diff --git a/README.md b/README.md
index 4a9036c396d..23afbbe5300 100644
--- a/README.md
+++ b/README.md
@@ -223,6 +223,13 @@ docker run -it \
 ```
 > This image has used `uv` to install all GPU-related dependencies of Trinity-RFT. The virtual environment will be automatically activated upon entering the container (you can also manually activate it via `source /opt/venv/bin/activate` if needed). You can use `uv pip install` to add extra packages as necessary.
 
+**Using the Installation Script (uv)**
+
+```bash
+bash scripts/install/install_trinity.sh
+source .venv/bin/activate
+```
+
 **Using Conda**
 
 ```bash
@@ -235,7 +242,7 @@ pip install -e ".[vllm,flash_attn]"
 # pip install -e ".[tinker]"
 
 # If you encounter issues when installing flash-attn, try:
-# pip install flash-attn==2.8.1 --no-build-isolation
+# pip install "flash-attn>=2.8.3" --no-build-isolation
 pip install -e ".[dev]"  # for development like linting and debugging
 ```
 
@@ -251,34 +258,38 @@ pip install -e ".[vllm,flash_attn]"
 # pip install -e ".[tinker]"
 
 # If you encounter issues when installing flash-attn, try:
-# pip install flash-attn==2.8.1 --no-build-isolation
+# pip install "flash-attn>=2.8.3" --no-build-isolation
 
 pip install -e ".[dev]"  # for development like linting and debugging
 ```
 
-**Using uv**
+**Using uv manually**
 
 ```bash
-uv sync --extra vllm --extra dev --extra flash_attn
+uv venv .venv
+
+uv pip install \
+  --python .venv/bin/python \
+  --overrides scripts/install/trinity-overrides.txt \
+  -e ".[vllm,dev]"
+
+.venv/bin/python scripts/install/install_flash_attn.py --uv
 
 # If you have no GPU, try to use Tinker instead:
-# uv sync --extra tinker --extra dev
+# uv pip install --python .venv/bin/python -e ".[tinker,dev]"
+
+source .venv/bin/activate
 ```
 
+If conflicts remain, add or adjust versions in `scripts/install/trinity-overrides.txt`.
+
 #### Via PyPI
 
 If you just want to use the package without modifying the code:
 
 ```bash
 pip install trinity-rft
-pip install flash-attn==2.8.1
-```
-
-Or with `uv`:
-
-```bash
-uv pip install trinity-rft
-uv pip install flash-attn==2.8.1
+pip install "flash-attn>=2.8.3" --no-build-isolation
 ```
 
 > For training with **Megatron-LM**, please refer to [Megatron-LM Backend](https://agentscope-ai.github.io/Trinity-RFT/en/main/tutorial/example_megatron.html).
@@ -363,9 +374,10 @@ export WANDB_API_KEY=<your_api_key>
 wandb login
 ```
 
-For command-line users, run the RFT process:
+For command-line users, activate the environment and run the RFT process:
 
 ```shell
+source .venv/bin/activate  # If using venv or uv
 trinity run --config <config_path>
 ```
 
diff --git a/README_zh.md b/README_zh.md
index 663fa027be8..f9f08a1b509 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -245,6 +245,13 @@ docker run -it \
 
 > 该镜像已经通过 `uv` 安装了 Trinity-RFT 以及所有 GPU 相关依赖，且会自动激活虚拟环境（也可通过 `source /opt/venv/bin/activate` 手动激活）。必要时可使用 `uv pip install` 添加额外的包。
 
+**使用安装脚本（uv）**
+
+```bash
+bash scripts/install/install_trinity.sh
+source .venv/bin/activate
+```
+
 **使用 Conda**
 
 ```bash
@@ -257,7 +264,7 @@ pip install -e ".[vllm,flash_attn]"
 # pip install -e ".[tinker]"
 
 # 如果安装 flash-attn 时遇到问题，可尝试：
-# pip install flash-attn==2.8.1 --no-build-isolation
+# pip install "flash-attn>=2.8.3" --no-build-isolation
 
 pip install -e ".[dev]"  # 用于调试和开发
 ```
@@ -274,36 +281,41 @@ pip install -e ".[vllm,flash_attn]"
 # pip install -e ".[tinker]"
 
 # 如果安装 flash-attn 时遇到问题，可尝试：
-# pip install flash-attn==2.8.1 --no-build-isolation
+# pip install "flash-attn>=2.8.3" --no-build-isolation
 
 pip install -e ".[dev]"  # 用于调试和开发
 ```
 
-**使用 uv**
+**手动使用 uv**
 
 [`uv`](https://github.com/astral-sh/uv) 是现代的 Python 包管理工具。
 
 ```bash
-uv sync --extra vllm --extra dev --extra flash_attn
+uv venv .venv
+
+uv pip install \
+  --python .venv/bin/python \
+  --overrides scripts/install/trinity-overrides.txt \
+  -e ".[vllm,dev]"
+
+.venv/bin/python scripts/install/install_flash_attn.py --uv
 
-# 如果没有GPU，可以改为使用Tinker：
-# uv sync --extra tinker --extra dev
+# 如果没有 GPU，可以改用 Tinker：
+# uv pip install --python .venv/bin/python -e ".[tinker,dev]"
+
+source .venv/bin/activate
 ```
 
+脚本会安装匹配当前环境的第三方 [flash-attn wheel](https://github.com/mjun0812/flash-attention-prebuild-wheels)。
+如果仍有依赖冲突，可在 `scripts/install/trinity-overrides.txt` 中新增或调整版本约束。
+
 #### 通过 PyPI 安装
 
 如果您只需使用 Trinity-RFT 而不打算修改代码：
 
 ```bash
 pip install trinity-rft
-pip install flash-attn==2.8.1
-```
-
-或使用 `uv`：
-
-```bash
-uv pip install trinity-rft
-uv pip install flash-attn==2.8.1
+pip install "flash-attn>=2.8.3" --no-build-isolation
 ```
 
 > 如需使用 **Megatron-LM** 进行训练，请参考 [Megatron-LM 支持](https://agentscope-ai.github.io/Trinity-RFT/zh/main/tutorial/example_megatron.html)
@@ -398,9 +410,10 @@ export WANDB_API_KEY=<your_api_key>
 wandb login
 ```
 
-对于命令行用户，运行 RFT 流程：
+对于命令行用户，先激活环境，再运行 RFT 流程：
 
 ```shell
+source .venv/bin/activate  # 使用 venv 或 uv 时
 trinity run --config <config_path>
 ```
 
diff --git a/docs/sphinx_doc/source/tutorial/faq.md b/docs/sphinx_doc/source/tutorial/faq.md
index 87626ae1cbb..adbe809a18f 100644
--- a/docs/sphinx_doc/source/tutorial/faq.md
+++ b/docs/sphinx_doc/source/tutorial/faq.md
@@ -71,7 +71,7 @@ File ".../flash_attn/flash_attn_interface.py", line 15, in ‹module>
 ImportError: ...
 ```
 
-**A:** The `flash-attn` module is not properly installed. Try to fix it by running `pip install flash-attn==2.8.1` or `pip install flash-attn==2.8.1 -v --no-build-isolation`.
+**A:** The `flash-attn` module is not properly installed. Try to fix it by running `pip install "flash-attn>=2.8.3"` or `pip install "flash-attn>=2.8.3" -v --no-build-isolation`.
 
 ---
 
diff --git a/docs/sphinx_doc/source/tutorial/trinity_installation.md b/docs/sphinx_doc/source/tutorial/trinity_installation.md
index 11ff36a5d89..b401dff387d 100644
--- a/docs/sphinx_doc/source/tutorial/trinity_installation.md
+++ b/docs/sphinx_doc/source/tutorial/trinity_installation.md
@@ -49,7 +49,7 @@ pip install -e ".[vllm,flash_attn]"
 # pip install -e ".[tinker]"
 
 # If you encounter issues when installing flash-attn, try:
-# pip install flash-attn==2.8.1 --no-build-isolation
+# pip install "flash-attn>=2.8.3" --no-build-isolation
 
 pip install -e ".[dev]"  # for development like linting and debugging
 ```
@@ -66,7 +66,7 @@ pip install -e ".[vllm,flash_attn]"
 # pip install -e ".[tinker]"
 
 # If you encounter issues when installing flash-attn, try:
-# pip install flash-attn==2.8.1 --no-build-isolation
+# pip install "flash-attn>=2.8.3" --no-build-isolation
 
 pip install -e ".[dev]"  # for development like linting and debugging
 ```
@@ -146,14 +146,14 @@ If you just want to use the package without modifying the code:
 
 ```bash
 pip install trinity-rft
-pip install flash-attn==2.8.1  --no-build-isolation
+pip install "flash-attn>=2.8.3" --no-build-isolation
 ```
 
 Or with `uv`:
 
 ```bash
 uv pip install trinity-rft
-uv pip install flash-attn==2.8.1  --no-build-isolation
+uv pip install "flash-attn>=2.8.3" --no-build-isolation
 ```
 
 ---
diff --git a/docs/sphinx_doc/source_zh/tutorial/faq.md b/docs/sphinx_doc/source_zh/tutorial/faq.md
index 67da2ee6459..578ab77da9e 100644
--- a/docs/sphinx_doc/source_zh/tutorial/faq.md
+++ b/docs/sphinx_doc/source_zh/tutorial/faq.md
@@ -70,7 +70,7 @@ File ".../flash_attn/flash_attn_interface.py", line 15, in ‹module>
 ImportError: ...
 ```
 
-**A:** `flash-attn` 模块未正确安装。请尝试运行 `pip install flash-attn==2.8.1` 或 `pip install flash-attn==2.8.1 -v --no-build-isolation` 进行修复。
+**A:** `flash-attn` 模块未正确安装。请尝试运行 `pip install "flash-attn>=2.8.3"` 或 `pip install "flash-attn>=2.8.3" -v --no-build-isolation` 进行修复。
 
 ---
 
diff --git a/docs/sphinx_doc/source_zh/tutorial/trinity_installation.md b/docs/sphinx_doc/source_zh/tutorial/trinity_installation.md
index 42730f003b2..7c51634f53b 100644
--- a/docs/sphinx_doc/source_zh/tutorial/trinity_installation.md
+++ b/docs/sphinx_doc/source_zh/tutorial/trinity_installation.md
@@ -49,7 +49,7 @@ pip install -e ".[vllm,flash_attn]"
 # pip install -e ".[tinker]"
 
 # 如果安装 flash-attn 时遇到问题，可尝试：
-# pip install flash-attn==2.8.1 --no-build-isolation
+# pip install "flash-attn>=2.8.3" --no-build-isolation
 
 pip install -e ".[dev]"  # 用于调试和开发
 ```
@@ -66,7 +66,7 @@ pip install -e ".[vllm,flash_attn]"
 # pip install -e ".[tinker]"
 
 # 如果安装 flash-attn 时遇到问题，可尝试：
-# pip install flash-attn==2.8.1 --no-build-isolation
+# pip install "flash-attn>=2.8.3" --no-build-isolation
 
 pip install -e ".[dev]"  # 用于调试和开发
 ```
@@ -147,14 +147,14 @@ docker run -it \
 
 ```bash
 pip install trinity-rft
-pip install flash-attn==2.8.1 --no-build-isolation
+pip install "flash-attn>=2.8.3" --no-build-isolation
 ```
 
 或使用 `uv`：
 
 ```bash
 uv pip install trinity-rft
-uv pip install flash-attn==2.8.1 --no-build-isolation
+uv pip install "flash-attn>=2.8.3" --no-build-isolation
 ```
 
 ---
diff --git a/pyproject.toml b/pyproject.toml
index 008c00ab23f..a5d4f0a6f0c 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -57,8 +57,7 @@ trinity = "trinity.cli.launcher:main"
 
 [project.optional-dependencies]
 vllm = [
-    # for routed_experts support, please install vllm 0.22.0 or above
-    "vllm>=0.19.1,<=0.23.0",
+    "vllm>=0.22.0,<=0.23.0",
 ]
 sglang = [
     "sglang==0.5.13",
diff --git a/scripts/install/install_flash_attn.py b/scripts/install/install_flash_attn.py
index 777fd883911..299adc3eae6 100644
--- a/scripts/install/install_flash_attn.py
+++ b/scripts/install/install_flash_attn.py
@@ -1,80 +1,125 @@
-"""This script is used to install flash-attn from a pre-built wheel hosted on an OSS bucket.
-Useful for mainland China users who have difficulty installing flash-attn from PyPI due to network issues.
-"""
+"""Install a compatible third-party flash-attn wheel from a community index."""
+
 import os
-import platform
+import re
 import subprocess
 import sys
 import tempfile
+from importlib.metadata import PackageNotFoundError
+from importlib.metadata import version as package_version
+from urllib.parse import unquote
 
 import torch
 import typer
+from packaging.specifiers import SpecifierSet
+from packaging.tags import sys_tags
+from packaging.utils import InvalidWheelFilename, parse_wheel_filename
+from packaging.version import InvalidVersion, Version
 
 app = typer.Typer()
-FLASH_VERSION = "2.8.1"
+FLASH_ATTN_RANGE = SpecifierSet(">=2.8.3")
+WHEEL_INDEX_URL = (
+    "https://raw.githubusercontent.com/mjun0812/"
+    "flash-attention-prebuild-wheels/main/doc/packages.md"
+)
+WHEEL_URL_PATTERN = re.compile(
+    r"(https://github\.com/mjun0812/flash-attention-prebuild-wheels/"
+    r"releases/download/[^ )]+/flash_attn-[^ )]+\.whl)"
+)
+
+
+def _get_cuda_torch_tag():
+    # torch.version.hip/cuda are runtime attributes not in type stubs
+    is_rocm = (
+        hasattr(torch.version, "hip")
+        and torch.version.hip is not None  # type: ignore[attr-defined]
+    )
+    if is_rocm:
+        raise RuntimeError("ROCm wheels are not supported.")
+
+    torch_cuda_version = torch.version.cuda  # type: ignore[attr-defined]
+    if torch_cuda_version is None:
+        raise RuntimeError("The installed PyTorch does not include CUDA support.")
+
+    torch_major, torch_minor = torch.__version__.split("+", 1)[0].split(".")[:2]
+    return f"cu{torch_cuda_version.replace('.', '')}torch{torch_major}.{torch_minor}"
 
 
 def check_flash_attn_installed():
     try:
-        import flash_attn
-
-        print(f"flash_attn version: {flash_attn.__version__}")
+        installed_version = package_version("flash-attn")
+        print(f"flash_attn version: {installed_version}")
+        version = Version(installed_version)
+        if version not in FLASH_ATTN_RANGE:
+            return False
+        if re.fullmatch(r"cu\d+torch\d+\.\d+", version.local or "") and (
+            version.local != _get_cuda_torch_tag()
+        ):
+            return False
+        __import__("flash_attn")
         return True
-    except ImportError:
+    except (ImportError, InvalidVersion, OSError, PackageNotFoundError, RuntimeError):
         return False
 
 
-def install_flash_attn(uv: bool = False, keep_wheel: bool = False):
-    # Get torch version
-    TORCH_VERSION_RAW = torch.__version__
-    torch_major, torch_minor = TORCH_VERSION_RAW.split(".")[:2]
-    torch_version = f"{torch_major}.{torch_minor}"
-
-    # Get python version
-    python_version = f"cp{sys.version_info.major}{sys.version_info.minor}"
+def find_wheel_url():
+    cuda_torch_tag = _get_cuda_torch_tag()
+
+    print(f"Detected: Python {sys.version_info.major}.{sys.version_info.minor} {cuda_torch_tag}")
+
+    wheel_index = subprocess.run(
+        ["curl", "-fL", "--retry", "5", "--retry-all-errors", WHEEL_INDEX_URL],
+        check=True,
+        capture_output=True,
+        text=True,
+    ).stdout
+    supported_tags = set(sys_tags())
+    matches = []
+    for wheel_url in WHEEL_URL_PATTERN.findall(wheel_index):
+        wheel_url = unquote(wheel_url)
+        try:
+            distribution, version, _, wheel_tags = parse_wheel_filename(
+                wheel_url.rsplit("/", 1)[-1]
+            )
+        except InvalidWheelFilename:
+            continue
+        if (
+            distribution == "flash-attn"
+            and version in FLASH_ATTN_RANGE
+            and version.local == cuda_torch_tag
+            and supported_tags.intersection(wheel_tags)
+        ):
+            matches.append((version, wheel_url))
+    if matches:
+        return max(matches, key=lambda item: item[0])[1]
+
+    raise RuntimeError(f"No matching flash-attn wheel found for {cuda_torch_tag}.")
 
-    # Get platform name
-    platform_name = platform.system().lower() + "_" + platform.machine()
 
-    # Get cxx11_abi
-    cxx11_abi = str(torch._C._GLIBCXX_USE_CXX11_ABI).upper()
-
-    # Is ROCM
-    # torch.version.hip/cuda are runtime attributes not in type stubs
-    IS_ROCM = hasattr(torch.version, "hip") and torch.version.hip is not None  # type: ignore[attr-defined]
-
-    if IS_ROCM:
-        print("We currently do not host ROCm wheels for flash-attn.")
-        sys.exit(1)
-    else:
-        torch_cuda_version = torch.version.cuda  # type: ignore[attr-defined]
-        cuda_major = torch_cuda_version.split(".")[0] if torch_cuda_version else None
-        if cuda_major != "12":
-            print("Only CUDA 12 wheels are hosted for flash-attn.")
-            sys.exit(1)
-        cuda_version = "12"
-        wheel_filename = (
-            f"flash_attn-{FLASH_VERSION}%2Bcu{cuda_version}torch{torch_version}"
-            f"cxx11abi{cxx11_abi}-{python_version}-{python_version}-{platform_name}.whl"
-        )
-        local_filename = (
-            f"flash_attn-{FLASH_VERSION}-{python_version}-{python_version}-{platform_name}.whl"
-        )
-
-    wheel_url = (
-        "https://dail-wlcb.oss-cn-wulanchabu.aliyuncs.com"
-        f"/AgentScope/download/flash-attn/{FLASH_VERSION}/{wheel_filename}"
-    )
+def install_flash_attn(uv: bool = False, keep_wheel: bool = False):
+    wheel_url = find_wheel_url()
+    local_filename = wheel_url.rsplit("/", 1)[-1]
 
     print(f"wheel_url: {wheel_url}")
-    print(f"target_local_file: {local_filename}")
 
     def _install_helper(local_path: str):
-        subprocess.run(["wget", wheel_url, "-O", local_path], check=True)
+        subprocess.run(
+            [
+                "curl",
+                "-fL",
+                "--retry",
+                "5",
+                "--retry-all-errors",
+                "-o",
+                local_path,
+                wheel_url,
+            ],
+            check=True,
+        )
         install_cmd = (
-            ["uv", "pip", "install", local_path]
+            ["uv", "pip", "install", "--python", sys.executable, "--no-deps", local_path]
             if uv
-            else [sys.executable, "-m", "pip", "install", local_path]
+            else [sys.executable, "-m", "pip", "install", "--no-deps", local_path]
         )
         subprocess.run(install_cmd, check=True)
 
@@ -99,7 +144,7 @@ def main(
         False, help="Keep the downloaded wheel file in current directory"
     ),
 ):
-    """Install flash-attn from a pre-built wheel."""
+    """Install flash-attn from a matching pre-built wheel."""
     if check_flash_attn_installed():
         print("flash_attn is already installed. Skipping installation.")
         return
diff --git a/scripts/install/install_trinity.sh b/scripts/install/install_trinity.sh
new file mode 100644
index 00000000000..cf907c046f4
--- /dev/null
+++ b/scripts/install/install_trinity.sh
@@ -0,0 +1,21 @@
+#!/usr/bin/env bash
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd -- "${SCRIPT_DIR}/../.." && pwd)"
+VENV="${REPO_ROOT}/.venv"
+PYTHON_BIN="${VENV}/bin/python"
+
+cd "${REPO_ROOT}"
+
+[[ -x "${PYTHON_BIN}" ]] || uv venv "${VENV}"
+
+uv pip install \
+    --python "${PYTHON_BIN}" \
+    --overrides "${SCRIPT_DIR}/trinity-overrides.txt" \
+    -e "${REPO_ROOT}[vllm,dev]"
+
+"${PYTHON_BIN}" "${SCRIPT_DIR}/install_flash_attn.py" --uv
+
+echo "Activate with: source ${VENV}/bin/activate"
diff --git a/scripts/install/trinity-overrides.txt b/scripts/install/trinity-overrides.txt
new file mode 100644
index 00000000000..e70ae0b6877
--- /dev/null
+++ b/scripts/install/trinity-overrides.txt
@@ -0,0 +1,2 @@
+numpy>=2
+transformers>=5.12.1