regisss HF staff jwieczorekhabana commited on
Commit
b66c277
1 Parent(s): 0c81a60

Remove hmp from README and config (#1)

Browse files

- Remove hmp from README and config (3a96bae9a72d522f1dc877b7af4e703a8112ca4f)
- Update README.md (8254c33326390682daa600746b37990defbe1dae)


Co-authored-by: Jan Wieczorek <jwieczorekhabana@users.noreply.huggingface.co>

Files changed (2) hide show
  1. README.md +5 -7
  2. gaudi_config.json +1 -24
README.md CHANGED
@@ -13,18 +13,15 @@ This model only contains the `GaudiConfig` file for running the [bert-large-unca
13
  **This model contains no model weights, only a GaudiConfig.**
14
 
15
  This enables to specify:
16
- - `use_habana_mixed_precision`: whether to use Habana Mixed Precision (HMP)
17
- - `hmp_opt_level`: optimization level for HMP, see [here](https://docs.habana.ai/en/latest/PyTorch/PyTorch_Mixed_Precision/PT_Mixed_Precision.html#configuration-options) for a detailed explanation
18
- - `hmp_bf16_ops`: list of operators that should run in bf16
19
- - `hmp_fp32_ops`: list of operators that should run in fp32
20
- - `hmp_is_verbose`: verbosity
21
  - `use_fused_adam`: whether to use Habana's custom AdamW implementation
22
  - `use_fused_clip_norm`: whether to use Habana's fused gradient norm clipping operator
 
23
 
24
  ## Usage
25
 
26
  The model is instantiated the same way as in the Transformers library.
27
- The only difference is that there are a few new training arguments specific to HPUs.
 
28
 
29
  [Here](https://github.com/huggingface/optimum-habana/blob/main/examples/question-answering/run_qa.py) is a question-answering example script to fine-tune a model on SQuAD. You can run it with BERT Large with the following command:
30
  ```bash
@@ -43,7 +40,8 @@ python run_qa.py \
43
  --output_dir /tmp/squad/ \
44
  --use_habana \
45
  --use_lazy_mode \
46
- --throughput_warmup_steps 2
 
47
  ```
48
 
49
  Check the [documentation](https://huggingface.co/docs/optimum/habana/index) out for more advanced usage and examples.
 
13
  **This model contains no model weights, only a GaudiConfig.**
14
 
15
  This enables to specify:
 
 
 
 
 
16
  - `use_fused_adam`: whether to use Habana's custom AdamW implementation
17
  - `use_fused_clip_norm`: whether to use Habana's fused gradient norm clipping operator
18
+ - `use_torch_autocast`: whether to use Torch Autocast for managing mixed precision
19
 
20
  ## Usage
21
 
22
  The model is instantiated the same way as in the Transformers library.
23
+ The only difference is that there are a few new training arguments specific to HPUs.\
24
+ It is strongly recommended to train this model doing bf16 mixed-precision training for optimal performance and accuracy.
25
 
26
  [Here](https://github.com/huggingface/optimum-habana/blob/main/examples/question-answering/run_qa.py) is a question-answering example script to fine-tune a model on SQuAD. You can run it with BERT Large with the following command:
27
  ```bash
 
40
  --output_dir /tmp/squad/ \
41
  --use_habana \
42
  --use_lazy_mode \
43
+ --throughput_warmup_steps 2 \
44
+ --bf16
45
  ```
46
 
47
  Check the [documentation](https://huggingface.co/docs/optimum/habana/index) out for more advanced usage and examples.
gaudi_config.json CHANGED
@@ -1,28 +1,5 @@
1
  {
2
- "use_habana_mixed_precision": true,
3
- "hmp_is_verbose": false,
4
  "use_fused_adam": true,
5
  "use_fused_clip_norm": true,
6
- "hmp_bf16_ops": [
7
- "add",
8
- "addmm",
9
- "bmm",
10
- "div",
11
- "dropout",
12
- "gelu",
13
- "iadd",
14
- "linear",
15
- "layer_norm",
16
- "matmul",
17
- "mm",
18
- "rsub",
19
- "softmax",
20
- "truediv"
21
- ],
22
- "hmp_fp32_ops": [
23
- "embedding",
24
- "nll_loss",
25
- "log_softmax",
26
- "cross_entropy"
27
- ]
28
  }
 
1
  {
 
 
2
  "use_fused_adam": true,
3
  "use_fused_clip_norm": true,
4
+ "user_torch_autocast": true
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5
  }