deepgeek/CosyVoice
2
0
Fork 0
mirror of https://github.com/FunAudioLLM/CosyVoice.git synced 2026-02-05 01:49:25 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix export_onnx.py

Browse Source
This commit is contained in:
lyuxiang.lx
2025-04-15 17:29:42 +08:00
parent efe1d15960
commit 2c193781cc
2 changed files with 9 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
cosyvoice/flow/decoder.py
View File

@@ -158,12 +158,9 @@ class CausalAttnProcessor2_0(AttnProcessor2_0):
key_cache = attn.to_k(encoder_hidden_states)
value_cache = attn.to_v(encoder_hidden_states)
# NOTE here we judge cache.size(0) instead of cache.size(1), because init_cache has size (2, 0, 512, 2)
if cache.size(0) != 0:
key = torch.concat([cache[:, :, :, 0], key_cache], dim=1)
value = torch.concat([cache[:, :, :, 1], value_cache], dim=1)
else:
key, value = key_cache, value_cache
# NOTE always concat cache for interface compatibility
key = torch.concat([cache[:, :, :, 0], key_cache], dim=1)
value = torch.concat([cache[:, :, :, 1], value_cache], dim=1)
cache = torch.stack([key_cache, value_cache], dim=3)
inner_dim = key.shape[-1]
@@ -799,6 +796,7 @@ class CausalConditionalDecoder(ConditionalDecoder):
output = self.final_proj(x * mask_up)
return output * mask
@torch.inference_mode()
def forward_chunk(self, x, mask, mu, t, spks=None, cond=None,
down_blocks_conv_cache: torch.Tensor = torch.zeros(0, 0, 0, 0),
down_blocks_kv_cache: torch.Tensor = torch.zeros(0, 0, 0, 0, 0, 0),