fix: pass lm_head to LogitsProcessor instead of calling forward()

In vLLM v0.6.2, ParallelLMHead.forward() raises RuntimeError since
its weights should be used through LogitsProcessor.linear_method.apply().
Pass lm_head as first arg to LogitsProcessor which handles the
hidden_states -> logits projection internally.

.gitignore

@@ -0,0 +1,240 @@
+# version file generated by setuptools-scm
+/vllm/_version.py
+
+# vllm-flash-attn built from source
+vllm/vllm_flash_attn/*
+
+# OpenAI triton kernels copied from source
+vllm/third_party/triton_kernels/*
+
+# FlashMLA interface copied from source
+vllm/third_party/flashmla/flash_mla_interface.py
+
+# triton jit
+.triton
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+cmake-build-*/
+CMakeUserPresets.json
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+/.deps/
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# generated files
+**/generated/**
+
+# uv
+uv.lock
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+docs/argparse
+docs/examples/*
+!docs/examples/README.md
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+.idea/
+
+# VSCode
+.vscode/
+
+# Claude
+CLAUDE.md
+.claude/
+
+# Codex
+AGENTS.md
+.codex/
+
+# Cursor
+.cursor/
+
+# DS Store
+.DS_Store
+
+# Results
+*.csv
+
+# Python pickle files
+*.pkl
+
+# Sphinx documentation
+_build/
+
+# vim swap files
+*.swo
+*.swp
+
+# hip files generated by PyTorch
+*.hip
+*_hip*
+hip_compat.h
+
+# Benchmark dataset
+benchmarks/**/*.json
+
+# Linting
+actionlint
+shellcheck*/
+
+# Ignore moe/marlin_moe gen code
+csrc/moe/marlin_moe_wna16/kernel_*
+
+# Ignore ep_kernels_workspace folder
+ep_kernels_workspace/
+
+# Allow tracked library source folders under submodules (e.g., benchmarks/lib)
+!vllm/benchmarks/lib/
+
+# Generated gRPC protobuf files (compiled at build time from vllm_engine.proto)
+vllm/grpc/vllm_engine_pb2.py
+vllm/grpc/vllm_engine_pb2_grpc.py
+vllm/grpc/vllm_engine_pb2.pyi

README.md

@@ -3,6 +3,7 @@
 # 寒武纪 mlu370 文本生成
 该模型测试框架在寒武纪mlu370 （X8/X4）加速卡上，基于vllm 推理引擎，适配了 Qwen1.5-1.8B-Chat 模型。
 
+
 * Qwen1.5-1.8B-Chat 是通义千问系列中一款约18亿参数、轻量级的中英文对话大模型，专为高效推理和多场景聊天交互设计。
 * Llama-2-7b-chat-hf：Meta 发布的 LLaMA 2 系列中 70 亿参数的对话优化版开源大模型，适合多轮聊天与通用任务。
 * ChatGLM3-6B：智谱 AI 推出的第 3 代 ChatGLM 系列中 60 亿参数的中英双语对话大模型，支持推理、代码和多任务能力。

vllm-v0.6.2/vllm/config.py

@@ -274,7 +274,13 @@ class ModelConfig:
         self, limit_mm_per_prompt: Optional[Mapping[str, int]]
     ) -> Optional["MultiModalConfig"]:
         architectures = getattr(self.hf_config, "architectures", [])
-        if ModelRegistry.is_multimodal_model(architectures):
+        if ModelRegistry.is_multimodal_model(
+            architectures,
+            model_path=self.model,
+            revision=self.revision,
+            trust_remote_code=self.trust_remote_code,
+            hf_config=self.hf_config,
+        ):
             return MultiModalConfig(limit_per_prompt=limit_mm_per_prompt or {})
 
         if limit_mm_per_prompt:
@@ -308,11 +314,23 @@ class ModelConfig:
 
     def _init_attention_free(self) -> bool:
         architectures = getattr(self.hf_config, "architectures", [])
-        return ModelRegistry.is_attention_free_model(architectures)
+        return ModelRegistry.is_attention_free_model(
+            architectures,
+            model_path=self.model,
+            revision=self.revision,
+            trust_remote_code=self.trust_remote_code,
+            hf_config=self.hf_config,
+        )
 
     def _init_has_inner_state(self) -> bool:
         architectures = getattr(self.hf_config, "architectures", [])
-        return ModelRegistry.model_has_inner_state(architectures)
+        return ModelRegistry.model_has_inner_state(
+            architectures,
+            model_path=self.model,
+            revision=self.revision,
+            trust_remote_code=self.trust_remote_code,
+            hf_config=self.hf_config,
+        )
 
     def _verify_tokenizer_mode(self) -> None:
         tokenizer_mode = self.tokenizer_mode.lower()
@@ -335,8 +353,20 @@ class ModelConfig:
         task_support: Dict[_Task, bool] = {
             # NOTE: Listed from highest to lowest priority,
             # in case the model supports multiple of them
-            "generate": ModelRegistry.is_text_generation_model(architectures),
+            "generate": ModelRegistry.is_text_generation_model(
-            "embedding": ModelRegistry.is_embedding_model(architectures),
+                architectures,
+                model_path=self.model,
+                revision=self.revision,
+                trust_remote_code=self.trust_remote_code,
+                hf_config=hf_config,
+            ),
+            "embedding": ModelRegistry.is_embedding_model(
+                architectures,
+                model_path=self.model,
+                revision=self.revision,
+                trust_remote_code=self.trust_remote_code,
+                hf_config=hf_config,
+            ),
         }
         supported_tasks_lst: List[_Task] = [
             task for task, is_supported in task_support.items() if is_supported

vllm-v0.6.2/vllm/model_executor/layers/linear.py

@@ -146,6 +146,7 @@ class LinearBase(torch.nn.Module):
         skip_bias_add: If true, skip adding bias but instead return it.
         params_dtype: Data type for the parameters.
         quant_config: Quantization configure.
+        return_bias: If False, return only output tensor instead of (output, bias) tuple.
     """
 
     def __init__(
@@ -156,6 +157,7 @@ class LinearBase(torch.nn.Module):
         params_dtype: Optional[torch.dtype] = None,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
+        return_bias: bool = True,
     ):
         super().__init__()
 
@@ -163,6 +165,7 @@ class LinearBase(torch.nn.Module):
         self.input_size = input_size
         self.output_size = output_size
         self.skip_bias_add = skip_bias_add
+        self.return_bias = return_bias
         if params_dtype is None:
             params_dtype = torch.get_default_dtype()
         self.params_dtype = params_dtype
@@ -198,13 +201,15 @@ class ReplicatedLinear(LinearBase):
                  skip_bias_add: bool = False,
                  params_dtype: Optional[torch.dtype] = None,
                  quant_config: Optional[QuantizationConfig] = None,
-                 prefix: str = ""):
+                 prefix: str = "",
+                 return_bias: bool = True):
         super().__init__(input_size,
                          output_size,
                          skip_bias_add,
                          params_dtype,
                          quant_config,
-                         prefix=prefix)
+                         prefix=prefix,
+                         return_bias=return_bias)
 
         # All the linear layer supports quant method.
         assert self.quant_method is not None
@@ -238,6 +243,9 @@ class ReplicatedLinear(LinearBase):
         bias = self.bias if not self.skip_bias_add else None
         assert self.quant_method is not None
         output = self.quant_method.apply(self, x, bias)
+        
+        if not self.return_bias:
+            return output
         output_bias = self.bias if self.skip_bias_add else None
         return output, output_bias
 
@@ -281,9 +289,10 @@ class ColumnParallelLinear(LinearBase):
                  params_dtype: Optional[torch.dtype] = None,
                  quant_config: Optional[QuantizationConfig] = None,
                  output_sizes: Optional[List[int]] = None,
-                 prefix: str = ""):
+                 prefix: str = "",
+                 return_bias: bool = True):
         super().__init__(input_size, output_size, skip_bias_add, params_dtype,
-                         quant_config, prefix)
+                         quant_config, prefix, return_bias=return_bias)
 
         self.gather_output = gather_output
 
@@ -375,6 +384,9 @@ class ColumnParallelLinear(LinearBase):
             output = tensor_model_parallel_all_gather(output_parallel)
         else:
             output = output_parallel
+        
+        if not self.return_bias:
+            return output
         output_bias = self.bias if self.skip_bias_add else None
         return output, output_bias
 
@@ -418,7 +430,8 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
                  skip_bias_add: bool = False,
                  params_dtype: Optional[torch.dtype] = None,
                  quant_config: Optional[QuantizationConfig] = None,
-                 prefix: str = ""):
+                 prefix: str = "",
+                 return_bias: bool = True):
         self.output_sizes = output_sizes
         tp_size = get_tensor_model_parallel_world_size()
         assert all(output_size % tp_size == 0 for output_size in output_sizes)
@@ -429,7 +442,8 @@ class MergedColumnParallelLinear(ColumnParallelLinear):
                          skip_bias_add=skip_bias_add,
                          params_dtype=params_dtype,
                          quant_config=quant_config,
-                         prefix=prefix)
+                         prefix=prefix,
+                         return_bias=return_bias)
 
     def weight_loader(self,
                       param: Parameter,
@@ -653,7 +667,8 @@ class QKVParallelLinear(ColumnParallelLinear):
                  skip_bias_add: bool = False,
                  params_dtype: Optional[torch.dtype] = None,
                  quant_config: Optional[QuantizationConfig] = None,
-                 prefix: str = ""):
+                 prefix: str = "",
+                 return_bias: bool = True):
         self.hidden_size = hidden_size
         self.head_size = head_size
         self.total_num_heads = total_num_heads
@@ -686,7 +701,8 @@ class QKVParallelLinear(ColumnParallelLinear):
                          skip_bias_add=skip_bias_add,
                          params_dtype=params_dtype,
                          quant_config=quant_config,
-                         prefix=prefix)
+                         prefix=prefix,
+                         return_bias=return_bias)
 
     def _get_shard_offset_mapping(self, loaded_shard_id: str):
         shard_offset_mapping = {
@@ -980,9 +996,10 @@ class RowParallelLinear(LinearBase):
                  params_dtype: Optional[torch.dtype] = None,
                  reduce_results: bool = True,
                  quant_config: Optional[QuantizationConfig] = None,
-                 prefix: str = ""):
+                 prefix: str = "",
+                 return_bias: bool = True):
         super().__init__(input_size, output_size, skip_bias_add, params_dtype,
-                         quant_config, prefix)
+                         quant_config, prefix, return_bias=return_bias)
 
         self.input_is_parallel = input_is_parallel
         self.reduce_results = reduce_results
@@ -1086,8 +1103,9 @@ class RowParallelLinear(LinearBase):
         else:
             output = output_parallel
 
+        if not self.return_bias:
+            return output
         output_bias = self.bias if self.skip_bias_add else None
-
         return output, output_bias
 
     def extra_repr(self) -> str:

vllm-v0.6.2/vllm/model_executor/model_loader/utils.py

@@ -32,7 +32,13 @@ def get_model_architecture(
             and "MixtralForCausalLM" in architectures):
         architectures = ["QuantMixtralForCausalLM"]
 
-    return ModelRegistry.resolve_model_cls(architectures)
+    return ModelRegistry.resolve_model_cls(
+        architectures,
+        model_path=model_config.model,
+        revision=model_config.revision,
+        trust_remote_code=model_config.trust_remote_code,
+        hf_config=model_config.hf_config,
+    )
 
 
 def get_architecture_class_name(model_config: ModelConfig) -> str:

vllm-v0.6.2/vllm/model_executor/models/registry.py

@@ -16,9 +16,11 @@ from typing import (AbstractSet, Callable, Dict, List, Optional, Tuple, Type,
 
 import cloudpickle
 import torch.nn as nn
+import transformers
 
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
+from vllm.transformers_utils.dynamic_module import try_get_class_from_dynamic_module
 
 from .interfaces import (has_inner_state, is_attention_free,
                          supports_multimodal, supports_pp)
@@ -157,6 +159,13 @@ _SPECULATIVE_DECODING_MODELS = {
     "MedusaModel": ("medusa", "Medusa"),
     "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
 }
+
+# Transformers backend models - wrapper classes for custom HuggingFace models
+# These provide the vLLM interface for models loaded via auto_map
+_TRANSFORMERS_BACKEND_MODELS = {
+    # Text generation models
+    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
+}
 # yapf: enable
 
 _VLLM_MODELS = {
@@ -164,6 +173,7 @@ _VLLM_MODELS = {
     **_EMBEDDING_MODELS,
     **_MULTIMODAL_MODELS,
     **_SPECULATIVE_DECODING_MODELS,
+    **_TRANSFORMERS_BACKEND_MODELS,
 }
 
 # Models not supported by ROCm.
@@ -369,6 +379,91 @@ class _ModelRegistry:
 
         return _try_inspect_model_cls(model_arch, self.models[model_arch])
 
+    def _try_resolve_transformers(
+        self,
+        architecture: str,
+        model_path: str,
+        revision: Optional[str],
+        trust_remote_code: bool,
+        hf_config: Optional[object] = None,
+    ) -> Optional[str]:
+        """
+        Try to resolve a model architecture using the Transformers backend.
+        This allows loading custom models that define their own implementation
+        via the `auto_map` field in config.json.
+        
+        Returns the vLLM wrapper architecture name (e.g. "TransformersForCausalLM")
+        if the model can be loaded via auto_map, None otherwise.
+        """
+        # If architecture is already a transformers backend model, return it
+        if architecture in _TRANSFORMERS_BACKEND_MODELS:
+            return architecture
+        
+        # Check if architecture exists in transformers library
+        model_module = getattr(transformers, architecture, None)
+        if model_module is not None:
+            # Model exists in transformers, can use TransformersForCausalLM wrapper
+            logger.info(
+                "Architecture %s found in transformers library, "
+                "using TransformersForCausalLM wrapper",
+                architecture
+            )
+            return "TransformersForCausalLM"
+        
+        # Get auto_map from hf_config
+        auto_map: Dict[str, str] = {}
+        if hf_config is not None:
+            auto_map = getattr(hf_config, "auto_map", None) or {}
+        
+        if not auto_map:
+            return None
+        
+        # Try to load from auto_map to verify it works
+        # First, ensure config class is loaded
+        for name, module in auto_map.items():
+            if name.startswith("AutoConfig"):
+                try_get_class_from_dynamic_module(
+                    module,
+                    model_path,
+                    trust_remote_code=trust_remote_code,
+                    revision=revision,
+                    warn_on_fail=False,
+                )
+        
+        # Check if auto_map has a model class we can use
+        # Priority: AutoModelForCausalLM > AutoModelForSeq2SeqLM > AutoModel
+        auto_model_keys = sorted(
+            [k for k in auto_map.keys() if k.startswith("AutoModel")],
+            key=lambda x: (0 if "ForCausalLM" in x else (1 if "ForSeq2Seq" in x else 2))
+        )
+        
+        for name in auto_model_keys:
+            module = auto_map[name]
+            model_cls = try_get_class_from_dynamic_module(
+                module,
+                model_path,
+                trust_remote_code=trust_remote_code,
+                revision=revision,
+                warn_on_fail=True,
+            )
+            if model_cls is not None:
+                # Only log once per model class to avoid spam
+                log_key = f"{model_cls.__name__}_{name}"
+                if not hasattr(self, '_logged_custom_models'):
+                    self._logged_custom_models = set()
+                if log_key not in self._logged_custom_models:
+                    logger.info(
+                        "Found custom model class %s from auto_map[%s], "
+                        "using TransformersForCausalLM wrapper",
+                        model_cls.__name__,
+                        name
+                    )
+                    self._logged_custom_models.add(log_key)
+                # Return the wrapper architecture, not the actual class
+                return "TransformersForCausalLM"
+        
+        return None
+
     def _normalize_archs(
         self,
         architectures: Union[str, List[str]],
@@ -383,6 +478,10 @@ class _ModelRegistry:
     def inspect_model_cls(
         self,
         architectures: Union[str, List[str]],
+        model_path: Optional[str] = None,
+        revision: Optional[str] = None,
+        trust_remote_code: bool = False,
+        hf_config: Optional[object] = None,
     ) -> _ModelInfo:
         architectures = self._normalize_archs(architectures)
 
@@ -391,11 +490,27 @@ class _ModelRegistry:
             if model_info is not None:
                 return model_info
 
+        # Fallback: try to resolve using transformers backend (auto_map)
+        if model_path and trust_remote_code and hf_config:
+            for arch in architectures:
+                wrapper_arch = self._try_resolve_transformers(
+                    arch, model_path, revision, trust_remote_code, hf_config
+                )
+                if wrapper_arch is not None:
+                    # Use the wrapper architecture's ModelInfo
+                    model_info = self._try_inspect_model_cls(wrapper_arch)
+                    if model_info is not None:
+                        return model_info
+
         return self._raise_for_unsupported(architectures)
 
     def resolve_model_cls(
         self,
         architectures: Union[str, List[str]],
+        model_path: Optional[str] = None,
+        revision: Optional[str] = None,
+        trust_remote_code: bool = False,
+        hf_config: Optional[object] = None,
     ) -> Tuple[Type[nn.Module], str]:
         architectures = self._normalize_archs(architectures)
 
@@ -404,39 +519,91 @@ class _ModelRegistry:
             if model_cls is not None:
                 return (model_cls, arch)
 
+        # Fallback: try to resolve using transformers backend (auto_map)
+        if model_path and trust_remote_code and hf_config:
+            for arch in architectures:
+                wrapper_arch = self._try_resolve_transformers(
+                    arch, model_path, revision, trust_remote_code, hf_config
+                )
+                if wrapper_arch is not None:
+                    model_cls = self._try_load_model_cls(wrapper_arch)
+                    if model_cls is not None:
+                        # Return wrapper class but keep original architecture name
+                        return (model_cls, arch)
+
         return self._raise_for_unsupported(architectures)
 
     def is_text_generation_model(
         self,
         architectures: Union[str, List[str]],
+        model_path: Optional[str] = None,
+        revision: Optional[str] = None,
+        trust_remote_code: bool = False,
+        hf_config: Optional[object] = None,
     ) -> bool:
-        return self.inspect_model_cls(architectures).is_text_generation_model
+        return self.inspect_model_cls(
+            architectures, model_path, revision, trust_remote_code, hf_config
+        ).is_text_generation_model
 
     def is_embedding_model(
         self,
         architectures: Union[str, List[str]],
+        model_path: Optional[str] = None,
+        revision: Optional[str] = None,
+        trust_remote_code: bool = False,
+        hf_config: Optional[object] = None,
     ) -> bool:
-        return self.inspect_model_cls(architectures).is_embedding_model
+        return self.inspect_model_cls(
+            architectures, model_path, revision, trust_remote_code, hf_config
+        ).is_embedding_model
 
     def is_multimodal_model(
         self,
         architectures: Union[str, List[str]],
+        model_path: Optional[str] = None,
+        revision: Optional[str] = None,
+        trust_remote_code: bool = False,
+        hf_config: Optional[object] = None,
     ) -> bool:
-        return self.inspect_model_cls(architectures).supports_multimodal
+        return self.inspect_model_cls(
+            architectures, model_path, revision, trust_remote_code, hf_config
+        ).supports_multimodal
 
     def is_pp_supported_model(
         self,
         architectures: Union[str, List[str]],
+        model_path: Optional[str] = None,
+        revision: Optional[str] = None,
+        trust_remote_code: bool = False,
+        hf_config: Optional[object] = None,
     ) -> bool:
-        return self.inspect_model_cls(architectures).supports_pp
+        return self.inspect_model_cls(
+            architectures, model_path, revision, trust_remote_code, hf_config
+        ).supports_pp
 
-    def model_has_inner_state(self, architectures: Union[str,
+    def model_has_inner_state(
-                                                         List[str]]) -> bool:
+        self,
-        return self.inspect_model_cls(architectures).has_inner_state
+        architectures: Union[str, List[str]],
+        model_path: Optional[str] = None,
+        revision: Optional[str] = None,
+        trust_remote_code: bool = False,
+        hf_config: Optional[object] = None,
+    ) -> bool:
+        return self.inspect_model_cls(
+            architectures, model_path, revision, trust_remote_code, hf_config
+        ).has_inner_state
 
-    def is_attention_free_model(self, architectures: Union[str,
+    def is_attention_free_model(
-                                                           List[str]]) -> bool:
+        self,
-        return self.inspect_model_cls(architectures).is_attention_free
+        architectures: Union[str, List[str]],
+        model_path: Optional[str] = None,
+        revision: Optional[str] = None,
+        trust_remote_code: bool = False,
+        hf_config: Optional[object] = None,
+    ) -> bool:
+        return self.inspect_model_cls(
+            architectures, model_path, revision, trust_remote_code, hf_config
+        ).is_attention_free
 
 
 ModelRegistry = _ModelRegistry({

vllm-v0.6.2/vllm/model_executor/models/transformers/__init__.py

@@ -0,0 +1,127 @@
+# SPDX-License-Identifier: Apache-2.0
+# Copyright 2024 The vLLM team.
+"""Wrapper around `transformers` models for vLLM v0.6.2.
+
+This module provides the Transformers modeling backend that wraps
+any HuggingFace model with the vLLM interface, enabling support for custom
+models that define their implementation via `auto_map` in config.json.
+
+Architecture (following latest vLLM patterns):
+- Base: Core functionality (meta init, PP/TP support, module replacement, attention, weight loading)
+- CausalMixin: Causal LM specific (lm_head, compute_logits, sample)
+- EmbeddingMixin: Embedding/pooling specific (pooler, pooling)
+- SequenceClassificationMixin: Classification specific (classifier, pooling)
+
+Composed model classes:
+- TransformersForCausalLM = CausalMixin + Base
+- TransformersForEmbedding = EmbeddingMixin + Base
+- TransformersForSequenceClassification = SequenceClassificationMixin + Base
+
+Key optimizations:
+- Meta device initialization for memory efficiency
+- Pipeline Parallel support (PPMissingLayer)
+- Tensor Parallel support (tp_plan based module replacement)
+- Module replacement (Linear, RMSNorm, Embedding) with vLLM optimized versions
+- vLLM Attention instances for proper KV cache allocation
+- AutoWeightsLoader for efficient weight loading with name mapping
+"""
+
+from vllm.model_executor.models.transformers.base import (
+    Base,
+    set_attention_context,
+    clear_attention_context,
+    get_attention_context,
+    vllm_flash_attention_forward,
+)
+from vllm.model_executor.models.transformers.causal import CausalMixin
+from vllm.model_executor.models.transformers.pooling import (
+    EmbeddingMixin,
+    SequenceClassificationMixin,
+)
+from vllm.model_executor.models.transformers.legacy import LegacyMixin
+from vllm.model_executor.models.transformers.utils import (
+    init_on_device_without_buffers,
+    replace_linear_class,
+    replace_rms_norm_class,
+    log_replacement,
+    maybe_prefix,
+)
+
+
+# ============================================================================
+# Composed Model Classes (Mixin + Base pattern)
+# ============================================================================
+
+class TransformersForCausalLM(CausalMixin, Base):
+    """
+    Transformers backend wrapper for causal language models.
+    
+    Combines CausalMixin (lm_head, compute_logits, sample) with
+    Base (meta init, PP/TP support, module replacement, attention, weight loading).
+    
+    Supports any HuggingFace model with auto_map in config.json.
+    """
+    pass
+
+
+class TransformersForEmbedding(EmbeddingMixin, Base):
+    """
+    Transformers backend wrapper for embedding/sentence similarity models.
+    
+    Combines EmbeddingMixin (pooler, pooling) with
+    Base (meta init, PP/TP support, module replacement, attention, weight loading).
+    
+    Supports embedding models like BERT, sentence-transformers, etc.
+    """
+    pass
+
+
+class TransformersForSequenceClassification(SequenceClassificationMixin, Base):
+    """
+    Transformers backend wrapper for sequence classification models.
+    
+    Combines SequenceClassificationMixin (classifier, pooling) with
+    Base (meta init, PP/TP support, module replacement, attention, weight loading).
+    
+    Supports cross-encoders and classification models.
+    """
+    pass
+
+
+class TransformersForLegacy(LegacyMixin, EmbeddingMixin, Base):
+    """
+    Transformers backend wrapper for legacy/encoder models.
+    
+    Combines LegacyMixin (BERT/RoBERTa weight mapping, position handling) with
+    EmbeddingMixin (pooler) and Base (core functionality).
+    
+    Supports BERT, RoBERTa, and similar encoder models.
+    """
+    pass
+
+
+__all__ = [
+    # Main wrapper classes
+    "TransformersForCausalLM",
+    "TransformersForEmbedding",
+    "TransformersForSequenceClassification",
+    "TransformersForLegacy",
+    # Base class for extension
+    "Base",
+    # Mixin classes for custom combinations
+    "CausalMixin",
+    "EmbeddingMixin",
+    "SequenceClassificationMixin",
+    "LegacyMixin",
+    # Attention context management
+    "set_attention_context",
+    "clear_attention_context",
+    "get_attention_context",
+    "vllm_flash_attention_forward",
+    # Utility functions
+    "init_on_device_without_buffers",
+    "replace_linear_class",
+    "replace_rms_norm_class",
+    "log_replacement",
+    "maybe_prefix",
+]

vllm-v0.6.2/vllm/model_executor/models/transformers/base.py

@@ -0,0 +1,704 @@
+# SPDX-License-Identifier: Apache-2.0
+# Copyright 2024 The vLLM team.
+"""Transformers modeling backend base class for v0.6.2.
+
+This module provides the Base class following latest vLLM architecture:
+- Meta device initialization for memory efficiency
+- Pipeline parallel support (PPMissingLayer)
+- Tensor parallel support (tp_plan based module replacement)
+- Module replacement (Linear, RMSNorm) with vLLM optimized versions
+- VocabParallelEmbedding for input embeddings
+- Attention instances for KV cache allocation
+- Weight loading with AutoWeightsLoader and WeightsMapper
+"""
+
+import re
+from typing import TYPE_CHECKING, Dict, Iterable, List, Optional, Set, Tuple
+
+import torch
+import torch.nn as nn
+
+from vllm.config import VllmConfig
+from vllm.distributed import get_pp_group, get_tp_group
+from vllm.distributed.utils import get_pp_indices
+from vllm.logger import init_logger
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
+from vllm.model_executor.models.utils import (
+    AutoWeightsLoader,
+    PPMissingLayer,
+    WeightsMapper,
+    make_empty_intermediate_tensors_factory,
+)
+from vllm.attention.layer import Attention
+from vllm.sequence import IntermediateTensors
+
+from .utils import (
+    init_on_device_without_buffers,
+    replace_linear_class,
+    replace_rms_norm_class,
+    log_replacement,
+    maybe_prefix,
+)
+
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel
+    from vllm.attention import AttentionMetadata
+
+logger = init_logger(__name__)
+
+
+# ============================================================================
+# Attention Context Management (for vLLM attention integration)
+# ============================================================================
+
+_current_attn_metadata = None
+_current_kv_caches = None
+
+
+def set_attention_context(attn_metadata, kv_caches):
+    """Set the current attention context for vLLM attention functions."""
+    global _current_attn_metadata, _current_kv_caches
+    _current_attn_metadata = attn_metadata
+    _current_kv_caches = kv_caches
+
+
+def clear_attention_context():
+    """Clear the current attention context after forward pass."""
+    global _current_attn_metadata, _current_kv_caches
+    _current_attn_metadata = None
+    _current_kv_caches = None
+
+
+def get_attention_context():
+    """Get the current attention context."""
+    return _current_attn_metadata, _current_kv_caches
+
+
+# ============================================================================
+# vLLM Attention Function for Transformers Integration
+# ============================================================================
+
+def vllm_flash_attention_forward(
+    module: torch.nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: torch.Tensor,
+    scaling: float = None,
+    attention_instances: Dict[int, Attention] = None,
+    **kwargs,
+):
+    """
+    vLLM's optimized attention function for transformers integration.
+    
+    In v0.6.2, Attention.forward signature is:
+        (query, key, value, kv_cache, attn_metadata)
+    """
+    layer_idx = getattr(module, 'layer_idx', 0)
+    
+    if attention_instances is None or layer_idx not in attention_instances:
+        return _standard_attention(query, key, value, attention_mask, scaling)
+    
+    self_attn = attention_instances[layer_idx]
+    attn_metadata, kv_caches = get_attention_context()
+    
+    if attn_metadata is None or kv_caches is None:
+        return _standard_attention(query, key, value, attention_mask, scaling)
+    
+    if scaling is not None:
+        self_attn.impl.scale = float(scaling)
+    
+    # Reshape: [batch, heads, seq, head_dim] -> [seq, heads * head_dim]
+    hidden = query.shape[-2]
+    query, key, value = (x.transpose(1, 2) for x in (query, key, value))
+    query, key, value = (x.reshape(hidden, -1) for x in (query, key, value))
+    
+    kv_cache = kv_caches[layer_idx] if layer_idx < len(kv_caches) else None
+    output = self_attn.forward(query, key, value, kv_cache, attn_metadata)
+    
+    return output, None
+
+
+def _standard_attention(query, key, value, attention_mask, scaling):
+    """Standard scaled dot-product attention fallback."""
+    attn_weights = torch.matmul(query, key.transpose(-2, -1))
+    if scaling is not None:
+        attn_weights = attn_weights * scaling
+    if attention_mask is not None:
+        attn_weights = attn_weights + attention_mask
+    attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1)
+    attn_output = torch.matmul(attn_weights, value)
+    return attn_output, None
+
+
+# Register vLLM attention to transformers
+_vllm_attention_registered = False
+try:
+    from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
+    ALL_ATTENTION_FUNCTIONS["vllm"] = vllm_flash_attention_forward
+    _vllm_attention_registered = True
+    logger.info("Registered vLLM attention function to transformers")
+except (ImportError, AttributeError) as e:
+    logger.warning("Could not register vLLM attention: %s", e)
+
+
+# ============================================================================
+# Base Class with Pipeline Parallel and Tensor Parallel Support
+# ============================================================================
+
+class Base(nn.Module):
+    """
+    Base class for Transformers backend models with full parallel support.
+    
+    Features:
+    - Pipeline Parallel: PPMissingLayer for distributed layers
+    - Tensor Parallel: tp_plan based module replacement
+    - Meta device initialization
+    - Module replacement (Linear → vLLM Linear, RMSNorm → vLLM RMSNorm)
+    - VocabParallelEmbedding for input embeddings
+    - Attention instances for KV cache allocation
+    """
+    
+    # For vLLM's weight loader
+    embedding_modules = ["embed_tokens"]
+    
+    # Weight name mapping following latest vLLM pattern
+    hf_to_vllm_mapper = WeightsMapper(
+        orig_to_new_prefix={
+            # Add `model.` prefix for base model checkpoints,
+            # handling the case where it is already present
+            "": "model.",
+            "model.model.": "model.",
+            # Heads will be adjacent to `model` (pooling included because of adapters)
+            "model.lm_head.": "lm_head.",
+            "model.score.": "classifier.",
+            "model.classifier.": "classifier.",
+        }
+    )
+    
+    # Note: __init_subclass__ with WeightsMapper merging is not supported in v0.6.2
+    # because WeightsMapper doesn't implement __or__/__ior__ operators.
+    # Each Mixin should define its own hf_to_vllm_mapper if needed.
+    
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
+        super().__init__()
+        
+        logger.info("Using Transformers modeling backend.")
+        
+        # Store configuration
+        self.config = vllm_config.model_config.hf_config
+        self.text_config = getattr(self.config, "text_config", self.config)
+        self.model_config = vllm_config.model_config
+        self.cache_config = vllm_config.cache_config
+        self.device_config = vllm_config.device_config
+        self.parallel_config = vllm_config.parallel_config
+        self.quant_config = vllm_config.quant_config
+        self.prefix = prefix
+        
+        # Parallel groups
+        self.pp_group = get_pp_group()
+        self.tp_group = get_tp_group()
+        
+        # Model dimensions
+        self.hidden_size = getattr(self.text_config, "hidden_size", 4096)
+        self.vocab_size = getattr(self.text_config, "vocab_size", 32000)
+        
+        # Weight loading configuration
+        self.skip_prefixes: List[str] = []
+        self.ignore_unexpected_prefixes: List[str] = []
+        
+        # Configure attention backend
+        self._configure_attention_backend()
+        
+        # Create model on meta device
+        self._init_model_on_meta()
+        
+        # Apply pipeline parallel
+        self._apply_pipeline_parallel()
+        
+        # Replace modules (with tensor parallel support)
+        self._replace_modules()
+        
+        # Fix attention head_dim in case config was incorrect
+        self._fix_attention_head_dim()
+        
+        # Add debug hook to first attention module to capture tensor shapes
+        self._add_attention_debug_hook()
+        
+        # Replace input embeddings
+        self._replace_input_embeddings()
+        
+        # Create attention instances
+        self.attention_instances = self._create_attention_instances()
+        
+        # Initialize parameters on target device
+        self._init_parameters()
+        
+        # Pipeline parallel intermediate tensors
+        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
+            ["hidden_states"], self.hidden_size
+        )
+    
+    def _configure_attention_backend(self) -> None:
+        """Configure vLLM attention backend."""
+        # Note: attention implementation is set in _init_model_on_meta
+        # This method is kept for potential platform-specific configuration
+        pass
+    
+    def _init_model_on_meta(self) -> None:
+        """Create model structure on meta device."""
+        from transformers import AutoModel
+        
+        logger.info("Creating model structure on meta device...")
+        
+        # Set attention implementation to vLLM's
+        self.text_config._attn_implementation = "vllm"
+        
+        # Ensure head_dim is correctly set in BOTH config and text_config
+        # Transformers models use config.head_dim to compute attention dimensions
+        # Some models may have incorrect head_dim, so we compute and set it
+        if hasattr(self.text_config, "num_attention_heads") and hasattr(self.text_config, "hidden_size"):
+            correct_head_dim = self.text_config.hidden_size // self.text_config.num_attention_heads
+            
+            # Check and fix head_dim in text_config
+            if hasattr(self.text_config, "head_dim"):
+                if self.text_config.head_dim != correct_head_dim:
+                    logger.warning(
+                        "Correcting head_dim in text_config: %d -> %d",
+                        self.text_config.head_dim, correct_head_dim
+                    )
+                    self.text_config.head_dim = correct_head_dim
+            else:
+                self.text_config.head_dim = correct_head_dim
+            
+            # Also set in self.config (which is passed to AutoModel.from_config)
+            if hasattr(self.config, "head_dim"):
+                if self.config.head_dim != correct_head_dim:
+                    logger.warning(
+                        "Correcting head_dim in config: %d -> %d",
+                        self.config.head_dim, correct_head_dim
+                    )
+                    self.config.head_dim = correct_head_dim
+            else:
+                self.config.head_dim = correct_head_dim
+            
+            # Some models also need _attn_implementation in config
+            self.config._attn_implementation = "vllm"
+        
+        with init_on_device_without_buffers("meta"):
+            self.model: "PreTrainedModel" = AutoModel.from_config(
+                self.config,
+                torch_dtype=self.model_config.dtype,
+                trust_remote_code=self.model_config.trust_remote_code,
+            )
+        
+        self.model.eval()
+        for param in self.model.parameters():
+            param.requires_grad = False
+    
+    def _apply_pipeline_parallel(self) -> None:
+        """
+        Apply pipeline parallelization plan.
+        
+        For models that don't explicitly support pp_plan, we do a best-effort
+        approach by splitting layers based on num_hidden_layers.
+        """
+        if self.pp_group.world_size <= 1:
+            return
+        
+        logger.info("Applying pipeline parallel (world_size=%d, rank=%d)",
+                   self.pp_group.world_size, self.pp_group.rank_in_group)
+        
+        num_layers = getattr(self.text_config, "num_hidden_layers",
+                            getattr(self.text_config, "num_layers", 32))
+        
+        start_layer, end_layer = get_pp_indices(
+            num_layers,
+            self.pp_group.rank_in_group,
+            self.pp_group.world_size,
+        )
+        
+        # Find and process layer modules
+        layers_module = self._find_layers_module()
+        if layers_module is not None:
+            layers = list(layers_module.children())
+            for i, layer in enumerate(layers):
+                if not (start_layer <= i < end_layer):
+                    # Replace layers not on this rank with PPMissingLayer
+                    setattr(layers_module, str(i), PPMissingLayer())
+        
+        # Handle embeddings (only on first rank)
+        if not self.pp_group.is_first_rank:
+            input_embeddings = self.model.get_input_embeddings()
+            if input_embeddings is not None:
+                # Keep a reference but mark as missing for forward
+                self._has_embeddings = False
+        else:
+            self._has_embeddings = True
+        
+        # Handle final norm and lm_head (only on last rank)  
+        if not self.pp_group.is_last_rank:
+            # Mark lm_head as missing
+            if hasattr(self.model, 'lm_head'):
+                self.model.lm_head = PPMissingLayer()
+        
+        logger.info("Pipeline parallel applied: layers %d-%d on this rank",
+                   start_layer, end_layer)
+    
+    def _find_layers_module(self) -> Optional[nn.Module]:
+        """Find the ModuleList containing transformer layers."""
+        # Common layer container names
+        layer_names = ['layers', 'h', 'blocks', 'layer', 'encoder.layer', 'decoder.layers']
+        
+        def _search_layers(module: nn.Module, prefix: str = "") -> Optional[nn.Module]:
+            for name, child in module.named_children():
+                if name in ['layers', 'h', 'blocks', 'layer'] and isinstance(child, nn.ModuleList):
+                    return child
+                # Recursively search in model backbone
+                if name in ['model', 'transformer', 'encoder', 'decoder']:
+                    result = _search_layers(child, f"{prefix}.{name}" if prefix else name)
+                    if result is not None:
+                        return result
+            return None
+        
+        return _search_layers(self.model)
+    
+    def _get_tp_plan(self) -> Dict[str, str]:
+        """
+        Get tensor parallel plan for module replacement.
+        
+        This maps module name patterns to parallelization styles:
+        - "colwise": Column parallel (split output dim)
+        - "rowwise": Row parallel (split input dim)
+        - "replicate": Replicated (no split)
+        
+        Returns a dict mapping regex patterns to styles.
+        """
+        # Check if model has explicit tp_plan
+        if hasattr(self.model, 'tp_plan') and self.model.tp_plan:
+            return {maybe_prefix("model", k): v for k, v in self.model.tp_plan.items()}
+        
+        # Default tp_plan for common LLM architectures
+        # Based on typical transformer structure
+        return {
+            r".*\.q_proj$": "colwise",
+            r".*\.k_proj$": "colwise", 
+            r".*\.v_proj$": "colwise",
+            r".*\.o_proj$": "rowwise",
+            r".*\.gate_proj$": "colwise",
+            r".*\.up_proj$": "colwise",
+            r".*\.down_proj$": "rowwise",
+            r".*\.query$": "colwise",
+            r".*\.key$": "colwise",
+            r".*\.value$": "colwise",
+            r".*\.dense$": "rowwise",
+            r".*\.fc1$": "colwise",
+            r".*\.fc2$": "rowwise",
+        }
+    
+    def _replace_modules(self) -> None:
+        """
+        Replace modules with vLLM optimized versions.
+        
+        Uses tp_plan for tensor parallel style selection.
+        Note: lm_head is NOT replaced here - it's created at wrapper level by CausalMixin.
+        """
+        logger.info("Replacing modules with vLLM optimized versions...")
+        replaced_count = 0
+        
+        # Get tensor parallel plan
+        tp_plan = self._get_tp_plan() if self.tp_group.world_size > 1 else {}
+        
+        # Modules to skip replacement (handled at wrapper level)
+        skip_modules = {"lm_head", "score", "classifier"}
+        
+        def _recursive_replace(module: nn.Module, prefix: str = ""):
+            nonlocal replaced_count
+            
+            for name, child in list(module.named_children()):
+                # Skip PPMissingLayer
+                if isinstance(child, PPMissingLayer):
+                    continue
+                
+                # Skip modules that are handled at wrapper level
+                if name in skip_modules:
+                    logger.debug("Skipping %s (handled at wrapper level)", name)
+                    continue
+                    
+                qual_name = maybe_prefix(prefix, name)
+                new_module = None
+                
+                if isinstance(child, nn.Linear):
+                    # Determine parallelization style from tp_plan
+                    style = "replicate"
+                    for pattern, plan_style in tp_plan.items():
+                        if re.match(pattern, qual_name):
+                            style = plan_style
+                            break
+                    
+                    new_module = replace_linear_class(
+                        child,
+                        style=style,
+                        quant_config=self.quant_config,
+                        prefix=qual_name,
+                    )
+                    replaced_count += 1
+                    
+                elif child.__class__.__name__.endswith("RMSNorm") and \
+                        not isinstance(child, RMSNorm):
+                    new_module = replace_rms_norm_class(child, self.hidden_size)
+                    replaced_count += 1
+                
+                if new_module is not None:
+                    setattr(module, name, new_module)
+                    log_replacement(qual_name, child, new_module)
+                else:
+                    _recursive_replace(child, qual_name)
+        
+        _recursive_replace(self.model, "model")
+        logger.info("Replaced %d modules", replaced_count)
+    
+    def _add_attention_debug_hook(self) -> None:
+        """No-op. Debug hooks removed after root cause identified."""
+        pass
+    
+    def _fix_attention_head_dim(self) -> None:
+        """
+        Fix head_dim in attention modules and rotary embeddings after model creation.
+        
+        Some models may have incorrect head_dim in config, which causes
+        Transformers attention modules and RoPE to use wrong dimensions.
+        This method corrects head_dim in all attention modules and recreates
+        rotary embeddings if needed.
+        """
+        correct_head_dim = self.hidden_size // getattr(
+            self.text_config, "num_attention_heads", 32
+        )
+        
+        fixed_count = 0
+        
+        for name, module in self.model.named_modules():
+            module_name = module.__class__.__name__
+            
+            # Fix head_dim in Attention modules
+            if "Attention" in module_name:
+                if hasattr(module, "head_dim"):
+                    if module.head_dim != correct_head_dim:
+                        logger.warning(
+                            "Fixing head_dim in %s: %d -> %d",
+                            name, module.head_dim, correct_head_dim
+                        )
+                        module.head_dim = correct_head_dim
+                        fixed_count += 1
+            
+            # Fix rotary embeddings - recreate inv_freq buffer if needed
+            if "RotaryEmbedding" in module_name:
+                if hasattr(module, "inv_freq"):
+                    current_dim = module.inv_freq.shape[0] * 2
+                    if current_dim != correct_head_dim:
+                        logger.warning(
+                            "Recreating rotary embedding %s: dim %d -> %d",
+                            name, current_dim, correct_head_dim
+                        )
+                        base = getattr(module.config, 'rope_theta', 10000.0)
+                        if hasattr(module.config, 'rope_parameters'):
+                            base = module.config.rope_parameters.get('rope_theta', base)
+                        device = module.inv_freq.device
+                        inv_freq = 1.0 / (
+                            base ** (
+                                torch.arange(0, correct_head_dim, 2, dtype=torch.int64)
+                                .to(device=device, dtype=torch.float) / correct_head_dim
+                            )
+                        )
+                        module.register_buffer("inv_freq", inv_freq, persistent=False)
+                        if hasattr(module, "original_inv_freq"):
+                            module.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)
+        
+        if fixed_count > 0:
+            logger.info("Fixed head_dim in %d attention modules", fixed_count)
+    
+    def _replace_input_embeddings(self) -> None:
+        """Replace input embeddings with VocabParallelEmbedding."""
+        input_embeddings = self.model.get_input_embeddings()
+        if input_embeddings is None or isinstance(input_embeddings, PPMissingLayer):
+            return
+        
+        if hasattr(input_embeddings, "embedding_dim"):
+            embedding_dim = input_embeddings.embedding_dim
+        elif hasattr(input_embeddings, "weight"):
+            embedding_dim = input_embeddings.weight.shape[1]
+        else:
+            embedding_dim = self.hidden_size
+        
+        self.embed_scale = getattr(input_embeddings, "embed_scale", None)
+        
+        logger.info("Replacing input embeddings (vocab=%d, dim=%d)",
+                   self.vocab_size, embedding_dim)
+        
+        new_embeddings = VocabParallelEmbedding(
+            self.vocab_size,
+            embedding_dim,
+            org_num_embeddings=self.vocab_size,
+            quant_config=self.quant_config,
+        )
+        self.model.set_input_embeddings(new_embeddings)
+    
+    def _create_attention_instances(self) -> Dict[int, Attention]:
+        """Create Attention instances for KV cache allocation."""
+        num_layers = getattr(self.text_config, "num_hidden_layers",
+                            getattr(self.text_config, "num_layers", 32))
+        num_heads = getattr(self.text_config, "num_attention_heads", 32)
+        head_size = self.hidden_size // num_heads
+        num_kv_heads = getattr(self.text_config, "num_key_value_heads", num_heads)
+        
+        # Get PP layer range
+        pp_rank = self.pp_group.rank_in_group
+        pp_size = self.pp_group.world_size
+        start_layer, end_layer = get_pp_indices(num_layers, pp_rank, pp_size)
+        
+        logger.info("Creating attention instances for layers %d-%d "
+                   "(heads=%d, head_size=%d, kv_heads=%d)",
+                   start_layer, end_layer, num_heads, head_size, num_kv_heads)
+        
+        attention_instances: Dict[int, Attention] = {}
+        for layer_idx in range(start_layer, end_layer):
+            per_layer_sliding_window = None
+            if hasattr(self.config, "layer_types"):
+                layer_types = self.config.layer_types
+                if layer_idx < len(layer_types) and layer_types[layer_idx] == "sliding_attention":
+                    per_layer_sliding_window = getattr(self.config, "sliding_window", None)
+            
+            attention = Attention(
+                num_heads=num_heads,
+                head_size=head_size,
+                scale=1.0 / (head_size ** 0.5),
+                num_kv_heads=num_kv_heads,
+                cache_config=self.cache_config,
+                quant_config=self.quant_config,
+                prefix=f"model.layers.{layer_idx}.self_attn",
+            )
+            attention_instances[layer_idx] = attention
+        
+        return attention_instances
+    
+    def _init_parameters(self) -> None:
+        """Initialize parameters from meta device to target device."""
+        device = self.device_config.device
+        if device is None:
+            device = torch.device("cpu")
+        dtype = self.model_config.dtype
+        
+        def _init_params(module: nn.Module):
+            if isinstance(module, PPMissingLayer):
+                return
+            for name, param in list(module.named_parameters(recurse=False)):
+                if param.device == torch.device("meta"):
+                    new_param = nn.Parameter(
+                        torch.empty_like(param.data, dtype=dtype, device=device),
+                        requires_grad=False,
+                    )
+                    setattr(module, name, new_param)
+            for child in module.children():
+                _init_params(child)
+        
+        _init_params(self.model)
+        logger.info("Parameters initialized on %s", device)
+    
+    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
+        """Get embeddings for input IDs."""
+        inputs_embeds = self.model.get_input_embeddings()(input_ids)
+        if self.embed_scale is not None:
+            inputs_embeds = inputs_embeds * self.embed_scale
+        return inputs_embeds
+    
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: "AttentionMetadata",
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        """Forward pass with pipeline parallel support."""
+        # Handle intermediate tensors for PP
+        if not self.pp_group.is_first_rank:
+            assert intermediate_tensors is not None
+            input_ids = None
+            inputs_embeds = intermediate_tensors["hidden_states"]
+        
+        set_attention_context(attn_metadata, kv_caches)
+        
+        try:
+            # Prepare inputs
+            if inputs_embeds is not None:
+                if inputs_embeds.dim() == 2:
+                    inputs_embeds = inputs_embeds.unsqueeze(0)
+                model_inputs = {"inputs_embeds": inputs_embeds}
+            else:
+                if input_ids is not None and input_ids.dim() == 1:
+                    input_ids = input_ids.unsqueeze(0)
+                model_inputs = {"input_ids": input_ids}
+            
+            if positions is not None:
+                if positions.dim() == 1:
+                    positions = positions.unsqueeze(0)
+                model_inputs["position_ids"] = positions
+            
+            # Apply embed_scale if needed
+            if (
+                self.embed_scale is not None
+                and input_ids is not None
+                and inputs_embeds is None
+            ):
+                inputs_embeds = self.embed_input_ids(model_inputs["input_ids"])
+                model_inputs = {"inputs_embeds": inputs_embeds}
+                if positions is not None:
+                    model_inputs["position_ids"] = positions
+            
+            # Forward through model
+            # Note: return_dict=False returns tuple, first element is last hidden state
+            with torch.no_grad():
+                outputs = self.model(
+                    **model_inputs,
+                    use_cache=False,
+                    return_dict=False,
+                    attention_instances=self.attention_instances,
+                )
+            
+            # Get hidden states from model output
+            # For models using return_dict=False, outputs is a tuple
+            # outputs[0] is usually the last hidden state
+            if isinstance(outputs, tuple):
+                hidden_states = outputs[0]
+            else:
+                hidden_states = outputs
+            
+            # Remove batch dimension
+            if hidden_states.dim() == 3 and hidden_states.size(0) == 1:
+                hidden_states = hidden_states.squeeze(0)
+            
+            # Return intermediate tensors for PP
+            if not self.pp_group.is_last_rank:
+                return IntermediateTensors({"hidden_states": hidden_states})
+            
+            return hidden_states
+            
+        finally:
+            clear_attention_context()
+    
+    def load_weights(
+        self,
+        weights: Iterable[Tuple[str, torch.Tensor]],
+    ) -> Set[str]:
+        """Load weights using AutoWeightsLoader with name mapping."""
+        loader = AutoWeightsLoader(
+            self,
+            skip_prefixes=self.skip_prefixes,
+            ignore_unexpected_prefixes=self.ignore_unexpected_prefixes,
+        )
+        loaded = loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
+        logger.info("Loaded %d weight tensors", len(loaded))
+        return set(loaded)

vllm-v0.6.2/vllm/model_executor/models/transformers/causal.py

@@ -0,0 +1,142 @@
+# SPDX-License-Identifier: Apache-2.0
+# Copyright 2024 The vLLM team.
+"""Transformers modeling backend mixin for causal language models.
+
+This module provides CausalMixin that adds causal language model specific
+functionality (lm_head, compute_logits, sample) to the Base class.
+
+Following latest vLLM architecture:
+- TransformersForCausalLM = CausalMixin + Base
+- lm_head is created at the wrapper level (not inside self.model)
+"""
+
+from typing import TYPE_CHECKING, Optional
+
+import torch
+
+from vllm.logger import init_logger
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
+from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
+from vllm.model_executor.models.utils import PPMissingLayer, maybe_prefix
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+
+if TYPE_CHECKING:
+    from vllm.config import VllmConfig
+
+logger = init_logger(__name__)
+
+
+class CausalMixin:
+    """
+    Mixin class that adds causal language model functionality.
+    
+    This mixin provides:
+    - ParallelLMHead for language model head (created at wrapper level)
+    - LogitsProcessor for logits computation
+    - Sampler for token sampling
+    - compute_logits method for VllmModelForTextGeneration protocol
+    - sample method for VllmModelForTextGeneration protocol
+    
+    Following latest vLLM architecture:
+    - lm_head is a direct attribute of TransformersForCausalLM (not inside self.model)
+    - hf_to_vllm_mapper maps "model.lm_head." -> "lm_head." to handle this
+    - For tied embeddings, lm_head weight loading is skipped and weights are tied
+    
+    Should be used with Base class:
+        class TransformersForCausalLM(CausalMixin, Base): ...
+    """
+    
+    def __init__(self, *, vllm_config: "VllmConfig", prefix: str = "") -> None:
+        # Call next class in MRO (should be Base)
+        super().__init__(vllm_config=vllm_config, prefix=prefix)
+        
+        # Handle tied word embeddings - skip loading lm_head weights
+        tie_word_embeddings = getattr(self.text_config, "tie_word_embeddings", False)
+        if tie_word_embeddings:
+            self.skip_prefixes.append("lm_head.")
+            logger.info("Model has tied word embeddings, will tie lm_head weights")
+        
+        # Create lm_head at wrapper level (following latest vLLM architecture)
+        # This is outside self.model, so weights map "model.lm_head." -> "lm_head."
+        if self.pp_group.is_last_rank:
+            self.lm_head = ParallelLMHead(
+                self.vocab_size,
+                self.hidden_size,
+                quant_config=self.quant_config,
+                prefix=maybe_prefix(prefix, "lm_head"),
+            )
+            
+            # Tie weights if needed
+            if tie_word_embeddings:
+                input_embeddings = self.model.get_input_embeddings()
+                if input_embeddings is not None:
+                    self.lm_head = self.lm_head.tie_weights(input_embeddings)
+                    logger.info("Tied lm_head weights with input embeddings")
+            
+            # Setup logits processor
+            logit_scale = getattr(self.text_config, "logit_scale", 1.0)
+            self.logits_processor = LogitsProcessor(
+                self.vocab_size,
+                logits_as_input=False,
+                scale=logit_scale,
+            )
+            
+            logger.info("CausalMixin initialized (vocab_size=%d, hidden_size=%d, logit_scale=%s)",
+                       self.vocab_size, self.hidden_size, logit_scale)
+        else:
+            # For non-last PP ranks, use PPMissingLayer
+            self.lm_head = PPMissingLayer()
+            self.logits_processor = None
+            logger.info("CausalMixin initialized (PP non-last rank, using PPMissingLayer)")
+        
+        # Setup sampler
+        self.sampler = Sampler()
+    
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[torch.Tensor]:
+        """
+        Compute logits from hidden states.
+        
+        This method conforms to the VllmModelForTextGeneration protocol.
+        
+        Args:
+            hidden_states: Hidden states from the model [seq_len, hidden_size]
+            sampling_metadata: Sampling metadata
+        
+        Returns:
+            Logits tensor or None
+        """
+        if self.logits_processor is None:
+            # Non-last PP rank
+            return None
+        
+        # In v0.6.2, LogitsProcessor handles the lm_head projection internally
+        # via lm_head.linear_method.apply(). Pass lm_head as the first arg.
+        logits = self.logits_processor(self.lm_head, hidden_states,
+                                       sampling_metadata)
+        
+        return logits
+    
+    def sample(
+        self,
+        logits: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[SamplerOutput]:
+        """
+        Sample tokens from logits.
+        
+        This method conforms to the VllmModelForTextGeneration protocol.
+        
+        Args:
+            logits: Logits tensor
+            sampling_metadata: Sampling metadata
+        
+        Returns:
+            SamplerOutput with sampled tokens
+        """
+        next_tokens = self.sampler(logits, sampling_metadata)
+        return next_tokens

vllm-v0.6.2/vllm/model_executor/models/transformers/legacy.py

@@ -0,0 +1,118 @@
+# SPDX-License-Identifier: Apache-2.0
+# Copyright 2024 The vLLM team.
+"""Transformers modeling backend mixin for legacy models.
+
+This module provides LegacyMixin for BERT-like encoder models that have
+different weight naming conventions and special position handling.
+
+Following latest vLLM architecture patterns adapted for v0.6.2.
+"""
+
+from typing import TYPE_CHECKING, List, Optional
+
+import torch
+
+from vllm.logger import init_logger
+from vllm.model_executor.models.utils import WeightsMapper
+from vllm.sequence import IntermediateTensors
+
+if TYPE_CHECKING:
+    from vllm.config import VllmConfig
+
+logger = init_logger(__name__)
+
+
+class LegacyMixin:
+    """
+    Mixin class for legacy/encoder models like BERT, RoBERTa.
+    
+    This mixin provides:
+    - Weight name mapping for legacy suffix conventions (.gamma/.beta)
+    - Prefix mapping for BERT-like model structures
+    - RoBERTa-specific position handling
+    - Skip prefixes for unsupported output layers
+    
+    Should be used with Base class:
+        class TransformersForLegacy(LegacyMixin, Base): ...
+    """
+    
+    # Weight name mapping for legacy models
+    hf_to_vllm_mapper = WeightsMapper(
+        # These are applied in order, so the order matters!
+        orig_to_new_prefix={
+            # Handle BERT-like models
+            "roberta": "model",
+            "bert": "model",
+        },
+        orig_to_new_suffix={
+            # Replace legacy suffixes used for norms
+            ".gamma": ".weight",
+            ".beta": ".bias",
+        },
+    )
+    
+    def __init__(self, *, vllm_config: "VllmConfig", prefix: str = "") -> None:
+        # Call next class in MRO (should be Base)
+        super().__init__(vllm_config=vllm_config, prefix=prefix)
+        
+        # Skip unsupported/unwanted output embeddings layers
+        self.skip_prefixes.extend([
+            "model.lm_head.",
+            "model.predictions.",
+            "model.qa_outputs.",
+            "model.embeddings_project.",
+            "model.discriminator_predictions.",
+        ])
+        
+        # v0.6.2 doesn't have skip_substrs, so we handle it differently
+        # Store patterns to skip during weight loading
+        self._legacy_skip_patterns: List[str] = [
+            "position_ids",  # Some encoder models have position_ids buffer
+            "score.bias",    # Final classifier bias not used by vLLM
+        ]
+        
+        # RoBERTa-like models have extra padding in positions
+        model_type = getattr(self.text_config, "model_type", "").lower()
+        self.is_roberta = "roberta" in model_type
+        self.padding_idx = getattr(self.text_config, "pad_token_id", 1)
+        
+        if self.is_roberta:
+            logger.info("LegacyMixin detected RoBERTa model, enabling position padding")
+        
+        logger.info("LegacyMixin initialized for legacy/encoder model")
+    
+    def _should_skip_weight(self, name: str) -> bool:
+        """Check if a weight should be skipped during loading."""
+        for pattern in self._legacy_skip_patterns:
+            if pattern in name:
+                return True
+        return False
+    
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor],
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        """
+        Forward pass with RoBERTa position handling.
+        
+        RoBERTa models require positions to be offset by padding_idx + 1.
+        """
+        if self.is_roberta and positions is not None:
+            # RoBERTa-specific positions padding
+            positions = positions + self.padding_idx + 1
+        
+        return super().forward(
+            input_ids=input_ids,
+            positions=positions,
+            kv_caches=kv_caches,
+            attn_metadata=attn_metadata,
+            intermediate_tensors=intermediate_tensors,
+            inputs_embeds=inputs_embeds,
+            **kwargs,
+        )

vllm-v0.6.2/vllm/model_executor/models/transformers/pooling.py

@@ -0,0 +1,170 @@
+# SPDX-License-Identifier: Apache-2.0
+# Copyright 2024 The vLLM team.
+"""Transformers modeling backend mixins for pooling/embedding models.
+
+This module provides mixins for embedding and sequence classification models:
+- EmbeddingMixin: For embedding/sentence similarity models
+- SequenceClassificationMixin: For sequence classification/cross-encoding
+
+Following latest vLLM architecture patterns adapted for v0.6.2.
+"""
+
+from typing import TYPE_CHECKING, List, Optional
+
+import torch
+import torch.nn as nn
+
+from vllm.logger import init_logger
+from vllm.model_executor.layers.pooler import Pooler, PoolingType
+from vllm.model_executor.pooling_metadata import PoolingMetadata
+from vllm.sequence import PoolerOutput
+
+if TYPE_CHECKING:
+    from vllm.config import VllmConfig
+
+logger = init_logger(__name__)
+
+
+class EmbeddingMixin:
+    """
+    Mixin class that adds embedding/pooling functionality.
+    
+    This mixin provides:
+    - Pooler layer for extracting embeddings
+    - pooling method for VllmModelForPooling protocol
+    
+    Should be used with Base class:
+        class TransformersForEmbedding(EmbeddingMixin, Base): ...
+    """
+    
+    # Default pooling configuration
+    default_pooling_type: PoolingType = PoolingType.CLS
+    default_normalize: bool = True
+    default_softmax: bool = False
+    
+    def __init__(self, *, vllm_config: "VllmConfig", prefix: str = "") -> None:
+        # Call next class in MRO (should be Base)
+        super().__init__(vllm_config=vllm_config, prefix=prefix)
+        
+        # Get pooler config from model config
+        pooler_config = vllm_config.model_config.pooler_config
+        
+        # Setup pooler
+        self.pooler = Pooler.from_config_with_defaults(
+            pooler_config=pooler_config,
+            pooling_type=self.default_pooling_type,
+            normalize=self.default_normalize,
+            softmax=self.default_softmax,
+        )
+        
+        if self.pooler is None:
+            # Create default pooler if config doesn't specify
+            self.pooler = Pooler(
+                pooling_type=self.default_pooling_type,
+                normalize=self.default_normalize,
+                softmax=self.default_softmax,
+            )
+        
+        logger.info("EmbeddingMixin initialized (pooling_type=%s, normalize=%s)",
+                   self.pooler.pooling_type.name, self.pooler.normalize)
+    
+    def pooling(
+        self,
+        hidden_states: torch.Tensor,
+        pooling_metadata: PoolingMetadata,
+    ) -> Optional[PoolerOutput]:
+        """
+        Apply pooling to hidden states.
+        
+        Args:
+            hidden_states: Hidden states from the model [seq_len, hidden_size]
+            pooling_metadata: Pooling metadata
+        
+        Returns:
+            PoolerOutput with pooled embeddings
+        """
+        return self.pooler(hidden_states, pooling_metadata)
+
+
+class SequenceClassificationMixin(EmbeddingMixin):
+    """
+    Mixin class that adds sequence classification functionality.
+    
+    This mixin provides:
+    - Classifier layer for sequence classification
+    - pooling method with classification logits
+    
+    Should be used with Base class:
+        class TransformersForSequenceClassification(SequenceClassificationMixin, Base): ...
+    """
+    
+    default_pooling_type: PoolingType = PoolingType.CLS
+    default_normalize: bool = False
+    default_softmax: bool = True
+    
+    def __init__(self, *, vllm_config: "VllmConfig", prefix: str = "") -> None:
+        # Call EmbeddingMixin.__init__ -> Base.__init__
+        super().__init__(vllm_config=vllm_config, prefix=prefix)
+        
+        # Find and setup classifier layer
+        self.classifier = self._find_classifier()
+        
+        if self.classifier is not None:
+            # Initialize classifier parameters on device
+            self._init_classifier_params()
+            logger.info("SequenceClassificationMixin initialized with classifier")
+        else:
+            logger.warning("Could not find classifier layer")
+    
+    def _find_classifier(self) -> Optional[nn.Module]:
+        """Find the classifier layer in the model."""
+        # Common classifier layer names
+        classifier_names = ['classifier', 'score', 'fc', 'head']
+        
+        for name in classifier_names:
+            if hasattr(self.model, name):
+                return getattr(self.model, name)
+        
+        return None
+    
+    def _init_classifier_params(self) -> None:
+        """Initialize classifier parameters on target device."""
+        device = self.device_config.device
+        if device is None:
+            device = torch.device("cpu")
+        dtype = self.model_config.dtype
+        
+        for name, param in list(self.classifier.named_parameters()):
+            if param.device == torch.device("meta"):
+                new_param = nn.Parameter(
+                    torch.empty_like(param.data, dtype=dtype, device=device),
+                    requires_grad=False,
+                )
+                setattr(self.classifier, name.split('.')[-1], new_param)
+    
+    def pooling(
+        self,
+        hidden_states: torch.Tensor,
+        pooling_metadata: PoolingMetadata,
+    ) -> Optional[PoolerOutput]:
+        """
+        Apply pooling and classification to hidden states.
+        
+        Args:
+            hidden_states: Hidden states from the model [seq_len, hidden_size]
+            pooling_metadata: Pooling metadata
+        
+        Returns:
+            PoolerOutput with classification logits
+        """
+        # First apply base pooling
+        pooled = self.pooler(hidden_states, pooling_metadata)
+        
+        # Apply classifier if available
+        if self.classifier is not None and pooled is not None:
+            # Apply classifier to each pooled output
+            for i, output in enumerate(pooled.outputs):
+                if hasattr(output, 'data'):
+                    output.data = self.classifier(output.data)
+        
+        return pooled

vllm-v0.6.2/vllm/model_executor/models/transformers/utils.py

@@ -0,0 +1,247 @@
+# SPDX-License-Identifier: Apache-2.0
+# Copyright 2024 The vLLM team.
+"""Transformers modeling backend utilities for v0.6.2.
+
+This module provides utility functions for the Transformers backend,
+including context managers for meta device initialization and
+module replacement functions.
+"""
+
+from contextlib import contextmanager
+from typing import TYPE_CHECKING, Literal, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from vllm.logger import init_logger
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (
+    ColumnParallelLinear,
+    ReplicatedLinear,
+    RowParallelLinear,
+)
+
+if TYPE_CHECKING:
+    from vllm.model_executor.layers.quantization.base_config import (
+        QuantizationConfig,
+    )
+
+logger = init_logger(__name__)
+
+
+@contextmanager
+def init_on_device_without_buffers(device: Union[str, torch.device]):
+    """
+    A context manager under which models are initialized with all
+    parameters on the specified device. However buffers are not
+    initialized on specified device.
+
+    This is useful for creating model structure without allocating
+    GPU memory, which is essential for memory efficiency.
+
+    Args:
+        device: Device to initialize all parameters on (e.g., "meta").
+    
+    Example:
+        with init_on_device_without_buffers("meta"):
+            model = AutoModel.from_config(config)
+        # Now model is on meta device, no GPU memory allocated
+    """
+    if isinstance(device, str):
+        device = torch.device(device)
+
+    old_register_parameter = nn.Module.register_parameter
+
+    def register_empty_parameter(module, name, param):
+        old_register_parameter(module, name, param)
+        if param is not None:
+            param_cls = type(module._parameters[name])
+            kwargs = module._parameters[name].__dict__
+            kwargs["requires_grad"] = param.requires_grad
+            module._parameters[name] = param_cls(
+                module._parameters[name].to(device), **kwargs
+            )
+
+    try:
+        nn.Module.register_parameter = register_empty_parameter
+        yield
+    finally:
+        nn.Module.register_parameter = old_register_parameter
+
+
+# Linear replacement styles
+Style = Literal["colwise", "colwise_rep", "rowwise", "rowwise_rep", "replicate"]
+
+
+def replace_linear_class(
+    linear: nn.Linear,
+    style: Style = "replicate",
+    quant_config: Optional["QuantizationConfig"] = None,
+    prefix: str = "",
+) -> Union[ColumnParallelLinear, RowParallelLinear, ReplicatedLinear]:
+    """
+    Replace nn.Linear with one of vLLM's tensor parallel linear classes.
+
+    This replacement provides:
+    - Memory efficiency through proper tensor allocation
+    - Support for quantization
+    - Tensor parallel support (when using ColumnParallel/RowParallel)
+
+    Args:
+        linear: `nn.Linear` to be replaced.
+        style: Tensor parallel style of the new linear:
+            - "colwise": Column parallel (split output dim)
+            - "colwise_rep": Column parallel with gather output
+            - "rowwise": Row parallel (split input dim)
+            - "rowwise_rep": Row parallel without parallel input
+            - "replicate": Replicated (no parallelism)
+        quant_config: Quantization config for the new linear.
+        prefix: The name of the layer for weight loading.
+    
+    Returns:
+        The new vLLM linear layer.
+    """
+    if not isinstance(style, str):
+        raise ValueError(f"Unsupported parallel style type {type(style)}, expected str")
+
+    vllm_linear_cls, vllm_linear_kwargs = {
+        "colwise": (ColumnParallelLinear, {}),
+        "colwise_rep": (ColumnParallelLinear, {"gather_output": True}),
+        "rowwise": (RowParallelLinear, {}),
+        "rowwise_rep": (RowParallelLinear, {"input_is_parallel": False}),
+        "replicate": (ReplicatedLinear, {}),
+    }.get(style, (ReplicatedLinear, {}))
+
+    return vllm_linear_cls(
+        input_size=linear.in_features,
+        output_size=linear.out_features,
+        bias=linear.bias is not None,
+        quant_config=quant_config,
+        prefix=prefix,
+        return_bias=False,  # Return tensor only, not (tensor, bias) tuple
+        **vllm_linear_kwargs,
+    )
+
+
+class TransformersRMSNorm(RMSNorm):
+    """
+    vLLM RMSNorm subclass that preserves tensor dimensions.
+    
+    vLLM's RMSNorm (especially the MLU backend) flattens input to 2D
+    (e.g., [batch, seq, hidden] -> [batch*seq, hidden]), but transformers
+    expects the batch dimension to be preserved. This subclass wraps
+    the parent forward methods to save and restore the original tensor shape.
+    
+    Since this inherits from RMSNorm directly, weight loading via
+    named_parameters() works correctly (weight path stays the same).
+    """
+    
+    def forward_native(
+        self,
+        x: torch.Tensor,
+        residual: Optional[torch.Tensor] = None,
+    ):
+        orig_shape = x.shape
+        result = super().forward_native(x, residual)
+        return self._restore_shape(result, orig_shape)
+    
+    def forward_cuda(
+        self,
+        x: torch.Tensor,
+        residual: Optional[torch.Tensor] = None,
+    ):
+        orig_shape = x.shape
+        result = super().forward_cuda(x, residual)
+        return self._restore_shape(result, orig_shape)
+    
+    def forward_mlu(
+        self,
+        x: torch.Tensor,
+        residual: Optional[torch.Tensor] = None,
+    ):
+        orig_shape = x.shape
+        result = super().forward_mlu(x, residual)
+        return self._restore_shape(result, orig_shape)
+    
+    def forward_xpu(
+        self,
+        x: torch.Tensor,
+        residual: Optional[torch.Tensor] = None,
+    ):
+        orig_shape = x.shape
+        result = super().forward_xpu(x, residual)
+        return self._restore_shape(result, orig_shape)
+    
+    def forward_hpu(
+        self,
+        x: torch.Tensor,
+        residual: Optional[torch.Tensor] = None,
+    ):
+        orig_shape = x.shape
+        result = super().forward_hpu(x, residual)
+        return self._restore_shape(result, orig_shape)
+    
+    @staticmethod
+    def _restore_shape(result, orig_shape: Tuple):
+        """Restore original tensor shape if it was changed."""
+        if isinstance(result, tuple):
+            restored = []
+            for t in result:
+                if t is not None and t.shape != orig_shape:
+                    t = t.view(orig_shape)
+                restored.append(t)
+            return tuple(restored)
+        else:
+            if result.shape != orig_shape:
+                result = result.view(orig_shape)
+            return result
+
+
+def replace_rms_norm_class(
+    rms_norm: nn.Module,
+    hidden_size: int,
+) -> nn.Module:
+    """
+    Replace a Transformers RMSNorm with vLLM's optimized RMSNorm,
+    wrapped to preserve tensor dimensions.
+
+    vLLM's RMSNorm provides:
+    - Fused CUDA kernels for better performance
+    - Support for fused add + norm operations
+
+    The wrapper ensures that the original tensor shape (including batch
+    dimension) is preserved, which is required by transformers' model
+    forward methods.
+
+    Args:
+        rms_norm: The RMSNorm module to replace.
+        hidden_size: The hidden size of the model.
+    
+    Returns:
+        The new vLLM RMSNorm layer wrapped for shape preservation.
+    """
+    # Try to get epsilon from various attribute names
+    eps = getattr(rms_norm, "eps", None)
+    if eps is None:
+        eps = getattr(rms_norm, "variance_epsilon", None)
+    if eps is None:
+        eps = 1e-6
+
+    # Check if weight exists and get its size
+    weight = getattr(rms_norm, "weight", None)
+    if weight is not None:
+        hidden_size = weight.size(0)
+
+    return TransformersRMSNorm(hidden_size=hidden_size, eps=eps)
+
+
+def log_replacement(name: str, old_module: nn.Module, new_module: nn.Module):
+    """Log module replacement for debugging."""
+    logger.debug("Replaced %s: %s -> %s", name, type(old_module).__name__, type(new_module).__name__)
+
+
+def maybe_prefix(prefix: str, name: str) -> str:
+    """Combine prefix and name with a dot separator."""
+    if prefix:
+        return f"{prefix}.{name}"
+    return name

vllm-v0.6.2/vllm/transformers_utils/dynamic_module.py

@@ -0,0 +1,76 @@
+"""
+Dynamic module loading utilities for custom HuggingFace models.
+Ported from latest vLLM to support auto_map in model config.
+"""
+import os
+from typing import Dict, Optional, Type, Union
+
+from transformers.dynamic_module_utils import (
+    get_class_from_dynamic_module,
+    resolve_trust_remote_code,
+)
+
+import vllm.envs as envs
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+def try_get_class_from_dynamic_module(
+    class_reference: str,
+    pretrained_model_name_or_path: str,
+    trust_remote_code: bool,
+    cache_dir: Optional[Union[str, os.PathLike]] = None,
+    force_download: bool = False,
+    resume_download: Optional[bool] = None,
+    proxies: Optional[Dict[str, str]] = None,
+    token: Optional[Union[bool, str]] = None,
+    revision: Optional[str] = None,
+    local_files_only: bool = False,
+    repo_type: Optional[str] = None,
+    code_revision: Optional[str] = None,
+    warn_on_fail: bool = True,
+    **kwargs,
+) -> Optional[Type]:
+    """
+    As `transformers.dynamic_module_utils.get_class_from_dynamic_module`,
+    but ignoring any errors.
+    
+    This allows vLLM to load custom models that define their own
+    model classes via the `auto_map` field in config.json.
+    """
+    try:
+        resolve_trust_remote_code(
+            trust_remote_code,
+            pretrained_model_name_or_path,
+            has_local_code=False,
+            has_remote_code=True,
+        )
+
+        return get_class_from_dynamic_module(
+            class_reference,
+            pretrained_model_name_or_path,
+            cache_dir=cache_dir,
+            force_download=force_download,
+            resume_download=resume_download,
+            proxies=proxies,
+            token=token,
+            revision=revision,
+            local_files_only=local_files_only,
+            repo_type=repo_type,
+            code_revision=code_revision,
+            **kwargs,
+        )
+    except Exception:
+        location = "ModelScope" if envs.VLLM_USE_MODELSCOPE else "HF Hub"
+
+        if warn_on_fail:
+            logger.warning(
+                "Unable to load %s from %s on %s.",
+                class_reference,
+                pretrained_model_name_or_path,
+                location,
+                exc_info=True,
+            )
+
+        return None

vllm-v0.6.2/vllm_mlu/vllm_mlu/model_executor/models/llama.py

@@ -74,18 +74,22 @@ def vllm__module_executor__models__llama__LlamaAttention__forward(
     smooth_quant_scale: Optional[torch.Tensor] = None,
 ) -> torch.Tensor:
     qkv, _ = self.qkv_proj(hidden_states, smooth_quant_scale)
-    q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
     '''
     =============================
     Modify by vllm_mlu
     =============================
     @brief: pack q & k to fit tmo.apply_rotary
+    @optimization: avoid redundant split operation
     '''
     if self.rope_scaling is not None and self.rope_scaling["rope_type"] == "longrope":
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
         q, k = self.rotary_emb(positions, q, k)
     else:
-        qk, _ = qkv.split([self.q_size + self.kv_size, self.kv_size], dim=-1)
+        # Optimized: split qkv into [qk, v] directly, avoiding redundant split
+        qk, v = qkv.split([self.q_size + self.kv_size, self.kv_size], dim=-1)
         self.rotary_emb(positions, qk.view(-1, self.num_heads + self.num_kv_heads, self.head_dim))
+        # Split qk into q and k after rotary embedding
+        q, k = qk.split([self.q_size, self.kv_size], dim=-1)
     '''
     ==================
     End of MLU Hijack