Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

fix test_evaluation and AutoTuner new issue (#1417)

Xin He520214a8edda
diff --git a/auto_round_extension/triton/__init__.py b/auto_round_extension/triton/__init__.py
index 0d2740cdb..d932ae6b2 100644
--- a/auto_round_extension/triton/__init__.py
+++ b/auto_round_extension/triton/__init__.py
@@ -11,3 +11,24 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
+
+
+def patch_triton_autotuner_if_needed():
+ """Patch Autotuner if it's missing _cache_lock (compatibility fix for older triton versions)."""
+ try:
+ import threading
+ from triton.runtime.autotuner import Autotuner
+
+ if not hasattr(Autotuner, "_cache_lock"):
+ Autotuner._cache_lock = threading.RLock()
+ if not hasattr(Autotuner, "_cache_futures"):
+ Autotuner._cache_futures = {}
+ if not hasattr(Autotuner, "_cache"):
+ Autotuner._cache = {}
+ if not hasattr(Autotuner, "cache"):
+ Autotuner.cache = {}
+ except (ImportError, AttributeError):
+ pass
+
+
+patch_triton_autotuner_if_needed()
diff --git a/auto_round_extension/triton/triton_utils/dequant.py b/auto_round_extension/triton/triton_utils/dequant.py
index d599289db..e1564daa3 100644
--- a/auto_round_extension/triton/triton_utils/dequant.py
+++ b/auto_round_extension/triton/triton_utils/dequant.py
@@ -50,18 +50,6 @@ def make_dequant_configs(block_sizes, num_warps):
DEFAULT_DEQUANT_CONFIGS = make_dequant_configs([128, 256, 512, 1024], [4, 8])
-# Patch Autotuner if it's missing _cache_lock (compatibility fix for older triton versions)
-try:
- from triton.runtime.autotuner import Autotuner
-
- if not hasattr(Autotuner, "_cache_lock"):
- # Add missing attributes to make autotune work
- Autotuner._cache_lock = threading.RLock()
- if not hasattr(Autotuner, "cache"):
- Autotuner.cache = {}
-except (ImportError, AttributeError):
- pass
-
@triton.autotune(DEFAULT_DEQUANT_CONFIGS, key=["numels"])
@triton.jit
diff --git a/auto_round_extension/triton/triton_utils_zp/dequant.py b/auto_round_extension/triton/triton_utils_zp/dequant.py
index 288d27ea8..b4de5fb8f 100644
--- a/auto_round_extension/triton/triton_utils_zp/dequant.py
+++ b/auto_round_extension/triton/triton_utils_zp/dequant.py
@@ -50,18 +50,6 @@ def make_dequant_configs(block_sizes, num_warps):
DEFAULT_DEQUANT_CONFIGS = make_dequant_configs([128, 256, 512, 1024], [4, 8])
-# Patch Autotuner if it's missing _cache_lock (compatibility fix for older triton versions)
-try:
- from triton.runtime.autotuner import Autotuner
-
- if not hasattr(Autotuner, "_cache_lock"):
- # Add missing attributes to make autotune work
- Autotuner._cache_lock = threading.RLock()
- if not hasattr(Autotuner, "cache"):
- Autotuner.cache = {}
-except (ImportError, AttributeError):
- pass
-
@triton.autotune(DEFAULT_DEQUANT_CONFIGS, key=["numels"])
@triton.jit
diff --git a/test/test_cuda/advanced/test_evaluation.py b/test/test_cuda/advanced/test_evaluation.py
index 8f67633ab..16d991cb9 100644
--- a/test/test_cuda/advanced/test_evaluation.py
+++ b/test/test_cuda/advanced/test_evaluation.py
@@ -45,6 +45,7 @@ class TestVllmEvaluation:
python_path = sys.executable
os.environ["VLLM_SKIP_WARMUP"] = "true"
+ os.environ["NCCL_ASYNC_ERROR_HANDLING"] = "1"
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
# Test with custom vllm_args
@@ -59,6 +60,7 @@ class TestVllmEvaluation:
python_path = sys.executable
os.environ["VLLM_SKIP_WARMUP"] = "true"
+ os.environ["NCCL_ASYNC_ERROR_HANDLING"] = "1"
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
cmd = f"{python_path} -m auto_round --model {opt_name_or_path} --iters 0 --disable_opt_rtn --tasks lambada_openai --eval_bs 8 --eval_backend vllm --limit 100"