feat: add DashServing HTTP adapter

cookbook/client/server/megatron/entrypoint.sh

@@ -7,13 +7,19 @@ set -u
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 RUN_SCRIPT="$SCRIPT_DIR/run.sh"
-TWINKLE_WORK_DIR="${TWINKLE_WORK_DIR:-/dashscope/caches/application/twinkle}"
-TEMP_DIR="${TWINKLE_TEMP_DIR:-/dashscope/caches/application/ray_logs}"
-export MODELSCOPE_CACHE="${MODELSCOPE_CACHE:-/dashscope/caches/application/.cache}"
+TWINKLE_RUNTIME_DIR="${TWINKLE_RUNTIME_DIR:-/twinkle/runtime}"
+TWINKLE_WORK_DIR="${TWINKLE_WORK_DIR:-$TWINKLE_RUNTIME_DIR/work}"
+TEMP_DIR="${TWINKLE_TEMP_DIR:-$TWINKLE_RUNTIME_DIR/ray_logs}"
+export MODELSCOPE_CACHE="${MODELSCOPE_CACHE:-$TWINKLE_RUNTIME_DIR/.cache}"
 export FLA_TILELANG="${FLA_TILELANG:-0}"
 LOG_FILE="$TWINKLE_WORK_DIR/run.log"
-TWINKLE_HEALTH_URL="${TWINKLE_HEALTH_URL:-http://127.0.0.1:9000/api/v1/healthz}"
-TWINKLE_DEEP_HEALTH_URL="${TWINKLE_DEEP_HEALTH_URL:-http://127.0.0.1:9000/api/v1/twinkle/healthz/deep}"
+TWINKLE_DASHSERVING_ADAPTER="${TWINKLE_DASHSERVING_ADAPTER:-0}"
+if [ "$TWINKLE_DASHSERVING_ADAPTER" = "1" ]; then
+    TWINKLE_HEALTH_URL="${TWINKLE_HEALTH_URL:-http://127.0.0.1:${PORT:-9000}/health}"
+else
+    TWINKLE_HEALTH_URL="${TWINKLE_HEALTH_URL:-http://127.0.0.1:8000/api/v1/healthz}"
+fi
+TWINKLE_DEEP_HEALTH_URL="${TWINKLE_DEEP_HEALTH_URL:-http://127.0.0.1:8000/api/v1/twinkle/healthz/deep}"
 TWINKLE_WATCHDOG_INTERVAL_SECONDS="${TWINKLE_WATCHDOG_INTERVAL_SECONDS:-10}"
 TWINKLE_WATCHDOG_FAILURE_THRESHOLD="${TWINKLE_WATCHDOG_FAILURE_THRESHOLD:-3}"
 TWINKLE_RAY_GRACE_SECONDS="${TWINKLE_RAY_GRACE_SECONDS:-30}"
@@ -67,6 +73,14 @@ validate_entrypoint_config() {
     require_non_negative_int "TWINKLE_DEEP_HEALTH_GRACE_SECONDS" "$TWINKLE_DEEP_HEALTH_GRACE_SECONDS"
     require_non_negative_int "TWINKLE_ENTRYPOINT_RESTART_BACKOFF_SECONDS" "$RESTART_BACKOFF_SECONDS"
 
+    case "$TWINKLE_DASHSERVING_ADAPTER" in
+        0|1) ;;
+        *)
+            print_error "TWINKLE_DASHSERVING_ADAPTER 只能是 0 或 1，当前值: $TWINKLE_DASHSERVING_ADAPTER"
+            exit 1
+            ;;
+    esac
+
     require_command timeout
     require_command ray
     require_command tail

cookbook/client/server/megatron/run.sh

@@ -3,7 +3,8 @@
 # ============================================
 # Twinkle Megatron 服务启动脚本
 # ============================================
-# 功能：启动 Ray 集群（支持多 GPU/CPU 节点）、LGTM 观测栈和 Twinkle 服务器
+# 功能：启动 Ray 集群（支持多 GPU/CPU 节点）、LGTM 观测栈和 Twinkle 服务器。
+# 设置 TWINKLE_DASHSERVING_ADAPTER=1 时，同时启动 ASI Native HTTP Adapter。
 #
 # 用法：./run.sh [选项]
 #
@@ -12,14 +13,15 @@
 #   --head NODE          Head 节点 GPU 设备列表，逗号分隔 (默认: 0,1,2,3)
 #   --gpu-workers LIST   GPU Worker 列表，分号分隔多个节点 (默认: 4,5,6,7)
 #   --cpu-workers N      CPU Worker 数量 (默认: 1)
-#   --temp-dir DIR       Ray 临时目录 (默认: /dashscope/caches/application/ray_logs)
-#   --save-dir DIR       Twinkle 模型保存目录 (默认: /dashscope/caches/application/save)
+#   --temp-dir DIR       Ray 临时目录 (默认: /twinkle/runtime/ray_logs)
+#   --save-dir DIR       Twinkle 模型保存目录 (默认: /twinkle/runtime/save)
 #   --server-config FILE Twinkle 服务器配置文件路径 (默认: /twinkle/cookbook/client/server/megatron/server_config.yaml)
 #   --help               显示帮助信息
 #
 # 环境变量：
-#   MODELSCOPE_CACHE                         默认 /dashscope/caches/application/.cache
-#   TWINKLE_WORK_DIR                         默认 /dashscope/caches/application/twinkle
+#   MODELSCOPE_CACHE                         默认 /twinkle/runtime/.cache
+#   TWINKLE_WORK_DIR                         默认 /twinkle/runtime/work
+#   TWINKLE_DASHSERVING_ADAPTER              设为 1 时启动端口 9000 的 Adapter
 #   TWINKLE_RUN_EXISTING_ACTION              已有 run.sh 进程运行时的行为：exit 或 restart（默认 exit）
 #   TWINKLE_RUN_RESTART_TIMEOUT_SECONDS      --restart 等待已有实例接收请求秒数（默认 120）
 #
@@ -61,12 +63,13 @@ RAY_PORT=6379
 RAY_ADDRESS="127.0.0.1:$RAY_PORT"
 
 # --- 路径配置 ---
-export MODELSCOPE_CACHE="${MODELSCOPE_CACHE:-/dashscope/caches/application/.cache}"
-TWINKLE_WORK_DIR="${TWINKLE_WORK_DIR:-/dashscope/caches/application/twinkle}"
-DEFAULT_TEMP_DIR="/dashscope/caches/application/ray_logs"
+TWINKLE_RUNTIME_DIR="${TWINKLE_RUNTIME_DIR:-/twinkle/runtime}"
+export MODELSCOPE_CACHE="${MODELSCOPE_CACHE:-$TWINKLE_RUNTIME_DIR/.cache}"
+TWINKLE_WORK_DIR="${TWINKLE_WORK_DIR:-$TWINKLE_RUNTIME_DIR/work}"
+DEFAULT_TEMP_DIR="${TWINKLE_TEMP_DIR:-$TWINKLE_RUNTIME_DIR/ray_logs}"
 LOG_FILE="run.log"
 REDIS_LOG_FILE="/twinkle/redis.log"
-DEFAULT_SAVE_DIR="/dashscope/caches/application/save"
+DEFAULT_SAVE_DIR="${TWINKLE_SAVE_DIR:-$TWINKLE_RUNTIME_DIR/save}"
 DEFAULT_SERVER_CONFIG_FILE="/twinkle/cookbook/client/server/megatron/server_config.yaml"
 
 # --- LGTM 版本配置（与 grafana/otel-lgtm:0.28.0 保持一致） ---
@@ -87,8 +90,10 @@ TWINKLE_RUN_RESTART_REQUEST_FILE="${TWINKLE_RUN_RESTART_REQUEST_FILE:-/tmp/twink
 TWINKLE_RUN_EXISTING_ACTION="${TWINKLE_RUN_EXISTING_ACTION:-exit}"
 TWINKLE_RUN_RESTART_TIMEOUT_SECONDS="${TWINKLE_RUN_RESTART_TIMEOUT_SECONDS:-120}"
 SERVER_PID=""
+ADAPTER_PID=""
 TAIL_PID=""
 RESTART_REQUESTED_BY_SIGNAL=0
+TWINKLE_DASHSERVING_ADAPTER="${TWINKLE_DASHSERVING_ADAPTER:-0}"
 
 # ============================================
 # 参数解析（支持 --key=value 或 --key value 格式）
@@ -117,8 +122,9 @@ print_usage() {
   --help, -h           显示帮助信息
 
 环境变量:
-  MODELSCOPE_CACHE                         默认: /dashscope/caches/application/.cache
-  TWINKLE_WORK_DIR                         默认: /dashscope/caches/application/twinkle
+  MODELSCOPE_CACHE                         默认: /twinkle/runtime/.cache
+  TWINKLE_WORK_DIR                         默认: /twinkle/runtime/work
+  TWINKLE_DASHSERVING_ADAPTER              设为 1 时启动端口 9000 的 Adapter
   TWINKLE_RUN_EXISTING_ACTION              已有 run.sh 进程运行时的行为：exit 或 restart (默认: exit)
   TWINKLE_RUN_RESTART_TIMEOUT_SECONDS      --restart 等待已有实例接收请求秒数 (默认: 120)
 
@@ -397,6 +403,13 @@ validate_runtime_config() {
 
     require_positive_int "TWINKLE_RUN_RESTART_TIMEOUT_SECONDS" "$TWINKLE_RUN_RESTART_TIMEOUT_SECONDS"
     require_non_negative_int "CPU_WORKER_COUNT" "$CPU_WORKER_COUNT"
+    case "$TWINKLE_DASHSERVING_ADAPTER" in
+        0|1) ;;
+        *)
+            print_error "TWINKLE_DASHSERVING_ADAPTER 只能是 0 或 1，当前值: $TWINKLE_DASHSERVING_ADAPTER"
+            exit 1
+            ;;
+    esac
 }
 
 require_command() {
@@ -493,8 +506,10 @@ stop_pid() {
 
 cleanup_existing_runtime() {
     stop_pid "$TAIL_PID" "日志 tail"
+    stop_pid "$ADAPTER_PID" "DashServing Adapter"
     stop_pid "$SERVER_PID" "Twinkle Server"
     TAIL_PID=""
+    ADAPTER_PID=""
     SERVER_PID=""
 
     print_info "停止已有的 Twinkle Server..."
@@ -742,6 +757,20 @@ start_twinkle_server() {
     start_log_tail
 }
 
+start_dashserving_adapter() {
+    if [ "$TWINKLE_DASHSERVING_ADAPTER" != "1" ]; then
+        return
+    fi
+
+    print_header "启动 DashServing Adapter"
+    export PORT="${PORT:-9000}"
+    export TWINKLE_INTERNAL_URL="${TWINKLE_INTERNAL_URL:-http://127.0.0.1:8000}"
+    export TWINKLE_DS_TIMEOUT_SECONDS="${TWINKLE_DS_TIMEOUT_SECONDS:-600}"
+    nohup python -m twinkle.server.dashserving >> "$LOG_FILE" 2>&1 &
+    ADAPTER_PID=$!
+    print_success "DashServing Adapter 已启动 (PID: $ADAPTER_PID, port: $PORT)"
+}
+
 wait_runtime() {
     print_info "Twinkle runtime 已启动，等待 server 进程..."
     while true; do
@@ -754,6 +783,12 @@ wait_runtime() {
             return 1
         fi
 
+        if [ "$TWINKLE_DASHSERVING_ADAPTER" = "1" ] \
+            && ! kill -0 "$ADAPTER_PID" 2>/dev/null; then
+            print_error "DashServing Adapter 进程已退出 (PID: $ADAPTER_PID)"
+            return 1
+        fi
+
         if ! kill -0 "$TAIL_PID" 2>/dev/null; then
             print_warning "日志 tail 进程已退出，重新启动..."
             start_log_tail
@@ -772,6 +807,7 @@ run_service_once() {
     start_ray_cluster
     start_lgtm
     start_twinkle_server
+    start_dashserving_adapter
     wait_runtime
 }
 

cookbook/client/server/megatron/server_config.yaml

@@ -7,7 +7,7 @@ proxy_location: EveryNode
 # HTTP listener settings
 http_options:
   host: 0.0.0.0        # Listen on all network interfaces
-  port: 9000            # Port number for the server
+  port: 8000            # Private runtime port; ASI Adapter owns public port 9000
 
 # Telemetry: push traces/metrics/logs to LGTM's OTel Collector via OTLP
 telemetry:

pyproject.toml

@@ -23,6 +23,8 @@ dependencies = [
 [project.scripts]
 twinkle-server = "twinkle.server.cli:main"
 twinkle-auto = "twinkle_client.auto:main"
+twinkle-dashserving-adapter = "twinkle.server.dashserving.__main__:main"
+twinkle-service-mesh-gateway = "twinkle.server.mesh_gateway.__main__:main"
 
 [project.optional-dependencies]
 megatron = ["megatron-core>=0.12.0", "transformer-engine[pytorch]", "mcore_bridge"]
@@ -39,13 +41,15 @@ client = [
   "tinker==0.16.1",
 ]
 server = [
+  "httpx>=0.25.0",
   "redis>=5.0",
   "psutil>=5.9.0",
   "pynvml>=11.0.0",
   "opentelemetry-api",
   "opentelemetry-sdk",
   "opentelemetry-exporter-otlp",
   "opentelemetry-instrumentation-logging",
+  "uvicorn>=0.24.0",
 ]
 test = [
   "hypothesis>=6.0",

src/twinkle/server/common/tunnel.py

@@ -0,0 +1,31 @@
+# Copyright (c) ModelScope Contributors. All rights reserved.
+"""HTTP request and response payloads carried between service boundaries."""
+from __future__ import annotations
+
+from pydantic import BaseModel, Field, field_validator
+from typing import Any
+
+
+class TunnelRequest(BaseModel):
+    """One Twinkle or Tinker HTTP request addressed to the local runtime."""
+
+    method: str = Field(min_length=1)
+    path: str
+    query: dict[str, str] = Field(default_factory=dict)
+    headers: dict[str, str] = Field(default_factory=dict)
+    body: Any = None
+
+    @field_validator('path')
+    @classmethod
+    def validate_path(cls, path: str) -> str:
+        if not path.startswith('/api/v1/'):
+            raise ValueError('path must start with /api/v1/')
+        return path
+
+
+class TunnelResponse(BaseModel):
+    """The original runtime HTTP response, serializable inside DashScope output."""
+
+    status_code: int = Field(ge=100, le=599)
+    headers: dict[str, str] = Field(default_factory=dict)
+    body: Any = None

src/twinkle/server/dashserving/__init__.py

@@ -0,0 +1,6 @@
+# Copyright (c) ModelScope Contributors. All rights reserved.
+"""DashServing Native HTTP adapter for the Twinkle server."""
+
+from .app import create_dashserving_app
+
+__all__ = ['create_dashserving_app']

src/twinkle/server/dashserving/__main__.py

@@ -0,0 +1,24 @@
+# Copyright (c) ModelScope Contributors. All rights reserved.
+"""Standalone launcher for the Twinkle DashServing adapter."""
+from __future__ import annotations
+
+import os
+
+
+def main() -> None:
+    import uvicorn
+
+    from .app import create_dashserving_app
+
+    upstream_url = os.getenv('TWINKLE_INTERNAL_URL', 'http://127.0.0.1:8000')
+    port = int(os.getenv('PORT', '9000'))
+    timeout_seconds = float(os.getenv('TWINKLE_DS_TIMEOUT_SECONDS', '600'))
+    app = create_dashserving_app(
+        upstream_url=upstream_url,
+        timeout_seconds=timeout_seconds,
+    )
+    uvicorn.run(app, host='0.0.0.0', port=port)
+
+
+if __name__ == '__main__':
+    main()