Bypass legacy GLM DSA layer types validation (#29454)

python/sglang/srt/utils/hf_transformers/config.py

@@ -51,6 +51,44 @@ def _apply_deepseek_ocr_overrides(config, model):
     config._name_or_path = model
 
 
+def _is_legacy_glm_moe_dsa_layer_types_error(error: Exception) -> bool:
+    error_msg = str(error)
+    return (
+        "validate_layer_type" in error_msg and "deepseek_sparse_attention" in error_msg
+    )
+
+
+def _load_glm_moe_dsa_config_without_legacy_layer_types(
+    model,
+    revision: Optional[str] = None,
+    **kwargs,
+):
+    from transformers import PretrainedConfig
+    from transformers.models.auto.configuration_auto import CONFIG_MAPPING
+
+    raw_config, unused_kwargs = PretrainedConfig.get_config_dict(
+        model, revision=revision, **kwargs
+    )
+    if raw_config.get("model_type") != "glm_moe_dsa" or raw_config.get(
+        "architectures"
+    ) != ["GlmMoeDsaForCausalLM"]:
+        return None
+
+    layer_types = raw_config.get("layer_types")
+    if not isinstance(layer_types, list) or any(
+        layer_type != "deepseek_sparse_attention" for layer_type in layer_types
+    ):
+        return None
+
+    raw_config = dict(raw_config)
+    raw_config.pop("layer_types", None)
+    config = CONFIG_MAPPING[raw_config["model_type"]].from_dict(
+        raw_config, **unused_kwargs
+    )
+    config._name_or_path = model
+    return config
+
+
 @register_model_config_parser("hf")
 class HfModelConfigParser(ModelConfigParserBase):
     def parse(
@@ -60,12 +98,23 @@ class HfModelConfigParser(ModelConfigParserBase):
         revision: Optional[str] = None,
         **kwargs,
     ):
-        config = AutoConfig.from_pretrained(
-            model,
-            trust_remote_code=trust_remote_code,
-            revision=revision,
-            **kwargs,
-        )
+        try:
+            config = AutoConfig.from_pretrained(
+                model,
+                trust_remote_code=trust_remote_code,
+                revision=revision,
+                **kwargs,
+            )
+        except Exception as e:
+            config = (
+                _load_glm_moe_dsa_config_without_legacy_layer_types(
+                    model, revision, **kwargs
+                )
+                if _is_legacy_glm_moe_dsa_layer_types_error(e)
+                else None
+            )
+            if config is None:
+                raise
 
         if (
             config.architectures is not None

python/sglang/srt/utils/hf_transformers/tokenizer.py

@@ -159,6 +159,33 @@ def _resolve_tokenizer_name(tokenizer_name, kwargs):
     return tokenizer_name
 
 
+# TODO: Remove after bumping huggingface transformers to v5.12
+def _retry_auto_tokenizer_with_glm_moe_dsa_config(
+    tokenizer_name, args, common_kwargs, error
+):
+    from .config import (
+        _is_legacy_glm_moe_dsa_layer_types_error,
+        _load_glm_moe_dsa_config_without_legacy_layer_types,
+    )
+
+    if not _is_legacy_glm_moe_dsa_layer_types_error(error):
+        return None
+
+    config = _load_glm_moe_dsa_config_without_legacy_layer_types(
+        tokenizer_name, revision=common_kwargs.get("tokenizer_revision")
+    )
+    if config is None:
+        return None
+
+    tokenizer = AutoTokenizer.from_pretrained(
+        tokenizer_name, *args, **{**common_kwargs, "config": config}
+    )
+    logging.getLogger(tokenizer.__class__.__module__).addFilter(
+        TokenizerWarningsFilter()
+    )
+    return tokenizer
+
+
 def _auto_tokenizer_from_pretrained(tokenizer_name, *args, **common_kwargs):
     """Call ``AutoTokenizer.from_pretrained`` with error handling."""
     try:
@@ -170,6 +197,11 @@ def _auto_tokenizer_from_pretrained(tokenizer_name, *args, **common_kwargs):
         )
         return tokenizer
     except TypeError as e:
+        tokenizer = _retry_auto_tokenizer_with_glm_moe_dsa_config(
+            tokenizer_name, args, common_kwargs, e
+        )
+        if tokenizer is not None:
+            return tokenizer
         err_msg = (
             "Failed to load the tokenizer. If you are using a LLaMA V1 model "
             f"consider using '{_FAST_LLAMA_TOKENIZER}' instead of the "
@@ -177,6 +209,11 @@ def _auto_tokenizer_from_pretrained(tokenizer_name, *args, **common_kwargs):
         )
         raise RuntimeError(err_msg) from e
     except ValueError as e:
+        tokenizer = _retry_auto_tokenizer_with_glm_moe_dsa_config(
+            tokenizer_name, args, common_kwargs, e
+        )
+        if tokenizer is not None:
+            return tokenizer
         # MistralCommon tokenizers reject standard HF kwargs like
         # trust_remote_code, use_fast etc. Retry without them.
         if "are not supported by" in str(e) and "MistralCommon" in str(e):
@@ -197,6 +234,13 @@ def _auto_tokenizer_from_pretrained(tokenizer_name, *args, **common_kwargs):
             )
             raise RuntimeError(err_msg) from e
         raise
+    except Exception as e:
+        tokenizer = _retry_auto_tokenizer_with_glm_moe_dsa_config(
+            tokenizer_name, args, common_kwargs, e
+        )
+        if tokenizer is not None:
+            return tokenizer
+        raise
 
 
 def _resolve_tokenizers_backend(tokenizer_name, *args, **common_kwargs):
@@ -218,7 +262,16 @@ def _resolve_tokenizers_backend(tokenizer_name, *args, **common_kwargs):
         tokenizer = AutoTokenizer.from_pretrained(
             tokenizer_name, *args, **common_kwargs
         )
-    except (ValueError, TypeError, OSError, ImportError, RuntimeError) as e:
+    except Exception as e:
+        tokenizer = _retry_auto_tokenizer_with_glm_moe_dsa_config(
+            tokenizer_name, args, common_kwargs, e
+        )
+        if tokenizer is not None:
+            return tokenizer
+        if not isinstance(
+            e, (ValueError, TypeError, OSError, ImportError, RuntimeError)
+        ):
+            raise
         raise RuntimeError(
             f"Retry with use_fast=False for {tokenizer_name} also failed "
             f"(initial load returned TokenizersBackend): {e}"