[Deps] Bump transformers to 5.12.1 (#29393)

3rdparty/amd/wheel/sglang/pyproject.toml

@@ -32,6 +32,7 @@ runtime_common = [
   "huggingface_hub",
   "interegular",
   "llguidance>=0.7.11,<0.8.0",
+  "mistral_common>=1.11.5",
   "modelscope",
   "msgspec",
   "ninja",
@@ -59,7 +60,7 @@ runtime_common = [
   "timm==1.0.16",
   "torchao==0.9.0",
   "tqdm",
-  "transformers==5.8.1",
+  "transformers==5.12.1",
   "uvicorn",
   "uvloop",
   "xgrammar==0.2.1",

python/pyproject.toml

@@ -38,7 +38,7 @@ dependencies = [
   "IPython",
   "kernels>=0.14.1,<0.15",
   "llguidance>=0.7.11,<0.8.0",
-  "mistral_common>=1.11.0",
+  "mistral_common>=1.11.5",
   "modelscope",
   "msgspec",
   "ninja",
@@ -80,7 +80,7 @@ dependencies = [
   "torchcodec==0.11.1 ; sys_platform != 'linux' or (sys_platform == 'linux' and platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')", # torchcodec 0.11.1 for torch 2.11.x (0.10 is ABI-incompatible: references the pre-2.11 c10::MessageLogger ctor signature). Not available on Linux ARM.
   "torchvision",
   "tqdm",
-  "transformers==5.8.1",
+  "transformers==5.12.1",
   "uvicorn",
   "uvloop",
   "watchfiles",

python/pyproject_cpu.toml

@@ -31,7 +31,7 @@ dependencies = [
   "interegular",
   "IPython",
   "llguidance>=0.7.11,<0.8.0",
-  "mistral_common>=1.11.0",
+  "mistral_common>=1.11.5",
   "modelscope",
   "msgspec",
   "ninja",
@@ -64,7 +64,7 @@ dependencies = [
   "torchaudio==2.11.0",
   "torchvision==0.27.0",
   "tqdm",
-  "transformers==5.8.1",
+  "transformers==5.12.1",
   "triton==3.7.0",
   "uvicorn",
   "uvloop",

python/pyproject_npu.toml

@@ -33,7 +33,7 @@ dependencies = [
   "interegular",
   "IPython",
   "llguidance>=0.7.11,<0.8.0",
-  "mistral_common>=1.11.0",
+  "mistral_common>=1.11.5",
   "modelscope",
   "msgspec",
   "ninja",
@@ -62,7 +62,7 @@ dependencies = [
   "timm==1.0.16",
   "torchao==0.9.0",
   "tqdm",
-  "transformers==5.8.1",
+  "transformers==5.12.1",
   "uvicorn",
   "uvloop",
   "xgrammar==0.2.1",

python/pyproject_other.toml

@@ -33,7 +33,7 @@ runtime_common = [
   "interegular",
   "IPython",
   "llguidance>=0.7.11,<0.8.0",
-  "mistral_common>=1.11.0",
+  "mistral_common>=1.11.5",
   "modelscope",
   "msgspec",
   "ninja",
@@ -62,7 +62,7 @@ runtime_common = [
   "timm==1.0.16",
   "torchao==0.9.0",
   "tqdm",
-  "transformers==5.8.1",
+  "transformers==5.12.1",
   "uvicorn",
   "uvloop",
   "xgrammar==0.2.1",

python/pyproject_xpu.toml

@@ -31,7 +31,7 @@ dependencies = [
   "interegular",
   "IPython",
   "llguidance>=0.7.11,<0.8.0",
-  "mistral_common>=1.11.0",
+  "mistral_common>=1.11.5",
   "modelscope",
   "msgspec",
   "ninja",
@@ -65,7 +65,7 @@ dependencies = [
   "torchcodec==0.12.0 ; sys_platform != 'linux' or (sys_platform == 'linux' and platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')", # torch==2.12.0 on XPU uses torchcodec 0.12.0
   "torchvision==0.27.0+xpu",
   "tqdm",
-  "transformers==5.8.1",
+  "transformers==5.12.1",
   "uvicorn",
   "uvloop",
   # "xgrammar==0.2.1", xgrammar depends on CUDA PyTorch and Triton only

python/sglang/multimodal_gen/runtime/models/encoders/mistral_3.py

@@ -416,7 +416,6 @@ class MistralModel(MistralPreTrainedModel):
                 "config": self.config,
                 _CREATE_CAUSAL_MASK_ARG: inputs_embeds,
                 "attention_mask": attention_mask,
-                "cache_position": cache_position,
                 "past_key_values": past_key_values,
                 "position_ids": position_ids,
             }

python/sglang/multimodal_gen/runtime/models/encoders/qwen2_5vl.py

@@ -547,7 +547,6 @@ class Qwen2_5_VLTextModel(nn.Module):
                 "config": self.config,
                 "inputs_embeds": inputs_embeds,
                 "attention_mask": attention_mask,
-                "cache_position": cache_position,
                 "past_key_values": past_key_values,
                 "position_ids": text_position_ids,
             }

python/sglang/multimodal_gen/runtime/models/encoders/qwen3vl.py

@@ -584,9 +584,8 @@ class Qwen3VLTextModel(nn.Module):
 
         attention_mask = create_causal_mask(
             config=self.config,
-            input_embeds=inputs_embeds,
+            inputs_embeds=inputs_embeds,
             attention_mask=attention_mask,
-            cache_position=cache_position,
             past_key_values=past_key_values,
             position_ids=text_position_ids,
         )

python/sglang/multimodal_gen/test/single_test_file/component_accuracy/engine.py

@@ -17,13 +17,13 @@ from transformers import (
     UMT5EncoderModel,
 )
 
+# `AutoModelForVision2Seq` is not available in transformers 5.x.
+# `AutoModelForImageTextToText` is the supported name, so the legacy fallback
+# is gone. `AutoVisionTextModel` stays None-guarded below.
 try:
     from transformers import AutoModelForImageTextToText as AutoVisionTextModel
 except ImportError:
-    try:
+    AutoVisionTextModel = None
-        from transformers import AutoModelForVision2Seq as AutoVisionTextModel
-    except ImportError:
-        AutoVisionTextModel = None
 
 import sglang.multimodal_gen.runtime.managers.forward_context as fc_mod
 from sglang.multimodal_gen.runtime.distributed.parallel_state import (

python/sglang/srt/models/mimo_vl.py

@@ -377,8 +377,11 @@ class MiMoVisionTransformer(nn.Module):
 
             pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
         pos_ids = torch.cat(pos_ids, dim=0)
-        max_grid_size = grid_thw[:, 1:].max()
+        max_grid_size = int(grid_thw[:, 1:].max())
-        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
+        # transformers 5.12's rotary forward takes 1-D position_ids on the input device (grid_thw is CPU).
+        rotary_pos_emb_full = self.rotary_pos_emb(
+            torch.arange(max_grid_size, device=self.device)
+        )
         rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
         return rotary_pos_emb
 

python/sglang/srt/models/minicpmo.py

@@ -57,7 +57,10 @@ from sglang.srt.models.qwen2 import Qwen2ForCausalLM
 from sglang.srt.utils import get_device, logger
 
 try:
-    from transformers import LogitsWarper
+    # `LogitsWarper` was removed in transformers v4.48 (merged into
+    # `LogitsProcessor`, same `__call__(input_ids, scores)` contract). Alias it
+    # so the annotations and `_tts_deps` below stay valid.
+    from transformers.generation import LogitsProcessor as LogitsWarper
     from vector_quantize_pytorch import GroupedResidualFSQ
 
     _tts_deps = True

python/sglang/srt/models/moss_vl.py

@@ -321,8 +321,11 @@ class MossVLVisionModel(nn.Module):
             wpos_ids = wpos_ids.permute(0, 2, 1, 3).flatten()
             pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
         pos_ids = torch.cat(pos_ids, dim=0)
-        max_grid_size = grid_thw[:, 1:].max()
+        max_grid_size = int(grid_thw[:, 1:].max())
-        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
+        # transformers 5.12's rotary forward takes 1-D position_ids on the input device (grid_thw is CPU).
+        rotary_pos_emb_full = self.rotary_pos_emb(
+            torch.arange(max_grid_size, device=self.device)
+        )
         rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
         return rotary_pos_emb
 

python/sglang/srt/models/qwen2_5_vl.py

@@ -390,8 +390,11 @@ class Qwen2_5_VisionTransformer(nn.Module, RotaryPosMixin):
             pos_ids.append(base if t == 1 else base.repeat(t, 1))
 
         pos_ids = torch.cat(pos_ids, dim=0)
-        max_grid_size = grid_thw[:, 1:].max()
+        max_grid_size = int(grid_thw[:, 1:].max())
-        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
+        # transformers 5.12's rotary forward takes 1-D position_ids on the input device (grid_thw is CPU).
+        rotary_pos_emb_full = self.rotary_pos_emb(
+            torch.arange(max_grid_size, device=self.device)
+        )
         rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
         return rotary_pos_emb
 

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

@@ -51,44 +51,6 @@ 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(
@@ -98,43 +60,12 @@ class HfModelConfigParser(ModelConfigParserBase):
         revision: Optional[str] = None,
         **kwargs,
     ):
-        try:
+        config = AutoConfig.from_pretrained(
-            config = AutoConfig.from_pretrained(
+            model,
-                model,
+            trust_remote_code=trust_remote_code,
-                trust_remote_code=trust_remote_code,
+            revision=revision,
-                revision=revision,
+            **kwargs,
-                **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
-            and config.architectures[0] == "GlmMoeDsaForCausalLM"
-        ):
-            # GlmMoeDsaConfig drops/clobbers raw checkpoint fields the DSA path
-            # needs, so re-read them from config.json and restore. Fixed upstream
-            # by https://github.com/huggingface/transformers/pull/46338; remove
-            # this block once SGLang requires transformers >= 5.10.
-            from transformers import PretrainedConfig
-
-            raw_config, _ = PretrainedConfig.get_config_dict(model, revision=revision)
-            for key in (
-                "qk_rope_head_dim",
-                "index_topk_freq",
-            ):
-                if key in raw_config:
-                    setattr(config, key, raw_config[key])
-            if hasattr(config, "qk_head_dim") and hasattr(config, "qk_nope_head_dim"):
-                config.qk_head_dim = config.qk_nope_head_dim + config.qk_rope_head_dim
 
         if (
             config.architectures is not None

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

@@ -464,6 +464,17 @@ def patch_mistral_common_tokenizer(tokenizer):
     if not hasattr(tokenizer, "get_added_vocab"):
         tokenizer.get_added_vocab = lambda: {}
 
+    # Keep the old no-op pad add working on transformers 5.12 MistralCommon.
+    _orig_add_special_tokens = tokenizer.add_special_tokens
+
+    def _safe_add_special_tokens(special_tokens_dict, *args, **kwargs):
+        if set(special_tokens_dict) == {"pad_token"}:
+            tokenizer.pad_token = special_tokens_dict["pad_token"]
+            return 0
+        return _orig_add_special_tokens(special_tokens_dict, *args, **kwargs)
+
+    tokenizer.add_special_tokens = _safe_add_special_tokens
+
     # Set a chat_template containing "audio" so that sglang's content format
     # detector returns "openai" (which preserves audio_url extraction).
     if not hasattr(tokenizer, "chat_template") or tokenizer.chat_template is None:
@@ -495,24 +506,99 @@ def patch_mistral_common_tokenizer(tokenizer):
         tokenizer.batch_decode, ["spaces_between_special_tokens"]
     )
 
+    if hasattr(tokenizer, "_text_to_ids"):
+        _orig_text_to_ids = tokenizer._text_to_ids
+        marker_to_id = {
+            "[IMG]": tokenizer.convert_tokens_to_ids("[IMG]"),
+            "[IMG_BREAK]": tokenizer.convert_tokens_to_ids("[IMG_BREAK]"),
+            "[IMG_END]": tokenizer.convert_tokens_to_ids("[IMG_END]"),
+        }
+
+        def _text_to_ids_with_pixtral_markers(text, add_special_tokens):
+            if not isinstance(text, str) or not any(
+                marker in text for marker in marker_to_id
+            ):
+                return _orig_text_to_ids(text, add_special_tokens)
+
+            ids = []
+            pos = 0
+            while pos < len(text):
+                next_marker = None
+                next_idx = len(text)
+                for marker in marker_to_id:
+                    marker_idx = text.find(marker, pos)
+                    if marker_idx != -1 and marker_idx < next_idx:
+                        next_marker = marker
+                        next_idx = marker_idx
+
+                if next_marker is None:
+                    ids.extend(_orig_text_to_ids(text[pos:], False))
+                    break
+                if next_idx > pos:
+                    ids.extend(_orig_text_to_ids(text[pos:next_idx], False))
+                ids.append(marker_to_id[next_marker])
+                pos = next_idx + len(next_marker)
+
+            if add_special_tokens:
+                return tokenizer.build_inputs_with_special_tokens(ids)
+            return ids
+
+        tokenizer._text_to_ids = _text_to_ids_with_pixtral_markers
+
     tokenizer._orig_apply_chat_template = tokenizer.apply_chat_template
 
+    def _adapt_placeholder_content_for_mistral_common(content):
+        if not isinstance(content, list):
+            return content
+
+        rendered_parts = []
+        has_placeholder = False
+        for part in content:
+            if not isinstance(part, dict):
+                return content
+            part_type = part.get("type")
+            if part_type in ("text", "input_text"):
+                rendered_parts.append(part.get("text", ""))
+            elif part_type == "image" and not any(
+                key in part for key in ("url", "path", "base64")
+            ):
+                has_placeholder = True
+                rendered_parts.append("[IMG]")
+            elif part_type in ("audio", "video") and not any(
+                key in part for key in ("url", "path", "base64")
+            ):
+                has_placeholder = True
+                continue
+            else:
+                return content
+
+        return "".join(rendered_parts) if has_placeholder else content
+
+    def _adapt_placeholder_messages_for_mistral_common(messages):
+        if not isinstance(messages, (list, tuple)):
+            return messages
+
+        adapted = []
+        for msg in messages:
+            if isinstance(msg, (list, tuple)):
+                adapted.append(_adapt_placeholder_messages_for_mistral_common(msg))
+            elif isinstance(msg, dict):
+                adapted.append(
+                    {
+                        **msg,
+                        "content": _adapt_placeholder_content_for_mistral_common(
+                            msg.get("content", "")
+                        ),
+                    }
+                )
+            else:
+                adapted.append(msg)
+        return adapted
+
     def _safe_apply_chat_template(messages, **kwargs):
         kwargs.pop("add_generation_prompt", None)
-        cleaned = []
+        messages = _adapt_placeholder_messages_for_mistral_common(messages)
-        for msg in messages:
+        return tokenizer._orig_apply_chat_template(messages, **kwargs)
-            if isinstance(msg, dict):
-                content = msg.get("content", "")
-                if isinstance(content, list):
-                    text_parts = [
-                        p.get("text", "")
-                        for p in content
-                        if isinstance(p, dict) and p.get("type") == "text"
-                    ]
-                    msg = {**msg, "content": " ".join(text_parts) if text_parts else ""}
-                cleaned.append(msg)
-            else:
-                cleaned.append(msg)
-        return tokenizer._orig_apply_chat_template(cleaned, **kwargs)
 
     tokenizer.apply_chat_template = _safe_apply_chat_template
+    return tokenizer

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

@@ -159,33 +159,6 @@ 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:
@@ -197,11 +170,6 @@ 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 "
@@ -209,11 +177,6 @@ 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):
@@ -234,13 +197,6 @@ 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):
@@ -262,16 +218,7 @@ def _resolve_tokenizers_backend(tokenizer_name, *args, **common_kwargs):
         tokenizer = AutoTokenizer.from_pretrained(
             tokenizer_name, *args, **common_kwargs
         )
-    except Exception as e:
+    except (ValueError, TypeError, OSError, ImportError, RuntimeError) 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}"