Yina Chen

Orca TF2 estimator cannot load tf savedmodel directly

1

I load tf savedmodel using ``` def model_creator(config): item_model = keras.models.load_model("/opt/work/guoqiong/friesian_embed/Training/Embedding/2tower_models/exp5/item-model", custom_objects={'tf':tf}) optimizer = tf.keras.optimizers.Adam(config["lr"]) item_model.compile(optimizer=optimizer, loss='binary_crossentropy', metrics=['binary_accuracy', 'binary_crossentropy', 'AUC']) return item_model est = Estimator.from_keras(model_creator=model_creator, verbose=False, config=config) ``` And every...

## Description Add Friesian Nearline User Guide ### 3. How to test? - [ ] Document test

- [ ] pip install bigdl-nano[pytorch] will install 0.14.0xxxx instead of 2.1.0xxxx - [ ] (PR #5106) unsetenv is not supported on windows, https://github.com/intel-analytics/BigDL/blob/main/python/nano/src/bigdl/nano/pytorch/__init__.py#L22 ```bash if 'KMP_INIT_AT_FORK' in os.environ: del...

## Description Fix nano pytorch import issue on windows since unsetenv is not supported on windows. Related issue: #5102

Documentation: - [x] Architecture of the serving pipelines - [x] Schema of the parquet files - [x] Schema of the redis tables and configs - [x] List of the services...

Friesian FeatureTable missing OPs

13

Operations supported by NVTabular: - [ ] Normalize (mean std, continuous columns) - [x] NormalizeMinMax (min max method) - [x] FillMedian (replaces missing values with the median value for the...

SparkXShards to_spark_df only process the first element in one partition

7

If a partition of an XShards has more than one pandas dataframe, the `to_spark_df` function will only process the first pdf. Error code: https://github.com/intel-analytics/BigDL/blob/main/python/orca/src/bigdl/orca/data/shard.py#L583

## Description Spec batch based on Xiangyu's commit. **PVC: only support llama now** ### 2. User API changes Llama: ``` tokenizer = LlamaTokenizer.from_pretrained(model_path, padding_side='left') tokenizer.pad_token_id = 0 with torch.inference_mode(): input_ids...

## Description ### 1. Why the change? ### 2. User API changes ### 3. Summary of the change ### 4. How to test? - [ ] N/A - [ ]...

The loss cannot converge when finetuning Llama2-7b-GPTQ on 4090

11

I finetuned https://huggingface.co/TheBloke/Llama-2-7B-GPTQ on 4090 using the code from this repo and modified the group_size in peft_utils.py, but it seems cannot converge. Only pass the learning rate = 3e-05 to...