feat: change windows dependencies

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

@@ -58,7 +58,7 @@ class ModelRunner:
         print(f"[debug]dist_port: {dist_port}")
         # 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)
+        dist.init_process_group(backend, f"tcp://127.0.0.1:{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

@@ -12,10 +12,10 @@ description = "a lightweight vLLM implementation built from scratch"
 requires-python = ">=3.10,<3.13"
 dependencies = [
     "torch>=2.4.0",
-    "triton-windows>=3.0.0; sys_platform == 'win32'",
+    "triton-windows>=3.0.0,<3.4; sys_platform == 'win32'",
     "triton>=3.0.0; sys_platform != 'win32'",
     "transformers>=4.51.0",
-    "flash-attn @ https://github.com/sdbds/flash-attention-for-windows/releases/download/2.8.3/flash_attn-2.8.3+cu128torch2.8.0cxx11abiFALSEfullbackward-cp311-cp311-win_amd64.whl ; sys_platform == 'win32' and python_version == '3.11' and platform_machine == 'AMD64'",
+    "flash-attn @ https://github.com/sdbds/flash-attention-for-windows/releases/download/2.8.2/flash_attn-2.8.2+cu128torch2.7.1cxx11abiFALSEfullbackward-cp311-cp311-win_amd64.whl ; sys_platform == 'win32' and python_version == '3.11' and platform_machine == 'AMD64'",
     "flash-attn; sys_platform != 'win32'",
     "xxhash",
 ]

pyproject.toml

@@ -7,9 +7,9 @@ requires-python = ">=3.11, <3.12"
 license = {text = "Apache-2.0"}
 dependencies = [
     # PyTorch for Linux/Windows with CUDA
-    "torch>=2.9.1; sys_platform != 'darwin'",
+    "torch==2.7.1; sys_platform != 'darwin'",
     "torchvision; sys_platform != 'darwin'",
-    "torchaudio>=2.9.1; sys_platform != 'darwin'",
+    "torchaudio==2.7.1; sys_platform != 'darwin'",
     # PyTorch for macOS (CPU / MPS)
     "torch>=2.9.1; sys_platform == 'darwin'",
     "torchvision; sys_platform == 'darwin'",
@@ -57,10 +57,3 @@ dev = []
 
 [tool.hatch.build.targets.wheel]
 packages = ["acestep"]
-
-[tool.uv.extra-build-dependencies]
-flash-attn = [
-  "packaging",
-  "wheel",
-  { requirement = "torch", match-runtime = true },
-]

requirements.txt

@@ -1,12 +1,35 @@
-torch
-transformers
+
+# PyTorch with CUDA 12.8 (for Windows/Linux)
+--extra-index-url https://download.pytorch.org/whl/cu128
+torch==2.7.1; sys_platform == 'win32'
+torchaudio==2.7.1; sys_platform == 'win32'
+torchvision; sys_platform == 'win32'
+torch>=2.9.1; sys_platform != 'win32'
+torchaudio>=2.9.1; sys_platform != 'win32'
+torchvision; sys_platform != 'win32'
+
+# Core dependencies
+transformers>=4.51.0
 diffusers
 gradio
-soundfile
-loguru
-einops
-accelerator
-vector-quantize-pytorch
-psutil
-fastapi
-uvicorn
+matplotlib>=3.7.5
+scipy>=1.10.1
+soundfile>=0.13.1
+loguru>=0.7.3
+einops>=0.8.1
+accelerate>=1.12.0
+fastapi>=0.110.0
+uvicorn[standard]>=0.27.0
+numba>=0.63.1
+vector-quantize-pytorch>=1.27.15
+torchcodec>=0.9.1
+
+# nano-vllm dependencies
+triton-windows>=3.0.0,<3.4; sys_platform == 'win32'
+triton>=3.0.0; sys_platform != 'win32'
+flash-attn @ https://github.com/sdbds/flash-attention-for-windows/releases/download/2.8.2/flash_attn-2.8.2+cu128torch2.7.1cxx11abiFALSEfullbackward-cp311-cp311-win_amd64.whl ; sys_platform == 'win32' and python_version == '3.11' and platform_machine == 'AMD64'
+flash-attn; sys_platform != 'win32'
+xxhash
+
+# Local package - install with: pip install -e acestep/third_parts/nano-vllm
+# nano-vllm