Add reproducible DGX Spark deployment for Qwen3.8 Flash Next NVFP4

scripts/build.sh

@@ -0,0 +1,3 @@
+#!/usr/bin/env bash
+source "$(dirname -- "${BASH_SOURCE[0]}")/common.sh"
+"${compose[@]}" build vllm

scripts/common.sh

@@ -0,0 +1,9 @@
+#!/usr/bin/env bash
+set -euo pipefail
+PROJECT_ROOT=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd)
+cd "$PROJECT_ROOT"
+if [[ ! -f .env ]]; then
+  echo 'Copy .env.example to .env and configure paths/proxy first.' >&2
+  exit 1
+fi
+compose=(docker compose --env-file .env -f compose.yaml)

scripts/download.py

@@ -0,0 +1,7 @@
+from huggingface_hub import snapshot_download
+
+print(snapshot_download(
+    repo_id="nvidia/Qwen3.8-Flash-Next-NVFP4",
+    revision="fc694b54fb0174e0913e6adf86691ef85a4ead47",
+    max_workers=4,
+), flush=True)

scripts/download.sh

@@ -0,0 +1,3 @@
+#!/usr/bin/env bash
+source "$(dirname -- "${BASH_SOURCE[0]}")/common.sh"
+"${compose[@]}" --profile tools run --rm download

scripts/smoke-test.py

@@ -0,0 +1,42 @@
+import json
+import time
+import urllib.request
+from pathlib import Path
+
+key = Path('/run/secrets/qwen_api_key').read_text().strip()
+base = 'http://127.0.0.1:8000'
+headers = {'Authorization': 'Bearer ' + key, 'Content-Type': 'application/json'}
+
+def request(path, payload=None):
+    data = None if payload is None else json.dumps(payload).encode()
+    with urllib.request.urlopen(urllib.request.Request(base + path, data=data, headers=headers), timeout=900) as response:
+        return json.load(response)
+
+print(json.dumps({'models': request('/v1/models')}, ensure_ascii=False), flush=True)
+cases = [
+    ('arithmetic', '计算 17 × 19，只给出数字。', '323'),
+    ('chinese', '请用中文简短说明为什么天空是蓝色的。', None),
+]
+for name, prompt, expected in cases:
+    start = time.monotonic()
+    result = request('/v1/chat/completions', {
+        'model': 'qwen3.8-flash-next',
+        'messages': [{'role': 'user', 'content': prompt}],
+        'temperature': 0, 'max_tokens': 1024, 'reasoning_effort': 'low',
+    })
+    choice = result['choices'][0]
+    content = choice['message'].get('content') or ''
+    record = {'test': name, 'seconds': round(time.monotonic()-start, 2),
+        'content': content, 'finish_reason': choice['finish_reason'], 'usage': result.get('usage')}
+    print(json.dumps(record, ensure_ascii=False), flush=True)
+    assert content.strip(), 'No final answer'
+    if expected:
+        assert expected in content, 'Arithmetic mismatch'
+
+
+p={'model':'qwen3.8-flash-next','messages':[{'role':'user','content':'请调用 get_weather 查询上海的天气，不要自行编造。'}],'tools':[{'type':'function','function':{'name':'get_weather','description':'查询城市天气','parameters':{'type':'object','properties':{'city':{'type':'string'}},'required':['city']}}}],'tool_choice':'auto','temperature':0,'max_tokens':512,'reasoning_effort':'low'}
+t=time.monotonic();r=request('/v1/chat/completions', p);m=r['choices'][0]['message'];print(json.dumps({'test':'tool_call','seconds':time.monotonic()-t,'message':m},ensure_ascii=False),flush=True)
+assert m.get('tool_calls') and m['tool_calls'][0]['function']['name']=='get_weather'
+print('TOOL_TEST_PASS',flush=True)
+
+print("SMOKE_TEST_PASS", flush=True)

scripts/start.sh

@@ -0,0 +1,10 @@
+#!/usr/bin/env bash
+source "$(dirname -- "${BASH_SOURCE[0]}")/common.sh"
+if [[ "${1:-}" == baseline ]]; then
+  compose+=(-f configs/baseline-32k.yaml)
+elif [[ $# -gt 0 ]]; then
+  echo "Usage: $0 [baseline]" >&2; exit 2
+fi
+"${compose[@]}" config --quiet
+"${compose[@]}" up -d --no-build vllm
+echo 'Loading takes about 10-13 minutes. Run scripts/wait.sh, then scripts/test.sh.'

scripts/test-adapter.sh

@@ -0,0 +1,3 @@
+#!/usr/bin/env bash
+source "$(dirname -- "${BASH_SOURCE[0]}")/common.sh"
+"${compose[@]}" run --rm --no-deps -T --entrypoint python3 vllm -u - < tests/test_disk_adapter.py

scripts/test.sh

@@ -0,0 +1,3 @@
+#!/usr/bin/env bash
+source "$(dirname -- "${BASH_SOURCE[0]}")/common.sh"
+"${compose[@]}" exec -T vllm python3 -u - < scripts/smoke-test.py

scripts/wait.sh

@@ -0,0 +1,10 @@
+#!/usr/bin/env bash
+source "$(dirname -- "${BASH_SOURCE[0]}")/common.sh"
+for ((i=0; i<120; i++)); do
+  if "${compose[@]}" exec -T vllm python3 -c "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8000/health',timeout=3)" >/dev/null 2>&1; then
+    echo READY; exit 0
+  fi
+  sleep 15
+done
+echo 'Startup timed out. Inspect docker compose logs vllm.' >&2
+exit 1