Download models/scripts/zero2_fp8.json from litwell/EVA-qwen: direct link, hf CLI and curl.
- Browser
- Download file 573 Bytes
-
https://huggingface.co/litwell/EVA-qwen/resolve/main/models/scripts/zero2_fp8.json
- Command line
-
hf download hf://litwell/EVA-qwen/models/scripts/zero2_fp8.json
-
curl -L -o zero2_fp8.json https://huggingface.co/litwell/EVA-qwen/resolve/main/models/scripts/zero2_fp8.json
573 Bytes
| { | |
| "fp16": { | |
| "enabled": "auto", | |
| "loss_scale": 0, | |
| "loss_scale_window": 1000, | |
| "initial_scale_power": 16, | |
| "hysteresis": 2, | |
| "min_loss_scale": 1 | |
| }, | |
| "bf16": { | |
| "enabled": "auto" | |
| }, | |
| "msamp": { | |
| "enabled": true, | |
| "opt_level": "O3", | |
| "use_te": false | |
| }, | |
| "train_micro_batch_size_per_gpu": "auto", | |
| "train_batch_size": "auto", | |
| "gradient_accumulation_steps": "auto", | |
| "zero_optimization": { | |
| "stage": 2, | |
| "overlap_comm": true, | |
| "contiguous_gradients": true, | |
| "sub_group_size": 1e9, | |
| "reduce_bucket_size": "auto" | |
| } | |
| } | |