Support ByteDance-Seed/BAGEL-7B-MoT quantization in w4a16 format

auto_round/compressors/base.py

@@ -285,7 +285,7 @@ def __init__(
             config: Optional[AutoConfig] = None
             try:
                 config = AutoConfig.from_pretrained(model, trust_remote_code=self.trust_remote_code)
-            except (OSError, EnvironmentError) as e:
+            except (OSError, EnvironmentError, ValueError) as e:
                 logger.debug(
                     "Failed to load config via AutoConfig.from_pretrained for %s: %s. "
                     "Proceeding without config-based checks.",
@@ -399,6 +399,10 @@ def __init__(
             torch.use_deterministic_algorithms(True, warn_only=True)
 
         self.to_quant_block_names = to_quant_block_names
+        if self.to_quant_block_names is None:
+            _hint = getattr(model, "_autoround_to_quant_block_names", None)
+            if _hint is not None:
+                self.to_quant_block_names = _hint
         if not hasattr(self, "quant_block_list"):
             all_blocks = get_block_names(model)
             self.quant_block_list = find_matching_blocks(model, all_blocks, self.to_quant_block_names)

auto_round/modeling/unfused_moe/__init__.py

@@ -145,7 +145,10 @@ def get_file_path_via_model_name(model_or_path: str, file_name):
 
 def pre_check_config(model_name: str | torch.nn.Module, trust_remote_code: bool = True):
     if isinstance(model_name, str):
-        config = AutoConfig.from_pretrained(model_name, trust_remote_code=trust_remote_code)
+        try:
+            config = AutoConfig.from_pretrained(model_name, trust_remote_code=trust_remote_code)
+        except (OSError, EnvironmentError, ValueError):
+            return False
     elif isinstance(model_name, torch.nn.Module):
         config = getattr(model_name, "config", None)
         if config is None:

auto_round/special_model_handler.py

@@ -48,6 +48,7 @@
     "qwen2_5_omni",
     "qwen3_omni_moe",
     "gemma3",
+    "bagel",
 ]
 
 NOT_SUPPORT_ONLY_TEXT_MODELS = ["mllama", "mistral3_2"]
@@ -198,11 +199,35 @@ def _get_glm_image_multimodal_block(model, quant_vision=False):
     return block_names
 
 
+def _get_bagel_multimodal_block(model, quant_vision=False):
+    """Get block names for BAGEL not (Mixture of Transformers) model.
+
+    BAGEL model structure:
+    - language_model.model.layers: Qwen2-based LLM with not dual paths
+    - vit_model: SigLIP vision encoder (not quantized by default)
+    - connector: Vision-language MLP connector
+    - encoder/decoder: VAE autoencoder
+    - time_embedder, vae2llm, llm2vae: bridge modules
+
+    By default, only the language_model layers are quantized.
+    """
+    block_names = []
+
+    if hasattr(model, "language_model") and hasattr(model.language_model, "model"):
+        if hasattr(model.language_model.model, "layers"):
+            block_names.append(
+                [f"language_model.model.layers.{i}" for i in range(len(model.language_model.model.layers))]
+            )
+
+    return block_names
+
+
 SPECIAL_MULTIMODAL_BLOCK = {
     "deepseek_vl_v2": _get_deepseek_vl2_multimodal_block,
     "qwen2_5_omni": _get_qwen2_5_omni_multimodal_block,
     "qwen3_omni_moe": _get_qwen3_omni_moe_multimodal_block,
     "glm_image": _get_glm_image_multimodal_block,
+    "bagel": _get_bagel_multimodal_block,
 }
 
 
@@ -575,6 +600,34 @@ def get_glm_flash_ignore_layers(model) -> list[str]:
 )
 
 
+def get_bagel_ignore_layers(model) -> list[str]:
+    """Keep BAGEL generation-path modules in FP16.
+
+    BAGEL uses `*_moe_gen` modules for the image-generation path. Quantizing
+    them causes quality to collapse during the iterative denoising loop.
+    The shared attention projections are also highly sensitive.
+    """
+    ignore_layers = [
+        "moe_gen",
+        "self_attn.q_proj",
+        "self_attn.k_proj",
+        "self_attn.v_proj",
+        "self_attn.o_proj",
+    ]
+
+    return ignore_layers
+
+
+register_ignore_layers(
+    matchers=[
+        ModelTypeMatcher(r"bagel", mode="full"),
+    ],
+    ignore_layers=[
+        get_bagel_ignore_layers,
+    ],
+)
+
+
 def get_predefined_ignore_layers(model: torch.nn.Module) -> list[str]:
     layers = []
     for rule in _PRE_DEFINED_IGNORE_LAYERS: