Autotune off by default (#90)

* Autotune off by default * rework tests
ROCm · Oct 31, 2024 · 6345da6 · 6345da6
1 parent 4e1d6ef
commit 6345da6
Show file tree

Hide file tree

Showing 2 changed files with 8 additions and 5 deletions.
diff --git a/.github/workflows/amd_tests.yml b/.github/workflows/amd_tests.yml
@@ -66,14 +66,17 @@ jobs:
       - name: Flash Attention Tests
         run: |
           export FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE"
-          export FLASH_ATTENTION_TRITON_AMD_AUTOTUNE=0
           pytest tests/test_flash_attn_triton_amd.py
-      - name: AMD Kernel Tests
+      - name: AMD Tests
         run: |
           export FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE"
-          export FLASH_ATTENTION_TRITON_AMD_AUTOTUNE=0
           pytest -v -s flash_attn/flash_attn_triton_amd/test.py
-      - name: AMD Kernel Bench
+      - name: AMD Bench
         run: |
           export FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE"
+          python flash_attn/flash_attn_triton_amd/bench.py
+      - name: AMD Bench with Autotune
+        run: |
+          export FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE"
+          export FLASH_ATTENTION_TRITON_AMD_AUTOTUNE=1
           python flash_attn/flash_attn_triton_amd/bench.py
diff --git a/flash_attn/flash_attn_triton_amd/utils.py b/flash_attn/flash_attn_triton_amd/utils.py
@@ -3,7 +3,7 @@
 import os
 import triton
 
-AUTOTUNE = os.environ.get('FLASH_ATTENTION_TRITON_AMD_AUTOTUNE', '1').lower() in ('1', 'true', 'yes')
+AUTOTUNE = os.environ.get('FLASH_ATTENTION_TRITON_AMD_AUTOTUNE', '0').lower() in ('1', 'true', 'yes')
 DEBUG = os.environ.get('FLASH_ATTENTION_TRITON_AMD_DEBUG', '0').lower() in ('1', 'true', 'yes')
 PERF = os.environ.get('FLASH_ATTENTION_TRITON_AMD_PERF', '0').lower() in ('1', 'true', 'yes')