Fix compile.yml syntax (#12)

And logic in that file:
 - Enable it on all pushes, not just one modifying the file
 - Use `pushd`/`popd`
 - Fix typo in `output-path`
 - FIx model_path definition
 - Download tinystories/tokenizer from the right location
 - Setup Python-3.8
 - Add numpy to requirements
 - Skip export (unsupported in torch-2.2)
 - Break job into several smaller steps

Author: malfet

.github/workflows/compile.yml

@@ -4,42 +4,47 @@ on:
   push:
     branches:
       - main
-    paths:
-      - .github/workflows/compile.yml
   pull_request:
   workflow_dispatch:
 
 jobs:
-  do_compile:
+  run-tinystories:
     runs-on: macos-12
     # runs-on: self-hosted
     steps:
       - name: Checkout repo
         uses: actions/checkout@v2
-      - name: Compile and run
-        # shell: arch -arch arm64 zsh {0}
+      - name: Setup Python
+        uses: actions/setup-python@v2
+        with:
+          python-version: 3.8
+      - name: Print machine info
         run: |
           uname -a
           sysctl machdep.cpu.brand_string
           sysctl machdep.cpu.core_count
-	  pip install -r requirements.txt
-	  mkdir -p checkpoints/stories15M
-	  cd checkpoints/stories15M
-	  wget https://huggingface.co/karpathy/tinyllamas/raw/main/stories15M.pt
-	  wget https://huggingface.co/karpathy/tinyllamas/raw/main/stories15M.bin
-	  wget https://github.com/karpathy/llama2.c/blob/master/tokenizer.model
-	  cd ../..
-	  export MODEL_REPO=stories15M
-	  python aoti_export.py --checkpoint_path checkpoints/$MODEL_REPO/ --output-path ./${MODEL_REPO}.so
-	  python generate.py --checkpoint_path checkpoints/$MODEL_REPO/ --temperature 0 |& tee output_eager
-	  python generate.py --compile --checkpoint_path checkpoints/$MODEL_REPO/ --temperature 0 |& tee output_compiled
-	  python generate.py --checkpoint_path checkpoints/$MODEL_REPO/ --temperature 0 --dso ./${MODEL_REPO}.so |& tee output_aoti
-	  echo "******************************************"
-	  echo "********* EAGER vs TORCH.COMPILE *********"
-	  echo "******************************************"
-	  diff output_eager output_compiled
-	  echo "******************************************"
-	  echo "********* EAGER vs AOT INDUCTOR  *********
-	  echo "******************************************"
-	  diff output_eager output_aoti
-	  
+      - name: Install requirements
+        run: |
+          pip install -r requirements.txt
+      - name: Download checkpoints
+        run: |
+          mkdir -p checkpoints/stories15M
+          pushd checkpoints/stories15M
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories15M.pt
+          wget https://github.com/karpathy/llama2.c/raw/master/tokenizer.model
+          popd
+      - name: Run inference
+        run: |          
+          export MODEL_PATH=checkpoints/stories15M/stories15M.pt
+          # python aoti_export.py --checkpoint_path ${MODEL_PATH} --output_path ./${MODEL_REPO}.so
+          python generate.py --checkpoint_path ${MODEL_PATH} --temperature 0 --device cpu | tee output_eager
+          python generate.py --compile --checkpoint_path ${MODEL_PATH} --temperature 0 --device cpu | tee output_compiled
+          # python generate.py --checkpoint_path ${MODEL_PATH} --temperature 0 --dso ./${MODEL_REPO}.so |& tee output_aoti
+          # echo "******************************************"
+          # echo "********* EAGER vs TORCH.COMPILE *********"
+          # echo "******************************************"
+          # diff output_eager output_compiled
+          # echo "******************************************"
+          # echo "********* EAGER vs AOT INDUCTOR  *********"
+          # echo "******************************************"
+          # diff output_eager output_aoti

requirements.txt

@@ -1,2 +1,3 @@
 torch
 sentencepiece
+numpy