Aquila support multi devices (#552)

* checkout aquila files * add Aquila * add readme * fix * fix * fix * update readme * update readme * support cuda sft * update readme * format * isort check * update * update readme * format
Oneflow-Inc · Sep 18, 2024 · 8d35c08 · 8d35c08
1 parent 13056f4
commit 8d35c08
Show file tree

Hide file tree

Showing 9 changed files with 1,559 additions and 0 deletions.
diff --git a/projects/Aquila/README.md b/projects/Aquila/README.md
@@ -0,0 +1,58 @@
+
+
+## Aquila
+### 推理
+- cuda
+
+```bash
+python projects/Aquila/pipeline.py --model_path=/root/models/Aquila-7B --mode=huggingface
+```
+
+- npu
+
+```bash
+python projects/Aquila/pipeline.py --model_path=/data0/hf_models/Aquila-7B --mode=huggingface --device=npu
+```
+
+- xpu
+
+```bash
+python projects/Aquila/pipeline.py --model_path=/root/models/Aquila-7B --mode=huggingface --device=xpu
+```
+
+### 训练 
+- data preparation
+```bash
+python projects/Aquila/utils/data_prepare.py
+```
+- cuda
+```bash
+export NUM_GPUS=4
+python3 -m oneflow.distributed.launch \
+    --nproc_per_node ${NUM_GPUS} \
+    --nnodes 1 \
+    --node_rank 0 \
+    --master_addr 127.0.0.1 \
+    --master_port 12345 \
+        tools/train_net.py --config-file=projects/Aquila/configs/aquila_sft.py \
+            graph.enabled=True \
+            train.input_placement_device="cuda" \
+            train.dist.device_type="cuda" \
+            train.dist.pipeline_parallel_size=${NUM_GPUS}
+```
+
+- xpu
+```bash
+export NUM_GPUS=1
+python3 -m oneflow.distributed.launch \
+    --nproc_per_node ${NUM_GPUS} \
+    --nnodes 1 \
+    --node_rank 0 \
+    --master_addr 127.0.0.1 \
+    --master_port 12345 \
+        tools/train_net.py --config-file=projects/Aquila/configs/aquila_sft.py \
+            graph.enabled=False \
+            train.input_placement_device="xpu" \
+            train.dist.device_type="xpu" \
+            train.dist.pipeline_parallel_size=${NUM_GPUS}
+```