I tried to reproduce MDI-LLM on Jetson orin NX, but the inference result I got was garbled：
Produced output:
Sample 1:
<|system|>
You are a friendly chatbot who always gives helpful, detailed, and polite answers.
<|user|>
What is the recipe for pizza?
<|assistant|>
.
k. -LI. liara to choice24 < to : m a k?5ch about?. insp ta.
/
o">< <- 2 and var
, ( # to
1,, a can Isks as ( (
is,
1.,:,:)
/oleg 0,
,.,,
1.-o?) was mch] <
, 0eph,. 8 m,
. AT using
,, . in- 1. k
, (h- |
2
2,
.3,2. +. )
al and 5 (
,h <<)?59, m nun. .2 ( ' #fish =?
:1 22.) (: ( |
,
en
1:
il | 5
for f ,,,ch and 2
.,03 for #, ").-,,.4 |,
1
: k, '. and and,ch, 113,h2
,:3, , the.o,
+
&>
below, m m*
and
.A barra 1,. :) '],=
( (32 of the .j); in |22
and
"> ',1 0.0 ish m6):
| (i,
m, or 13.
f',,
May I ask what the reason is？
I tried to reproduce MDI-LLM on Jetson orin NX, but the inference result I got was garbled：
Produced output:
Sample 1:
<|system|>
You are a friendly chatbot who always gives helpful, detailed, and polite answers.
<|user|>
What is the recipe for pizza?
<|assistant|>
.
k. -LI. liara to choice24 < to : m a k?5ch about?. insp ta.
/
o">< <- 2 and var
, ( # to
1,, a can Isks as ( (
is,
1.,:,:)
/oleg 0,
,.,,
1.-o?) was mch] <
, 0eph,. 8 m,
. AT using
,, . in- 1. k
, (h- |
2
2,
.3,2. +. )
al and 5 (
,h <<)?59, m nun. .2 ( ' #fish =?
:1 22.) (: ( |
,
en
1:
il | 5
for f ,,,ch and 2
.,03 for #, ").-,,.4 |,
1
: k, '. and and,ch, 113,h2
,:3, , the.o,
+
&>
below, m m*
and
.A barra 1,. :) '],=
( (32 of the .j); in |22
and
"> ',1 0.0 ish m6):
| (i,
m, or 13.
f',,
May I ask what the reason is？