@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 0e40ac9b7b
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 16, 15:45 UTC

[ci][build] fix vllm-flash-attn (#8699)

youkaichao committed Sep 21, 2024 at 23:24 UTC 0e40ac9b7b5d953dfe38933bc7d2fb0a6c8da53c
3 files changed +18
CMakeLists.txt
+3
@@ -382,6 +382,9 @@ endif()
382 # Set the parent build flag so that the vllm-flash-attn library does not redo compile flag and arch initialization.
383 set(VLLM_PARENT_BUILD ON)
384
385 +# Ensure the vllm/vllm_flash_attn directory exists before installation
386 +install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/vllm_flash_attn\")" COMPONENT vllm_flash_attn_c)
387 +
388 # Make sure vllm-flash-attn install rules are nested under vllm/
389 install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY FALSE)" COMPONENT vllm_flash_attn_c)
390 install(CODE "set(OLD_CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}\")" COMPONENT vllm_flash_attn_c)
setup.py
+15
@@ -258,6 +258,21 @@ class cmake_build_ext(build_ext):
258 ]
259 subprocess.check_call(install_args, cwd=self.build_temp)
260
261 + def run(self):
262 + # First, run the standard build_ext command to compile the extensions
263 + super().run()
264 +
265 + # copy vllm/vllm_flash_attn/*.py from self.build_lib to current
266 + # directory so that they can be included in the editable build
267 + import glob
268 + files = glob.glob(
269 + os.path.join(self.build_lib, "vllm", "vllm_flash_attn", "*.py"))
270 + for file in files:
271 + dst_file = os.path.join("vllm/vllm_flash_attn",
272 + os.path.basename(file))
273 + print(f"Copying {file} to {dst_file}")
274 + self.copy_file(file, dst_file)
275 +
276
277 def _no_device() -> bool:
278 return VLLM_TARGET_DEVICE == "empty"
vllm/vllm_flash_attn/.gitkeep