[XPU] upgrade sglang xpu backend to PyTorch 2.13 (#31751)
Co-authored-by: MingxuZh <109504044+MingxuZh@users.noreply.github.com> Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>
This commit is contained in:
yuchengliu1
co-authored by
MingxuZh
Ma Mingfei
MingxuZh
Ma Mingfei
parent
c82e928fe5
commit
744740dbea
@@ -1,7 +1,7 @@
|
||||
# docker build -t sglang:xpu -f xpu.Dockerfile --build-arg http_proxy=${http_proxy} --build-arg https_proxy=${https_proxy} --build-arg no_proxy=${no_proxy} --no-cache .
|
||||
|
||||
# Use Intel deep learning essentials base image with Ubuntu 24.04
|
||||
FROM intel/deep-learning-essentials:2025.3.2-0-devel-ubuntu24.04
|
||||
FROM intel/deep-learning-essentials:2026.0.0-devel-ubuntu24.04
|
||||
|
||||
# Avoid interactive prompts during package install
|
||||
ENV DEBIAN_FRONTEND=noninteractive
|
||||
@@ -19,9 +19,10 @@ USER root
|
||||
|
||||
# Pin Level-Zero UMD + IGC (rolling PPA once faulted libze on B580; see sgl-kernel-xpu#296).
|
||||
# Keep in lockstep with the host xe KMD; override via --build-arg.
|
||||
ARG COMPUTE_RUNTIME_VERSION=26.05.37020.3
|
||||
ARG IGC_VERSION=2.28.4+20760
|
||||
ARG GMM_VERSION=22.9.0
|
||||
ARG COMPUTE_RUNTIME_VERSION=26.18.38308.1
|
||||
ARG IGC_VERSION=2.34.4+21428
|
||||
ARG GMM_VERSION=22.10.0
|
||||
|
||||
RUN apt-get update && apt-get install -y software-properties-common curl && \
|
||||
add-apt-repository -y ppa:kobuk-team/intel-graphics && \
|
||||
apt-get update && \
|
||||
@@ -68,14 +69,14 @@ ENV PATH="$VIRTUAL_ENV/bin:$PATH"
|
||||
|
||||
WORKDIR /sgl-workspace
|
||||
|
||||
RUN pip install --no-cache-dir msgspec blake3 py-cpuinfo compressed_tensors gguf partial_json_parser einops tabulate --root-user-action=ignore && \
|
||||
pip install --no-cache-dir torch==2.12.0+xpu torchvision==0.27.0+xpu torchaudio==2.11.0+xpu --index-url https://download.pytorch.org/whl/xpu
|
||||
RUN pip install --no-cache-dir torch==2.13.0+xpu torchvision==0.28.0+xpu torchaudio==2.11.0+xpu --index-url https://download.pytorch.org/whl/xpu && \
|
||||
pip install --no-cache-dir msgspec blake3 py-cpuinfo compressed_tensors gguf partial_json_parser einops tabulate --root-user-action=ignore
|
||||
|
||||
RUN echo "Cloning ${SG_LANG_BRANCH} from ${SG_LANG_REPO}" && \
|
||||
git clone --branch ${SG_LANG_BRANCH} --single-branch ${SG_LANG_REPO} sglang && \
|
||||
cd sglang && cd python && \
|
||||
cp pyproject_xpu.toml pyproject.toml && \
|
||||
pip install --no-cache-dir . --extra-index-url https://download.pytorch.org/whl/xpu && \
|
||||
pip install --no-cache-dir ".[dev,diffusion]" --extra-index-url https://download.pytorch.org/whl/xpu && \
|
||||
pip install --no-cache-dir --no-deps xgrammar==0.1.33
|
||||
|
||||
CMD ["bash", "-c", "source /opt/intel/oneapi/setvars.sh --force && exec bash"]
|
||||
|
||||
@@ -53,7 +53,7 @@ conda create -n sgl-xpu python=3.12 -y
|
||||
conda activate sgl-xpu
|
||||
|
||||
# Set PyTorch XPU as primary pip install channel to avoid installing the larger CUDA-enabled version and prevent potential runtime issues.
|
||||
pip3 install torch==2.12.0+xpu torchvision==0.27.0+xpu torchaudio==2.11.0+xpu --index-url https://download.pytorch.org/whl/xpu
|
||||
pip3 install torch==2.13.0+xpu torchvision==0.28.0+xpu torchaudio==2.11.0+xpu --index-url https://download.pytorch.org/whl/xpu
|
||||
pip3 install xgrammar --no-deps # xgrammar will introduce CUDA-enabled triton which might conflict with XPU
|
||||
pip3 install apache-tvm-ffi # xgrammar requires apache-tvm-ffi
|
||||
|
||||
|
||||
@@ -87,11 +87,7 @@ pip install -e "python[all_musa]"
|
||||
|
||||
## Platform-Specific: Intel XPU
|
||||
|
||||
For Intel Data Center GPU Max or Arc GPUs, follow the [XPU installation guide](../hardware-platforms/xpu) to set up the base environment, then install diffusion dependencies:
|
||||
|
||||
```bash Command
|
||||
pip install -e "python[diffusion]"
|
||||
```
|
||||
For Intel Data Center GPU Max or Arc GPUs, follow the installation with docker in [XPU installation guide](../hardware-platforms/xpu). The dockerfile already include diffusion dependencies:
|
||||
|
||||
## Platform-Specific: Ascend NPU
|
||||
|
||||
|
||||
@@ -59,10 +59,10 @@ dependencies = [
|
||||
"soundfile==0.13.1",
|
||||
"tiktoken",
|
||||
"timm==1.0.16",
|
||||
"torch==2.12.0+xpu",
|
||||
"torch==2.13.0+xpu",
|
||||
"torchaudio==2.11.0+xpu",
|
||||
"torchcodec==0.12.0 ; sys_platform != 'linux' or (sys_platform == 'linux' and platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')", # torch==2.12.0 on XPU uses torchcodec 0.12.0
|
||||
"torchvision==0.27.0+xpu",
|
||||
"torchcodec==0.13.0 ; sys_platform != 'linux' or (sys_platform == 'linux' and platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')", # torch==2.13.0 on XPU uses torchcodec 0.13.0
|
||||
"torchvision==0.28.0+xpu",
|
||||
"tqdm",
|
||||
"transformers==5.12.1",
|
||||
"uvicorn",
|
||||
|
||||
@@ -1,7 +1,10 @@
|
||||
import unittest
|
||||
from typing import Optional
|
||||
|
||||
import torch
|
||||
|
||||
torch.use_deterministic_algorithms(True)
|
||||
|
||||
from sglang.srt.layers.moe.topk import (
|
||||
biased_grouped_topk_gpu,
|
||||
)
|
||||
@@ -33,6 +36,39 @@ def _scatter_by_expert(
|
||||
return dense
|
||||
|
||||
|
||||
def assert_equal(
|
||||
score: torch.Tensor,
|
||||
indices_ref: torch.Tensor,
|
||||
indices_our: torch.Tensor,
|
||||
bs: int,
|
||||
k: int,
|
||||
seq_len: int,
|
||||
topk_indices_offset: Optional[torch.Tensor] = None,
|
||||
max_permit_error: int = 0,
|
||||
):
|
||||
indices_our_cpu = indices_our.cpu().tolist()
|
||||
indices_ref_cpu = indices_ref.cpu().tolist()
|
||||
|
||||
wrong_values = 0
|
||||
for i in range(bs):
|
||||
indices_ref_set_i = set(indices_ref_cpu[i])
|
||||
indices_our_set_i = set(indices_our_cpu[i])
|
||||
more = indices_our_set_i - indices_ref_set_i
|
||||
less = indices_ref_set_i - indices_our_set_i
|
||||
offset = topk_indices_offset[i].item() if topk_indices_offset is not None else 0
|
||||
if len(more) > 0 or len(less) > 0:
|
||||
# check whether more values are the same with less values
|
||||
# if so, either one is acceptable, since their values are the same
|
||||
more_values = sorted(score[i, idx - offset].item() for idx in more)
|
||||
less_values = sorted(score[i, idx - offset].item() for idx in less)
|
||||
if more_values != less_values:
|
||||
wrong_values += len(more)
|
||||
print(
|
||||
f"{bs=}, {k=}, {seq_len=}, {i=}, {more=}, {less=} failed, with {more_values=}, {less_values=}"
|
||||
)
|
||||
assert wrong_values <= max_permit_error, f"{wrong_values=}, {max_permit_error=}"
|
||||
|
||||
|
||||
# Nemotron-3 uses biased_grouped_topk
|
||||
class TestBiasedGroupedTopK(CustomTestCase):
|
||||
def _run_single_test(
|
||||
@@ -226,15 +262,13 @@ class TestBiasedGroupedTopK(CustomTestCase):
|
||||
routed_scaling_factor,
|
||||
)
|
||||
|
||||
torch.testing.assert_close(
|
||||
_scatter_by_expert(
|
||||
topk_weights[:, :topk_routed], topk_ids[:, :topk_routed], E_num
|
||||
),
|
||||
_scatter_by_expert(
|
||||
ref_topk_weights[:, :topk_routed],
|
||||
assert_equal(
|
||||
gating_output,
|
||||
ref_topk_ids[:, :topk_routed],
|
||||
E_num,
|
||||
),
|
||||
topk_ids[:, :topk_routed],
|
||||
bs=len(bs),
|
||||
k=topk_value,
|
||||
seq_len=seq_len,
|
||||
)
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user