fix cuda ut bug of use_deterministic_algorithms (#805)

(cherry picked from commit 886d657)

Author: n1ck-guo

auto_round/autoround.py

@@ -224,7 +224,8 @@ def __init__(
         to_quant_block_names: Union[str, list, None] = kwargs.pop("to_quant_block_names", None)
         enable_norm_bias_tuning: bool = kwargs.pop("enable_norm_bias_tuning", False)
         enable_quanted_input: bool = kwargs.pop("enable_quanted_input", True)
-        disable_deterministic_algorithms = kwargs.pop("disable_deterministic_algorithms", False)
+        disable_deterministic_algorithms = kwargs.pop("disable_deterministic_algorithms", True)
+        enable_deterministic_algorithms = kwargs.pop("enable_deterministic_algorithms", False)
         static_kv_dtype = kwargs.pop("static_kv_dtype", None)
         device = kwargs.pop("device", None)
         self.quant_lm_head = kwargs.pop("quant_lm_head", False)
@@ -234,11 +235,19 @@ def __init__(
 
         if kwargs:
             logger.warning(f"unrecognized keys {list(kwargs.keys())} were passed. Please check them.")
+        if "CUBLAS_WORKSPACE_CONFIG" not in os.environ:
+            os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
+        # deprecated, default not to use torch.use_deterministic_algorithms
+        if not disable_deterministic_algorithms or enable_deterministic_algorithms:
+            if not disable_deterministic_algorithms:
+                logger.warning(
+                    "default not use deterministic_algorithms. disable_deterministic_algorithms is deprecated,"
+                    " please use enable_deterministic_algorithms instead. "
+                )
 
-        if not disable_deterministic_algorithms:
-            if "CUBLAS_WORKSPACE_CONFIG" not in os.environ:
-                os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
             torch.use_deterministic_algorithms(True, warn_only=False)
+        else:
+            torch.use_deterministic_algorithms(True, warn_only=True)
 
         if device is not None:
             logger.warning("`device` is deprecated, please use `device_map` instead")

auto_round/script/llm.py

@@ -206,7 +206,12 @@ def __init__(self, *args, **kwargs):
         self.add_argument("--enable_alg_ext", action="store_true", help="whether to enable probably better algorithm")
 
         self.add_argument(
-            "--disable_deterministic_algorithms", action="store_true", help="disable torch deterministic algorithms."
+            "--disable_deterministic_algorithms",
+            action="store_true",
+            help="deprecated, disable torch deterministic algorithms.",
+        )
+        self.add_argument(
+            "--enable_deterministic_algorithms", action="store_true", help="enable torch deterministic algorithms."
         )
 
         self.add_argument(
@@ -543,6 +548,11 @@ def tune(args):
     scheme = args.scheme.upper()
     if scheme not in PRESET_SCHEMES:
         raise ValueError(f"{scheme} is not supported. only {PRESET_SCHEMES.keys()} are supported ")
+    if args.disable_deterministic_algorithms:
+        logger.warning(
+            "default not use deterministic_algorithms. disable_deterministic_algorithms is deprecated,"
+            " please use enable_deterministic_algorithms instead. "
+        )
     autoround = round(
         model=model,
         tokenizer=tokenizer,
@@ -580,7 +590,7 @@ def tune(args):
         super_group_size=args.super_group_size,
         super_bits=args.super_bits,
         disable_opt_rtn=args.disable_opt_rtn,
-        disable_deterministic_algorithms=args.disable_deterministic_algorithms,
+        enable_deterministic_algorithms=args.enable_deterministic_algorithms,
         enable_alg_ext=args.enable_alg_ext,
         **mllm_kwargs,
     )

auto_round/script/mllm.py

@@ -173,7 +173,13 @@ def __init__(self, *args, **kwargs):
         self.add_argument("--enable_torch_compile", action="store_true", help="whether to enable torch compile")
 
         self.add_argument(
-            "--disable_deterministic_algorithms", action="store_true", help="disable torch deterministic algorithms."
+            "--disable_deterministic_algorithms",
+            action="store_true",
+            help="deprecated, disable torch deterministic algorithms.",
+        )
+
+        self.add_argument(
+            "--enable_deterministic_algorithms", action="store_true", help="enable torch deterministic algorithms."
         )
 
         ## ======================= VLM =======================
@@ -435,7 +441,11 @@ def tune(args):
     scheme = args.scheme.upper()
     if scheme not in PRESET_SCHEMES:
         raise ValueError(f"{scheme} is not supported. only {PRESET_SCHEMES.keys()} are supported ")
-
+    if args.disable_deterministic_algorithms:
+        logger.warning(
+            "default not use deterministic_algorithms. disable_deterministic_algorithms is deprecated,"
+            " please use enable_deterministic_algorithms instead. "
+        )
     autoround = round(
         model,
         tokenizer,
@@ -473,7 +483,7 @@ def tune(args):
         model_kwargs=model_kwargs,
         data_type=args.data_type,
         disable_opt_rtn=args.disable_opt_rtn,
-        disable_deterministic_algorithms=args.disable_deterministic_algorithms,
+        enable_deterministic_algorithms=args.enable_deterministic_algorithms,
     )
 
     model_name = args.model.rstrip("/")