[Bugfix] Fix GLM4 model (#16618)
Signed-off-by: intervitens <intervitens@tutanota.com>
This commit is contained in:
@@ -82,7 +82,7 @@ class Glm4Attention(nn.Module):
|
||||
partial_rotary_factor = getattr(config, "partial_rotary_factor", 0.5)
|
||||
self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
|
||||
self.head_dim = head_dim or hidden_size // self.total_num_heads
|
||||
self.rotary_dim = int(partial_rotary_factor * self.head_dim)
|
||||
self.rotary_dim = self.head_dim
|
||||
self.q_size = self.num_heads * self.head_dim
|
||||
self.kv_size = self.num_kv_heads * self.head_dim
|
||||
self.scaling = self.head_dim**-0.5
|
||||
@@ -110,6 +110,7 @@ class Glm4Attention(nn.Module):
|
||||
base=self.rope_theta,
|
||||
rope_scaling=rope_scaling,
|
||||
partial_rotary_factor=partial_rotary_factor,
|
||||
is_neox_style=False,
|
||||
)
|
||||
self.attn = Attention(self.num_heads,
|
||||
self.head_dim,
|
||||
@@ -197,13 +198,12 @@ class Glm4DecoderLayer(nn.Module):
|
||||
)
|
||||
|
||||
hidden_states = self.post_self_attn_layernorm(hidden_states)
|
||||
hidden_states = residual + hidden_states
|
||||
|
||||
# Fully Connected
|
||||
hidden_states = self.post_attention_layernorm(hidden_states, residual)
|
||||
hidden_states, residual = self.post_attention_layernorm(
|
||||
hidden_states, residual)
|
||||
hidden_states = self.mlp(hidden_states)
|
||||
hidden_states = self.post_mlp_layernorm(hidden_states)
|
||||
hidden_states = residual + hidden_states
|
||||
|
||||
return hidden_states, residual
|
||||
|
||||
|
||||
Reference in New Issue
Block a user