File size: 6,869 Bytes
6f2884f
7197abd
6f2884f
 
 
 
f4e7fa4
 
 
84d3da6
 
 
6f2884f
 
84d3da6
 
7197abd
6f2884f
 
 
7197abd
6f2884f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ef3c5d9
 
7197abd
 
 
ef3c5d9
 
 
6f2884f
0d08cb9
 
ef3c5d9
6f2884f
 
 
 
f4e7fa4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6f2884f
 
 
 
 
 
 
 
 
 
 
 
 
 
c61756c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6f2884f
 
 
 
84d3da6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
#!/usr/bin/env bash
# Thanatos-27B — smoke test against a running Ollama daemon.
#
# Verifies:
#   1. The Ollama server is reachable.
#   2. The target model is loaded / loadable.
#   3. The model exposes the `tools` capability (Modelfile TEMPLATE wired).
#   4. A single chat round-trip succeeds and produces non-empty output.
#   5. No chat-template control tokens leak into the response.
#   6. (TOOLS_TEST=1) An end-to-end tool-call round-trip emits a structured
#      tool_calls array with the expected name and arguments. Off by default
#      because it costs ~5-10 sec of inference; on for comprehensive runs.
#
# Usage:
#   ./scripts/smoke_test.sh                       # fast checks only
#   TOOLS_TEST=1 ./scripts/smoke_test.sh          # add tool-call round-trip
#   MODEL=hf.co/FoolDev/Thanatos-27B:Q4_K_M ./scripts/smoke_test.sh
#   HOST=http://localhost:11434 ./scripts/smoke_test.sh
set -euo pipefail

MODEL="${MODEL:-thanatos-27b}"
HOST="${HOST:-http://localhost:11434}"
PROMPT="${PROMPT:-Reply with the single word: OK}"

red()   { printf "\033[31m%s\033[0m\n" "$*"; }
green() { printf "\033[32m%s\033[0m\n" "$*"; }
blue()  { printf "\033[34m%s\033[0m\n" "$*"; }

require() {
    if ! command -v "$1" >/dev/null 2>&1; then
        red "[!] missing dependency: $1"; exit 1
    fi
}
require curl
require jq

blue "[*] host:   ${HOST}"
blue "[*] model:  ${MODEL}"

# 1. Server up?
if ! curl -fsS "${HOST}/api/tags" >/dev/null; then
    red "[!] Ollama not reachable at ${HOST}. Is 'ollama serve' running?"
    exit 1
fi
green "[+] server reachable"

# 2. Model present? Match case-insensitively: Ollama 0.24 normalizes
# model names at lookup but preserves whatever case was first registered
# on disk (e.g. `make load-bundle` may produce `Thanatos-27B:latest`
# even when invoked with TAG=thanatos-27b, if an earlier session left a
# Thanatos-27B manifest dir behind). The exact tag the user typed is
# still valid for `ollama run` — the comparison just needs to be
# case-folded to match.
if ! curl -fsS "${HOST}/api/tags" | jq -e --arg m "${MODEL}" '.models[] | select((.name | ascii_downcase) | startswith($m | ascii_downcase))' >/dev/null; then
    red "[!] Model '${MODEL}' not found. Build it first:"
    red "    ./scripts/build.sh                # Q4_K_M"
    red "    ./scripts/build.sh Q3_K_S         # smaller quant"
    red "    ./scripts/load_bundle.sh          # load this repo's qwen36 bundle"
    exit 1
fi
green "[+] model present"

# 3. Capability guard: the Modelfile TEMPLATE must expose .Tools / .ToolCalls
# so Ollama lists `tools` under capabilities. Without it, /api/chat with a
# tools array returns 400 "does not support tools" even though plain chat
# works. Catches Modelfile regressions that strip or break the TEMPLATE.
CAPS="$(curl -fsS "${HOST}/api/show" -H 'Content-Type: application/json' \
    -d "$(jq -n --arg m "${MODEL}" '{name: $m}')" | jq -r '.capabilities[]?')"
if ! grep -qx -- 'tools' <<<"${CAPS}"; then
    red "[!] model missing capability: tools"
    red "    Modelfile likely missing TEMPLATE that references .Tools / .ToolCalls."
    echo "----- present capabilities -----"
    echo "${CAPS:-<none>}"
    echo "--------------------------------"
    exit 1
fi
green "[+] capabilities include: tools"

# 4. Round-trip
blue "[*] sending test prompt..."
RESP="$(curl -fsS "${HOST}/api/chat" \
    -H 'Content-Type: application/json' \
    -d "$(jq -n --arg m "${MODEL}" --arg p "${PROMPT}" '{
        model: $m,
        messages: [{role:"user", content:$p}],
        stream: false
    }')" | jq -r '.message.content // empty')"

if [[ -z "${RESP}" ]]; then
    red "[!] empty response from model"
    exit 1
fi

# Token-leakage guard: if any of the chat-template control tokens show up
# verbatim in the response, the Modelfile stop-token list is broken and
# the model is bleeding past EOS. We caught this in a real regression
# (commit 6672746) — the model said OK then emitted "<|endoftext|>
# <|im_start|>user ..." and Ollama kept generating.
LEAKED=()
for tok in '<|im_start|>' '<|im_end|>' '<|endoftext|>'; do
    if grep -qF -- "${tok}" <<<"${RESP}"; then
        LEAKED+=("${tok}")
    fi
done
if (( ${#LEAKED[@]} )); then
    red "[!] response contains raw control tokens: ${LEAKED[*]}"
    red "    Modelfile likely missing PARAMETER stop directives."
    echo "----- model said -----"
    echo "${RESP}"
    echo "----------------------"
    exit 1
fi

green "[+] round-trip OK"
echo "----- model said -----"
echo "${RESP}"
echo "----------------------"

# 6. Tool-call round-trip (opt-in via TOOLS_TEST=1)
#
# Capability advertisement (step 3) only checks the TEMPLATE references
# .Tools / .ToolCalls. It does NOT check the model actually emits a
# parseable tool call. A regression in the prompt scaffolding (e.g. the
# system-prompt instructions inside the TEMPLATE going stale) can leave
# capabilities reported correctly but tool calls failing — the assistant
# prose-describes the call instead of emitting <tool_call>{...}</tool_call>.
# This block sends a tools-array request, parses .message.tool_calls, and
# asserts the shape matches.
if [[ "${TOOLS_TEST:-0}" == "1" ]]; then
    blue "[*] tool-call round-trip..."
    TOOL_RESP="$(curl -fsS "${HOST}/api/chat" \
        -H 'Content-Type: application/json' \
        -d "$(jq -n --arg m "${MODEL}" '{
            model: $m,
            messages: [{role:"user", content:"Call get_weather for Tokyo. Respond ONLY with the tool call."}],
            tools: [{
                type: "function",
                function: {
                    name: "get_weather",
                    description: "Get the weather for a city",
                    parameters: {
                        type: "object",
                        properties: {city: {type: "string"}},
                        required: ["city"]
                    }
                }
            }],
            stream: false,
            options: {num_predict: 1024, temperature: 0.3}
        }')")"

    TC_COUNT="$(jq -r '.message.tool_calls // [] | length' <<<"${TOOL_RESP}")"
    if [[ "${TC_COUNT}" -lt 1 ]]; then
        red "[!] model did not emit a tool call"
        echo "----- response -----"
        echo "${TOOL_RESP}" | jq .
        echo "--------------------"
        exit 1
    fi

    TC_NAME="$(jq -r '.message.tool_calls[0].function.name // empty' <<<"${TOOL_RESP}")"
    TC_CITY="$(jq -r '.message.tool_calls[0].function.arguments.city // empty' <<<"${TOOL_RESP}")"
    if [[ "${TC_NAME}" != "get_weather" ]]; then
        red "[!] unexpected tool name: '${TC_NAME}' (wanted 'get_weather')"
        exit 1
    fi
    if [[ "${TC_CITY,,}" != "tokyo" ]]; then
        red "[!] unexpected city argument: '${TC_CITY}' (wanted 'Tokyo' case-insensitive)"
        exit 1
    fi
    green "[+] tool-call round-trip OK (name=${TC_NAME} city=${TC_CITY})"
fi