speech_recognition/FunASR
1
0
Fork 0
mirror of https://github.com/modelscope/FunASR synced 2025-09-15 14:48:36 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

decoding

Browse Source
This commit is contained in:
游雁 2024-06-13 16:59:40 +08:00
parent a1e18f808e
commit 664c400545
1 changed files with 5 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
funasr/models/llm_asr/model.py
View File

@ -449,9 +449,9 @@ class LLMASR2(nn.Module):
for name, param in model.named_parameters(): for name, param in model.named_parameters():
param.requires_grad = False param.requires_grad = False
model.eval() model.eval()
self.llm = model
llm_dim = model.get_input_embeddings().weight.shape[-1]
self.llm_dtype = llm_conf.get("llm_dtype", "fp32") self.llm_dtype = llm_conf.get("llm_dtype", "fp32")
self.llm = model.to(dtype_map[self.llm_dtype])
llm_dim = model.get_input_embeddings().weight.shape[-1]
# adaptor # adaptor
adaptor_class = tables.adaptor_classes.get(audio_adaptor) adaptor_class = tables.adaptor_classes.get(audio_adaptor)
@ -536,7 +536,9 @@ class LLMASR2(nn.Module):
labels_ids[labels_ids == -1] = -100 labels_ids[labels_ids == -1] = -100
attention_mask[attention_mask < 0] = 0 attention_mask[attention_mask < 0] = 0
model_outputs = self.llm( model_outputs = self.llm(
inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels_ids inputs_embeds=inputs_embeds.to(dtype_map[self.llm_dtype]),
attention_mask=attention_mask,
labels=labels_ids,
) )
loss = model_outputs.loss loss = model_outputs.loss