[BugFix] Disable TP for Qwen3-Omni audio encoder when heads % TP != 0 (#50858)

Signed-off-by: CalvinXKY <xky@users.noreply.github.com>
Co-authored-by: CalvinXKY <xky@users.noreply.github.com>
This commit is contained in:
kaiyuanxie
2026-08-29 11:18:04 +08:00
committed by GitHub
co-authored by CalvinXKY
parent 46a83642f6
commit 738bc8811a
@@ -196,8 +196,13 @@ class Qwen3OmniMoeAudioAttention(nn.Module):
self.embed_dim = config.d_model
self.num_heads = config.encoder_attention_heads
self.head_dim = self.embed_dim // self.num_heads
# Audio encoder uses 20 heads. Shard across TP when divisible
# (e.g. TP=2/4); otherwise keep unreplicated (e.g. TP=8).
tp_size = get_tensor_model_parallel_world_size()
self.num_local_heads = self.num_heads // tp_size
self.disable_tp = self.num_heads % tp_size != 0
self.num_local_heads = (
self.num_heads if self.disable_tp else self.num_heads // tp_size
)
if (self.head_dim * self.num_heads) != self.embed_dim:
raise ValueError(
@@ -214,6 +219,7 @@ class Qwen3OmniMoeAudioAttention(nn.Module):
total_num_kv_heads=self.num_heads,
bias=True,
prefix=f"{prefix}.qkv_proj",
disable_tp=self.disable_tp,
)
self.out_proj = RowParallelLinear(
@@ -221,6 +227,7 @@ class Qwen3OmniMoeAudioAttention(nn.Module):
output_size=self.embed_dim,
bias=True,
prefix=f"{prefix}.out_proj",
disable_tp=self.disable_tp,
)
self.attn = MMEncoderAttention(