added abstraction for target model backend (#278)

* added abstraction for target model backend

Co-authored-by: canghua <[email protected]>
Co-authored-by: zyksir <[email protected]>

* polish

---------

Co-authored-by: canghua <[email protected]>
Co-authored-by: zyksir <[email protected]>

Author: 3 people

.github/workflows/test.yaml

@@ -24,11 +24,29 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v4
 
+      - name: Restore cache
+        run: |
+          if [ -d /github/home/sf ] && [ ! -z "$(ls -A /github/home/sf/)" ]; then
+            cp -p -r /github/home/sf/* ./
+          fi
+
       - name: Install dependencies
+        shell: bash
         run: |
-          pip install -e .
+          # if sf venv does not exist, create it
+          if [ ! -d sf ]; then
+            uv venv sf -p 3.11
+          fi
+          source sf/bin/activate
+          uv pip install -v . --prerelease=allow
 
       - name: Run test
         timeout-minutes: 30
+        shell: bash
         run: |
+          source sf/bin/activate
           python -m unittest discover -s ./tests -p "test_*.py" -v
+
+      - name: Save cache
+        run: |
+          cp -p -r sf /github/home/

examples/run_llama3_eagle3_online.sh

@@ -11,12 +11,14 @@ torchrun \
     $ROOT_DIR/scripts/train_eagle3_online.py \
     --target-model-path meta-llama/Meta-Llama-3.1-8B-Instruct \
     --draft-model-config $ROOT_DIR/configs/llama3-8B-eagle3.json \
-    --train-data-path $ROOT_DIR/cache/dataset/sharegpt.jsonl \
+    --train-data-path $ROOT_DIR/cache/dataset/sharegpt_train.jsonl \
     --output-dir $ROOT_DIR/outputs/llama3-8b-eagle3 \
     --num-epochs 2 \
     --batch-size 2 \
     --learning-rate 1e-4 \
     --max-length 2048 \
     --chat-template llama3 \
     --cache-dir $ROOT_DIR/cache \
-    --attention-backend flex_attention
+    --attention-backend sdpa \
+    --target-model-backend sglang \
+    --log-interval 10