Skip to content

Fix/amd wheel jit kernel support #22661

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
akao-amd wants to merge 3 commits into
base: main
Choose a base branch
from
+61 −59
Open

Fix/amd wheel jit kernel support #22661

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
82ff9e4
[AMD] Enable JIT kernel support for wheel installations
akao-amd Apr 9, 2026
63daa07
[AMD] Sync AMD wheel dependencies with main pyproject_other.toml
akao-amd Apr 9, 2026
75ff225
[AMD] Remove MUSA and HPU sections from AMD wheel config
akao-amd Apr 9, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
23 changes: 23 additions & 0 deletions 3rdparty/amd/wheel/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,29 @@ pip install "amd-sglang[all-hip,rocm720]" -i https://pypi.amd.com/rocm-7.2.0/sim

Note: You must resolve the two dependencies, AITER and triton, below. Others are optional depending on your applications.

## JIT Kernel Support

The amd-sglang wheel includes JIT (Just-In-Time) kernel compilation support. JIT kernels allow for dynamic compilation of optimized CUDA/HIP kernels at runtime.

### Requirements

JIT kernel compilation requires:
1. **apache-tvm-ffi** - Included in the `runtime_common` dependencies (installed with `amd-sglang[all-hip,...]`)
2. **System compiler toolchain** - A C++ compiler compatible with your ROCm installation
- For ROCm environments, this is typically provided by the ROCm installation
- Ensure `hipcc` is available in your PATH

The JIT kernel source files (`.cuh`, `.cu` headers) are bundled with the wheel and will be available at runtime for compilation.

### Verification

To verify JIT kernel support is working:
```python
from sglang.jit_kernel.utils import KERNEL_PATH
print(f"JIT kernel path: {KERNEL_PATH}")
# Should print the path to site-packages/sglang/jit_kernel
```

## Manual Dependency Resolution

### Resolving AITER
Expand Down
97 changes: 38 additions & 59 deletions 3rdparty/amd/wheel/sglang/pyproject.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,17 +19,16 @@ dependencies = ["aiohttp", "requests", "tqdm", "numpy", "IPython", "setproctitle
runtime_common = [
"IPython",
"aiohttp",
"apache-tvm-ffi>=0.1.5,<0.2",
"anthropic>=0.20.0",
"blobfile==3.0.0",
"av",
"build",
"compressed-tensors",
"decord2",
"datasets",
"einops",
"fastapi",
"gguf",
"hf_transfer",
"huggingface_hub",
"interegular",
"llguidance>=0.7.11,<0.8.0",
"modelscope",
Expand Down Expand Up @@ -59,7 +58,8 @@ runtime_common = [
"timm==1.0.16",
"torchao==0.9.0",
"tqdm",
"transformers==4.57.1",
"mistral_common>=1.9.0",
"transformers==5.3.0",
"uvicorn",
"uvloop",
"xgrammar==0.1.32",
Expand All @@ -76,7 +76,7 @@ rocm700 = [
"torchaudio @ https://repo.radeon.com/rocm/manylinux/rocm-rel-7.0.2/torchaudio-2.9.0%2Brocm7.0.2.gite3c6ee2b-cp310-cp310-linux_x86_64.whl",
"torchvision @ https://repo.radeon.com/rocm/manylinux/rocm-rel-7.0.2/torchvision-0.24.0%2Brocm7.0.2.gitb919bd0c-cp310-cp310-linux_x86_64.whl",
"mooncake-transfer-engine-non-cuda==0.3.8.post1",
"sglang-kernel @ https://github.com/sgl-project/whl/releases/download/v0.4.0/sglang_kernel-0.4.0+rocm700-cp310-abi3-manylinux2014_x86_64.whl",
"sglang-kernel @ https://github.com/sgl-project/whl/releases/download/v0.4.1/sglang_kernel-0.4.1+rocm700-cp310-abi3-manylinux2014_x86_64.whl",
]

rocm720 = [
Expand All @@ -85,50 +85,10 @@ rocm720 = [
"torchaudio @ https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/torchaudio-2.9.0%2Brocm7.2.0.gite3c6ee2b-cp310-cp310-linux_x86_64.whl",
"torchvision @ https://repo.radeon.com/rocm/manylinux/rocm-rel-7.2/torchvision-0.24.0%2Brocm7.2.0.gitb919bd0c-cp310-cp310-linux_x86_64.whl",
"mooncake-transfer-engine-non-cuda==0.3.8.post1",
"sglang-kernel @ https://github.com/sgl-project/whl/releases/download/v0.4.0/sglang_kernel-0.4.0+rocm720-cp310-abi3-manylinux2014_x86_64.whl",
"sglang-kernel @ https://github.com/sgl-project/whl/releases/download/v0.4.1/sglang_kernel-0.4.1+rocm720-cp310-abi3-manylinux2014_x86_64.whl",
]

# HIP (Heterogeneous-computing Interface for Portability) for AMD
# Install with one of:
# pip install "amd-sglang[srt_hip,rocm700]"
# pip install "amd-sglang[srt_hip,rocm720]"
srt_hip = [
"amd-sglang[runtime_common]",
"petit_kernel==0.0.2",
"wave-lang==3.8.2",
]

diffusion_hip = [
"PyYAML==6.0.1",
"cloudpickle",
"diffusers==0.37.0",
"imageio==2.36.0",
"imageio-ffmpeg==0.5.1",
"moviepy>=2.0.0",
"opencv-python-headless==4.10.0.84",
"remote-pdb",
"st_attn==0.0.7",
"vsa==0.0.4",
"runai_model_streamer>=0.15.5",
"cache-dit==1.1.8",
"addict",
]

# For Intel Gaudi(device : hpu) follow the installation guide
# https://docs.vllm.ai/en/latest/getting_started/gaudi-installation.html
srt_hpu = ["sglang[runtime_common]"]

# https://docs.sglang.io/platforms/mthreads_gpu.md
srt_musa = [
"sglang[runtime_common]",
"torch",
"torch_musa",
"torchada>=0.1.45",
"mthreads-ml-py",
"numpy<2.0",
]

diffusion_musa = [
diffusion_common = [
"PyYAML==6.0.1",
"cloudpickle",
"diffusers==0.37.0",
Expand All @@ -137,11 +97,10 @@ diffusion_musa = [
"moviepy>=2.0.0",
"opencv-python-headless==4.10.0.84",
"remote-pdb",
"st_attn==0.0.7",
"vsa==0.0.4",
"runai_model_streamer>=0.15.5",
"cache-dit==1.1.8",
"addict",
"scikit-image==0.25.2",
"trimesh>=4.0.0",
"xatlas",
]

tracing = [
Expand All @@ -151,33 +110,53 @@ tracing = [
"opentelemetry-exporter-otlp-proto-grpc",
]

# HIP (Heterogeneous-computing Interface for Portability) for AMD
# Install with one of:
# pip install "amd-sglang[srt_hip,rocm700]"
# pip install "amd-sglang[srt_hip,rocm720]"
srt_hip = [
"amd-sglang[runtime_common]",
"petit_kernel==0.0.2",
"wave-lang==3.8.2",
]

diffusion_hip = [
"amd-sglang[diffusion_common]",
"peft>=0.18.0",
"st_attn==0.0.7",
"vsa==0.0.4",
"runai_model_streamer>=0.15.7",
"cache-dit==1.3.0",
]

test = [
"accelerate",
"addict",
"bitsandbytes",
"expecttest",
"gguf",
"jsonlines",
"matplotlib",
"pandas",
"peft",
"parameterized",
"peft>=0.18.0",
"polars",
"pytest",
"pytest-cov",
"diff-cover",
"sentence_transformers",
"tabulate",
]

all_hip = ["amd-sglang[srt_hip]", "amd-sglang[diffusion_hip]"]
all_hpu = ["sglang[srt_hpu]"]
all_musa = ["sglang[srt_musa]", "sglang[diffusion_musa]"]

all_hip = ["amd-sglang[srt_hip]", "amd-sglang[diffusion_hip]", "amd-sglang[tracing]"]
dev_hip = ["amd-sglang[all_hip]", "amd-sglang[test]"]
dev_hpu = ["sglang[all_hpu]", "sglang[test]"]
dev_musa = ["sglang[all_musa]", "sglang[test]"]

[project.urls]
"Homepage" = "https://github.com/sgl-project/sglang"
"Bug Tracker" = "https://github.com/sgl-project/sglang/issues"

[project.scripts]
sglang = "sglang.cli.main:main"
killall_sglang = "sglang.cli.killall:main"

[tool.setuptools.package-data]
"sglang" = [
Expand Down
Loading