fix: coordinate autotune_warmup sequence (#1156)

Co-authored-by: niushengxiao <[email protected]>
Co-authored-by: hiworldwzj <[email protected]>

Author: 3 people

docker/cuda_version_12.8.0/Dockerfile.nixl

@@ -81,7 +81,7 @@ RUN apt-get update && apt-get install -y cmake automake autotools-dev  libtool l
 RUN apt-get update && apt-get install -y  pkg-config tmux net-tools libaio-dev ;  \
     cd /usr/local/src; \
     pip install --upgrade meson pybind11 patchelf; \
-    git clone https://github.com/ai-dynamo/nixl.git -b main && \
+    git clone https://github.com/ai-dynamo/nixl.git -b 0.8.0 && \
     cd nixl && \
     rm -rf build && \
     mkdir build && \

docker/cuda_version_12.8.0/Dockerfile.nixl.deepep

@@ -108,7 +108,7 @@ RUN apt-get update && apt-get install -y cmake automake autotools-dev  libtool l
 RUN apt-get update && apt-get install -y  pkg-config tmux net-tools libaio-dev ;  \
     cd /usr/local/src; \
     pip install --upgrade meson pybind11 patchelf; \
-    git clone https://github.com/ai-dynamo/nixl.git -b main && \
+    git clone https://github.com/ai-dynamo/nixl.git -b 0.8.0 && \
     cd nixl && \
     rm -rf build && \
     mkdir build && \

docker/cuda_version_12.8.0/Dockerfile.nixl.deepep.cache

@@ -110,7 +110,7 @@ RUN apt-get update && apt-get install -y cmake automake autotools-dev  libtool l
 RUN apt-get update && apt-get install -y  pkg-config tmux net-tools libaio-dev ;  \
     cd /usr/local/src; \
     pip install --upgrade meson pybind11 patchelf; \
-    git clone https://github.com/ai-dynamo/nixl.git -b main && \
+    git clone https://github.com/ai-dynamo/nixl.git -b 0.8.0 && \
     cd nixl && \
     rm -rf build && \
     mkdir build && \

lightllm/common/basemodel/basemodel.py

@@ -112,10 +112,10 @@ def __init__(self, kvargs):
         self._init_some_value()
         self._init_custom()
         self._init_inferstate_cls()
-        self._autotune_warmup()
-        self._init_padded_req()
         # wait必须在init cudagraph 之前，避免错误捕获
         self._wait_other_modules_ready()
+        self._autotune_warmup()
+        self._init_padded_req()
         self._init_cudagraph()
         self._init_prefill_cuda_graph()
         self._check_max_len_infer()

requirements.txt

@@ -92,4 +92,5 @@ torchvision==0.23.0
 interegular==0.3.3
 partial_json_parser==0.2.1.1.post6
 websockets==15.0.1
-cupy-cuda12x==13.6.0
+cupy-cuda12x==13.6.0
+nixl==0.8.0