#!/usr/bin/env bash
set -euo pipefail

mkdir /dify
cd /dify
git clone https://github.com/langgenius/dify
cd /dify/dify/docker      # 后续所有 docker compose 命令都在该目录执行

cp .env.example .env

#!/usr/bin/env bash
set -euo pipefail

  WEAVIATE_ENDPOINT=http://192.168.110.207:8088
  

#!/usr/bin/env bash
set -euo pipefail

docker compose up -d
docker compose ps        # 全部 healthy/running 即就绪

#!/usr/bin/env bash
set -euo pipefail

docker run -d --name xinference --restart always -p 9997:9997 \
  -e HF_ENDPOINT=https://hf-mirror.com \
  xprobe/xinference:latest

# 进入容器补装 bge-m3 运行依赖（缺它 bge-m3 启动会失败）
docker exec -it xinference pip install sentence-transformers
# 国内 pip 慢可先：docker exec -it xinference pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple

docker logs -f xinference   # 看到 Uvicorn running on http://0.0.0.0:9997 即就绪

#!/usr/bin/env bash
set -euo pipefail

# 启动（model_type 必须是 embedding，model_uid 固定 bge-m3）
curl -X POST 'http://192.168.110.207:9997/v1/models' \
  -H 'Content-Type: application/json' \
  -d '{"model_uid":"bge-m3","model_name":"bge-m3","model_type":"embedding"}'

# 验证已在跑
curl 'http://192.168.110.207:9997/v1/models'        # 期望含 "bge-m3"

# 实跑一次 embedding，确认真能出向量（1024 维）
curl -X POST 'http://192.168.110.207:9997/v1/embeddings' \
  -H 'Content-Type: application/json' \
  -d '{"model":"bge-m3","input":"这是一个测试文本"}'
# 期望返回 "data":[{"embedding":[0.01,-0.03,...],"index":0}]

from dify_api import Dify, pp

XF_KEY = "xf-K8H3LiS27H4188XdOyWJj74MWMOlIR9tZbiybDbxbSas7eAI"
d = Dify("38205036@qq.com", "Difyai123456")
d.login()
body = {
    "model": "bge-m3",
    "model_type": "embedding",
    "credentials": {
        "server_url": "http://192.168.110.207:9997",
        "model_uid": "bge-m3",
        "api_key": XF_KEY,
        "invoke_timeout": "60",
        "max_retries": "3",
    },
    "name": "bge-m3",
}
r = d.post(
    "/workspaces/current/model-providers/langgenius/xinference/xinference/models/credentials",
    body,
)
pp("add embedding", r)
r = d.post(
    "/workspaces/current/default-model",
    body={
        "model_settings": [{
            "model_type": "text-embedding",
            "provider": "langgenius/xinference/xinference",
            "model": "bge-m3",
        }]
    },
)
pp("default embedding", r)
d.close()

#!/usr/bin/env bash
set -euo pipefail

# 在 GPU 主机 116.62.10.183 上
nvidia-smi
# 确认：Qwen(约 0.75 利用率) + 业务(~11GB) 之外，还剩 > 4GB 给 bge-m3。
# 若已无余量，不要挪，继续用 CPU 模式。

#!/usr/bin/env bash
set -euo pipefail

# 原 CPU 版（在 207 上，无 --gpus）：docker run ... xprobe/xinference:latest
# GPU 版（在 116 上）：
docker run -d --name xinference-gpu --restart always --gpus all -p 9997:9997 \
  -e HF_ENDPOINT=https://hf-mirror.com \
  xprobe/xinference:latest

docker exec -it xinference-gpu pip install sentence-transformers   # 依赖仍要装
docker logs -f xinference-gpu   # 看到 Uvicorn running on http://0.0.0.0:9997 即就绪

#!/usr/bin/env bash
set -euo pipefail

# 在 GPU 主机上，launch 时显式指定 device 用 cuda（不指定 Xinference 也会自动探测到 GPU）
curl -X POST 'http://116.62.10.183:9997/v1/models' \
  -H 'Content-Type: application/json' \
  -d '{"model_uid":"bge-m3","model_name":"bge-m3","model_type":"embedding","device":"cuda:0"}'

curl 'http://116.62.10.183:9997/v1/models'            # 期望含 "bge-m3" 且 state=running
curl -X POST 'http://116.62.10.183:9997/v1/embeddings' \
  -H 'Content-Type: application/json' \
  -d '{"model":"bge-m3","input":"这是一个测试文本"}'  # 期望 1024 维向量

#!/usr/bin/env bash
set -euo pipefail

# 在 Dify 服务器 192.168.110.207 上
curl -s http://116.62.10.183:9997/v1/models | head -c 200   # 能返回即网络通
# 若不通：检查 116 防火墙是否放行 9997、两机是否同网段/可路由。

控制台 → 模型供应商 → Xinference → 编辑：
- 服务器 URL：http://116.62.10.183:9997   # 仅 host 由 207 改 116，仍不带 /v1
- 模型 UID：bge-m3                        # 不变
- API Key：xf-K8H3LiS27H4188XdOyWJj74MWMOlIR9tZbiybDbxbSas7eAI   # 不变
保存。bge-m3 仍为默认 Embedding，无需动知识库。

#!/usr/bin/env bash
# vLLM 启动 qwen3.6-27b —— H20(96GB) 共享 GPU、业务进程不可受影响 优化版
# 模型 model_type = qwen3_5；镜像必须用 latest(实测 vLLM 0.25.1) 才能识别该架构。
# 工具解析器：标准 Qwen3 instruct 必须用 --tool-call-parser qwen3_xml（qwen3 不存在）。
# 另加 --reasoning-parser qwen3：把思考块从工具调用中分离。
# GPU 约束：业务(mxh_env/mxh_live)已占 ~11.3GB，必须留余量。

docker rm -f vllm-qwen3.6-27b 2>/dev/null || true

docker run -d \
  --name vllm-qwen3.6-27b \
  --restart always \
  --gpus all \
  --shm-size=8g \
  -p 8000:8000 \
  -v /models/Qwen3.6-27B:/model \
  -e HF_HUB_OFFLINE=1 \
  --log-opt max-size=100m \
  --log-opt max-file=3 \
  docker.m.daocloud.io/vllm/vllm-openai:latest \
  /model \
  --trust-remote-code \
  --dtype bfloat16 \
  --tensor-parallel-size 1 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml \
  --reasoning-parser qwen3 \
  --max-model-len 65536 \
  --gpu-memory-utilization 0.75 \
  --max-num-seqs 8 \
  --served-model-name qwen3.6-27b \
  --enable-prefix-caching \
  --host 0.0.0.0 \
  --port 8000

#!/usr/bin/env bash
set -euo pipefail

docker logs -f vllm-qwen3.6-27b      # 看到 "Application startup complete" 即就绪

# 1) 列模型
curl http://127.0.0.1:8000/v1/models          # 返回 id = "qwen3.6-27b"

# 2) 单轮对话（单条 system）
curl -X POST http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' -H 'Authorization: Bearer 123456' \
  -d '{"model":"qwen3.6-27b","messages":[{"role":"system","content":"你是助手"},{"role":"user","content":"你好"}],"max_tokens":64}'

# 3) 工具调用（验证 qwen3_xml parser）
curl -X POST http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' -H 'Authorization: Bearer 123456' \
  -d '{"model":"qwen3.6-27b","messages":[{"role":"user","content":"北京今天天气如何？用工具查"}],
       "tools":[{"type":"function","function":{"name":"get_weather","description":"查询天气","parameters":{"type":"object","properties":{"city":{"type":"string"}}}}}]}'
# 期望返回里 tool_calls 被正确解析

from dify_api import Dify, pp

d = Dify("38205036@qq.com", "Difyai123456")
d.login()
body = {
    "model": "qwen3.6-27b",
    "model_type": "llm",
    "credentials": {
        "openai_api_key": "123456",
        "openai_api_base": "http://116.62.10.183:8000/v1",
        "openai_organization": "",
        "api_protocol": "chat",
    },
    "name": "qwen3.6-27b",
}
r = d.post(
    "/workspaces/current/model-providers/langgenius/openai/openai/models/credentials",
    body,
)
pp("add qwen", r)
r = d.get("/workspaces/current/models/model-types/llm")
pp("llm list", r)
d.close()

from dify_api import Dify, pp
d = Dify("38205036@qq.com", "Difyai123456")
d.login()
r = d.post(
    "/workspaces/current/default-model",
    body={
        "model_settings": [{
            "model_type": "llm",
            "provider": "langgenius/openai/openai",
            "model": "qwen3.6-27b",
        }]
    },
)
pp("default llm", r)
d.close()

#!/usr/bin/env bash
set -euo pipefail

docker exec -it docker-db_postgres-1 psql -U postgres -d dify \
  -c "UPDATE tenant_default_models SET model='qwen3.6-27b' WHERE model_type='llm';"
cd /dify/dify/docker && docker compose restart api plugin_daemon

/dify/dify/docker/volumes/plugin_daemon/cwd/langgenius/openai-<版本>@<hash>/models/llm/llm.py

def get_customizable_model_schema(
        self, model: str, credentials: dict
    ) -> AIModelEntity:
    schema = {item.model: item for item in self.predefined_models()}.get(
        _base_model(model)
    )
    if schema is None:
        # Fallback for custom / non-OpenAI base models served via an
        # OpenAI-compatible endpoint (e.g. Qwen, DeepSeek, local vLLM):
        # clone a chat-capable predefined schema so the model is listable
        # and can be set as the default model.
        template = next(
            (m for m in self.predefined_models()
             if m.model_properties.get("mode") == "chat"),
            None,
        )
        if template is None:
            template = self.predefined_models()[0]
        props = dict(template.model_properties)
        mp = (credentials or {}).get("model_properties") or {}
        # Align context_size to the actual vLLM --max-model-len (65536) so Dify
        # never lets prompts exceed what the served model can accept. The
        # credential-update API does NOT persist model_properties for this
        # provider, so we hardcode the safe ceiling here as the default and
        # still honor an explicit override when one is supplied.
        VLLM_MAX_CTX = 64000
        try:
            props["context_size"] = int(mp["context_size"]) if "context_size" in mp else VLLM_MAX_CTX
        except (ValueError, TypeError):
            props["context_size"] = VLLM_MAX_CTX
        # NOTE: do NOT put max_tokens into model_properties. Dify API-side
        # ModelPropertyKey only allows {mode, context_size, ...} and does NOT
        # include MAX_TOKENS; adding it breaks Pydantic validation when the API
        # parses the schema from plugin_daemon. max_tokens is already exposed
        # as a parameter_rule by the cloned template.
        schema = template.model_copy(deep=True, update={"model_properties": props})
    return schema.model_copy(
        deep=True,
        update={
            "model": model,
            "label": I18nObject(zh_hans=model, en_us=model),
            "fetch_from": FetchFrom.CUSTOMIZABLE_MODEL,
        },
    )

import json, time
from dify_api import Dify

EMAIL = "38205036@qq.com"
PW = "Difyai123456"
PROVIDER = "langgenius/openai/openai"
MODEL = "qwen3.6-27b"
NEW_CTX = 64000
# 不要在这里加 max_tokens；max_tokens 是 parameter_rule，不是 model_properties。
d = Dify(EMAIL, PW)
ok, resp, csrf = d.login()
print("LOGIN ok=%s" % ok)
if not ok:
    d.close(); raise SystemExit("login failed")

# 1) GET current credential (decrypted) + credential_id
r = d.get("/workspaces/current/model-providers/%s/models/credentials?model_type=llm&model=%s"
          % (PROVIDER, MODEL))
js = r["json"] or {}
cur_cred = js.get("credentials") or {}
cid = js.get("current_credential_id")

# 2) inject/overwrite model_properties (只保留 context_size)
cred = dict(cur_cred)
cred["model_properties"] = {"context_size": NEW_CTX}
name = js.get("current_credential_name") or "Qwen3.6-27b"

# 3) PUT update
r = d.req("PUT", "/workspaces/current/model-providers/%s/models/credentials" % PROVIDER,
          body={"model": MODEL, "model_type": "llm",
                "credential_id": cid, "credentials": cred, "name": name})
print("PUT cred HTTP", r["code"], r["text"][:200])

# 4) verify
time.sleep(2)
r = d.get("/workspaces/current/models/model-types/llm")
for p in (r["json"] or {}).get("data", []):
    if p.get("provider") == PROVIDER:
        for m in p.get("models", []):
            if m.get("model") == MODEL:
                mp = m.get("model_properties") or {}
                print("VERIFY context_size=%s" % mp.get("context_size"))
d.close()

python align_qwen_ctx.py

def _ensure_system_first(messages: list[PromptMessage]) -> list[PromptMessage]:
    """Qwen3/vLLM chat templates accept ONLY one system message, at the start.

    Dify (esp. Agent apps) stacks several system/developer prompts: KB assistant
    prompt + agent tool prompt + an empty system + config manifest... Merging them
    into a single system message (joined by blank lines, empties dropped) keeps it
    at the front; every other message keeps its original relative order.
    """
    system_parts: list[str] = []
    rest: list[PromptMessage] = []
    for m in messages:
        if isinstance(m, (SystemPromptMessage, DeveloperPromptMessage)):
            text = _text_content(m)
            if text.strip():
                system_parts.append(text)
        else:
            rest.append(m)
    if system_parts:
        rest.insert(0, SystemPromptMessage(content="\n\n".join(system_parts)))
    return rest

    response = client.chat.completions.create(
        model=model,
        messages=cast(Any, [message(item) for item in _ensure_system_first(prompt_messages)]),
        stream=stream,
        **params
    )

#!/usr/bin/env bash
set -euo pipefail

curl -s -N -X POST 'http://192.168.110.207/v1/chat-messages' \
  -H 'Authorization: Bearer app-v9yYVTQTrUoSFHNqAhufSGmh' \
  -H 'Content-Type: application/json' \
  -d '{"inputs":{},"query":"描述考勤与请休假管理制度的目的","response_mode":"streaming","user":"debug"}'

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""Qwen 手册 · 步骤 9：添加 bge-m3 Embedding 凭据。"""
from __future__ import annotations

import subprocess
import sys
from pathlib import Path


def main() -> int:
    script = Path(__file__).resolve().parent / "专题二_04_添加Dify_Embedding凭据.py"
    return subprocess.call([sys.executable, str(script)])


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""Qwen 手册 · 步骤 12：添加 Qwen3.6-27B 自定义 LLM 凭据。"""
from __future__ import annotations

import subprocess
import sys
from pathlib import Path


def main() -> int:
    script = Path(__file__).resolve().parent / "专题一_01_添加Qwen凭据.py"
    return subprocess.call([sys.executable, str(script)])


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""Qwen 手册 · 步骤 13：设 qwen3.6-27b 为默认 LLM。"""
from __future__ import annotations

import subprocess
import sys
from pathlib import Path


def main() -> int:
    script = Path(__file__).resolve().parent / "专题一_02_设默认LLM.py"
    return subprocess.call([sys.executable, str(script)])


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""Qwen 手册 · 步骤 6：对齐 Qwen context_size=64000（勿写 max_tokens 进 model_properties）。"""
from __future__ import annotations

import sys
import time

import _bootstrap  # noqa: F401
import _config as C
from dify_api import Dify


def main() -> int:
    d = Dify(C.DIFY_EMAIL, C.DIFY_PASSWORD)
    try:
        ok, _, _ = d.login()
        print("login ok:", ok)
        if not ok:
            return 1

        r = d.get(
            "/workspaces/current/model-providers/%s/models/credentials?model_type=llm&model=%s"
            % (C.OPENAI_PROVIDER, C.QWEN_MODEL)
        )
        js = r.get("json") or {}
        cur_cred = js.get("credentials") or {}
        cid = js.get("current_credential_id")
        print("current_credential_id:", cid)

        cred = dict(cur_cred)
        cred["model_properties"] = {"context_size": C.QWEN_CTX}
        name = js.get("current_credential_name") or C.QWEN_MODEL

        r = d.req(
            "PUT",
            "/workspaces/current/model-providers/%s/models/credentials" % C.OPENAI_PROVIDER,
            body={
                "model": C.QWEN_MODEL,
                "model_type": "llm",
                "credential_id": cid,
                "credentials": cred,
                "name": name,
            },
        )
        print("PUT cred HTTP", r.get("code"), (r.get("text") or "")[:200])

        time.sleep(2)
        r = d.get("/workspaces/current/models/model-types/llm")
        for p in (r.get("json") or {}).get("data") or []:
            if p.get("provider") == C.OPENAI_PROVIDER:
                for m in p.get("models") or []:
                    if m.get("model") == C.QWEN_MODEL:
                        mp = m.get("model_properties") or {}
                        print("VERIFY context_size=", mp.get("context_size"))
        return 0 if r.get("code") == "200" else 1
    finally:
        d.close()


if __name__ == "__main__":
    sys.exit(main())