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, 03:15 UTC
[Hardware][AMD]: Replace HIPCC version with more precise ROCm version (#11515)
Signed-off-by: hjwei <hjwei_xd@163.com>
hj-wei committed
Dec 28, 2024 at 19:17 UTC
59d6bb4c863e511e58799efac847065c28c52c8b
1 file changed
+29
-23
setup.py
+29
-23
@@ -1,3 +1,4 @@
1
+import ctypes
2
import importlib.util
3
import logging
4
import os
@@ -13,7 +14,7 @@ from packaging.version import Version, parse
14
from setuptools import Extension, find_packages, setup
15
from setuptools.command.build_ext import build_ext
16
from setuptools_scm import get_version
16
-from torch.utils.cpp_extension import CUDA_HOME
17
+from torch.utils.cpp_extension import CUDA_HOME, ROCM_HOME
18
19
20
def load_module_from_path(module_name, path):
@@ -379,25 +380,31 @@ def _build_custom_ops() -> bool:
380
return _is_cuda() or _is_hip() or _is_cpu()
381
382
382
-def get_hipcc_rocm_version():
383
- # Run the hipcc --version command
384
- result = subprocess.run(['hipcc', '--version'],
385
- stdout=subprocess.PIPE,
386
- stderr=subprocess.STDOUT,
387
- text=True)
383
+def get_rocm_version():
384
+ # Get the Rocm version from the ROCM_HOME/bin/librocm-core.so
385
+ # see https://github.com/ROCm/rocm-core/blob/d11f5c20d500f729c393680a01fa902ebf92094b/rocm_version.cpp#L21
386
+ try:
387
+ librocm_core_file = Path(ROCM_HOME) / "lib" / "librocm-core.so"
388
+ if not librocm_core_file.is_file():
389
+ return None
390
+ librocm_core = ctypes.CDLL(librocm_core_file)
391
+ VerErrors = ctypes.c_uint32
392
+ get_rocm_core_version = librocm_core.getROCmVersion
393
+ get_rocm_core_version.restype = VerErrors
394
+ get_rocm_core_version.argtypes = [
395
+ ctypes.POINTER(ctypes.c_uint32),
396
+ ctypes.POINTER(ctypes.c_uint32),
397
+ ctypes.POINTER(ctypes.c_uint32),
398
+ ]
399
+ major = ctypes.c_uint32()
400
+ minor = ctypes.c_uint32()
401
+ patch = ctypes.c_uint32()
402
389
- # Check if the command was executed successfully
390
- if result.returncode != 0:
391
- print("Error running 'hipcc --version'")
403
+ if (get_rocm_core_version(ctypes.byref(major), ctypes.byref(minor),
404
+ ctypes.byref(patch)) == 0):
405
+ return "%d.%d.%d" % (major.value, minor.value, patch.value)
406
return None
393
-
394
- # Extract the version using a regular expression
395
- match = re.search(r'HIP version: (\S+)', result.stdout)
396
- if match:
397
- # Return the version string
398
- return match.group(1)
399
- else:
400
- print("Could not find HIP version in the output")
407
+ except Exception:
408
return None
409
410
@@ -479,11 +486,10 @@ def get_vllm_version() -> str:
486
if "sdist" not in sys.argv:
487
version += f"{sep}cu{cuda_version_str}"
488
elif _is_hip():
482
- # Get the HIP version
483
- hipcc_version = get_hipcc_rocm_version()
484
- if hipcc_version != MAIN_CUDA_VERSION:
485
- rocm_version_str = hipcc_version.replace(".", "")[:3]
486
- version += f"{sep}rocm{rocm_version_str}"
489
+ # Get the Rocm Version
490
+ rocm_version = get_rocm_version() or torch.version.hip
491
+ if rocm_version and rocm_version != MAIN_CUDA_VERSION:
492
+ version += f"{sep}rocm{rocm_version.replace('.', '')[:3]}"
493
elif _is_neuron():
494
# Get the Neuron version
495
neuron_version = str(get_neuronxcc_version())