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

Support Wan2.2 S2V Quantization (#1924)

Wang, Changca37879959dc
diff --git a/auto_round/export/export_to_autoround/export_to_fp8.py b/auto_round/export/export_to_autoround/export_to_fp8.py
index 467df50ef..c40f5a33a 100644
--- a/auto_round/export/export_to_autoround/export_to_fp8.py
+++ b/auto_round/export/export_to_autoround/export_to_fp8.py
@@ -280,8 +280,10 @@ def save_quantized_as_autoround(
immediate_saving = is_immediate_saving_mode(model, serialization_dict)
if os.path.exists(output_dir) and not immediate_saving:
logger.warning(f"{output_dir} already exists, this may cause model conflict")
- if tokenizer is not None:
+ if tokenizer is not None and hasattr(tokenizer, "save_pretrained"):
tokenizer.save_pretrained(output_dir)
+ else:
+ print(f"Skip tokenizer save, tokenizer={tokenizer}")
if processor is not None:
processor.save_pretrained(output_dir)
diff --git a/auto_round/export/export_to_autoround/export_to_nvfp_mx.py b/auto_round/export/export_to_autoround/export_to_nvfp_mx.py
index c9607ae40..045f8f70f 100644
--- a/auto_round/export/export_to_autoround/export_to_nvfp_mx.py
+++ b/auto_round/export/export_to_autoround/export_to_nvfp_mx.py
@@ -269,8 +269,10 @@ def save_quantized_as_fp(
immediate_saving = is_immediate_saving_mode(model, serialization_dict)
if os.path.exists(output_dir) and not immediate_saving:
logger.warning(f"{output_dir} already exists, this may cause model conflict")
- if tokenizer is not None:
+ if tokenizer is not None and hasattr(tokenizer, "save_pretrained"):
tokenizer.save_pretrained(output_dir)
+ else:
+ print(f"Skip tokenizer save, tokenizer={tokenizer}")
if processor is not None:
processor.save_pretrained(output_dir)
diff --git a/auto_round/utils/model.py b/auto_round/utils/model.py
index a809ccd60..1cd28df7a 100644
--- a/auto_round/utils/model.py
+++ b/auto_round/utils/model.py
@@ -18,7 +18,7 @@ import os
import re
from collections import UserDict
from pathlib import Path
-from typing import TYPE_CHECKING, Union
+from typing import TYPE_CHECKING, Optional, Union
import psutil
import torch
@@ -949,10 +949,16 @@ _is_mllm_model_cache: dict = {}
_LLM_ONLY_MODEL_TYPES = {"bagel"}
+def get_model_name_or_path(model_or_path: Union[str, torch.nn.Module]) -> Optional[str]:
+ if isinstance(model_or_path, str):
+ return model_or_path
+ return getattr(model_or_path, "_name_or_path", None) or getattr(model_or_path, "name_or_path", None)
+
+
def is_mllm_model(model_or_path: Union[str, torch.nn.Module], platform: str = None):
from auto_round.utils.common import MM_KEYS
- model_path = model_or_path if isinstance(model_or_path, str) else model_or_path.name_or_path
+ model_path = get_model_name_or_path(model_or_path)
# Fast path: return cached result for already-seen paths
if model_path in _is_mllm_model_cache:
@@ -999,7 +1005,7 @@ def is_mllm_model(model_or_path: Union[str, torch.nn.Module], platform: str = No
break
# Cache by the original path key (model_path may have been resolved above)
- original_key = model_or_path if isinstance(model_or_path, str) else model_or_path.name_or_path
+ original_key = get_model_name_or_path(model_or_path)
_is_mllm_model_cache[original_key] = result
return result