feat: windows support

acestep/third_parts/nano-vllm/nanovllm/engine/model_runner.py

@@ -3,6 +3,7 @@ import torch
 import torch.distributed as dist
 from multiprocessing.synchronize import Event
 from multiprocessing.shared_memory import SharedMemory
+import sys
 
 from nanovllm.config import Config
 from nanovllm.engine.sequence import Sequence
@@ -55,7 +56,9 @@ class ModelRunner:
         self.event = event
         dist_port = find_available_port()
         print(f"[debug]dist_port: {dist_port}")
-        dist.init_process_group("nccl", f"tcp://localhost:{dist_port}", world_size=self.world_size, rank=rank)
+        # Use gloo backend on Windows, nccl on Linux/other platforms
+        backend = "gloo" if sys.platform == "win32" else "nccl"
+        dist.init_process_group(backend, f"tcp://localhost:{dist_port}", world_size=self.world_size, rank=rank)
         torch.cuda.set_device(rank)
         default_dtype = torch.get_default_dtype()
         # Use dtype instead of deprecated torch_dtype

acestep/third_parts/nano-vllm/pyproject.toml

@@ -6,14 +6,14 @@ build-backend = "setuptools.build_meta"
 name = "nano-vllm"
 version = "0.2.0"
 authors = [{ name = "Xingkai Yu" }]
-license = "MIT"
-license-files = ["LICENSE"]
+license = {text = "MIT"}
 readme = "README.md"
 description = "a lightweight vLLM implementation built from scratch"
 requires-python = ">=3.10,<3.13"
 dependencies = [
     "torch>=2.4.0",
-    "triton>=3.0.0",
+    "triton-windows>=3.0.0; sys_platform == 'win32'",
+    "triton>=3.0.0; sys_platform != 'win32'",
     "transformers>=4.51.0",
     "flash-attn",
     "xxhash",

pyproject.toml

@@ -3,7 +3,7 @@ name = "ace-step"
 version = "1.5.0"
 description = "ACE-Step 1.5"
 readme = "README.md"
-requires-python = ">=3.12,<3.13"
+requires-python = ">=3.11, <3.12"
 license = {text = "Apache-2.0"}
 dependencies = [
     # PyTorch for Linux/Windows with CUDA