Home - News ( United States | United Kingdom | Italy | Germany ) - Football scores

Showing content from https://github.com/jax-ml/jax/issues/12533 below:

Failed to initialize CUDA backend on multi-process distributed environments · Issue #12533 · jax-ml/jax · GitHub

The issue is happening when trying to initialize a multi-process multi-GPU environment with Slurm (but I think the problem is external to that).

Take the following simple script

import jax
import logging

logging.getLogger().setLevel(logging.DEBUG)

jax.distributed.initialize()

if jax.process_index() == 0:
  print(jax.devices())
  print(jax.device_count())    # total number of accelerator devices in the cluster
  print(jax.local_device_count())    # number of accelerator devices attached to this host

and executed with srun --gres=gpu:2 --ntasks=2 --nodes=1 python main.py and it return

INFO:absl:JAX distributed initialized with visible devices: 0
INFO:absl:JAX distributed initialized with visible devices: 1
INFO:absl:Starting JAX distributed service on ainode17:4192
INFO:absl:Connecting to JAX distributed service on ainode17:4192
INFO:absl:Connecting to JAX distributed service on ainode17:4192
DEBUG:absl:Initializing backend 'interpreter'
DEBUG:absl:Initializing backend 'interpreter'
DEBUG:absl:Backend 'interpreter' initialized
DEBUG:absl:Initializing backend 'cpu'
DEBUG:absl:Backend 'cpu' initialized
DEBUG:absl:Initializing backend 'tpu_driver'
INFO:absl:Unable to initialize backend 'tpu_driver': NOT_FOUND: Unable to find driver in registry given worker: 
DEBUG:absl:Initializing backend 'cuda'
DEBUG:absl:Backend 'interpreter' initialized
DEBUG:absl:Initializing backend 'cpu'
DEBUG:absl:Backend 'cpu' initialized
DEBUG:absl:Initializing backend 'tpu_driver'
INFO:absl:Unable to initialize backend 'tpu_driver': NOT_FOUND: Unable to find driver in registry given worker: 
DEBUG:absl:Initializing backend 'cuda'
2022-09-27 19:23:48.425044: E external/org_tensorflow/tensorflow/compiler/xla/status_macros.cc:57] INTERNAL: RET_CHECK failure (external/org_tensorflow/tensorflow/compiler/xla/pjrt/gpu_device.cc:345) local_device->device_ordinal() == local_topology.devices_size() 
*** Begin stack trace ***
	
	
	
	
	
	
	
	PyCFunction_Call
	_PyObject_MakeTpCall
	_PyEval_EvalFrameDefault
	_PyEval_EvalCodeWithName
	_PyFunction_Vectorcall
	PyObject_Call
	_PyEval_EvalFrameDefault
	_PyEval_EvalCodeWithName
	_PyFunction_Vectorcall
	_PyObject_FastCallDict
	
	_PyObject_MakeTpCall
	_PyEval_EvalFrameDefault
	_PyFunction_Vectorcall
	_PyEval_EvalFrameDefault
	_PyFunction_Vectorcall
	_PyEval_EvalFrameDefault
	_PyEval_EvalCodeWithName
	_PyFunction_Vectorcall
	_PyEval_EvalFrameDefault
	_PyEval_EvalCodeWithName
	_PyFunction_Vectorcall
	PyObject_Call
	
	_PyObject_MakeTpCall
	_PyEval_EvalFrameDefault
	
	_PyEval_EvalFrameDefault
	_PyEval_EvalCodeWithName
	PyEval_EvalCode
	
	
	
	PyRun_SimpleFileExFlags
	Py_RunMain
	Py_BytesMain
	__libc_start_main
	_start
*** End stack trace ***

INFO:absl:Unable to initialize backend 'cuda': INTERNAL: RET_CHECK failure (external/org_tensorflow/tensorflow/compiler/xla/pjrt/gpu_device.cc:345) local_device->device_ordinal() == local_topology.devices_size() 
DEBUG:absl:Initializing backend 'rocm'
INFO:absl:Unable to initialize backend 'rocm': NOT_FOUND: Could not find registered platform with name: "rocm". Available platform names are: CUDA Interpreter Host
DEBUG:absl:Initializing backend 'tpu'
INFO:absl:Unable to initialize backend 'tpu': module 'jaxlib.xla_extension' has no attribute 'get_tpu_client'
WARNING:absl:No GPU/TPU found, falling back to CPU. (Set TF_CPP_MIN_LOG_LEVEL=0 and rerun for more info.)

Recently (in 0.3.18) there has been an update to the interface for clusters (Slurm and TPUpods), but it doesn't look like it's due to that (i.e. manually setting coordinator_address, num_processes and process_id in distributed.initialize(...) has the same effect).

Am I doing something wrong?

jax==0.3.18, jaxlib==0.3.15+cuda11.cudnn82

GPUs

No response

No response

RetroSearch is an open source project built by @garambo | Open a GitHub Issue

Search and Browse the WWW like it's 1997 | Search results from DuckDuckGo

HTML: 3.2 | Encoding: UTF-8 | Version: 0.7.4