PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
support HPU memory dump in MemoryMonitor (#1605)
Xin He6de3f6ae7da5
diff --git a/auto_round/utils/device.py b/auto_round/utils/device.pyindex 5d615cf59..747a4eb2b 100644--- a/auto_round/utils/device.py+++ b/auto_round/utils/device.py@@ -606,7 +606,7 @@ class ClearMemory:from auto_round.utils.device import is_hpex_availableif is_hpex_available():- memory_monitor.update_cpu()+ memory_monitor.update_hpu(device_list)returnelse:if device_list is not None:@@ -683,7 +683,9 @@ def check_memory_availability(device, inputs, weight, org_seqlen, org_bs):free_space = total_memory - used_memoryelif "hpu" in device: # pragma: no covercurrent_hpu_index = torch.hpu.current_device()- free_space = torch.hpu.memory_reserved(current_hpu_index)+ total_memory = torch.hpu.memory_cached(current_hpu_index)+ used_memory = torch.hpu.memory_allocated(current_hpu_index)+ free_space = total_memory - used_memoryelse:return True, org_seqlen, org_bs@@ -1585,6 +1587,13 @@ class MemoryMonitor:device_list = list(range(torch.cuda.device_count()))elif torch.xpu.is_available():device_list = list(range(torch.xpu.device_count()))+ elif is_hpex_available():+ try:+ import habana_frameworks.torch.hpu as hthpu # pylint: disable=E0401++ device_list = list(range(hthpu.device_count()))+ except Exception:+ device_list = [0]for device in device_list:if str(device) == "cpu":@@ -1597,6 +1606,13 @@ class MemoryMonitor:current_vram = torch.xpu.memory_reserved(device) / 1024**3 # GBif device == "xpu":device = "0"+ elif is_hpex_available():+ try:+ current_vram = torch.hpu.memory_allocated(device) / 1024**3 # GB+ except Exception:+ current_vram = 0.0+ if device == "hpu":+ device = "0"else:return@@ -1614,6 +1630,40 @@ class MemoryMonitor:current_ram = process.memory_info().rss / 1024**3 # GBself.peak_ram = max(self.peak_ram, current_ram)+ def update_hpu(self, device_list=None):+ """Update memory usage for HPU devices."""+ if not self.enabled:+ return+ # Track RAM+ self.update_cpu()+ # Track HPU VRAM+ if not is_hpex_available():+ return+ if device_list is None:+ try:+ import habana_frameworks.torch.hpu as hthpu # pylint: disable=E0401++ device_list = list(range(hthpu.device_count()))+ except Exception:+ device_list = [0]+ elif not isinstance(device_list, (list, tuple)):+ device_list = [device_list]+ for device in device_list:+ if str(device) == "cpu":+ continue+ try:+ current_vram = torch.hpu.memory_allocated(device) / 1024**3 # GB+ except Exception:+ continue+ dev_key = str(device)+ if dev_key == "hpu":+ dev_key = "0"+ dev_key = dev_key.split(":")[-1]+ if current_vram > 0:+ if dev_key not in self.peak_vram:+ self.peak_vram[dev_key] = 0.0+ self.peak_vram[dev_key] = max(self.peak_vram[dev_key], current_vram)+def reset(self):"""Reset all statistics."""self.peak_ram = 0.0