Ke Wen
|
20142ab542
|
Simplify Tensor Parallel implementation with PyTorch TP (#34184)
* Simplify Tensor Parallel implementation with PyTorch TP
* Move tp_plan to config
* Lint
* Format and warning
* Disable copy-from check
* Conditionally get attr from config
* make fix-copies
* Move base_model_tp_plan to PretrainedConfig
* Move TP into from_pretrained
* Add device context for load
* Do not serialize
* Move _tp_plan setting to post_init
* Add has_tp_plan
* Add test_tp
* Add 'Multi-gpu inference' doc
* Add backward support for device type identification
* Auto-detect accelerator
* supports_tp_plan
* copyright year
* Fix copy
|
2024-11-18 19:51:49 +01:00 |
|