[Fix] Stop deferring the last layer's FFN all-reduce in five models (#40868)

This commit is contained in:
Cheng Wan
2026-09-23 14:41:14 -07:00
committed by GitHub
parent 701cf7e47e
commit e31cbc01a2
8 changed files with 217 additions and 4 deletions
+3 -2
View File
@@ -644,6 +644,7 @@ class BailingMoEBlock(nn.Module):
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
alt_stream: Optional[torch.cuda.Stream] = None,
is_nextn: bool = False,
):
super().__init__()
self.config = config
@@ -681,7 +682,7 @@ class BailingMoEBlock(nn.Module):
is_next_layer_sparse=is_next_layer_sparse,
)
self.is_last_layer = self.layer_id == config.num_hidden_layers - 1
self.is_last_layer = is_nextn or (self.layer_id == config.num_hidden_layers - 1)
if self.is_layer_sparse:
self.mlp = BailingMoESparseMoeBlock(
@@ -712,7 +713,7 @@ class BailingMoEBlock(nn.Module):
input_layernorm=self.input_layernorm,
post_attention_layernorm=self.post_attention_layernorm,
allow_reduce_scatter=True,
is_last_layer=(self.layer_id == self.config.num_hidden_layers - 1),
is_last_layer=self.is_last_layer,
)
def _is_layer_sparse(
@@ -835,6 +835,7 @@ class BailingMoELinearDecoderLayer(nn.Module):
input_layernorm=self.input_layernorm,
post_attention_layernorm=self.post_attention_layernorm,
allow_reduce_scatter=False,
is_last_layer=(is_nextn or layer_id == config.num_hidden_layers - 1),
qkv_latent_func=qkv_latent_func,
)
@@ -127,6 +127,7 @@ class BailingMoEModelNextN(nn.Module):
0,
quant_config=quant_config,
prefix=add_prefix("decoder", prefix),
is_nextn=True,
)
self.shared_head = nn.Module()
+1
View File
@@ -1361,6 +1361,7 @@ class MiniMaxM3DecoderLayer(nn.Module):
input_layernorm=self.input_layernorm,
post_attention_layernorm=self.post_attention_layernorm,
allow_reduce_scatter=True,
is_last_layer=(layer_id == config.num_hidden_layers - 1),
)
def forward(
+2
View File
@@ -561,6 +561,7 @@ class Qwen3HybridLinearDecoderLayer(nn.Module):
input_layernorm=self.input_layernorm,
post_attention_layernorm=self.post_attention_layernorm,
allow_reduce_scatter=True,
is_last_layer=(layer_id == config.num_hidden_layers - 1),
)
def forward(
@@ -733,6 +734,7 @@ class Qwen3HybridAttentionDecoderLayer(nn.Module):
input_layernorm=self.input_layernorm,
post_attention_layernorm=self.post_attention_layernorm,
allow_reduce_scatter=True,
is_last_layer=(layer_id == config.num_hidden_layers - 1),
)
self.alt_stream = alt_stream
+2 -1
View File
@@ -464,6 +464,7 @@ class Step3p5DecoderLayer(nn.Module):
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
alt_stream: Optional[torch.cuda.Stream] = None,
is_nextn: bool = False,
) -> None:
super().__init__()
self.hidden_size = config.hidden_size
@@ -575,7 +576,7 @@ class Step3p5DecoderLayer(nn.Module):
input_layernorm=self.input_layernorm,
post_attention_layernorm=self.post_attention_layernorm,
allow_reduce_scatter=True,
is_last_layer=(layer_id == config.num_hidden_layers - 1),
is_last_layer=(is_nextn or layer_id == config.num_hidden_layers - 1),
)
self.layer_id = layer_id
+4 -1
View File
@@ -80,7 +80,10 @@ class Step3p5AMultiTokenPredictor(nn.Module):
self.eh_proj = nn.Linear(config.hidden_size * 2, config.hidden_size, bias=False)
self.shared_head = SharedHead(config=config, quant_config=quant_config)
self.mtp_block = Step3p5DecoderLayer(
config=config, layer_id=layer_id, prefix=f"{prefix}.mtp_block"
config=config,
layer_id=layer_id,
prefix=f"{prefix}.mtp_block",
is_nextn=True,
)
self.lm_head = self.shared_head.head