| 23 | |
| 24 | # accelerator for Intel CPU |
| 25 | class CPU_Accelerator(DeepSpeedAccelerator): |
| 26 | |
| 27 | def __init__(self): |
| 28 | self._name = 'cpu' |
| 29 | self._compile_backend = "inductor" |
| 30 | if oneccl_imported_p: |
| 31 | self._communication_backend_name = 'ccl' |
| 32 | else: |
| 33 | # fallback to gloo if oneccl_binding_for_pytorch is not installed |
| 34 | self._communication_backend_name = 'gloo' |
| 35 | try: |
| 36 | import psutil |
| 37 | mem = psutil.Process().memory_info().rss |
| 38 | self.max_mem = mem |
| 39 | except ImportError as e: |
| 40 | self.max_mem = 0 |
| 41 | |
| 42 | def is_synchronized_device(self): |
| 43 | return True |
| 44 | |
| 45 | def use_host_timers(self): |
| 46 | return self.is_synchronized_device() |
| 47 | |
| 48 | def resolves_data_dependency(self): |
| 49 | return self.is_synchronized_device() |
| 50 | |
| 51 | def handles_memory_backpressure(self): |
| 52 | return self.is_synchronized_device() |
| 53 | |
| 54 | # Device APIs |
| 55 | def device_name(self, device_index=None): |
| 56 | return 'cpu' |
| 57 | |
| 58 | def device(self, device_index=None): |
| 59 | return None |
| 60 | |
| 61 | def set_device(self, device_index): |
| 62 | return |
| 63 | |
| 64 | def current_device(self): |
| 65 | return os.environ.get('LOCAL_RANK', 0) |
| 66 | |
| 67 | def current_device_name(self): |
| 68 | return 'cpu' |
| 69 | |
| 70 | def device_count(self): |
| 71 | device_count = int(os.environ.get('LOCAL_SIZE', 0)) |
| 72 | if device_count > 0: |
| 73 | return device_count |
| 74 | else: |
| 75 | from deepspeed.utils.numa import get_numa_cores |
| 76 | # Count NUMA node for number of cpu accelerators. On machine with HBM |
| 77 | # In flat mode, HBM is in separate NUMA node with no cores on this node. |
| 78 | # Ignore these NUMA nodes with no cores. |
| 79 | numa_core_lists = get_numa_cores() |
| 80 | if not numa_core_lists: |
| 81 | return 1 |
| 82 | numa_count = 0 |
no outgoing calls
no test coverage detected
searching dependent graphs…