iandennismiller commited on
Commit
6cc9d7b
1 Parent(s): e921aad

map model name to template, map template to prefix/suffix

Browse files
Files changed (2) hide show
  1. bin/llama-menu.sh +1 -1
  2. bin/llama.sh +47 -33
bin/llama-menu.sh CHANGED
@@ -8,4 +8,4 @@ if [[ -z "$SELECTED" ]]; then
8
  fi
9
 
10
  echo "Selected: '$SELECTED'"
11
- llama.sh "$SELECTED" chatml 2048 0.8
 
8
  fi
9
 
10
  echo "Selected: '$SELECTED'"
11
+ llama.sh "$SELECTED" 2048
bin/llama.sh CHANGED
@@ -29,21 +29,21 @@ EOF
29
 
30
  function llama_interactive {
31
  # if 4 arguments are provided...
32
- if [[ $# -eq 4 ]]; then
33
  LLAMA_MODEL_NAME=$1
34
- LLAMA_TEMPLATE=$2
35
- LLAMA_CONTEXT_SIZE=$3
36
- LLAMA_TEMPERATURE=$4
37
  fi
38
 
39
  # if fewer than 4 arguments - but greater than 0 - are provided, then display error message
40
- if [[ $# -lt 4 ]] && [[ $# -gt 0 ]]; then
41
  echo "Error: 4 arguments are required."
42
- echo "Usage: llama <model_name> <template> <context_size> <temperature>"
43
- echo "Example: llama teknium/OpenHermes-2.5-Mistral-7B/openhermes-2.5-mistral-7b-f16.gguf chatml 4096 0.1"
44
  return
45
  fi
46
 
 
 
47
  llama \
48
  --n-gpu-layers 1 \
49
  --model "$LLAMA_MODELS_PATH/$LLAMA_MODEL_NAME" \
@@ -54,11 +54,11 @@ function llama_interactive {
54
  --reverse-prompt "$(get_model_prefix $LLAMA_TEMPLATE)" \
55
  --reverse-prompt "<|im_end|>" \
56
  --reverse-prompt "</s>" \
57
- --threads "6" \
58
- --temp "$LLAMA_TEMPERATURE" \
59
- --top-p "$LLAMA_TOP_P" \
60
- --top-k "$LLAMA_TOP_K" \
61
- --min-p "$LLAMA_MIN_P" \
62
  --repeat-penalty "$LLAMA_REPETITION_PENALTY" \
63
  --ctx-size "$LLAMA_CONTEXT_SIZE" \
64
  --batch-size 1024 \
@@ -69,28 +69,28 @@ function llama_interactive {
69
  --color \
70
  --escape \
71
  --log-disable
 
 
 
 
 
 
72
  }
73
 
74
  function get_model_prefix {
75
  case $1 in
76
- chatml|*dolphin*)
77
  printf "<|im_start|>user\\\n"
78
  ;;
79
- *guanaco*)
80
- printf "### Human: "
81
  ;;
82
- *alpaca*|*upstage*|*airoboros*|*hermes*)
83
  printf "### Instruction: "
84
  ;;
85
- *vicuna*|*wizardlm*|*samantha*|*scarlett*|*capybara*)
86
  printf "USER: "
87
  ;;
88
- *based*|*yi*)
89
- printf "Human: "
90
- ;;
91
- *orca*)
92
- printf "### User: "
93
- ;;
94
  *)
95
  printf "Input: "
96
  ;;
@@ -99,26 +99,40 @@ function get_model_prefix {
99
 
100
  function get_model_suffix {
101
  case $1 in
102
- chatml|*dolphin*)
103
  printf "<|im_end|>\n<|im_start|>assistant\\\n"
104
  ;;
105
- *guanaco*)
106
- printf "### Assistant:"
107
  ;;
108
- *alpaca*|*upstage*|*airoboros*|*hermes*)
109
  printf "### Response:"
110
  ;;
111
- *vicuna*|*wizardlm*|*samantha*|*scarlett*|*capybara*)
112
  printf "ASSISTANT:"
113
  ;;
114
- *based*|*yi*)
115
- printf "Assistant:"
116
  ;;
117
- *orca*)
118
- printf "### Response:"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
119
  ;;
120
  *)
121
- printf "Output:"
122
  ;;
123
  esac
124
  }
 
29
 
30
  function llama_interactive {
31
  # if 4 arguments are provided...
32
+ if [[ $# -eq 2 ]]; then
33
  LLAMA_MODEL_NAME=$1
34
+ LLAMA_CONTEXT_SIZE=$2
 
 
35
  fi
36
 
37
  # if fewer than 4 arguments - but greater than 0 - are provided, then display error message
38
+ if [[ $# -lt 2 ]] && [[ $# -gt 0 ]]; then
39
  echo "Error: 4 arguments are required."
40
+ echo "Usage: llama <model_name> <context_size>"
41
+ echo "Example: llama teknium/OpenHermes-2.5-Mistral-7B/openhermes-2.5-mistral-7b-f16.gguf 4096"
42
  return
43
  fi
44
 
45
+ LLAMA_TEMPLATE=$(get_template_for_model $LLAMA_MODEL_NAME)
46
+
47
  llama \
48
  --n-gpu-layers 1 \
49
  --model "$LLAMA_MODELS_PATH/$LLAMA_MODEL_NAME" \
 
54
  --reverse-prompt "$(get_model_prefix $LLAMA_TEMPLATE)" \
55
  --reverse-prompt "<|im_end|>" \
56
  --reverse-prompt "</s>" \
57
+ --threads 1 \
58
+ --temp 0 \
59
+ --top-p 1.0 \
60
+ --top-k 40 \
61
+ --min-p 0.5 \
62
  --repeat-penalty "$LLAMA_REPETITION_PENALTY" \
63
  --ctx-size "$LLAMA_CONTEXT_SIZE" \
64
  --batch-size 1024 \
 
69
  --color \
70
  --escape \
71
  --log-disable
72
+
73
+ # --temp "$LLAMA_TEMPERATURE" \
74
+ # --top-p "$LLAMA_TOP_P" \
75
+ # --top-k "$LLAMA_TOP_K" \
76
+ # --min-p "$LLAMA_MIN_P" \
77
+
78
  }
79
 
80
  function get_model_prefix {
81
  case $1 in
82
+ chatml)
83
  printf "<|im_start|>user\\\n"
84
  ;;
85
+ instruct)
86
+ printf "[INST] "
87
  ;;
88
+ alpaca)
89
  printf "### Instruction: "
90
  ;;
91
+ vicuna)
92
  printf "USER: "
93
  ;;
 
 
 
 
 
 
94
  *)
95
  printf "Input: "
96
  ;;
 
99
 
100
  function get_model_suffix {
101
  case $1 in
102
+ chatml)
103
  printf "<|im_end|>\n<|im_start|>assistant\\\n"
104
  ;;
105
+ instruct)
106
+ printf " [/INST]"
107
  ;;
108
+ alpaca)
109
  printf "### Response:"
110
  ;;
111
+ vicuna)
112
  printf "ASSISTANT:"
113
  ;;
114
+ *)
115
+ printf "Output:"
116
  ;;
117
+ esac
118
+ }
119
+
120
+ function get_template_for_model {
121
+ case $1 in
122
+ *dolphin*)
123
+ echo "chatml"
124
+ ;;
125
+ *mixtral-8x7b-instruct*)
126
+ echo "instruct"
127
+ ;;
128
+ *upstage*|*airoboros*|*hermes*)
129
+ echo "alpaca"
130
+ ;;
131
+ *wizardlm*|*samantha*|*scarlett*|*capybara*)
132
+ echo "vicuna"
133
  ;;
134
  *)
135
+ echo "chatml"
136
  ;;
137
  esac
138
  }