最近MindIE开始支持DeepSeek MTP（multi token prediction）特性了，用于推理加速。但是有些开发者打开MTP开关后，没有发现明显的性能提升。这篇文章提供一种定位策略。
原理很简单，就是看一下每次MTP推理后，模型是输出1个token还是多个token。由于MTP的token处理算法是用python实现的，所以可以在镜像的python代码中添加日志，可以在2个地方加日志查看MTP的采信率（也就是verify的成功比例）。
首先可以在MindIE镜像的/usr/local/lib/python3.11/site-packages/mindie_llm/text_generator/plugins/mtp/mtp_plugin.py 路径中找到verify_greedy_one_batch()函数，然后打印相关参数。
@staticmethod
def verify_greedy_one_batch(verify_guess_tokens, next_guess_tokens):
gg = 0
for eg, guess_tokens in enumerate(verify_guess_tokens):
correct = next_guess_tokens[eg]
guess = guess_tokens
if guess != correct:
break
gg += 1
或者在 /usr/local/lib/python3.11/site-packages/mindie_llm/text_generator/plugins/plugin_manager.py里面的init函数和generate_token函数增加如下代码：
from ...utils.log.logging import logger # 新增
class PluginManager:
def __init__(...):
self.all_token_num = 0
self.all_decode_count = 0
self.all_prefill_count = 0
...
@timer.track_time_async('generate_token')
def generate_token(self, input_metadata: InputMetadata):
...
span_end(prof)
if not input_metadata.is_dummy_batch:
if not input_metadata.is_prefill:
for i in range(input_metadata.batch_size):
next_tokens = generation_output.token_ids[i]
if -1 in next_tokens:
first_neg_one_index = np.argmax(next_tokens == -1)
next_tokens = next_tokens[:first_neg_one_index]
self.all_token_num += len(next_tokens)
self.all_decode_count += 1
logger.error(f"self.all_token_num is {self.all_token_num}, self.all_decode_count is {self.all_decode_count}, self.all_prefill_count is {self.all_prefill_count}. Ratio is {self.all_token_num / self.all_decode_count}")
if input_metadata.is_prefill:
for _ in range(input_metadata.batch_size):
self.all_prefill_count += 1
generation_output.trace_ids = trace_ids
