专栏算法工具链compile 一直没有生成hbm模型

compile 一直没有生成hbm模型

3小时前
9
3

芯片型号:J6m

开发包:3.8.1-py310_20260326

完成qat_train 后执行compile_perf_hbir.py 生成出compile_hbir.bc 后一直卡主,不能生成出hbm模型

测试了12个小时一直没有导出,内存没有变化

2026-04-29 16:05:36,844 INFO {'type': 'HbirCompiler', 'march': 'nash-m', 'opt': 0, 'jobs': 144, 'debug': False, 'max_time_per_fc': 0.0, 'only_export': False, 'out_path': './lidar_data/ztao/port_train/pointpillars_5class_swmotion_v4/compile/model.hbm', 'progress_bar': True, 'advice': 1000000, 'input_source': ['ddr'], 'transpose_dim': defaultdict(<class 'dict'>, {}), 'split_dim': defaultdict(<class 'dict'>, {}), 'layer_details': False, 'hbdk3_compatible_mode': False, 'enable_vpu': True, 'model_path': './lidar_data/ztao/port_train/pointpillars_5class_swmotion_v4/qat.bc'}

2026-04-29 16:05:37,868 WARNING wrap usage has been changed, please pass necessary args

/open_explorer/scripts/models/hat/models/task_modules/maptr/sparse_head.py:157: FutureWarning: torch.cuda.amp.autocast(args...) is deprecated. Please use torch.amp.autocast('cuda', args...) instead.

@autocast(enabled=False)

/open_explorer/scripts/models/hat/models/task_modules/maptr/sparse_decoder.py:121: FutureWarning: torch.cuda.amp.autocast(args...) is deprecated. Please use torch.amp.autocast('cuda', args...) instead.

@autocast(enabled=False)

/open_explorer/scripts/models/hat/models/task_modules/maptr/sparse_decoder.py:139: FutureWarning: torch.cuda.amp.autocast(args...) is deprecated. Please use torch.amp.autocast('cuda', args...) instead.

@autocast(enabled=False)

2026-04-29 16:05:38,095 WARNING The pytree support of plugin is deprecated and will be removed soon, please use native pytree to get better experience.

2026-04-29 16:05:38,095 INFO build_from_registry(HbirCompiler) took 1.25s (HBIR load + constructor runs here)

2026-04-29 16:05:38,095 INFO HBIR already loaded from ./lidar_data/ztao/port_train/pointpillars_5class_swmotion_v4/qat.bc (load took 0.11s). Stages: graph edits → convert → save compile_hbir.bc → compile()→HBM.

2026-04-29 16:05:38,103 INFO Input node : points, compiled with ddr!!transpose_dim: None

2026-04-29 16:05:38,104 INFO Start convert hbir

2026-04-29 16:05:41,221 INFO Convert Succeed

2026-04-29 16:05:41,222 INFO convert(hbir → march=nash-m, enable_vpu=True) took 3.12s

Remove op type b30vpu.dequantize successfully, delete node type list ['Quantize', 'Dequantize']

Remove op type b30vpu.dequantize successfully, delete node type list ['Quantize', 'Dequantize']

Remove op type b30vpu.dequantize successfully, delete node type list ['Quantize', 'Dequantize']

2026-04-29 16:05:41,521 INFO remove_io_quant + save(compile_hbir.bc) took 0.30s

2026-04-29 16:05:41,521 INFO Compiled hbir saved: ./lidar_data/ztao/port_train/pointpillars_5class_swmotion_v4/compile/compile_hbir.bc

2026-04-29 16:05:41,521 INFO Starting hbdk compile() → ./lidar_data/ztao/port_train/pointpillars_5class_swmotion_v4/compile/model.hbm (opt=0, jobs=144, progress_bar=True, advice=1000000us). This step often takes minutes on large models; use progress_bar or a positive advice threshold for more toolchain output.

2026-04-29 16:05:41,521 INFO [hbdk compile] pid=1477730 — heartbeat every 30s (disable with --compile_heartbeat_sec 0).

[================- ]33%

附件:
算法工具链
技术深度解析征程6
评论1
0/1000
  • zyhang
    Lv.1

    你好,如果方便可以把你的的编译脚本和bc模型一起分享一下呢,我们交叉验证一下,以防开发环境问题~

    1小时前
    0
    2
    • 回复zyhang:

      可以提供一下邮箱吗

      1小时前
      0
    • 回复zyhang:

      附件中的compile 文件和脚本可以复现我的问题,麻烦支持一下,感谢

      刚刚
      0