)
|
||||
return output, residual
|
||||
|
||||
|
||||
|
||||
@@ -567,11 +567,7 @@ class Zamba2MambaDecoderLayer(nn.Module):
|
||||
hidden_states = self.input_layernorm(hidden_states)
|
||||
|
||||
# Process through Mamba mixer
|
||||
output = torch.empty_like(hidden_states)
|
||||
self.mamba(
|
||||
hidden_states,
|
||||
output,
|
||||
)
|
||||
output = self.mamba(hidden_states)
|
||||
|
||||
# residual connection after mamba
|
||||
hidden_states = residual + output
|
||||
|
||||
Reference in New Issue
Block a user