[Fix] GigaChat 3.5: apply the sandwich norms to the complete sums (#42310)

This commit is contained in:
Cheng Wan
2026-10-03 02:21:10 -07:00
committed by GitHub
parent 6faccf3e84
commit f85c2c4923
2 changed files with 20 additions and 26 deletions
+4 -2
View File
@@ -2721,9 +2721,11 @@ class DeepseekV2DecoderLayer(nn.Module):
post_attention_layernorm: nn.Module,
qkv_latent_func: Optional[Callable],
output=None,
attn_output=None,
):
"""The stage boundaries for this layer's norms; they choose their
steps from them at construction."""
steps from them at construction. ``attn_output`` and ``output``: the
attention's and the FFN's output transforms, if any."""
fusions = None
if (
not get_parallel().enable_prefill_cp
@@ -2736,7 +2738,7 @@ class DeepseekV2DecoderLayer(nn.Module):
fusions = CuteDSLFusion()
attn_boundary, ffn_boundary = make_stages(
(
declare_attn(),
declare_attn(output_transform=attn_output),
input_layernorm,
{"qkv_latent_func": qkv_latent_func, "fusions": fusions},
),
+16 -24
View File
@@ -130,7 +130,8 @@ def build_norm(
class GigaChat35PassthroughNorm(nn.Module):
"""Identity norm with the fused residual-add convention.
"""Identity norm with the norm calling convention: ``forward(x)`` returns
``x``, ``forward(x, residual)`` returns ``(residual + x, residual + x)``.
Used as the ``input_layernorm`` slot when a layer has no pre-norm
(``layernorm_type="post"``); it only folds the residual so the
@@ -146,29 +147,21 @@ class GigaChat35PassthroughNorm(nn.Module):
if post_residual_addition is not None:
x = x + post_residual_addition
if residual is None:
return x, x
return x
merged = x + residual
return merged, merged
class GigaChat35MlpPrepNorm(nn.Module):
"""Pre-MLP norm wrapper that folds the post-attention (sandwich) norm.
Serves as the FFN boundary's norm so the
optional ``post_self_attn_layernorm`` (applied to the attention output,
before the residual add) and the pre-MLP ``post_attention_layernorm`` are
both run during FFN input preparation, matching the per-layer math of the
``pre_post`` sandwich exactly.
"""Pre-MLP norm wrapper: adds the residual, then runs the optional pre-MLP
``post_attention_layernorm`` on the sum, unfused. The attention's own
``post_self_attn_layernorm`` is its declared output transform, which the
FFN input runs before this.
"""
def __init__(
self,
pre_layernorm: Optional[nn.Module],
post_layernorm: Optional[nn.Module],
) -> None:
def __init__(self, pre_layernorm: Optional[nn.Module]) -> None:
super().__init__()
self.pre_layernorm = pre_layernorm
self.post_layernorm = post_layernorm
def forward(
self,
@@ -176,14 +169,11 @@ class GigaChat35MlpPrepNorm(nn.Module):
residual: Optional[torch.Tensor] = None,
post_residual_addition: Optional[torch.Tensor] = None,
):
if self.post_layernorm is not None:
x = self.post_layernorm(x)
if post_residual_addition is not None:
x = x + post_residual_addition
if residual is None:
merged = x
else:
merged = x + residual
return x if self.pre_layernorm is None else self.pre_layernorm(x)
merged = x + residual
if self.pre_layernorm is None:
return merged, merged
return self.pre_layernorm(merged), merged
@@ -438,17 +428,19 @@ class GigaChat35DecoderLayer(deepseek_v2.DeepseekV2DecoderLayer):
)
mlp_prepare_layernorm = GigaChat35MlpPrepNorm(
pre_layernorm=self.post_attention_layernorm if self._use_pre else None,
post_layernorm=self.post_self_attn_layernorm,
)
# The stage boundaries choose their fused steps from their norms when built.
self.attn_boundary, self.ffn_boundary = self._build_stages(
input_layernorm=attn_prepare_layernorm,
post_attention_layernorm=mlp_prepare_layernorm,
qkv_latent_func=qkv_latent_func,
attn_output=(
OutputTransform(self.post_self_attn_layernorm)
if self.post_self_attn_layernorm is not None
else None
),
output=(
OutputTransform(
self.post_feedforward_layernorm, before_reduce_scatter=True
)
OutputTransform(self.post_feedforward_layernorm)
if self.post_feedforward_layernorm is not None
else None
),