diff --git a/pyproject.toml b/pyproject.toml
index 809d1f53cd1..016b05d606a 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -25,27 +25,12 @@ classifiers = [
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
     "Topic :: Software Development :: Libraries :: Python Modules",
 ]
-
-
-dependencies = [
-    # Core runtime dependencies (required for actual usage)
-    "omegaconf>=2.3.0",
-    "librosa>=0.11.0",
-    "resampy>=0.4.3",
-    "diffusers>=0.36.0",
-    "accelerate==1.12.0",
-    "gradio==5.50",
-    "soundfile>=0.13.1",
-    "cache-dit==1.2.0",
-    "tqdm>=4.66.0",
-    "torchsde>=0.2.6",  # Required for Stable Audio scheduler
-    "fa3-fwd", # flash attention 3, maintained by @ZJY0516
-    "openai-whisper>=20250625",
-    "imageio[ffmpeg]>=2.37.2",
-    "onnxruntime>=1.19.0",
-    "sox>=1.5.0",
-    # "vllm==0.14.0",  # TODO: fix the entrypoints overwrite problem
-]
+# NOTE: dependencies are managed via setup.py for platform-aware routing
+# See requirements/common.txt for common dependencies
+# See requirements/{cuda,rocm,npu,xpu,cpu}.txt for platform-specific dependencies
+# Platform detection happens at install time via VLLM_OMNI_TARGET_DEVICE env var
+# or torch backend auto-detection
+dynamic = ["dependencies"]
 
 [project.optional-dependencies]
 
@@ -203,4 +188,5 @@ extend-ignore-identifiers-re = [
     ".*NOTHINK.*",
     ".*nin.*",
     "Ono_Anna",
+    ".*cann.*",
 ]
diff --git a/requirements/common.txt b/requirements/common.txt
new file mode 100644
index 00000000000..99d356199cf
--- /dev/null
+++ b/requirements/common.txt
@@ -0,0 +1,13 @@
+# Common dependencies (platform-independent)
+omegaconf>=2.3.0
+librosa>=0.11.0
+resampy>=0.4.3
+diffusers>=0.36.0
+accelerate==1.12.0
+gradio==5.50
+soundfile>=0.13.1
+cache-dit==1.2.0
+tqdm>=4.66.0
+torchsde>=0.2.6
+openai-whisper>=20250625
+imageio[ffmpeg]>=2.37.2
diff --git a/requirements/cpu.txt b/requirements/cpu.txt
new file mode 100644
index 00000000000..d5bb9e30474
--- /dev/null
+++ b/requirements/cpu.txt
@@ -0,0 +1,2 @@
+# CPU-specific dependencies
+# Add CPU-specific optimization libraries here
diff --git a/requirements/cuda.txt b/requirements/cuda.txt
new file mode 100644
index 00000000000..0c3c178f520
--- /dev/null
+++ b/requirements/cuda.txt
@@ -0,0 +1,3 @@
+fa3-fwd
+onnxruntime>=1.19.0
+sox>=1.5.0
diff --git a/requirements/npu.txt b/requirements/npu.txt
new file mode 100644
index 00000000000..363cb4eb8fc
--- /dev/null
+++ b/requirements/npu.txt
@@ -0,0 +1,4 @@
+# NPU-specific dependencies (Ascend)
+# Add NPU-specific acceleration libraries here
+onnxruntime-cann>=1.23.2
+sox>=1.5.0
diff --git a/requirements/rocm.txt b/requirements/rocm.txt
new file mode 100644
index 00000000000..65f4133a44f
--- /dev/null
+++ b/requirements/rocm.txt
@@ -0,0 +1,4 @@
+# ROCm-specific dependencies
+# Add AMD-specific acceleration libraries here
+onnxruntime-rocm>=1.22.2
+sox>=1.5.0
diff --git a/requirements/xpu.txt b/requirements/xpu.txt
new file mode 100644
index 00000000000..c103a8a89cf
--- /dev/null
+++ b/requirements/xpu.txt
@@ -0,0 +1,2 @@
+# XPU-specific dependencies (Intel)
+# Add Intel XPU-specific acceleration libraries here
diff --git a/setup.py b/setup.py
new file mode 100644
index 00000000000..a556df33e46
--- /dev/null
+++ b/setup.py
@@ -0,0 +1,127 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Platform-aware dependency routing for vLLM-Omni.
+
+This module implements install-time detection of the target hardware platform
+and automatically selects the appropriate platform-specific dependencies.
+
+Detection Priority:
+1. Explicit override via VLLM_OMNI_TARGET_DEVICE environment variable
+2. Torch backend detection (CUDA, ROCm, NPU, XPU)
+3. Fallback to common dependencies only (treated as CPU)
+
+Supported platforms: cuda, rocm, npu, xpu, cpu
+"""
+
+from __future__ import annotations
+
+import os
+from pathlib import Path
+from typing import Literal
+
+from setuptools import setup
+
+try:
+    import torch
+except Exception:  # pragma: no cover - torch may not be installed at build time
+    torch = None
+
+ROOT = Path(__file__).parent
+
+# Supported target devices
+TargetDevice = Literal["cuda", "rocm", "npu", "xpu", "cpu"]
+
+
+def _read_requirements(filename: str) -> list[str]:
+    """Read and resolve requirements from a file, handling -r includes."""
+    requirements_path = ROOT / "requirements" / filename
+    if not requirements_path.exists():
+        return []
+    lines = requirements_path.read_text().splitlines()
+    resolved: list[str] = []
+    for line in lines:
+        line = line.strip()
+        if not line or line.startswith("#"):
+            continue
+        if line.startswith("-r "):
+            resolved += _read_requirements(line.split()[1])
+        else:
+            resolved.append(line)
+    return resolved
+
+
+def _detect_target_device() -> TargetDevice | None:
+    """
+    Detect the target device platform for dependency selection.
+
+    Priority rules:
+    1. VLLM_OMNI_TARGET_DEVICE environment variable (highest priority)
+    2. Torch backend detection via torch.version.cuda/hip and device availability
+    3. None (fallback - only common dependencies will be installed)
+
+    Returns:
+        The detected target device, or None if no platform can be determined.
+    """
+    # Priority 1: Explicit environment variable override
+    env_target = os.getenv("VLLM_OMNI_TARGET_DEVICE", "").lower()
+    if env_target:
+        valid_devices = {"cuda", "rocm", "npu", "xpu", "cpu"}
+        if env_target in valid_devices:
+            return env_target  # type: ignore[return-value]
+        # Invalid value - log warning and continue with auto-detection
+        print(
+            f"Warning: Invalid VLLM_OMNI_TARGET_DEVICE='{env_target}'. "
+            f"Valid values: {valid_devices}. Falling back to auto-detection."
+        )
+
+    # Priority 2: Torch backend detection
+    if torch is not None:
+        # Check CUDA
+        if getattr(torch.version, "cuda", None) is not None:
+            return "cuda"
+
+        # Check ROCm
+        if getattr(torch.version, "hip", None) is not None:
+            return "rocm"
+
+        # Check NPU
+        try:
+            if hasattr(torch, "npu") and torch.npu.is_available():
+                return "npu"
+        except Exception:
+            pass
+
+        # Check XPU
+        try:
+            if hasattr(torch, "xpu") and torch.xpu.is_available():
+                return "xpu"
+        except Exception:
+            pass
+
+    # Priority 3: Fallback - no specific platform detected
+    return None
+
+
+def _get_install_requires() -> list[str]:
+    """
+    Build the complete list of install requirements based on detected platform.
+
+    Always includes common.txt, then adds platform-specific dependencies
+    based on the detected target device.
+    """
+    install_requires = _read_requirements("common.txt")
+
+    target_device = _detect_target_device()
+
+    if target_device is not None:
+        platform_requirements_file = f"{target_device}.txt"
+        platform_requirements = _read_requirements(platform_requirements_file)
+        install_requires += platform_requirements
+
+    return install_requires
+
+
+install_requires = _get_install_requires()
+
+setup(install_requires=install_requires)