Apply HF transformers patches from sglang init (#23103)

python/sglang/__init__.py

@@ -24,6 +24,11 @@ if _sys.platform == "darwin":
         pass
 del _sys
 
+from sglang.srt.utils.hf_transformers_patches import apply_all as _apply_hf_patches
+
+_apply_hf_patches()
+del _apply_hf_patches
+
 # Frontend Language APIs
 from sglang.global_config import global_config
 from sglang.lang.api import (

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

@@ -14,17 +14,16 @@
 """Hugging Face Transformers utilities.
 
 This package provides HF Transformers helpers, split into submodules
-(common, compat, config, tokenizer, processor, mistral_utils).
+(common, config, tokenizer, processor, mistral_utils).  Compatibility
+monkey-patches live in the sibling ``sglang.srt.utils.hf_transformers_patches``
+module and are applied at sglang import time.
 All public symbols are re-exported here for convenience.  The old import
 path ``sglang.srt.utils.hf_transformers_utils`` is preserved by a
 separate shim module.
 """
 
-from .compat import apply_all as _apply_compat
-
-_apply_compat()
-
-from .common import (  # noqa: E402
+from ..hf_transformers_patches import normalize_rope_scaling_compat
+from .common import (
     CONTEXT_LENGTH_KEYS,
     AutoConfig,
     attach_additional_stop_token_ids,
@@ -37,10 +36,9 @@ from .common import (  # noqa: E402
     get_sparse_attention_config,
     get_tokenizer_from_processor,
 )
-from .compat import normalize_rope_scaling_compat  # noqa: E402
-from .config import get_config  # noqa: E402
-from .processor import get_processor  # noqa: E402
-from .tokenizer import (  # noqa: E402
+from .config import get_config
+from .processor import get_processor
+from .tokenizer import (
     _fix_added_tokens_encoding,
     _fix_v5_add_bos_eos_token,
     get_tokenizer,

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

@@ -52,7 +52,7 @@ from sglang.srt.configs.deepseek_ocr import DeepseekVLV2Config
 from sglang.srt.configs.internvl import InternVLChatConfig
 from sglang.srt.utils import get_bool_env_var, logger, lru_cache_frozenset
 
-from .compat import normalize_rope_scaling_compat
+from ..hf_transformers_patches import normalize_rope_scaling_compat
 
 if get_bool_env_var("SGLANG_USE_MODELSCOPE"):
     from modelscope import AutoConfig, GenerationConfig

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

@@ -23,6 +23,7 @@ from sglang.srt.connector import create_remote_connector
 from sglang.srt.utils import is_remote_url, logger, lru_cache_frozenset
 from sglang.srt.utils.runai_utils import ObjectStorageModel, is_runai_obj_uri
 
+from ..hf_transformers_patches import _ensure_gguf_version
 from .common import (
     _CONFIG_REGISTRY,
     AutoConfig,
@@ -34,7 +35,6 @@ from .common import (
     check_gguf_file,
     get_hf_text_config,
 )
-from .compat import _ensure_gguf_version
 from .mistral_utils import is_mistral_model, load_mistral_config
 
 

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

@@ -30,12 +30,12 @@ from sglang.srt.utils import is_remote_url, logger
 from sglang.srt.utils.patch_tokenizer import patch_tokenizer
 from sglang.srt.utils.runai_utils import ObjectStorageModel, is_runai_obj_uri
 
+from ..hf_transformers_patches import _ensure_gguf_version
 from .common import (
     _resolve_local_or_cached_file,
     attach_additional_stop_token_ids,
     check_gguf_file,
 )
-from .compat import _ensure_gguf_version, patch_is_base_mistral_in_ci
 from .mistral_utils import (
     _MISTRAL_TOKENIZER_REDIRECTS,
     patch_mistral_common_tokenizer,
@@ -462,7 +462,6 @@ def get_tokenizer(
             kwargs["use_fast"] = True
 
     tokenizer_name = _resolve_tokenizer_name(tokenizer_name, kwargs)
-    patch_is_base_mistral_in_ci()
 
     common_kwargs = dict(
         trust_remote_code=trust_remote_code,

python/sglang/srt/utils/hf_transformers_patches.py

@@ -11,21 +11,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-"""Compatibility patches for transformers v5.x.
+"""Monkey-patches on transformers internals.
 
-This module applies monkey-patches to work around breaking changes in
-transformers v5.  Each patch is tagged with the upstream issue it works
-around so it can be removed once the upstream fix lands.
+Mix of backward-compat shims (re-add symbols removed in v5), workarounds
+for transformers v5 bugs, fixes for remote-model-code (trust_remote_code)
+that hasn't been updated for v5 yet, and CI-only patches (e.g. neutralize
+HF API calls to avoid rate limits).
 
 Import this module early (before any ``from_pretrained`` call) to activate
 all patches.  It is safe to import multiple times -- patches are idempotent.
-
-Patches fall into two categories:
-
-1. **Transformers bugs / regressions** -- issues in transformers itself.
-2. **Remote-model-code compat** -- remote model code (trust_remote_code)
-   that hasn't been updated for v5 yet.  These should be removed once
-   the model authors publish fixes.
 """
 
 import inspect
@@ -44,10 +38,19 @@ def apply_all():
     """Apply all transformers compatibility patches (idempotent).
 
     Call this once at import time.  It is safe to call multiple times.
+
+    No-op when the ``transformers`` package is not installed -- frontend-only
+    sglang users should not be forced to install transformers just to import
+    the top-level ``sglang`` package.
     """
     global _applied
     if _applied:
         return
+    try:
+        import transformers  # noqa: F401
+    except ImportError:
+        _applied = True
+        return
     _applied = True
 
     # v5.4 patches
@@ -62,6 +65,9 @@ def apply_all():
     _ensure_clean_up_tokenization_compat()
     _ensure_is_torch_fx_available_compat()
 
+    # CI-only: neutralize HF API calls inside tokenizer from_pretrained
+    patch_is_base_mistral_in_ci()
+
     logger.debug("transformers compatibility patches applied")