biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Hardware][PPC64LE] Enable V1 for ppc64le and ARM (#20554)

Browse Source 
Signed-off-by: Akash Kaothalkar <akash.kaothalkar@ibm.com>
Co-authored-by: Akash Kaothalkar <akash.kaothalkar@ibm.com>
Co-authored-by: Nikhil Gupta <nikhil.gupta2@arm.com>
This commit is contained in:
Akash kaothalkar
2025-07-09 08:30:41 +05:30
committed by GitHub
parent 977180c912
commit 6db31e7a27
4 changed files with 77 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
vllm/engine/arg_utils.py
View File

@@ -36,6 +36,7 @@ from vllm.config import (BlockSize, CacheConfig, CacheDType, CompilationConfig,
from vllm.executor.executor_base import ExecutorBase
from vllm.logger import init_logger
from vllm.model_executor.layers.quantization import QuantizationMethods
from vllm.platforms import CpuArchEnum, current_platform
from vllm.plugins import load_general_plugins
from vllm.reasoning import ReasoningParserManager
from vllm.test_utils import MODEL_WEIGHTS_S3_BUCKET, MODELS_ON_S3
@@ -1096,7 +1097,6 @@ class EngineArgs:
If VLLM_USE_V1 is specified by the user but the VllmConfig
is incompatible, we raise an error.
"""
from vllm.platforms import current_platform
current_platform.pre_register_and_update()
device_config = DeviceConfig(
@@ -1123,9 +1123,16 @@ class EngineArgs:
# Set default arguments for V0 or V1 Engine.
if use_v1:
self._set_default_args_v1(usage_context, model_config)
# Disable chunked prefill for POWER (ppc64le)/ARM CPUs in V1
if current_platform.is_cpu(
) and current_platform.get_cpu_architecture() in (
CpuArchEnum.POWERPC, CpuArchEnum.ARM):
logger.info(
"Chunked prefill is not supported for ARM and POWER CPUs; "
"disabling it for V1 backend.")
self.enable_chunked_prefill = False
else:
self._set_default_args_v0(model_config)
assert self.enable_chunked_prefill is not None
if envs.VLLM_ATTENTION_BACKEND in [STR_DUAL_CHUNK_FLASH_ATTN_VAL]:
@@ -1242,7 +1249,6 @@ class EngineArgs:
if self.enable_chunked_prefill and self.pipeline_parallel_size > 1:
raise ValueError("Multi-Step Chunked-Prefill is not supported "
"for pipeline-parallel-size > 1")
from vllm.platforms import current_platform
if current_platform.is_cpu():
logger.warning("Multi-Step (--num-scheduler-steps > 1) is "
"currently not supported for CPUs and has been "
@@ -1391,7 +1397,6 @@ class EngineArgs:
# Skip this check if we are running on a non-GPU platform,
# or if the device capability is not available
# (e.g. in a Ray actor without GPUs).
from vllm.platforms import current_platform
if (current_platform.is_cuda()
and current_platform.get_device_capability()
and current_platform.get_device_capability().major < 8):
@@ -1652,7 +1657,6 @@ class EngineArgs:
# as the platform that vLLM is running on (e.g. the case of scaling
# vLLM with Ray) and has no GPUs. In this case we use the default
# values for non-H100/H200 GPUs.
from vllm.platforms import current_platform
try:
device_memory = current_platform.get_device_total_memory()
device_name = current_platform.get_device_name().lower()
@@ -1755,7 +1759,6 @@ class AsyncEngineArgs(EngineArgs):
parser.add_argument('--disable-log-requests',
action='store_true',
help='Disable logging requests.')
from vllm.platforms import current_platform
current_platform.pre_register_and_update(parser)
return parser