From 5ebcba3d611cc8caca3c25ccbc937b60487f12fb Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期四, 13 六月 2024 19:16:19 +0800
Subject: [PATCH] decoding
---
funasr/models/llm_asr/model.py | 29 +++++++++++++++++------------
1 files changed, 17 insertions(+), 12 deletions(-)
diff --git a/funasr/models/llm_asr/model.py b/funasr/models/llm_asr/model.py
index 15969e3..2d229b9 100644
--- a/funasr/models/llm_asr/model.py
+++ b/funasr/models/llm_asr/model.py
@@ -410,19 +410,19 @@
audio_encoder_output_size = audio_encoder.output_size()
freeze = audio_encoder_conf.get("freeze", True)
freeze_layer_num = int(audio_encoder_conf.get("freeze_layer_num", -1))
- if freeze_layer_num > 0:
- freeze_layer_num = range(freeze_layer_num)
+ # if freeze_layer_num > 0:
+ # freeze_layer_num = range(freeze_layer_num)
if freeze:
for name, param in audio_encoder.named_parameters():
- if isinstance(freeze_layer_num, (list, tuple)):
+ if freeze_layer_num > 0:
idx = re.search(r"\.\d+\.", name)
if idx is not None:
beg, end = idx.regs[0]
layer_id = int(name[beg + 1 : end - 1])
- if layer_id in freeze_layer_num:
+ if layer_id < freeze_layer_num:
param.requires_grad = False
- else:
+ elif ".ln_post." not in name:
param.requires_grad = False
else:
param.requires_grad = False
@@ -449,9 +449,9 @@
for name, param in model.named_parameters():
param.requires_grad = False
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 = model.to(dtype_map[self.llm_dtype])
+ llm_dim = model.get_input_embeddings().weight.shape[-1]
# adaptor
adaptor_class = tables.adaptor_classes.get(audio_adaptor)
@@ -496,11 +496,14 @@
batch_size, frames, _ = speech.shape
- # audio encoder
- encoder_out, encoder_out_lens = self.audio_encoder(speech.permute(0, 2, 1), speech_lengths)
+ with torch.cuda.amp.autocast(enabled=False):
+ # audio encoder
+ encoder_out, encoder_out_lens = self.audio_encoder(
+ speech.permute(0, 2, 1), speech_lengths
+ )
- # audio_adaptor
- encoder_out, encoder_out_lens = self.audio_adaptor(encoder_out, encoder_out_lens)
+ # audio_adaptor
+ encoder_out, encoder_out_lens = self.audio_adaptor(encoder_out, encoder_out_lens)
input_ids[input_ids < 0] = 0
inputs_embeds = self.llm.model.get_input_embeddings()(input_ids)
@@ -536,7 +539,9 @@
labels_ids[labels_ids == -1] = -100
attention_mask[attention_mask < 0] = 0
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
--
Gitblit v1.9.1