Add files using upload-large-folder tool

.gitattributes

@@ -108,3 +108,4 @@ tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/_inductor/_
 .venv/lib/python3.11/site-packages/pillow.libs/liblcms2-e69eef39.so.2.0.16 filter=lfs diff=lfs merge=lfs -text
 .venv/lib/python3.11/site-packages/pillow.libs/libopenjp2-05423b53.so filter=lfs diff=lfs merge=lfs -text
 .venv/lib/python3.11/site-packages/pillow.libs/libxcb-b8a56d01.so.1.1.0 filter=lfs diff=lfs merge=lfs -text
+.venv/lib/python3.11/site-packages/vllm/worker/__pycache__/hpu_model_runner.cpython-311.pyc filter=lfs diff=lfs merge=lfs -text

.venv/lib/python3.11/site-packages/vllm/entrypoints/openai/tool_parsers/__init__.py

@@ -0,0 +1,18 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from .abstract_tool_parser import ToolParser, ToolParserManager
+from .granite_20b_fc_tool_parser import Granite20bFCToolParser
+from .granite_tool_parser import GraniteToolParser
+from .hermes_tool_parser import Hermes2ProToolParser
+from .internlm2_tool_parser import Internlm2ToolParser
+from .jamba_tool_parser import JambaToolParser
+from .llama_tool_parser import Llama3JsonToolParser
+from .mistral_tool_parser import MistralToolParser
+from .pythonic_tool_parser import PythonicToolParser
+
+__all__ = [
+    "ToolParser", "ToolParserManager", "Granite20bFCToolParser",
+    "GraniteToolParser", "Hermes2ProToolParser", "MistralToolParser",
+    "Internlm2ToolParser", "Llama3JsonToolParser", "JambaToolParser",
+    "PythonicToolParser"
+]

.venv/lib/python3.11/site-packages/vllm/entrypoints/openai/tool_parsers/abstract_tool_parser.py

@@ -0,0 +1,162 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import os
+from functools import cached_property
+from typing import Callable, Dict, List, Optional, Sequence, Type, Union
+
+from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
+                                              DeltaMessage,
+                                              ExtractedToolCallInformation)
+from vllm.logger import init_logger
+from vllm.transformers_utils.tokenizer import AnyTokenizer
+from vllm.utils import import_from_path, is_list_of
+
+logger = init_logger(__name__)
+
+
+class ToolParser:
+    """
+    Abstract ToolParser class that should not be used directly. Provided
+    properties and methods should be used in
+    derived classes.
+    """
+
+    def __init__(self, tokenizer: AnyTokenizer):
+        self.prev_tool_call_arr: List[Dict] = []
+        # the index of the tool call that is currently being parsed
+        self.current_tool_id: int = -1
+        self.current_tool_name_sent: bool = False
+        self.streamed_args_for_tool: List[str] = []
+
+        self.model_tokenizer = tokenizer
+
+    @cached_property
+    def vocab(self) -> Dict[str, int]:
+        # NOTE: Only PreTrainedTokenizerFast is guaranteed to have .vocab
+        # whereas all tokenizers have .get_vocab()
+        return self.model_tokenizer.get_vocab()
+
+    def adjust_request(
+            self, request: ChatCompletionRequest) -> ChatCompletionRequest:
+        """
+        Static method that used to adjust the request parameters.
+        """
+        return request
+
+    def extract_tool_calls(
+            self, model_output: str,
+            request: ChatCompletionRequest) -> ExtractedToolCallInformation:
+        """
+        Static method that should be implemented for extracting tool calls from
+        a complete model-generated string.
+        Used for non-streaming responses where we have the entire model response
+        available before sending to the client.
+        Static because it's stateless.
+        """
+        raise NotImplementedError(
+            "AbstractToolParser.extract_tool_calls has not been implemented!")
+
+    def extract_tool_calls_streaming(
+        self,
+        previous_text: str,
+        current_text: str,
+        delta_text: str,
+        previous_token_ids: Sequence[int],
+        current_token_ids: Sequence[int],
+        delta_token_ids: Sequence[int],
+        request: ChatCompletionRequest,
+    ) -> Union[DeltaMessage, None]:
+        """
+        Instance method that should be implemented for extracting tool calls
+        from an incomplete response; for use when handling tool calls and
+        streaming. Has to be an instance method because  it requires state -
+        the current tokens/diffs, but also the information about what has
+        previously been parsed and extracted (see constructor)
+        """
+        raise NotImplementedError(
+            "AbstractToolParser.extract_tool_calls_streaming has not been "
+            "implemented!")
+
+
+class ToolParserManager:
+    tool_parsers: Dict[str, Type] = {}
+
+    @classmethod
+    def get_tool_parser(cls, name) -> Type:
+        """
+        Get tool parser by name which is registered by `register_module`.
+
+        Raise a KeyError exception if the name is not registered.
+        """
+        if name in cls.tool_parsers:
+            return cls.tool_parsers[name]
+
+        raise KeyError(f"tool helper: '{name}' not found in tool_parsers")
+
+    @classmethod
+    def _register_module(cls,
+                         module: Type,
+                         module_name: Optional[Union[str, List[str]]] = None,
+                         force: bool = True) -> None:
+        if not issubclass(module, ToolParser):
+            raise TypeError(
+                f'module must be subclass of ToolParser, but got {type(module)}'
+            )
+        if module_name is None:
+            module_name = module.__name__
+        if isinstance(module_name, str):
+            module_name = [module_name]
+        for name in module_name:
+            if not force and name in cls.tool_parsers:
+                existed_module = cls.tool_parsers[name]
+                raise KeyError(f'{name} is already registered '
+                               f'at {existed_module.__module__}')
+            cls.tool_parsers[name] = module
+
+    @classmethod
+    def register_module(
+            cls,
+            name: Optional[Union[str, List[str]]] = None,
+            force: bool = True,
+            module: Union[Type, None] = None) -> Union[type, Callable]:
+        """
+        Register module with the given name or name list. it can be used as a
+        decoder(with module as None) or normal function(with module as not 
+        None).
+        """
+        if not isinstance(force, bool):
+            raise TypeError(f'force must be a boolean, but got {type(force)}')
+
+        # raise the error ahead of time
+        if not (name is None or isinstance(name, str)
+                or is_list_of(name, str)):
+            raise TypeError(
+                'name must be None, an instance of str, or a sequence of str, '
+                f'but got {type(name)}')
+
+        # use it as a normal method: x.register_module(module=SomeClass)
+        if module is not None:
+            cls._register_module(module=module, module_name=name, force=force)
+            return module
+
+        # use it as a decorator: @x.register_module()
+        def _register(module):
+            cls._register_module(module=module, module_name=name, force=force)
+            return module
+
+        return _register
+
+    @classmethod
+    def import_tool_parser(cls, plugin_path: str) -> None:
+        """
+        Import a user-defined tool parser by the path of the tool parser define
+        file.
+        """
+        module_name = os.path.splitext(os.path.basename(plugin_path))[0]
+
+        try:
+            import_from_path(module_name, plugin_path)
+        except Exception:
+            logger.exception("Failed to load module '%s' from %s.",
+                             module_name, plugin_path)
+            return

.venv/lib/python3.11/site-packages/vllm/entrypoints/openai/tool_parsers/jamba_tool_parser.py

@@ -0,0 +1,302 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import json
+import re
+from typing import Dict, List, Sequence, Union
+
+import partial_json_parser
+from partial_json_parser.core.options import Allow
+
+from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
+                                              DeltaFunctionCall, DeltaMessage,
+                                              DeltaToolCall,
+                                              ExtractedToolCallInformation,
+                                              FunctionCall, ToolCall)
+from vllm.entrypoints.openai.tool_parsers import ToolParser, ToolParserManager
+from vllm.entrypoints.openai.tool_parsers.utils import (
+    extract_intermediate_diff)
+from vllm.logger import init_logger
+from vllm.transformers_utils.tokenizer import AnyTokenizer
+from vllm.transformers_utils.tokenizers import MistralTokenizer
+from vllm.utils import random_uuid
+
+logger = init_logger(__name__)
+
+
+@ToolParserManager.register_module("jamba")
+class JambaToolParser(ToolParser):
+
+    def __init__(self, tokenizer: AnyTokenizer):
+        super().__init__(tokenizer)
+
+        if isinstance(self.model_tokenizer, MistralTokenizer):
+            raise ValueError(
+                "Detected a MistralTokenizer tokenizer when using a Jamba model"
+            )
+
+        self.current_tool_name_sent: bool = False
+        self.prev_tool_call_arr: List[Dict] = []
+        self.current_tool_id: int = -1
+        self.streamed_args_for_tool: List[str] = [
+        ]  # map what has been streamed for each tool so far to a list
+
+        self.tool_calls_start_token: str = "<tool_calls>"
+        self.tool_calls_end_token: str = "</tool_calls>"
+
+        self.tool_calls_regex = re.compile(
+            rf"{self.tool_calls_start_token}(.*?){self.tool_calls_end_token}",
+            re.DOTALL)
+
+        if not self.model_tokenizer:
+            raise ValueError(
+                "The model tokenizer must be passed to the ToolParser "
+                "constructor during construction.")
+        self.tool_calls_start_token_id = self.vocab.get(
+            self.tool_calls_start_token)
+        self.tool_calls_end_token_id = self.vocab.get(
+            self.tool_calls_end_token)
+        if (self.tool_calls_start_token_id is None
+                or self.tool_calls_end_token_id is None):
+            raise RuntimeError(
+                "Jamba Tool parser could not locate tool calls start/end "
+                "tokens in the tokenizer!")
+
+    def adjust_request(
+            self, request: ChatCompletionRequest) -> ChatCompletionRequest:
+        if request.tools and request.tool_choice != 'none':
+            # do not skip special tokens because jamba use the special
+            # tokens to indicate the start and end of the tool calls
+            # information.
+            request.skip_special_tokens = False
+        return request
+
+    def extract_tool_calls(
+            self, model_output: str,
+            request: ChatCompletionRequest) -> ExtractedToolCallInformation:
+
+        # sanity check; avoid unnecessary processing
+        if self.tool_calls_start_token not in model_output:
+            return ExtractedToolCallInformation(tools_called=False,
+                                                tool_calls=[],
+                                                content=model_output)
+
+        else:
+
+            try:
+                # use a regex to find the tool call between the tags
+                function_calls = self.tool_calls_regex.findall(model_output)[0]
+
+                # load the JSON, and then use it to build the Function and
+                # Tool Call
+                raw_function_calls = json.loads(function_calls)
+                tool_calls = [
+                    ToolCall(
+                        type="function",
+                        function=FunctionCall(
+                            name=function_call["name"],
+                            # function call args are JSON but as a string
+                            arguments=json.dumps(function_call["arguments"])))
+                    for function_call in raw_function_calls
+                ]
+
+                content = model_output[:model_output.
+                                       find(self.tool_calls_start_token)]
+                return ExtractedToolCallInformation(
+                    tools_called=True,
+                    tool_calls=tool_calls,
+                    content=content if
+                    (len(content) > 0 and content != " ") else None)
+
+            except Exception:
+                logger.exception(
+                    "Error in extracting tool call from response.")
+                return ExtractedToolCallInformation(tools_called=False,
+                                                    tool_calls=[],
+                                                    content=model_output)
+
+    def extract_tool_calls_streaming(
+        self,
+        previous_text: str,
+        current_text: str,
+        delta_text: str,
+        previous_token_ids: Sequence[int],
+        current_token_ids: Sequence[int],
+        delta_token_ids: Sequence[int],
+        request: ChatCompletionRequest,
+    ) -> Union[DeltaMessage, None]:
+
+        # if the tool call token is not in the tokens generated so far, append
+        # output to contents since it's not a tool
+        if self.tool_calls_start_token not in current_text:
+            return DeltaMessage(content=delta_text)
+
+        # if the tool call token ID IS in the tokens generated so far, that
+        # means we're parsing as tool calls now
+
+        # handle if we detected the start of tool calls token which means
+        # the start of tool calling
+        if (self.tool_calls_start_token_id in delta_token_ids
+                and len(delta_token_ids) == 1):
+            # if it's the only token, return None, so we don't send a chat
+            # completion and don't send a control token
+            return None
+
+        # bit mask flags for partial JSON parsing. If the name hasn't been
+        # sent yet, don't allow sending
+        # an incomplete string since OpenAI only ever (as far as I have
+        # seen) allows sending the entire tool/ function name at once.
+        flags = Allow.ALL if self.current_tool_name_sent \
+            else Allow.ALL & ~Allow.STR
+        try:
+
+            # Extract the tool calls between the special tool call tokens
+            parsable_arr = current_text.split(
+                self.tool_calls_start_token)[-1].split(
+                    self.tool_calls_end_token)[0]
+
+            # tool calls are generated in an array, so do partial JSON
+            # parsing on the entire array
+            try:
+                tool_call_arr: List[Dict] = partial_json_parser.loads(
+                    parsable_arr, flags)
+            except partial_json_parser.core.exceptions.MalformedJSON:
+                logger.debug('not enough tokens to parse into JSON yet')
+                return None
+
+            # select as the current tool call the one we're on the state at
+
+            current_tool_call: Dict = tool_call_arr[self.current_tool_id] \
+                if len(tool_call_arr) > 0 else {}
+
+            # case -- if no tokens have been streamed for the tool, e.g.
+            #   only the array brackets, stream nothing
+            if len(tool_call_arr) == 0:
+                return None
+
+            # case: we are starting a new tool in the array
+            #   -> array has > 0 length AND length has moved past cursor
+            elif (len(tool_call_arr) > 0
+                  and len(tool_call_arr) > self.current_tool_id + 1):
+
+                # if we're moving on to a new call, first make sure we
+                # haven't missed anything in the previous one that was
+                # auto-generated due to JSON completions, but wasn't
+                # streamed to the client yet.
+                if self.current_tool_id >= 0:
+                    diff: Union[str, None] = current_tool_call.get("arguments")
+
+                    if diff:
+                        diff = json.dumps(diff).replace(
+                            self.streamed_args_for_tool[self.current_tool_id],
+                            "")
+                        delta = DeltaMessage(tool_calls=[
+                            DeltaToolCall(index=self.current_tool_id,
+                                          function=DeltaFunctionCall(
+                                              arguments=diff).model_dump(
+                                                  exclude_none=True))
+                        ])
+                        self.streamed_args_for_tool[
+                            self.current_tool_id] += diff
+                    else:
+                        delta = None
+                else:
+                    delta = None
+                # re-set stuff pertaining to progress in the current tool
+                self.current_tool_id = len(tool_call_arr) - 1
+                self.current_tool_name_sent = False
+                self.streamed_args_for_tool.append("")
+                logger.debug("starting on new tool %d", self.current_tool_id)
+                return delta
+
+            # case: update an existing tool - this is handled below
+
+            # if the current tool name hasn't been sent, send if available
+            # - otherwise send nothing
+            if not self.current_tool_name_sent:
+                function_name = current_tool_call.get("name")
+                if function_name:
+
+                    delta = DeltaMessage(tool_calls=[
+                        DeltaToolCall(index=self.current_tool_id,
+                                      type="function",
+                                      id=f"chatcmpl-tool-{random_uuid()}",
+                                      function=DeltaFunctionCall(
+                                          name=function_name).model_dump(
+                                              exclude_none=True))
+                    ])
+                    self.current_tool_name_sent = True
+                else:
+                    delta = None
+
+            # now we know we're on the same tool call and we're streaming
+            # arguments
+            else:
+
+                prev_arguments = self.prev_tool_call_arr[
+                    self.current_tool_id].get("arguments")
+                cur_arguments = current_tool_call.get("arguments")
+
+                new_text = delta_text.replace("\'", "\"")
+
+                if not cur_arguments and not prev_arguments:
+
+                    delta = None
+                elif not cur_arguments and prev_arguments:
+                    logger.error(
+                        "INVARIANT - impossible to have arguments reset "
+                        "mid-arguments")
+                    delta = None
+                elif cur_arguments and not prev_arguments:
+                    cur_arguments_json = json.dumps(cur_arguments)
+                    logger.debug("finding %s in %s", new_text,
+                                 cur_arguments_json)
+
+                    arguments_delta = cur_arguments_json[:cur_arguments_json.
+                                                         index(new_text) +
+                                                         len(new_text)]
+                    logger.debug("First tokens in arguments received: %s",
+                                 arguments_delta)
+                    delta = DeltaMessage(tool_calls=[
+                        DeltaToolCall(index=self.current_tool_id,
+                                      function=DeltaFunctionCall(
+                                          arguments=arguments_delta).
+                                      model_dump(exclude_none=True))
+                    ])
+                    self.streamed_args_for_tool[
+                        self.current_tool_id] += arguments_delta
+
+                elif cur_arguments and prev_arguments:
+                    cur_args_json = json.dumps(cur_arguments)
+                    prev_args_json = json.dumps(prev_arguments)
+                    logger.debug("Searching for diff between \n%s\n%s",
+                                 cur_args_json, prev_args_json)
+
+                    argument_diff = extract_intermediate_diff(
+                        cur_args_json, prev_args_json)
+                    logger.debug("got arguments diff: %s", argument_diff)
+                    delta = DeltaMessage(tool_calls=[
+                        DeltaToolCall(index=self.current_tool_id,
+                                      function=DeltaFunctionCall(
+                                          arguments=argument_diff).model_dump(
+                                              exclude_none=True))
+                    ])
+                    self.streamed_args_for_tool[
+                        self.current_tool_id] += argument_diff
+                else:
+                    # try parsing it with regular JSON - if it works we're
+                    # at the end, and we need to send the difference between
+                    # tokens streamed so far and the valid JSON
+                    delta = None
+
+            # check to see if the name is defined and has been sent. if so,
+            # stream the name - otherwise keep waiting
+            # finish by setting old and returning None as base case
+            self.prev_tool_call_arr = tool_call_arr
+            return delta
+
+        except Exception:
+            logger.exception("Error trying to handle streaming tool call.")
+            logger.debug(
+                "Skipping chunk as a result of tool streaming extraction "
+                "error")
+            return None

.venv/lib/python3.11/site-packages/vllm/entrypoints/openai/tool_parsers/llama_tool_parser.py

@@ -0,0 +1,260 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import json
+import re
+from json import JSONDecoder
+from typing import Dict, List, Sequence, Union
+
+import partial_json_parser
+from partial_json_parser.core.options import Allow
+from transformers import PreTrainedTokenizerBase
+
+from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
+                                              DeltaFunctionCall, DeltaMessage,
+                                              DeltaToolCall,
+                                              ExtractedToolCallInformation,
+                                              FunctionCall, ToolCall)
+from vllm.entrypoints.openai.tool_parsers.abstract_tool_parser import (
+    ToolParser, ToolParserManager)
+from vllm.entrypoints.openai.tool_parsers.utils import (find_common_prefix,
+                                                        is_complete_json,
+                                                        partial_json_loads)
+from vllm.logger import init_logger
+from vllm.utils import random_uuid
+
+logger = init_logger(__name__)
+
+
+@ToolParserManager.register_module("llama3_json")
+class Llama3JsonToolParser(ToolParser):
+    """
+    Tool call parser for Llama 3.1 models intended for use with the
+    examples/tool_chat_template_llama.jinja template.
+
+    Used when --enable-auto-tool-choice --tool-call-parser llama3_json 
+    are all set
+    """
+
+    def __init__(self, tokenizer: PreTrainedTokenizerBase):
+        super().__init__(tokenizer)
+
+        # initialize properties used for state when parsing tool calls in
+        # streaming mode
+        self.prev_tool_call_arr: List[Dict] = []
+        self.current_tool_id: int = -1
+        self.current_tool_name_sent: bool = False
+        self.streamed_args_for_tool: List[str] = [
+        ]  # map what has been streamed for each tool so far to a list
+        self.bot_token = "<|python_tag|>"
+        self.bot_token_id = tokenizer.encode(self.bot_token,
+                                             add_special_tokens=False)[0]
+        self.tool_call_regex = re.compile(r"\[{.*?}\]", re.DOTALL)
+
+    def extract_tool_calls(
+            self, model_output: str,
+            request: ChatCompletionRequest) -> ExtractedToolCallInformation:
+        """
+        Extract the tool calls from a complete model response.
+        """
+        # case -- if a tool call token is not present, return a text response
+        if not (model_output.startswith(self.bot_token)
+                or model_output.startswith('{')):
+            return ExtractedToolCallInformation(tools_called=False,
+                                                tool_calls=[],
+                                                content=model_output)
+
+        try:
+            # load the JSON, and then use it to build the Function and
+            # Tool Call
+            dec = JSONDecoder()
+            function_call_arr = []
+
+            # depending on the prompt format the Llama model may or may not
+            # prefix the output with the <|python_tag|> token
+            start_idx = len(self.bot_token) if model_output.startswith(
+                self.bot_token) else 0
+            while start_idx < len(model_output):
+                (obj, end_idx) = dec.raw_decode(model_output[start_idx:])
+                start_idx += end_idx + len('; ')
+                function_call_arr.append(obj)
+
+            tool_calls: List[ToolCall] = [
+                ToolCall(
+                    type="function",
+                    function=FunctionCall(
+                        name=raw_function_call["name"],
+                        # function call args are JSON but as a string
+                        arguments=json.dumps(raw_function_call["arguments"] \
+                                if "arguments" in raw_function_call \
+                                else raw_function_call["parameters"])))
+                for raw_function_call in function_call_arr
+            ]
+
+            # get any content before  the tool call
+            ret = ExtractedToolCallInformation(tools_called=True,
+                                               tool_calls=tool_calls,
+                                               content=None)
+            return ret
+
+        except Exception:
+            logger.exception("Error in extracting tool call from response.")
+            # return information to just treat the tool call as regular JSON
+            return ExtractedToolCallInformation(tools_called=False,
+                                                tool_calls=[],
+                                                content=model_output)
+
+    def extract_tool_calls_streaming(
+        self,
+        previous_text: str,
+        current_text: str,
+        delta_text: str,
+        previous_token_ids: Sequence[int],
+        current_token_ids: Sequence[int],
+        delta_token_ids: Sequence[int],
+        request: ChatCompletionRequest,
+    ) -> Union[DeltaMessage, None]:
+
+        if not (current_text.startswith(self.bot_token)
+                or current_text.startswith('{')):
+            return DeltaMessage(content=delta_text)
+
+        # bit mask flags for partial JSON parsing. If the name hasn't been
+        # sent yet, don't allow sending
+        # an incomplete string since OpenAI only ever (as far as I have
+        # seen) allows sending the entire tool/ function name at once.
+        flags = Allow.ALL if self.current_tool_name_sent \
+            else Allow.ALL & ~Allow.STR
+        try:
+            tool_call_arr = []
+            is_complete = []
+            try:
+                # depending on the prompt format the Llama model may or may not
+                # prefix the output with the <|python_tag|> token
+                start_idx = len(self.bot_token) if current_text.startswith(
+                    self.bot_token) else 0
+                while start_idx < len(current_text):
+                    (obj,
+                     end_idx) = partial_json_loads(current_text[start_idx:],
+                                                   flags)
+                    is_complete.append(
+                        is_complete_json(current_text[start_idx:start_idx +
+                                                      end_idx]))
+                    start_idx += end_idx + len('; ')
+                    # depending on the prompt Llama can use
+                    # either arguments or parameters
+                    if "parameters" in obj:
+                        assert "arguments" not in obj, \
+                            "model generated both parameters and arguments"
+                        obj["arguments"] = obj["parameters"]
+                    tool_call_arr.append(obj)
+            except partial_json_parser.core.exceptions.MalformedJSON:
+                logger.debug('not enough tokens to parse into JSON yet')
+                return None
+
+            # select as the current tool call the one we're on the state at
+            current_tool_call: Dict = tool_call_arr[self.current_tool_id] \
+                if len(tool_call_arr) > 0 else {}
+
+            # case -- if no tokens have been streamed for the tool, e.g.
+            #   only the array brackets, stream nothing
+            if len(tool_call_arr) == 0:
+                return None
+
+            # case: we are starting a new tool in the array
+            #   -> array has > 0 length AND length has moved past cursor
+            elif (len(tool_call_arr) > 0
+                  and len(tool_call_arr) > self.current_tool_id + 1):
+
+                # if we're moving on to a new call, first make sure we
+                # haven't missed anything in the previous one that was
+                # auto-generated due to JSON completions, but wasn't
+                # streamed to the client yet.
+                if self.current_tool_id >= 0:
+                    cur_arguments = current_tool_call.get("arguments")
+                    if cur_arguments:
+                        cur_args_json = json.dumps(cur_arguments)
+                        sent = len(
+                            self.streamed_args_for_tool[self.current_tool_id])
+                        argument_diff = cur_args_json[sent:]
+
+                        logger.debug("got arguments diff: %s", argument_diff)
+                        delta = DeltaMessage(tool_calls=[
+                            DeltaToolCall(index=self.current_tool_id,
+                                          function=DeltaFunctionCall(
+                                              arguments=argument_diff).
+                                          model_dump(exclude_none=True))
+                        ])
+                        self.streamed_args_for_tool[
+                            self.current_tool_id] += argument_diff
+                    else:
+                        delta = None
+                else:
+                    delta = None
+                # re-set stuff pertaining to progress in the current tool
+                self.current_tool_id = len(tool_call_arr) - 1
+                self.current_tool_name_sent = False
+                self.streamed_args_for_tool.append("")
+                logger.debug("starting on new tool %d", self.current_tool_id)
+                return delta
+
+            # if the current tool name hasn't been sent, send if available
+            # - otherwise send nothing
+            elif not self.current_tool_name_sent:
+                function_name = current_tool_call.get("name")
+                if function_name:
+
+                    delta = DeltaMessage(tool_calls=[
+                        DeltaToolCall(index=self.current_tool_id,
+                                      type="function",
+                                      id=f"chatcmpl-tool-{random_uuid()}",
+                                      function=DeltaFunctionCall(
+                                          name=function_name).model_dump(
+                                              exclude_none=True))
+                    ])
+                    self.current_tool_name_sent = True
+                else:
+                    delta = None
+
+            # now we know we're on the same tool call and we're streaming
+            # arguments
+            else:
+                cur_arguments = current_tool_call.get("arguments")
+                delta = None
+
+                if cur_arguments:
+                    sent = len(
+                        self.streamed_args_for_tool[self.current_tool_id])
+                    cur_args_json = json.dumps(cur_arguments)
+                    prev_arguments = self.prev_tool_call_arr[
+                        self.current_tool_id].get("arguments")
+
+                    argument_diff = None
+                    if is_complete[self.current_tool_id]:
+                        argument_diff = cur_args_json[sent:]
+                    elif prev_arguments:
+                        prev_args_json = json.dumps(prev_arguments)
+                        if cur_args_json != prev_args_json:
+
+                            prefix = find_common_prefix(
+                                prev_args_json, cur_args_json)
+                            argument_diff = prefix[sent:]
+
+                    if argument_diff is not None:
+                        delta = DeltaMessage(tool_calls=[
+                            DeltaToolCall(index=self.current_tool_id,
+                                          function=DeltaFunctionCall(
+                                              arguments=argument_diff).
+                                          model_dump(exclude_none=True))
+                        ])
+                        self.streamed_args_for_tool[
+                            self.current_tool_id] += argument_diff
+
+            self.prev_tool_call_arr = tool_call_arr
+            return delta
+
+        except Exception:
+            logger.exception("Error trying to handle streaming tool call.")
+            logger.debug(
+                "Skipping chunk as a result of tool streaming extraction "
+                "error")
+            return None

.venv/lib/python3.11/site-packages/vllm/entrypoints/openai/tool_parsers/mistral_tool_parser.py

@@ -0,0 +1,324 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import json
+import re
+from random import choices
+from string import ascii_letters, digits
+from typing import Dict, List, Sequence, Union
+
+import partial_json_parser
+from partial_json_parser.core.options import Allow
+from pydantic import Field
+
+from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
+                                              DeltaFunctionCall, DeltaMessage,
+                                              DeltaToolCall,
+                                              ExtractedToolCallInformation,
+                                              FunctionCall, ToolCall)
+from vllm.entrypoints.openai.tool_parsers.abstract_tool_parser import (
+    ToolParser, ToolParserManager)
+from vllm.entrypoints.openai.tool_parsers.utils import (
+    extract_intermediate_diff)
+from vllm.logger import init_logger
+from vllm.transformers_utils.tokenizer import AnyTokenizer, MistralTokenizer
+
+logger = init_logger(__name__)
+
+ALPHANUMERIC = ascii_letters + digits
+
+
+class MistralToolCall(ToolCall):
+    id: str = Field(
+        default_factory=lambda: MistralToolCall.generate_random_id())
+
+    @staticmethod
+    def generate_random_id():
+        # Mistral Tool Call Ids must be alphanumeric with a maximum length of 9.
+        # https://github.com/mistralai/mistral-common/blob/21ee9f6cee3441e9bb1e6ed2d10173f90bd9b94b/src/mistral_common/protocol/instruct/validator.py#L299
+        return "".join(choices(ALPHANUMERIC, k=9))
+
+
+@ToolParserManager.register_module("mistral")
+class MistralToolParser(ToolParser):
+    """
+    Tool call parser for Mistral 7B Instruct v0.3, intended for use with the
+    examples/tool_chat_template_mistral.jinja template.
+
+    Used when --enable-auto-tool-choice --tool-call-parser mistral are all set
+    """
+
+    def __init__(self, tokenizer: AnyTokenizer):
+        super().__init__(tokenizer)
+
+        if not isinstance(self.model_tokenizer, MistralTokenizer):
+            logger.info("Non-Mistral tokenizer detected when using a Mistral "
+                        "model...")
+
+        # initialize properties used for state when parsing tool calls in
+        # streaming mode
+        self.prev_tool_call_arr: List[Dict] = []
+        self.current_tool_id: int = -1
+        self.current_tool_name_sent: bool = False
+        self.streamed_args_for_tool: List[str] = [
+        ]  # map what has been streamed for each tool so far to a list
+        self.bot_token = "[TOOL_CALLS]"
+        self.bot_token_id = self.vocab.get(self.bot_token)
+        self.tool_call_regex = re.compile(r"\[{.*}\]", re.DOTALL)
+        if self.bot_token_id is None:
+            raise RuntimeError(
+                "Mistral Tool Parser could not locate the tool call token in "
+                "the tokenizer!")
+
+    def extract_tool_calls(
+        self,
+        model_output: str,
+        request: ChatCompletionRequest,
+    ) -> ExtractedToolCallInformation:
+        """
+        Extract the tool calls from a complete model response. Requires
+        find-and-replacing single quotes with double quotes for JSON parsing,
+        make sure your tool call arguments don't ever include quotes!
+        """
+
+        # case -- if a tool call token is not present, return a text response
+        if self.bot_token not in model_output:
+            return ExtractedToolCallInformation(tools_called=False,
+                                                tool_calls=[],
+                                                content=model_output)
+
+        # first remove the BOT token
+        tool_content = model_output.replace(self.bot_token, "").strip()
+
+        try:
+
+            # we first try to directly load the json as parsing very nested
+            # jsons is difficult
+            try:
+                function_call_arr = json.loads(tool_content)
+            except json.JSONDecodeError:
+                # use a regex to find the part corresponding to the tool call.
+                # NOTE: This use case should not happen if the model is trained
+                # correctly. It's a easy possible fix so it's included, but
+                # can be brittle for very complex / highly nested tool calls
+                raw_tool_call = self.tool_call_regex.findall(tool_content)[0]
+                function_call_arr = json.loads(raw_tool_call)
+
+            # Tool Call
+            tool_calls: List[MistralToolCall] = [
+                MistralToolCall(
+                    type="function",
+                    function=FunctionCall(
+                        name=raw_function_call["name"],
+                        # function call args are JSON but as a string
+                        arguments=json.dumps(raw_function_call["arguments"],
+                                             ensure_ascii=False)))
+                for raw_function_call in function_call_arr
+            ]
+
+            # get any content before  the tool call
+            content = model_output.split(self.bot_token)[0]
+            return ExtractedToolCallInformation(
+                tools_called=True,
+                tool_calls=tool_calls,
+                content=content if len(content) > 0 else None)
+
+        except Exception:
+            logger.exception("Error in extracting tool call from response.")
+            # return information to just treat the tool call as regular JSON
+            return ExtractedToolCallInformation(tools_called=False,
+                                                tool_calls=[],
+                                                content=tool_content)
+
+    def extract_tool_calls_streaming(
+        self,
+        previous_text: str,
+        current_text: str,
+        delta_text: str,
+        previous_token_ids: Sequence[int],
+        current_token_ids: Sequence[int],
+        delta_token_ids: Sequence[int],
+        request: ChatCompletionRequest,
+    ) -> Union[DeltaMessage, None]:
+
+        # if the tool call token is not in the tokens generated so far, append
+        # output to contents since it's not a tool
+        if self.bot_token not in current_text:
+            return DeltaMessage(content=delta_text)
+
+        # if the tool call token ID IS in the tokens generated so far, that
+        # means we're parsing as tool calls now
+
+        # handle if we detected the BOT token which means the start of tool
+        # calling
+        if (self.bot_token_id in delta_token_ids
+                and len(delta_token_ids) == 1):
+            # if it's the only token, return None, so we don't send a chat
+            # completion any don't send a control token
+            return None
+
+        # bit mask flags for partial JSON parsing. If the name hasn't been
+        # sent yet, don't allow sending
+        # an incomplete string since OpenAI only ever (as far as I have
+        # seen) allows sending the entire tool/ function name at once.
+        flags = Allow.ALL if self.current_tool_name_sent \
+            else Allow.ALL & ~Allow.STR
+        try:
+
+            # replace BOT token with empty string, and convert single quotes
+            # to double to allow parsing as JSON since mistral uses single
+            # quotes instead of double for tool calls
+            parsable_arr = current_text.split(self.bot_token)[-1]
+
+            # tool calls are generated in an array, so do partial JSON
+            # parsing on the entire array
+            try:
+                tool_call_arr: List[Dict] = partial_json_parser.loads(
+                    parsable_arr, flags)
+            except partial_json_parser.core.exceptions.MalformedJSON:
+                logger.debug('not enough tokens to parse into JSON yet')
+                return None
+
+            # select as the current tool call the one we're on the state at
+
+            current_tool_call: Dict = tool_call_arr[self.current_tool_id] \
+                if len(tool_call_arr) > 0 else {}
+
+            # case -- if no tokens have been streamed for the tool, e.g.
+            #   only the array brackets, stream nothing
+            if len(tool_call_arr) == 0:
+                return None
+
+            # case: we are starting a new tool in the array
+            #   -> array has > 0 length AND length has moved past cursor
+            elif (len(tool_call_arr) > 0
+                  and len(tool_call_arr) > self.current_tool_id + 1):
+
+                # if we're moving on to a new call, first make sure we
+                # haven't missed anything in the previous one that was
+                # auto-generated due to JSON completions, but wasn't
+                # streamed to the client yet.
+                if self.current_tool_id >= 0:
+                    diff: Union[str, None] = current_tool_call.get("arguments")
+
+                    if diff:
+                        diff = json.dumps(diff, ensure_ascii=False).replace(
+                            self.streamed_args_for_tool[self.current_tool_id],
+                            "")
+                        delta = DeltaMessage(tool_calls=[
+                            DeltaToolCall(index=self.current_tool_id,
+                                          function=DeltaFunctionCall(
+                                              arguments=diff).model_dump(
+                                                  exclude_none=True))
+                        ])
+                        self.streamed_args_for_tool[
+                            self.current_tool_id] += diff
+                    else:
+                        delta = None
+                else:
+                    delta = None
+                # re-set stuff pertaining to progress in the current tool
+                self.current_tool_id = len(tool_call_arr) - 1
+                self.current_tool_name_sent = False
+                self.streamed_args_for_tool.append("")
+                logger.debug("starting on new tool %d", self.current_tool_id)
+                return delta
+
+            # case: update an existing tool - this is handled below
+
+            # if the current tool name hasn't been sent, send if available
+            # - otherwise send nothing
+            if not self.current_tool_name_sent:
+                function_name = current_tool_call.get("name")
+                if function_name:
+
+                    delta = DeltaMessage(tool_calls=[
+                        DeltaToolCall(index=self.current_tool_id,
+                                      type="function",
+                                      id=MistralToolCall.generate_random_id(),
+                                      function=DeltaFunctionCall(
+                                          name=function_name).model_dump(
+                                              exclude_none=True))
+                    ])
+                    self.current_tool_name_sent = True
+                else:
+                    delta = None
+
+            # now we know we're on the same tool call and we're streaming
+            # arguments
+            else:
+
+                prev_arguments = self.prev_tool_call_arr[
+                    self.current_tool_id].get("arguments")
+                cur_arguments = current_tool_call.get("arguments")
+
+                new_text = delta_text.replace("\'", "\"")
+                if ('"}' in new_text):
+                    new_text = new_text[:new_text.rindex('"}')]
+
+                if not cur_arguments and not prev_arguments:
+
+                    delta = None
+                elif not cur_arguments and prev_arguments:
+                    logger.error(
+                        "INVARIANT - impossible to have arguments reset "
+                        "mid-arguments")
+                    delta = None
+                elif cur_arguments and not prev_arguments:
+                    cur_arguments_json = json.dumps(cur_arguments,
+                                                    ensure_ascii=False)[:-2]
+                    logger.debug("finding %s in %s", new_text,
+                                 cur_arguments_json)
+
+                    if (new_text not in cur_arguments_json):
+                        return None
+                    arguments_delta = cur_arguments_json[:cur_arguments_json.
+                                                         rindex(new_text) +
+                                                         len(new_text)]
+                    logger.debug("First tokens in arguments received: %s",
+                                 arguments_delta)
+                    delta = DeltaMessage(tool_calls=[
+                        DeltaToolCall(index=self.current_tool_id,
+                                      function=DeltaFunctionCall(
+                                          arguments=arguments_delta).
+                                      model_dump(exclude_none=True))
+                    ])
+                    self.streamed_args_for_tool[
+                        self.current_tool_id] += arguments_delta
+
+                elif cur_arguments and prev_arguments:
+                    cur_args_json = json.dumps(cur_arguments,
+                                               ensure_ascii=False)
+                    prev_args_json = json.dumps(prev_arguments,
+                                                ensure_ascii=False)
+                    logger.debug("Searching for diff between \n%s\n%s",
+                                 cur_args_json, prev_args_json)
+
+                    argument_diff = extract_intermediate_diff(
+                        cur_args_json, prev_args_json)
+                    logger.debug("got arguments diff: %s", argument_diff)
+                    delta = DeltaMessage(tool_calls=[
+                        DeltaToolCall(index=self.current_tool_id,
+                                      function=DeltaFunctionCall(
+                                          arguments=argument_diff).model_dump(
+                                              exclude_none=True))
+                    ])
+                    self.streamed_args_for_tool[
+                        self.current_tool_id] += argument_diff
+                else:
+                    # try parsing it with regular JSON - if it works we're
+                    # at the end, and we need to send the difference between
+                    # tokens streamed so far and the valid JSON
+                    delta = None
+
+            # check to see if the name is defined and has been sent. if so,
+            # stream the name - otherwise keep waiting
+            # finish by setting old and returning None as base case
+            self.prev_tool_call_arr = tool_call_arr
+            return delta
+
+        except Exception:
+            logger.exception("Error trying to handle streaming tool call.")
+            logger.debug(
+                "Skipping chunk as a result of tool streaming extraction "
+                "error")
+            return None

.venv/lib/python3.11/site-packages/vllm/entrypoints/openai/tool_parsers/utils.py

@@ -0,0 +1,123 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import json
+from json import JSONDecodeError, JSONDecoder
+from typing import Any, List, Tuple
+
+import partial_json_parser
+from partial_json_parser.core.options import Allow
+
+
+def find_common_prefix(s1: str, s2: str) -> str:
+    """
+    Finds a common prefix that is shared between two strings, if there is one.
+    Order of arguments is NOT important.
+
+    This function is provided as a UTILITY for extracting information from JSON
+    generated by partial_json_parser, to help in ensuring that the right tokens
+    are returned in streaming, so that close-quotes, close-brackets and
+    close-braces are not returned prematurely.
+
+    e.g. find_common_prefix('{"fruit": "ap"}', '{"fruit": "apple"}') ->
+    '{"fruit": "ap'
+    """
+    prefix = ''
+    min_length = min(len(s1), len(s2))
+    for i in range(0, min_length):
+        if s1[i] == s2[i]:
+            prefix += s1[i]
+        else:
+            break
+    return prefix
+
+
+def find_common_suffix(s1: str, s2: str) -> str:
+    """
+    Finds a common suffix shared between two strings, if there is one. Order of
+    arguments is NOT important.
+    Stops when the suffix ends OR it hits an alphanumeric character
+
+    e.g. find_common_suffix('{"fruit": "ap"}', '{"fruit": "apple"}') -> '"}'
+    """
+    suffix = ''
+    min_length = min(len(s1), len(s2))
+    for i in range(1, min_length + 1):
+        if s1[-i] == s2[-i] and not s1[-i].isalnum():
+            suffix = s1[-i] + suffix
+        else:
+            break
+    return suffix
+
+
+def extract_intermediate_diff(curr: str, old: str) -> str:
+    """
+    Given two strings, extract the difference in the middle between two strings
+    that are known to have a common prefix and/or suffix.
+
+    This function is provided as a UTILITY for extracting information from JSON
+    generated by partial_json_parser, to help in ensuring that the right tokens
+    are returned in streaming, so that close-quotes, close-brackets and
+    close-braces are not returned prematurely. The order of arguments IS
+    important - the new version of the partially-parsed JSON must be the first
+    argument, and the secnod argument must be from the previous generation.
+
+    What it returns, is tokens that should be streamed to the client.
+
+    e.g. extract_intermediate_diff('{"fruit": "apple"}', '{"fruit": "ap"}')
+        -> 'ple'
+
+    """
+    suffix = find_common_suffix(curr, old)
+
+    old = old[::-1].replace(suffix[::-1], '', 1)[::-1]
+    prefix = find_common_prefix(curr, old)
+    diff = curr
+    if len(suffix):
+        diff = diff[::-1].replace(suffix[::-1], '', 1)[::-1]
+
+    if len(prefix):
+        # replace the prefix only once in case it's mirrored
+        diff = diff.replace(prefix, '', 1)
+
+    return diff
+
+
+def find_all_indices(string: str, substring: str) -> List[int]:
+    """
+    Find all (starting) indices of a substring in a given string. Useful for
+    tool call extraction
+    """
+    indices = []
+    index = -1
+    while True:
+        index = string.find(substring, index + 1)
+        if index == -1:
+            break
+        indices.append(index)
+    return indices
+
+
+# partial_json_parser doesn't support extra data and
+# JSONDecorder.raw_decode doesn't support partial JSON
+def partial_json_loads(input_str: str, flags: Allow) -> Tuple[Any, int]:
+    try:
+        return (partial_json_parser.loads(input_str, flags), len(input_str))
+    except JSONDecodeError as e:
+        if "Extra data" in e.msg:
+            dec = JSONDecoder()
+            return dec.raw_decode(input_str)
+        raise
+
+
+def is_complete_json(input_str: str) -> bool:
+    try:
+        json.loads(input_str)
+        return True
+    except JSONDecodeError:
+        return False
+
+
+def consume_space(i: int, s: str) -> int:
+    while i < len(s) and s[i].isspace():
+        i += 1
+    return i

.venv/lib/python3.11/site-packages/vllm/lora/fully_sharded_layers.py

@@ -0,0 +1,335 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# pylint: disable=unused-argument
+from typing import TYPE_CHECKING, List, Optional, Tuple, Union, cast
+
+import torch
+import torch.nn as nn
+from transformers import PretrainedConfig
+
+from vllm.config import LoRAConfig
+from vllm.distributed.communication_op import (
+    tensor_model_parallel_all_gather, tensor_model_parallel_all_reduce)
+from vllm.distributed.parallel_state import get_tensor_model_parallel_rank
+from vllm.lora.layers import (ColumnParallelLinearWithLoRA,
+                              MergedColumnParallelLinearWithLoRA,
+                              MergedQKVParallelLinearWithLora,
+                              QKVParallelLinearWithLora,
+                              RowParallelLinearWithLoRA)
+
+if TYPE_CHECKING:
+    pass
+
+
+def _fully_sharded_can_replace(can_replace):
+    """
+    decorator which adds the condition of fully sharded loras
+    intended to wrap can_replace_layer()
+    """
+
+    def dec(*args, **kwargs):
+        return (can_replace(*args, **kwargs)
+                and kwargs["lora_config"].fully_sharded_loras)
+
+    return dec
+
+
+def _mcp_apply(x, bias, layer: ColumnParallelLinearWithLoRA):
+    """ 
+    For `ColumnParallelLinearWithLoRA` or classes that inherit from 
+    `ColumnParallelLinearWithLoRA`, they share the same `apply` logic.
+    """
+    assert (layer.n_slices == len(layer.lora_a_stacked) == len(
+        layer.lora_b_stacked) == len(layer.output_slices))
+    if layer.lora_bias_stacked is not None:
+        assert layer.n_slices == len(layer.lora_bias_stacked)
+
+    output = layer.base_layer.quant_method.apply(layer.base_layer, x, bias)
+
+    x = x.view(-1, x.shape[-1])
+    output, out_orig_shape = output.view(-1, output.shape[-1]), output.shape
+
+    # Since communication is needed, the buffer is directly initialized as a
+    # tensor rather than a tuple of tensor.
+    buffers = torch.zeros(
+        (layer.n_slices, x.shape[0], layer.lora_a_stacked[0].shape[2]),
+        dtype=torch.float32,
+        device=x.device,
+    )
+
+    layer.punica_wrapper.add_shrink(buffers, x, layer.lora_a_stacked, 1.0)
+    buffers = tensor_model_parallel_all_gather(buffers)
+    layer.punica_wrapper.add_expand(output,
+                                    buffers,
+                                    layer.lora_b_stacked,
+                                    layer.lora_bias_stacked,
+                                    layer.output_slices,
+                                    offset_start=0,
+                                    add_input=True)
+
+    output = output.view(*out_orig_shape)
+    # now have column partitioned and packed output
+    return output
+
+
+# these layers are based on the tensor parallelism strategy given in
+# Y. Sheng et al., S-LoRA: Serving Thousands of Concurrent LoRA Adapters. 2023,
+# https://arxiv.org/abs/2311.03285.
+
+
+class ColumnParallelLinearWithShardedLoRA(ColumnParallelLinearWithLoRA):
+    """
+    Differs from ColumnParallelLinearWithLoRA by slicing LoRA A also.
+
+    Based on S-LoRA, slicing happens along the rank dim.
+    """
+
+    # For all LoRA layers where the `base_layer` is `ColumnParallelLinear`,
+    # their `lora_a` and `lora_b` have different sharding patterns. After
+    # completing the `lora_a` GEMM , a gather operation is performed.
+    # Therefore, the sharding of `lora_a` only needs to correspond with the
+    # gather operation.
+    def slice_lora_a(self, lora_a: torch.Tensor) -> torch.Tensor:
+        tp_rank = get_tensor_model_parallel_rank()
+        shard_size = self.lora_a_stacked[0].shape[2]
+        start_idx = tp_rank * shard_size
+        lora_a = lora_a[:, start_idx:start_idx + shard_size]
+        return lora_a
+
+    def apply(self,
+              x: torch.Tensor,
+              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+        return _mcp_apply(x, bias, self)
+
+    @classmethod
+    @_fully_sharded_can_replace
+    def can_replace_layer(
+        cls,
+        source_layer: nn.Module,
+        lora_config: LoRAConfig,
+        packed_modules_list: List,
+        model_config: Optional[PretrainedConfig],
+    ) -> bool:
+        # specifying kwargs so they can be easily accessed in decorator
+        return super().can_replace_layer(
+            source_layer=source_layer,
+            lora_config=lora_config,
+            packed_modules_list=packed_modules_list,
+            model_config=model_config,
+            decorate=False,
+        )
+
+
+class MergedColumnParallelLinearWithShardedLoRA(
+        MergedColumnParallelLinearWithLoRA):
+    """
+    Differs from MergedColumnParallelLinearWithLoRA by slicing the
+    LoRA A's also.
+
+    Based on S-LoRA, slicing happens along the rank dim.
+    """
+
+    def slice_lora_a(
+        self, lora_a: List[Union[torch.Tensor, None]]
+    ) -> List[Union[torch.Tensor, None]]:
+        #NOTE: lora_a contains 2 subloras, and each sublora could be None.
+        output_shard_size = self.lora_a_stacked[0].shape[2]
+        output_start_idx = self.tp_rank * output_shard_size
+        lora_a = [
+            lora_a[0][:, output_start_idx:output_start_idx +
+                      output_shard_size] if lora_a[0] is not None else None,
+            lora_a[1][:, output_start_idx:output_start_idx +
+                      output_shard_size] if lora_a[1] is not None else None,
+        ]
+        return lora_a
+
+    def apply(self,
+              x: torch.Tensor,
+              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+        return _mcp_apply(x, bias, self)
+
+    @classmethod
+    @_fully_sharded_can_replace
+    def can_replace_layer(
+        cls,
+        source_layer: nn.Module,
+        lora_config: LoRAConfig,
+        packed_modules_list: List,
+        model_config: Optional[PretrainedConfig],
+    ) -> bool:
+        # specifying kwargs so they can be easily accessed in decorator
+        return super().can_replace_layer(
+            source_layer=source_layer,
+            lora_config=lora_config,
+            packed_modules_list=packed_modules_list,
+            model_config=model_config,
+            decorate=False,
+        )
+
+
+class QKVParallelLinearWithShardedLora(QKVParallelLinearWithLora):
+    """
+    Differs from QKVParallelLinearWithLora by slicing the
+    LoRA A's also.
+
+    Based on S-LoRA, slicing happens along the rank dim.
+    """
+
+    def slice_lora_a(self, lora_a: torch.Tensor) -> torch.Tensor:
+        tp_rank = get_tensor_model_parallel_rank()
+        shard_size = self.lora_a_stacked[0].shape[2]
+        start_idx = tp_rank * shard_size
+        lora_a = lora_a[:, start_idx:start_idx + shard_size]
+        return lora_a
+
+    def apply(self,
+              x: torch.Tensor,
+              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+        return _mcp_apply(x, bias, self)
+
+    @classmethod
+    @_fully_sharded_can_replace
+    def can_replace_layer(cls, source_layer: nn.Module,
+                          lora_config: LoRAConfig, packed_modules_list: List,
+                          model_config: Optional[PretrainedConfig]) -> bool:
+        # specifying kwargs so they can be easily accessed in decorator
+        return super().can_replace_layer(
+            source_layer=source_layer,
+            lora_config=lora_config,
+            packed_modules_list=packed_modules_list,
+            model_config=model_config,
+            decorate=False,
+        )
+
+
+class MergedQKVParallelLinearWithShardedLora(MergedQKVParallelLinearWithLora):
+    """
+    Differs from MergedQKVParallelLinearWithLora by slicing the 
+    LoRA A's also.
+
+    Based on S-LoRA, slicing happens along the rank dim.
+    """
+
+    def slice_lora_a(
+        self, lora_a: List[Union[torch.Tensor, None]]
+    ) -> List[Union[torch.Tensor, None]]:
+        # NOTE: lora_a contains 3 subloras, and each sublora could be None.
+        shard_size = [self.lora_a_stacked[i].shape[2] for i in range(3)]
+        start_idx = [self.tp_rank * shard_size[i] for i in range(3)]
+        lora_a = [
+            lora_a[0][:, start_idx[0]:start_idx[0] +
+                      shard_size[0]] if lora_a[0] is not None else None,
+            lora_a[1][:, start_idx[1]:start_idx[1] +
+                      shard_size[1]] if lora_a[1] is not None else None,
+            lora_a[2][:, start_idx[2]:start_idx[2] +
+                      shard_size[2]] if lora_a[2] is not None else None,
+        ]
+        return lora_a
+
+    def apply(self,
+              x: torch.Tensor,
+              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+        return _mcp_apply(x, bias, self)
+
+    @classmethod
+    @_fully_sharded_can_replace
+    def can_replace_layer(
+        cls,
+        source_layer: nn.Module,
+        lora_config: LoRAConfig,
+        packed_modules_list: List,
+        model_config: Optional[PretrainedConfig],
+    ) -> bool:
+        # specifying kwargs so they can be easily accessed in decorator
+        return super().can_replace_layer(
+            source_layer=source_layer,
+            lora_config=lora_config,
+            packed_modules_list=packed_modules_list,
+            model_config=model_config,
+            decorate=False,
+        )
+
+
+class RowParallelLinearWithShardedLoRA(RowParallelLinearWithLoRA):
+    """
+    Differs from RowParallelLinearWithLoRA by slicing the
+    LoRA B's also.
+
+    Based on S-LoRA, slicing happens along the output dim.
+    This yields a combined partial sum from the row parallel base
+    layer and column partitioned output from the LoRA.
+    """
+
+    def slice_lora_b(self, lora_b: torch.Tensor) -> torch.Tensor:
+        shard_size = self.lora_b_stacked[0].shape[2]
+        start_idx = self.tp_rank * shard_size
+        end_idx = (self.tp_rank + 1) * shard_size
+        lora_b = lora_b[:, start_idx:end_idx]
+        return lora_b
+
+    def slice_bias(self, bias: torch.Tensor) -> torch.Tensor:
+        if bias is None:
+            return bias
+        self.lora_bias_stacked = cast(Tuple[torch.Tensor, ...],
+                                      self.lora_bias_stacked)
+        shard_size = self.lora_bias_stacked[0].shape[2]
+        start_idx = self.tp_rank * shard_size
+        end_idx = (self.tp_rank + 1) * shard_size
+        bias = bias[start_idx:end_idx]
+        return bias
+
+    def apply(self,
+              x: torch.Tensor,
+              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+        output = self.base_layer.quant_method.apply(self.base_layer, x)
+
+        x = x.view(-1, x.shape[-1])
+        output, out_orig_shape = output.view(-1,
+                                             output.shape[-1]), output.shape
+        buffer = torch.zeros(
+            (self.n_slices, x.shape[0], self.lora_a_stacked[0].shape[2]),
+            dtype=torch.float32,
+            device=x.device,
+        )
+
+        self.punica_wrapper.add_shrink(buffer, x, self.lora_a_stacked, 1.0)
+        buffer = tensor_model_parallel_all_reduce(buffer)
+
+        # following S-LoRA, allows the fusing of all_gather and all_reduce
+        # by adding the column partitioned lora output to a slice of output
+        # tensor, which is a partial sum due to row parallel. All that
+        # remains is a standard all_reduce. User should be aware though that
+        # the output is not the same as a normal row_parallel, it should be
+        # reduced before being used
+        # NOTE offset are based on the rank.
+        shard_size = self.lora_b_stacked[0].shape[2]
+        offset_start = self.tp_rank * shard_size
+        self.punica_wrapper.add_expand(
+            output,
+            buffer,
+            self.lora_b_stacked,
+            self.lora_bias_stacked,
+            self.output_slices,
+            offset_start=offset_start,
+            add_input=True,
+        )
+        output = output.view(*out_orig_shape)
+        return output
+
+    @classmethod
+    @_fully_sharded_can_replace
+    def can_replace_layer(
+        cls,
+        source_layer: nn.Module,
+        lora_config: LoRAConfig,
+        packed_modules_list: List,
+        model_config: Optional[PretrainedConfig],
+    ) -> bool:
+        # specifying kwargs so they can be easily accessed in decorator
+        return super().can_replace_layer(
+            source_layer=source_layer,
+            lora_config=lora_config,
+            packed_modules_list=packed_modules_list,
+            model_config=model_config,
+            decorate=False,
+        )

.venv/lib/python3.11/site-packages/vllm/lora/layers.py

@@ -0,0 +1,1206 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# pylint: disable=unused-argument
+import math
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union, cast
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import PretrainedConfig
+
+from vllm.adapter_commons.layers import AdapterMapping
+from vllm.config import LoRAConfig
+from vllm.distributed import (get_tensor_model_parallel_rank,
+                              get_tensor_model_parallel_world_size,
+                              split_tensor_along_last_dim,
+                              tensor_model_parallel_all_gather,
+                              tensor_model_parallel_all_reduce,
+                              tensor_model_parallel_gather)
+from vllm.distributed.utils import divide
+# yapf: disable
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               LinearBase,
+                                               MergedColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               ReplicatedLinear,
+                                               RowParallelLinear)
+# yapf: enable
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.rotary_embedding import (
+    LinearScalingRotaryEmbedding, RotaryEmbedding)
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    VocabParallelEmbedding)
+from vllm.platforms import current_platform
+
+if TYPE_CHECKING:
+    from vllm.lora.punica_wrapper import PunicaWrapperBase
+
+
+def _get_lora_device(base_layer: nn.Module) -> torch.device:
+    # code borrowed from https://github.com/fmmoret/vllm/blob/fm-support-lora-on-quantized-models/vllm/lora/layers.py#L34
+    """Returns the device for where to place the LoRA tensors."""
+    # unquantizedLinear
+    if hasattr(base_layer, "weight"):
+        return base_layer.weight.device
+    # Compressed Tensor
+    elif hasattr(base_layer, "weight_packed"):
+        return base_layer.weight_packed.device
+    # GPTQ/AWQ
+    elif hasattr(base_layer, "qweight"):
+        return base_layer.qweight.device
+    # marlin
+    elif hasattr(base_layer, "B"):
+        return base_layer.B.device
+    # HQQ marlin
+    elif hasattr(base_layer, "W_q"):
+        return base_layer.W_q.device
+    else:
+        raise ValueError(f"Unsupported base layer: {base_layer}")
+
+
+def _not_fully_sharded_can_replace(can_replace):
+    """
+    decorator which adds the condition of not using fully sharded loras
+    intended to wrap can_replace_layer()
+    """
+
+    def dec(*args, **kwargs):
+        decorate = kwargs.pop("decorate") if "decorate" in kwargs else True
+        condition = (not kwargs["lora_config"].fully_sharded_loras
+                     if decorate else True)
+        return can_replace(*args, **kwargs) and condition
+
+    return dec
+
+
+@dataclass
+class LoRAMapping(AdapterMapping):
+    is_prefill: bool = False
+
+
+class BaseLayerWithLoRA(nn.Module):
+
+    def slice_lora_a(
+        self, lora_a: Union[torch.Tensor, List[Union[torch.Tensor, None]]]
+    ) -> Union[torch.Tensor, List[Union[torch.Tensor, None]]]:
+        """Slice lora a if splitting for tensor parallelism."""
+        ...
+
+    def slice_lora_b(
+        self, lora_b: Union[torch.Tensor, List[Union[torch.Tensor, None]]]
+    ) -> Union[torch.Tensor, List[Union[torch.Tensor, None]]]:
+        """Slice lora b if splitting with tensor parallelism."""
+        ...
+
+    def create_lora_weights(
+        self,
+        max_loras: int,
+        lora_config: LoRAConfig,
+        model_config: Optional[PretrainedConfig] = None,
+    ) -> None:
+        """Initializes lora matrices."""
+        ...
+
+    def reset_lora(self, index: int):
+        """Resets the lora weights at index back to 0."""
+        ...
+
+    def set_lora(
+        self,
+        index: int,
+        lora_a: torch.Tensor,
+        lora_b: torch.Tensor,
+        embeddings_tensor: Optional[torch.Tensor],
+        bias: Optional[torch.Tensor] = None,
+    ):
+        """Overwrites lora tensors at index."""
+        ...
+
+    def set_mapping(
+        self,
+        punica_wrapper,
+    ):
+        self.punica_wrapper: PunicaWrapperBase = punica_wrapper
+
+    @classmethod
+    def can_replace_layer(
+        cls,
+        source_layer: nn.Module,
+        lora_config: LoRAConfig,
+        packed_modules_list: List,
+        model_config: Optional[PretrainedConfig],
+    ) -> bool:
+        """Returns True if the layer can be replaced by this LoRA layer."""
+        raise NotImplementedError
+
+
+class VocabParallelEmbeddingWithLoRA(BaseLayerWithLoRA):
+
+    def __init__(self, base_layer: VocabParallelEmbedding) -> None:
+        super().__init__()
+        self.base_layer = base_layer
+        self.embeddings_slice: Optional[Tuple[int, int]]
+        self.embeddings_weights: Optional[torch.Tensor]
+
+    def create_lora_weights(
+            self,
+            max_loras: int,
+            lora_config: LoRAConfig,
+            model_config: Optional[PretrainedConfig] = None) -> None:
+
+        if self.base_layer.num_added_embeddings_per_partition > 0:
+            # We can start adding lora weights
+            self.embeddings_weights = self.base_layer.weight.data[
+                self.base_layer.num_org_embeddings_per_partition:self.
+                base_layer.num_org_embeddings_per_partition +
+                self.base_layer.num_added_embeddings_per_partition]
+            self.embeddings_slice = (
+                self.base_layer.shard_indices.added_vocab_start_index -
+                self.base_layer.org_vocab_size,
+                self.base_layer.shard_indices.added_vocab_end_index -
+                self.base_layer.org_vocab_size)
+            self.base_layer.weight.data[
+                self.base_layer.num_org_embeddings_per_partition:].fill_(0)
+        else:
+            self.embeddings_slice = None
+            self.embeddings_weights = None
+
+        self.embeddings_tensors = torch.zeros(
+            (
+                max_loras,
+                lora_config.lora_extra_vocab_size,
+                self.base_layer.embedding_dim,
+            ),
+            dtype=self.base_layer.weight.dtype,
+            device=self.base_layer.weight.device,
+        )
+        self.lora_a_stacked = torch.zeros(
+            (
+                max_loras,
+                self.base_layer.org_vocab_size +
+                lora_config.lora_extra_vocab_size,
+                lora_config.max_lora_rank,
+            ),
+            dtype=lora_config.lora_dtype,
+            device=self.base_layer.weight.device,
+        )
+        self.lora_b_stacked = torch.zeros(
+            (
+                max_loras,
+                1,
+                self.base_layer.embedding_dim,
+                lora_config.max_lora_rank,
+            ),
+            dtype=lora_config.lora_dtype,
+            device=self.base_layer.weight.device,
+        )
+        self.lora_a_stacked_2d = self.lora_a_stacked.view(
+            self.lora_a_stacked.shape[0] * self.lora_a_stacked.shape[1],
+            self.lora_a_stacked.shape[2],
+        )
+
+    def reset_lora(self, index: int):
+        self.lora_a_stacked[index] = 0
+        self.lora_b_stacked[index] = 0
+        self.embeddings_tensors[index] = 0
+
+    def set_lora(
+        self,
+        index: int,
+        lora_a: torch.Tensor,
+        lora_b: torch.Tensor,
+        embeddings_tensor: Optional[torch.Tensor],
+        bias: Optional[torch.Tensor] = None,
+    ):
+        self.reset_lora(index)
+        self.lora_a_stacked[index, :lora_a.shape[0], :lora_a.shape[1]].copy_(
+            lora_a, non_blocking=True)
+        self.lora_b_stacked[index,
+                            0, :lora_b.shape[1], :lora_b.shape[0]].copy_(
+                                lora_b.T, non_blocking=True)
+        if embeddings_tensor is not None:
+            self.embeddings_tensors[
+                index,
+                :embeddings_tensor.shape[0],
+                :embeddings_tensor.shape[1],
+            ].copy_(embeddings_tensor, non_blocking=True)
+            if self.embeddings_slice is not None:
+                # TODO(yard1): Optimize this copy, we don't need to copy
+                # everything, just the modified part
+                embeddings = self.embeddings_tensors.view(
+                    self.embeddings_tensors.shape[0] *
+                    self.embeddings_tensors.shape[1],
+                    self.embeddings_tensors.shape[2],
+                )[self.embeddings_slice[0]:self.embeddings_slice[1]]
+                assert self.embeddings_weights is not None
+                self.embeddings_weights[:embeddings.shape[0]].copy_(embeddings)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        added_tokens_mask = x > self.base_layer.org_vocab_size - 1
+        embeddings_indices = self.punica_wrapper.embeddings_indices
+        indices = embeddings_indices[1].view_as(x)
+        full_lora_a_embeddings = F.embedding(
+            x + indices,
+            self.lora_a_stacked_2d,
+        )
+        indices = embeddings_indices[0].view_as(x)
+        full_output = self.base_layer.forward(
+            x.add_(indices * added_tokens_mask))
+
+        full_output_org = full_output
+        if full_output.ndim == 3:
+            full_output = full_output.view(
+                full_output.shape[0] * full_output.shape[1], -1)
+        if full_lora_a_embeddings.ndim == 3:
+            full_lora_a_embeddings = full_lora_a_embeddings.view(
+                full_lora_a_embeddings.shape[0] *
+                full_lora_a_embeddings.shape[1],
+                -1,
+            )
+        self.punica_wrapper.add_lora_embedding(full_output,
+                                               full_lora_a_embeddings,
+                                               self.lora_b_stacked,
+                                               add_input=True)
+        return full_output.view_as(full_output_org)
+
+    @classmethod
+    def can_replace_layer(
+        cls,
+        source_layer: nn.Module,
+        lora_config: LoRAConfig,
+        packed_modules_list: List,
+        model_config: Optional[PretrainedConfig],
+    ) -> bool:
+        return type(source_layer) is VocabParallelEmbedding
+
+
+class BaseLinearLayerWithLoRA(BaseLayerWithLoRA):
+
+    def __init__(self, base_layer: LinearBase):
+        super().__init__()
+        self.base_layer = base_layer
+        self.input_size = self.base_layer.input_size
+        self.device = _get_lora_device(self.base_layer)
+        self.lora_bias_stacked: Optional[Tuple[torch.Tensor, ...]] = None
+
+        self.output_slices: Tuple[int, ...]
+        self.tp_size: int
+        self.output_size: int
+        self.n_slices: int
+
+    def create_lora_weights(
+        self,
+        max_loras: int,
+        lora_config: LoRAConfig,
+        model_config: Optional[PretrainedConfig] = None,
+    ) -> None:
+        self.lora_config = lora_config
+        #
+        if isinstance(self.base_layer, ReplicatedLinear):
+            lora_a_out_size = lora_config.max_lora_rank
+            lora_b_out_size = self.output_size
+
+        elif isinstance(self.base_layer, ColumnParallelLinear):
+            lora_a_out_size = (lora_config.max_lora_rank if
+                               not lora_config.fully_sharded_loras else divide(
+                                   lora_config.max_lora_rank, self.tp_size))
+            lora_b_out_size = self.output_size
+
+        elif isinstance(self.base_layer, RowParallelLinear):
+            lora_a_out_size = lora_config.max_lora_rank
+            lora_b_out_size = (self.output_size if
+                               not lora_config.fully_sharded_loras else divide(
+                                   self.output_size, self.tp_size))
+        else:
+            raise NotImplementedError
+
+        self.lora_a_stacked = tuple(
+            torch.zeros(
+                max_loras,
+                1,
+                lora_a_out_size,
+                self.input_size,
+                dtype=lora_config.lora_dtype,
+                device=self.device,
+            ) for _ in range(self.n_slices))
+        self.lora_b_stacked = tuple(
+            torch.zeros(
+                max_loras,
+                1,
+                lora_b_out_size,
+                lora_config.max_lora_rank,
+                dtype=lora_config.lora_dtype,
+                device=self.device,
+            ) for _ in range(self.n_slices))
+        if lora_config.bias_enabled:
+            lora_bias_out_size = lora_b_out_size
+            self.lora_bias_stacked = tuple(
+                torch.zeros(
+                    max_loras,
+                    1,
+                    lora_bias_out_size,
+                    dtype=lora_config.lora_dtype,
+                    device=self.device,
+                ) for _ in range(self.n_slices))
+        self.output_slices = (self.lora_b_stacked[0].shape[2], )
+
+    def reset_lora(self, index: int):
+        for s_index in range(self.n_slices):
+            self.lora_a_stacked[s_index][index] = 0
+            self.lora_b_stacked[s_index][index] = 0
+            if self.lora_config.bias_enabled:
+                # Make mypy happy
+                self.lora_bias_stacked = cast(Tuple[torch.Tensor, ...],
+                                              self.lora_bias_stacked)
+                self.lora_bias_stacked[s_index][index] = 0
+
+    def set_lora(
+        self,
+        index: int,
+        lora_a: torch.Tensor,
+        lora_b: torch.Tensor,
+        embeddings_tensor: Optional[torch.Tensor],
+        lora_bias: Optional[torch.Tensor] = None,
+    ):
+        # Except for QKVParallelLinearWithLora and
+        # MergedColumnParallelLinearWithLoRA, all other linear LoRA layers
+        # store weights in a tuple of size 1. These two layers will
+        # override this function.
+        assert (len(self.lora_a_stacked) == len(self.lora_b_stacked) ==
+                self.n_slices == 1)
+
+        self.reset_lora(index)
+        if self.tp_size > 1:
+            lora_a = self.slice_lora_a(lora_a)
+            lora_b = self.slice_lora_b(lora_b)
+            if lora_bias is not None:
+                lora_bias = self.slice_bias(lora_bias)
+
+        self.lora_a_stacked[0][index,
+                               0, :lora_a.shape[1], :lora_a.shape[0]].copy_(
+                                   lora_a.T, non_blocking=True)
+        self.lora_b_stacked[0][index,
+                               0, :lora_b.shape[1], :lora_b.shape[0]].copy_(
+                                   lora_b.T, non_blocking=True)
+        if lora_bias is not None:
+
+            self.lora_bias_stacked = cast(Tuple[torch.Tensor, ...],
+                                          self.lora_bias_stacked)
+            assert len(self.lora_bias_stacked)
+            self.lora_bias_stacked[0][index, 0, :lora_bias.shape[0]].copy_(
+                lora_bias.T, non_blocking=True)
+
+    def apply(self,
+              x: torch.Tensor,
+              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+        output = self.base_layer.quant_method.apply(self.base_layer, x, bias)
+        self.punica_wrapper.add_lora_linear(output, x, self.lora_a_stacked,
+                                            self.lora_b_stacked,
+                                            self.lora_bias_stacked, 1.0,
+                                            self.output_slices)
+        return output
+
+
+class ReplicatedLinearWithLoRA(BaseLinearLayerWithLoRA):
+
+    def __init__(self, base_layer: ReplicatedLinear) -> None:
+        super().__init__(base_layer, )
+        # To ensure interface compatibility, set to 1 always.
+        self.tp_size = 1
+        self.output_size = self.base_layer.output_size
+        self.n_slices = 1
+
+    def forward(
+        self, input_: torch.Tensor
+    ) -> Tuple[Optional[torch.Tensor], Optional[torch.Tensor]]:
+        """Forward of ReplicatedLinearWithLoRA
+
+        Args:
+            input_: Tensor whose last dimension is `input_size`.
+
+        Returns:
+            - output
+            - bias
+        """
+        bias = (self.base_layer.bias
+                if not self.base_layer.skip_bias_add else None)
+
+        # Matrix multiply.
+        output = self.apply(input_, bias)
+
+        output_bias = (self.base_layer.bias
+                       if self.base_layer.skip_bias_add else None)
+        return output, output_bias
+
+    # ReplicatedLinear should always be replaced, regardless of the fully
+    # sharded LoRAs setting, because it is, by definition, copied per GPU.
+    @classmethod
+    def can_replace_layer(
+        cls,
+        source_layer: nn.Module,
+        lora_config: LoRAConfig,
+        packed_modules_list: List,
+        model_config: Optional[PretrainedConfig],
+    ) -> bool:
+        return type(source_layer) is ReplicatedLinear
+
+
+class ColumnParallelLinearWithLoRA(BaseLinearLayerWithLoRA):
+    """
+    LoRA on top of ColumnParallelLinear layer.
+    LoRA B is sliced for tensor parallelism.
+    There are two types for the `base_layer`:
+    1. ColumnParallelLinear, e.g.`dense_h_to_4h` in `FalconForCausalLM`.
+    2. MergedColumnParallelLinear, e.g.`gate_up_proj` in `Phi3ForCausalLM`.
+    """
+
+    def __init__(self, base_layer: ColumnParallelLinear) -> None:
+        super().__init__(base_layer)
+        # The base_layer type is ColumnParallelLinear or
+        # MergedColumnParallelLinear, their weight sharding logic is
+        # inconsistent when TP is greater than 1.
+        self.is_merged_col_linear = type(
+            base_layer) is MergedColumnParallelLinear
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.output_size = self.base_layer.output_size_per_partition
+        # There is only one LoRA layer
+        self.n_slices = 1
+
+    def slice_lora_a(self, lora_a: torch.Tensor) -> torch.Tensor:
+        return lora_a
+
+    def slice_lora_b(self, lora_b: torch.Tensor) -> torch.Tensor:
+        # Applicable to cases where the base_layer is
+        # MergedColumnParallelLinear.
+        if self.is_merged_col_linear:
+            tp_rank = get_tensor_model_parallel_rank()
+            shard_size = self.output_size // 2
+            offset = lora_b.shape[-1] // 2
+
+            left_weight = lora_b[:, tp_rank * shard_size:(tp_rank + 1) *
+                                 shard_size]
+            right_weight = lora_b[:, offset + tp_rank * shard_size:offset +
+                                  (tp_rank + 1) * shard_size]
+            lora_b = torch.cat([left_weight, right_weight], dim=1)
+        # Applicable to cases where the base_layer is
+        # ColumnParallelLinear.
+        else:
+            tensor_model_parallel_rank = get_tensor_model_parallel_rank()
+            shard_size = self.output_size
+            start_idx = tensor_model_parallel_rank * shard_size
+            end_idx = (tensor_model_parallel_rank + 1) * shard_size
+            lora_b = lora_b[:, start_idx:end_idx]
+        return lora_b
+
+    def slice_bias(self, bias: torch.Tensor) -> torch.Tensor:
+        # TODO: Fix the slicing logic of bias.
+        if bias is None:
+            return bias
+        tensor_model_parallel_rank = get_tensor_model_parallel_rank()
+        shard_size = self.output_size
+        start_idx = tensor_model_parallel_rank * shard_size
+        end_idx = (tensor_model_parallel_rank + 1) * shard_size
+        bias = bias[start_idx:end_idx]
+        return bias
+
+    def forward(
+        self, input_: torch.Tensor
+    ) -> Tuple[Optional[torch.Tensor], Optional[torch.Tensor]]:
+        """Forward of ColumnParallelLinear
+
+        Args:
+            input_: Tensor whose last dimension is `input_size`.
+
+        Returns:
+            - output
+            - bias
+        """
+        bias = (self.base_layer.bias
+                if not self.base_layer.skip_bias_add else None)
+
+        # Matrix multiply.
+        output_parallel = self.apply(input_, bias)
+        if self.base_layer.gather_output:
+            # All-gather across the partitions.
+            output = tensor_model_parallel_all_gather(output_parallel)
+        else:
+            output = output_parallel
+        output_bias = (self.base_layer.bias
+                       if self.base_layer.skip_bias_add else None)
+        return output, output_bias
+
+    @classmethod
+    @_not_fully_sharded_can_replace
+    def can_replace_layer(
+        cls,
+        source_layer: nn.Module,
+        lora_config: LoRAConfig,
+        packed_modules_list: List,
+        model_config: Optional[PretrainedConfig],
+    ) -> bool:
+        return type(source_layer) is ColumnParallelLinear or (
+            type(source_layer) is MergedColumnParallelLinear
+            and len(packed_modules_list) == 1)
+
+
+class MergedColumnParallelLinearWithLoRA(ColumnParallelLinearWithLoRA):
+    """ColumnParallelLinear layer that is composed of 2 sublayers (slices)
+    packed together (eg. gate_proj + up_proj -> gate_up_proj).
+
+    This means we have 2 LoRAs, each applied to one half of the layer.
+
+    Both slices must have the same size.
+    """
+
+    def __init__(
+        self, base_layer: Union[MergedColumnParallelLinear,
+                                QKVParallelLinear]) -> None:
+        super().__init__(base_layer)
+        # There are two LoRA layers
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.tp_rank = get_tensor_model_parallel_rank()
+        # the output_sizes in MergedColumnParallelLinear is not sharded by tp
+        # we need to divide it by the tp_size to get correct slices size
+        output_sizes = self.base_layer.output_sizes
+        self.output_slices = tuple(
+            divide(output_size, self.tp_size) for output_size in output_sizes)
+        self.n_slices = len(self.output_slices)
+        self.output_ids = (self.tp_rank, ) * self.n_slices
+
+    def create_lora_weights(
+        self,
+        max_loras: int,
+        lora_config: LoRAConfig,
+        model_config: Optional[PretrainedConfig] = None,
+    ) -> None:
+        """
+        The main reason for overriding this function is to enhance  code 
+        maintainability.
+        """
+        self.lora_config = lora_config
+
+        lora_a_output_size_per_partition = (
+            lora_config.max_lora_rank if not lora_config.fully_sharded_loras
+            else divide(lora_config.max_lora_rank, self.tp_size))
+
+        self.lora_a_stacked = tuple(
+            torch.zeros(
+                max_loras,
+                1,
+                lora_a_output_size_per_partition,
+                self.input_size,
+                dtype=lora_config.lora_dtype,
+                device=self.device,
+            ) for _ in range(self.n_slices))
+        self.lora_b_stacked = tuple(
+            torch.zeros(
+                max_loras,
+                1,
+                output_size,
+                lora_config.max_lora_rank,
+                dtype=lora_config.lora_dtype,
+                device=self.device,
+            ) for output_size in self.output_slices)
+        if lora_config.bias_enabled:
+            self.lora_bias_stacked = tuple(
+                torch.zeros(
+                    max_loras,
+                    1,
+                    output_size,
+                    dtype=lora_config.lora_dtype,
+                    device=self.device,
+                ) for output_size in self.output_slices)
+
+    def slice_lora_a(
+        self, lora_a: List[Union[torch.Tensor, None]]
+    ) -> List[Union[torch.Tensor, None]]:
+        return lora_a
+
+    def slice_lora_b(
+        self, lora_b: List[Union[torch.Tensor, None]]
+    ) -> List[Union[torch.Tensor, None]]:
+        for i, (shard_id, shard_size) in enumerate(
+                zip(self.output_ids, self.output_slices)):
+            if (lora_b_i := lora_b[i]) is not None:
+                lora_b[i] = lora_b_i[:, shard_size * shard_id:shard_size *
+                                     (shard_id + 1)]
+        return lora_b
+
+    def slice_bias(
+        self, bias: List[Union[torch.Tensor,
+                               None]]) -> List[Union[torch.Tensor, None]]:
+        for i, (shard_id, shard_size) in enumerate(
+                zip(self.output_ids, self.output_slices)):
+            if (bias_i := bias[i]) is not None:
+                bias[i] = bias_i[shard_size * shard_id:shard_size *
+                                 (shard_id + 1)]
+        return bias
+
+    def set_lora(
+        self,
+        index: int,
+        lora_a: torch.Tensor,
+        lora_b: torch.Tensor,
+        embeddings_tensor: Optional[torch.Tensor],
+        lora_bias: Optional[torch.Tensor] = None,
+    ):
+        self.reset_lora(index)
+
+        if self.tp_size > 1:
+            lora_a = self.slice_lora_a(lora_a)
+            lora_b = self.slice_lora_b(lora_b)
+            if lora_bias is not None:
+                lora_bias = self.slice_bias(lora_bias)
+
+        for i in range(self.n_slices):
+            if (lora_a_i := lora_a[i]) is not None:
+                self.lora_a_stacked[i][
+                    index, 0, :lora_a_i.shape[1], :lora_a_i.shape[0]].copy_(
+                        lora_a_i.T, non_blocking=True)
+            if (lora_b_i := lora_b[i]) is not None:
+                self.lora_b_stacked[i][
+                    index, 0, :lora_b_i.shape[1], :lora_b_i.shape[0]].copy_(
+                        lora_b_i.T, non_blocking=True)
+
+        if lora_bias is not None:
+            self.lora_bias_stacked = cast(Tuple[torch.Tensor, ...],
+                                          self.lora_bias_stacked)
+            for i in range(self.n_slices):
+                if (lora_bias_i := lora_bias[i]) is not None:
+                    self.lora_bias_stacked[i][index,
+                                              0, :lora_bias_i.shape[0]].copy_(
+                                                  lora_bias_i.T,
+                                                  non_blocking=True)
+
+    @classmethod
+    @_not_fully_sharded_can_replace
+    def can_replace_layer(
+        cls,
+        source_layer: nn.Module,
+        lora_config: LoRAConfig,
+        packed_modules_list: List,
+        model_config: Optional[PretrainedConfig],
+    ) -> bool:
+        return (type(source_layer) is MergedColumnParallelLinear
+                and len(packed_modules_list) == 2)
+
+
+class QKVParallelLinearWithLora(ColumnParallelLinearWithLoRA):
+    """
+    ColumnParallelLinear layer that is specifically designed for
+    qkv_proj. Certain models, such as chatglm3 and baichuan-7b,
+    only contains a single LoRA within their qkv_proj layer.
+
+    During inference with Tensor Parallel, the weights of lora_b
+    must be accurately partitioned according to the respective ranks.
+
+    Q slice may have different shape than K and V slices (which both have
+    the same shape).
+    """
+
+    def __init__(self, base_layer: QKVParallelLinear) -> None:
+        super().__init__(base_layer)
+        self.q_proj_total_size = (self.base_layer.total_num_heads *
+                                  self.base_layer.head_size)
+        self.q_proj_shard_size = (self.base_layer.num_heads *
+                                  self.base_layer.head_size)
+        self.kv_proj_shard_size = (self.base_layer.num_kv_heads *
+                                   self.base_layer.head_size)
+        self.kv_proj_total_size = (self.base_layer.total_num_kv_heads *
+                                   self.base_layer.head_size)
+        # There is only one LoRA layer
+        self.n_slices = 1
+
+    def slice_lora_b(self, lora_b: torch.Tensor) -> torch.Tensor:
+        tp_rank = get_tensor_model_parallel_rank()
+        self.q_shard_id = tp_rank
+        self.kv_shard_id = tp_rank // self.base_layer.num_kv_head_replicas
+        lora_b_q = lora_b[:, self.q_proj_shard_size *
+                          self.q_shard_id:self.q_proj_shard_size *
+                          (self.q_shard_id + 1)]
+        k_offset = self.q_proj_total_size
+        lora_b_k = lora_b[:, k_offset +
+                          self.kv_proj_shard_size * self.kv_shard_id:k_offset +
+                          self.kv_proj_shard_size * (self.kv_shard_id + 1)]
+        v_offset = k_offset + self.kv_proj_total_size
+        lora_b_v = lora_b[:, v_offset +
+                          self.kv_proj_shard_size * self.kv_shard_id:v_offset +
+                          self.kv_proj_shard_size * (self.kv_shard_id + 1)]
+        lora_b = torch.cat([lora_b_q, lora_b_k, lora_b_v], dim=1)
+        return lora_b
+
+    def slice_bias(self, bias: torch.Tensor) -> torch.Tensor:
+        bias_q = bias[self.q_proj_shard_size *
+                      self.q_shard_id:self.q_proj_shard_size *
+                      (self.q_shard_id + 1)]
+        k_offset = self.q_proj_total_size
+        bias_k = bias[k_offset +
+                      self.kv_proj_shard_size * self.kv_shard_id:k_offset +
+                      self.kv_proj_shard_size * (self.kv_shard_id + 1)]
+        v_offset = k_offset + self.kv_proj_total_size
+        bias_v = bias[v_offset +
+                      self.kv_proj_shard_size * self.kv_shard_id:v_offset +
+                      self.kv_proj_shard_size * (self.kv_shard_id + 1)]
+        bias = torch.cat([bias_q, bias_k, bias_v], dim=1)
+        return bias
+
+    @classmethod
+    @_not_fully_sharded_can_replace
+    def can_replace_layer(cls, source_layer: nn.Module,
+                          lora_config: LoRAConfig, packed_modules_list: List,
+                          model_config: Optional[PretrainedConfig]) -> bool:
+        return type(source_layer) is QKVParallelLinear and len(
+            packed_modules_list) == 1
+
+
+class MergedQKVParallelLinearWithLora(MergedColumnParallelLinearWithLoRA):
+    """MergedColumnParallelLinear layer that is composed of 3 sublayers (slices)
+    packed together in qkv proj fashion
+    (q_proj + k_proj + v_proj -> qkv_proj).
+
+    This means we have 3 LoRAs, each applied to one slice of the layer.
+
+    Q slice may have different shape than K and V slices (which both have
+    the same shape).
+    """
+
+    def __init__(self, base_layer: QKVParallelLinear) -> None:
+        super().__init__(base_layer)
+        # There are three LoRA layer.
+        self.n_slices = len(self.base_layer.output_sizes)
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.tp_rank = get_tensor_model_parallel_rank()
+
+        self.q_proj_shard_size = (self.base_layer.num_heads *
+                                  self.base_layer.head_size)
+        self.kv_proj_shard_size = (self.base_layer.num_kv_heads *
+                                   self.base_layer.head_size)
+        self.q_shard_id = self.tp_rank
+        self.kv_shard_id = self.tp_rank // self.base_layer.num_kv_head_replicas
+
+        self.output_slices = (
+            self.q_proj_shard_size,
+            self.kv_proj_shard_size,
+            self.kv_proj_shard_size,
+        )
+        self.output_ids = (
+            self.q_shard_id,
+            self.kv_shard_id,
+            self.kv_shard_id,
+        )
+
+    def create_lora_weights(
+        self,
+        max_loras: int,
+        lora_config: LoRAConfig,
+        model_config: Optional[PretrainedConfig] = None,
+    ) -> None:
+        """
+        The main reason for overloading this function is to handle inconsistent 
+        weight dimensions in qkv lora.
+        """
+        super().create_lora_weights(max_loras, lora_config, model_config)
+
+    @classmethod
+    @_not_fully_sharded_can_replace
+    def can_replace_layer(
+        cls,
+        source_layer: nn.Module,
+        lora_config: LoRAConfig,
+        packed_modules_list: List,
+        model_config: Optional[PretrainedConfig],
+    ) -> bool:
+        return (type(source_layer) is QKVParallelLinear
+                and len(packed_modules_list) == 3)
+
+
+class RowParallelLinearWithLoRA(BaseLinearLayerWithLoRA):
+
+    def __init__(self, base_layer: RowParallelLinear) -> None:
+        super().__init__(base_layer)
+
+        self.tp_size = get_tensor_model_parallel_world_size()
+        # reset input_size
+        self.input_size = self.base_layer.input_size_per_partition
+        self.output_size = self.base_layer.output_size
+
+        self.tp_rank = get_tensor_model_parallel_rank()
+        # There is only one LoRA layer.
+        self.n_slices = 1
+
+    def slice_lora_a(self, lora_a: torch.Tensor) -> torch.Tensor:
+
+        shard_size = self.input_size
+        start_idx = self.tp_rank * shard_size
+        end_idx = (self.tp_rank + 1) * shard_size
+        lora_a = lora_a[start_idx:end_idx, :]
+        return lora_a
+
+    def slice_lora_b(self, lora_b: torch.Tensor) -> torch.Tensor:
+        return lora_b
+
+    def slice_bias(self, bias: torch.Tensor) -> torch.Tensor:
+        return bias
+
+    def forward(
+        self, input_: torch.Tensor
+    ) -> Tuple[Optional[torch.Tensor], Optional[torch.Tensor]]:
+        """Forward of RowParallelLinear
+
+        Args:
+            input_: tensor whose last dimension is `input_size`. If
+                    `input_is_parallel` is set, then the last dimension
+                    is `input_size // tp_size`.
+
+        Returns:
+            - output
+            - bias
+        """
+        # Set up backprop all-reduce.
+        if self.base_layer.input_is_parallel:
+            input_parallel = input_
+        else:
+            # TODO: simplify code below
+            splitted_input = split_tensor_along_last_dim(
+                input_, num_partitions=self.base_layer.tp_size)
+            input_parallel = splitted_input[self.tp_rank].contiguous()
+
+        # Matrix multiply.
+        output_parallel = self.apply(input_parallel)
+        if self.base_layer.reduce_results and self.base_layer.tp_size > 1:
+            output_ = tensor_model_parallel_all_reduce(output_parallel)
+        else:
+            output_ = output_parallel
+
+        if not self.base_layer.skip_bias_add:
+            output = (output_ + self.base_layer.bias
+                      if self.base_layer.bias is not None else output_)
+            output_bias = None
+        else:
+            output = output_
+            output_bias = self.base_layer.bias
+        return output, output_bias
+
+    @property
+    def weight(self):
+        return (self.base_layer.weight if hasattr(self.base_layer, "weight")
+                else self.base_layer.qweight)
+
+    @classmethod
+    @_not_fully_sharded_can_replace
+    def can_replace_layer(
+        cls,
+        source_layer: nn.Module,
+        lora_config: LoRAConfig,
+        packed_modules_list: List,
+        model_config: Optional[PretrainedConfig],
+    ) -> bool:
+        return type(source_layer) is RowParallelLinear
+
+
+class LogitsProcessorWithLoRA(BaseLayerWithLoRA):
+    """
+    LoRA wrapper for LogitsProcessor, with extra logic to handle the
+    application of the LoRA adapter and added LoRA vocabulary.
+
+    Args:
+        base_layer: LogitsProcessor layer
+        hidden_size: hidden size of the model
+        dtype: data type of the model
+        device: device of the model
+        sharded_to_full_mapping: index mapping from sharded vocab to full vocab
+            received from base_layer.get_sharded_to_full_mapping(). If None,
+            no reindexing will be done.
+    """
+
+    def __init__(self, base_layer: LogitsProcessor, hidden_size: int,
+                 dtype: torch.dtype, device: torch.device,
+                 sharded_to_full_mapping: Optional[List[int]]) -> None:
+        super().__init__()
+        self.base_layer = base_layer
+        self.hidden_size = hidden_size
+        self.dtype = dtype
+        self.device = device
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.tp_rank = get_tensor_model_parallel_rank()
+        self.sharded_to_full_mapping = sharded_to_full_mapping
+
+    @property
+    def logits_as_input(self):
+        return self.base_layer.logits_as_input
+
+    @property
+    def vocab_size(self):
+        return self.base_layer.vocab_size
+
+    @property
+    def scale(self):
+        return self.base_layer.scale
+
+    @property
+    def soft_cap(self):
+        return self.base_layer.soft_cap
+
+    @property
+    def use_all_gather(self):
+        return self.base_layer.use_all_gather
+
+    @property
+    def org_vocab_size(self):
+        return self.base_layer.org_vocab_size
+
+    @property
+    def include_gpu_probs_tensor(self):
+        return self.base_layer.include_gpu_probs_tensor
+
+    @property
+    def should_modify_greedy_probs_inplace(self):
+        return self.base_layer.should_modify_greedy_probs_inplace
+
+    def create_lora_weights(
+        self,
+        max_loras: int,
+        lora_config: LoRAConfig,
+        model_config: Optional[PretrainedConfig] = None,
+    ) -> None:
+        # TODO: Verify if this condition can be further relaxed
+        if 32000 < self.base_layer.vocab_size > 257024:
+            raise ValueError("When using LoRA, vocab size must be "
+                             "32000 >= vocab_size <= 257024")
+        self.lora_a_stacked = torch.zeros(
+            (
+                max_loras,
+                1,
+                lora_config.max_lora_rank,
+                self.hidden_size,
+            ),
+            dtype=lora_config.lora_dtype,
+            device=self.device,
+        )
+        self.lora_b_stacked = torch.zeros(
+            (
+                max_loras,
+                1,
+                # Pad for kernel compatibility
+                math.ceil(self.base_layer.vocab_size /
+                          lora_config.lora_vocab_padding_size) *
+                lora_config.lora_vocab_padding_size,
+                lora_config.max_lora_rank,
+            ),
+            dtype=lora_config.lora_dtype,
+            device=self.device,
+        )
+        self.embeddings_tensors = torch.full(
+            (max_loras, lora_config.lora_extra_vocab_size, self.hidden_size),
+            fill_value=float("-inf"),
+            dtype=self.dtype,
+            device=self.device,
+        )
+        if self.sharded_to_full_mapping is not None:
+            self.sharded_to_full_mapping_gpu = torch.tensor(
+                self.sharded_to_full_mapping,
+                device=self.device,
+                dtype=torch.long)
+        else:
+            self.sharded_to_full_mapping_gpu = None
+
+    def reset_lora(self, index: int):
+        self.lora_a_stacked[index] = 0
+        self.lora_b_stacked[index] = 0
+        self.embeddings_tensors[index] = float("-inf")
+
+    def set_lora(
+        self,
+        index: int,
+        lora_a: torch.Tensor,
+        lora_b: torch.Tensor,
+        embeddings_tensor: Optional[torch.Tensor],
+        bias: Optional[torch.Tensor] = None,
+    ):
+        self.reset_lora(index)
+        self.lora_a_stacked[index,
+                            0, :lora_a.shape[1], :lora_a.shape[0]].copy_(
+                                lora_a.T, non_blocking=True)
+        self.lora_b_stacked[index,
+                            0, :lora_b.shape[1], :lora_b.shape[0]].copy_(
+                                lora_b.T, non_blocking=True)
+        if embeddings_tensor is not None:
+            self.embeddings_tensors[
+                index,
+                :embeddings_tensor.shape[0],
+                :embeddings_tensor.shape[1],
+            ] = embeddings_tensor
+
+    def _get_logits(
+        self,
+        hidden_states: torch.Tensor,
+        lm_head: VocabParallelEmbedding,
+        embedding_bias: Optional[torch.Tensor] = None,
+    ) -> Optional[torch.Tensor]:
+        # Get the logits for the next tokens.
+        logits = lm_head.linear_method.apply(lm_head, hidden_states)
+        if embedding_bias is not None:
+            logits += embedding_bias
+        logits = tensor_model_parallel_gather(logits)
+        if logits is None:
+            return None
+
+        if self.sharded_to_full_mapping_gpu is not None:
+            # Reindex full logits tensor to ensure 1:1 mapping between
+            # index and token_id
+            # Example for:
+            #   org_vocab_size = 4
+            #   added_vocab_size = 2
+            #   pad_to_size = 8
+            #   tp_size = 2
+
+            # indices:  [0, 1, 2,  3, 4, 5, 6,  7]
+            # token_id: [0, 1, 4, -1, 2, 3, 5, -1]
+
+            # Therefore, the mapping is expected to be:
+            # [0, 1, 4, 6, 2, 3, 5, 7] so that when we reindex,
+            # we get:
+            # indices:  [0, 1, 2, 3, 4, 5,  6,  7]
+            # token_id: [0, 1, 2, 3, 4, 5, -1, -1]
+            logits = logits[:, self.sharded_to_full_mapping_gpu]
+
+        lora_logits = torch.empty(
+            self.embeddings_tensors.shape[0] + 1,
+            self.embeddings_tensors.shape[1],
+            hidden_states.shape[0],
+            dtype=self.embeddings_tensors.dtype,
+            device=self.embeddings_tensors.device,
+        )
+        torch.matmul(self.embeddings_tensors,
+                     hidden_states.T,
+                     out=lora_logits[:-1])
+        lora_logits[-1] = float("-inf")
+        lora_logits = lora_logits.mT
+        indices_padded = self.punica_wrapper.sampler_indices_padded
+        lora_logits = (lora_logits.reshape(
+            lora_logits.shape[0] * lora_logits.shape[1],
+            lora_logits.shape[2],
+        ).index_select(0, indices_padded).nan_to_num_(nan=float("-inf"),
+                                                      posinf=float("inf"),
+                                                      neginf=float("-inf")))
+
+        # HPU needs special handling to prune out dummy samples.
+        if current_platform.is_hpu():
+            lora_logits = lora_logits[:logits.shape[0], :]
+
+        logits[:,
+               self.base_layer.org_vocab_size:self.base_layer.org_vocab_size +
+               lora_logits.shape[1]] = lora_logits
+
+        # LogitsProcessorWithLoRA always using bgmv
+        self.punica_wrapper.add_lora_logits(logits, hidden_states,
+                                            self.lora_a_stacked,
+                                            self.lora_b_stacked, 1.0)
+
+        # Remove paddings in vocab (if any).
+        logits = logits[:, :self.base_layer.vocab_size]
+        return logits
+
+    def forward(self, *args, **kwargs):
+        return type(self.base_layer).forward(self, *args, **kwargs)
+
+    @classmethod
+    def can_replace_layer(
+        cls,
+        source_layer: nn.Module,
+        lora_config: LoRAConfig,
+        packed_modules_list: List,
+        model_config: Optional[PretrainedConfig],
+    ) -> bool:
+        # Special handling for the LogitsProcessor.
+        return False
+
+
+class LinearScalingRotaryEmbeddingWithLora(BaseLayerWithLoRA):
+    """Implements RoPE-scaled embeddings with linear scaling for
+    multiple LoRA adapters with a specialized kernel.
+
+    Replace LinearScalingRotaryEmbedding with MultiLinearScalingRotaryEmbedding
+    which can handle multi lora adapters in a specialied kernel.
+    """
+
+    def __init__(self, base_layer: RotaryEmbedding) -> None:
+        super().__init__()
+        self.base_layer = base_layer
+
+    @property
+    def scaling_factors(self):
+        return self.base_layer.scaling_factors
+
+    @property
+    def rotary_dim(self):
+        return self.base_layer.rotary_dim
+
+    def create_lora_weights(
+        self,
+        max_loras: int,
+        lora_config: LoRAConfig,
+        model_config: Optional[PretrainedConfig] = None,
+    ) -> None:
+        scaling_factors = (list(lora_config.long_lora_scaling_factors)
+                           if lora_config.long_lora_scaling_factors else [])
+        base_scaling_factor = (self.base_layer.scaling_factor if isinstance(
+            self.base_layer, LinearScalingRotaryEmbedding) else 1.0)
+        scaling_factors = sorted(
+            list(set([base_scaling_factor] + scaling_factors)))
+        self.base_layer = LinearScalingRotaryEmbedding(
+            self.base_layer.head_size,
+            self.base_layer.rotary_dim,
+            self.base_layer.max_position_embeddings,
+            self.base_layer.base,
+            self.base_layer.is_neox_style,
+            scaling_factors,
+            self.base_layer.dtype,
+        )
+
+    def reset_lora(self, index: int):
+        ...
+
+    def set_lora(
+        self,
+        index: int,
+        lora_a: torch.Tensor,
+        lora_b: torch.Tensor,
+        embeddings_tensor: Optional[torch.Tensor],
+        bias: Optional[torch.Tensor] = None,
+    ):
+        ...
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        return self.base_layer(
+            positions,
+            query,
+            key,
+            offsets=self.punica_wrapper.long_lora_indices,
+        )
+
+    @property
+    def scaling_factor_to_offset(self) -> Dict[float, int]:
+        return self.base_layer.scaling_factor_to_offset
+
+    @classmethod
+    def can_replace_layer(
+        cls,
+        source_layer: nn.Module,
+        lora_config: LoRAConfig,
+        packed_modules_list: List,
+        model_config: Optional[PretrainedConfig],
+    ) -> bool:
+        """Returns True if the layer can be replaced by this LoRA layer."""
+        return (type(source_layer) is LinearScalingRotaryEmbedding
+                or type(source_layer) is RotaryEmbedding)
+
+    def extra_repr(self) -> str:
+        return self.base_layer.extra_repr()

.venv/lib/python3.11/site-packages/vllm/lora/lora.py

@@ -0,0 +1,198 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import List, Optional
+from typing import Sequence as GenericSequence
+
+import torch
+import torch.types
+
+from vllm.lora.peft_helper import PEFTHelper
+from vllm.utils import is_pin_memory_available
+
+
+class LoRALayerWeights:
+    """LoRA weights for a layer composed of two low rank matrixes."""
+
+    def __init__(
+        self,
+        module_name: str,
+        rank: int,
+        lora_alpha: int,
+        lora_a: torch.Tensor,
+        lora_b: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+        embeddings_tensor: Optional[torch.Tensor] = None,
+        scaling: Optional[float] = None,
+    ) -> None:
+        self.module_name = module_name
+        self.rank = rank
+        self.lora_alpha = lora_alpha
+        self.lora_a = lora_a
+        self.lora_b = lora_b
+        self.bias = bias
+        self.embeddings_tensor = embeddings_tensor
+
+        if scaling is None:
+            self.scaling = self.lora_alpha / self.rank
+        else:
+            self.scaling = scaling
+
+    def optimize(self) -> "LoRALayerWeights":
+        """Optimize the LoRA by merging the scaling into lora_b."""
+        if self.scaling == 1:
+            return self
+        self.lora_b *= self.scaling
+        self.scaling = 1
+        return self
+
+    @property
+    def input_dim(self) -> int:
+        return self.lora_a.shape[0]
+
+    @property
+    def output_dim(self) -> int:
+        return self.lora_b.shape[1]
+
+    @property
+    def is_packed(self) -> bool:
+        return False
+
+    @property
+    def extra_vocab_size(self) -> int:
+        return self.embeddings_tensor.shape[
+            0] if self.embeddings_tensor is not None else 0
+
+    @classmethod
+    def from_config(
+        cls,
+        module_name: str,
+        peft_helper: PEFTHelper,
+        embeddings_tensor: Optional[torch.Tensor] = None,
+    ) -> "LoRALayerWeights":
+        return cls(module_name, peft_helper.r, peft_helper.lora_alpha, None,
+                   None, None, embeddings_tensor,
+                   peft_helper.vllm_lora_scaling_factor)
+
+    @classmethod
+    def create_dummy_lora_weights(
+            cls,
+            module_name: str,
+            input_dim: int,
+            output_dim: int,
+            rank: int,
+            dtype: torch.dtype,
+            device: torch.types.Device,
+            embeddings_tensor_dim: Optional[int] = None,
+            bias_enabled: Optional[bool] = False) -> "LoRALayerWeights":
+        pin_memory = str(device) == "cpu" and is_pin_memory_available()
+        lora_a = torch.zeros([input_dim, rank],
+                             dtype=dtype,
+                             device=device,
+                             pin_memory=pin_memory)
+        lora_b = torch.zeros([rank, output_dim],
+                             dtype=dtype,
+                             device=device,
+                             pin_memory=pin_memory)
+        if bias_enabled:
+            bias = torch.zeros([output_dim],
+                               dtype=dtype,
+                               device=device,
+                               pin_memory=pin_memory)
+        else:
+            bias = None
+
+        embeddings_tensor = torch.rand(
+            10,
+            embeddings_tensor_dim,
+            dtype=dtype,
+            device=device,
+            pin_memory=pin_memory) if embeddings_tensor_dim else None
+        return cls(
+            module_name,
+            rank=rank,
+            lora_alpha=1,
+            lora_a=lora_a,
+            lora_b=lora_b,
+            bias=bias,
+            embeddings_tensor=embeddings_tensor,
+        )
+
+
+class PackedLoRALayerWeights(LoRALayerWeights):
+    """LoRA used for packed layers (eg. qkv_proj)."""
+
+    def __init__(
+        self,
+        module_name: str,
+        rank: int,
+        lora_alphas: List[Optional[int]],
+        lora_a: List[Optional[torch.Tensor]],
+        lora_b: List[Optional[torch.Tensor]],
+        bias: Optional[List[Optional[torch.Tensor]]] = None,
+        scaling: Optional[List[float]] = None,
+    ) -> None:
+        super().__init__(
+            module_name=module_name,
+            rank=rank,
+            lora_alpha=0,
+            lora_a=lora_a,
+            lora_b=lora_b,
+            bias=bias,
+            scaling=scaling,  # type: ignore
+            embeddings_tensor=None,
+        )
+        self.lora_alphas = lora_alphas
+        if scaling is None:
+            self.scaling = [  # type: ignore
+                lora_alpha / self.rank  # type: ignore # noqa
+                for lora_alpha in self.lora_alphas
+            ]
+
+    @classmethod
+    def pack(
+        cls, loras: GenericSequence[Optional["LoRALayerWeights"]]
+    ) -> "PackedLoRALayerWeights":
+        """Pack a list of LoRAs into a single LoRA.
+
+        If LoRA is None, it signifies that the submodule does not have a LoRA.
+        """
+        first_lora = next(lora for lora in loras if lora is not None)
+        for lora in loras:
+            if lora is None:
+                continue
+            lora.optimize()
+        rank = first_lora.rank
+        module_name = first_lora.module_name
+        obj = cls(
+            module_name,
+            rank,
+            [lora.lora_alpha if lora is not None else None for lora in loras],
+            [lora.lora_a if lora is not None else None for lora in loras],
+            [lora.lora_b if lora is not None else None for lora in loras],
+            [lora.bias if lora is not None else None for lora in loras],
+            scaling=[
+                1 if lora is not None else None  # type: ignore
+                for lora in loras
+            ])
+        return obj
+
+    def optimize(self) -> "PackedLoRALayerWeights":
+        """Optimize the LoRA by merging the scaling into lora_b."""
+        for i in range(len(self.lora_b)):
+            if self.scaling[i] == 1 or self.lora_b[i] is None:  # type: ignore
+                continue
+            self.lora_b[i] *= self.scaling[i]  # type: ignore
+            self.scaling[i] = 1  # type: ignore
+        return self
+
+    @property
+    def input_dim(self) -> int:
+        raise NotImplementedError()
+
+    @property
+    def output_dim(self) -> int:
+        raise NotImplementedError()
+
+    @property
+    def is_packed(self) -> bool:
+        return True

.venv/lib/python3.11/site-packages/vllm/lora/models.py

@@ -0,0 +1,763 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import copy
+import math
+import os
+import re
+from dataclasses import dataclass, field
+from typing import Any, Callable, Dict, List, Optional, Sequence, Type, Union
+
+import safetensors.torch
+import torch
+from torch import nn
+
+from vllm.adapter_commons.models import (AdapterLRUCache, AdapterModel,
+                                         AdapterModelManager)
+from vllm.adapter_commons.utils import (add_adapter, deactivate_adapter,
+                                        get_adapter, list_adapters,
+                                        remove_adapter, set_adapter_mapping)
+from vllm.config import LoRAConfig
+from vllm.logger import init_logger
+from vllm.lora.layers import (BaseLayerWithLoRA,
+                              LinearScalingRotaryEmbeddingWithLora,
+                              LoRAMapping)
+from vllm.lora.lora import LoRALayerWeights, PackedLoRALayerWeights
+from vllm.lora.peft_helper import PEFTHelper
+from vllm.lora.punica_wrapper import get_punica_wrapper
+from vllm.lora.utils import (from_layer, from_layer_logits_processor,
+                             is_regex_target_modules,
+                             parse_fine_tuned_lora_name, replace_submodule)
+from vllm.model_executor.models import SupportsLoRA, supports_multimodal
+from vllm.model_executor.models.module_mapping import MultiModelKeys
+from vllm.model_executor.models.utils import PPMissingLayer, WeightsMapper
+from vllm.utils import is_pin_memory_available
+
+logger = init_logger(__name__)
+
+_GLOBAL_LORA_ID = 0
+
+
+@dataclass
+class LongContextLoRAContext:
+    """Context for lora adapters that support long context."""
+    # The scaling factors to support long context lora fine tuned models.
+    scaling_factors: List[float]
+    # dimension to apply rotary embedding.
+    rot_dim: int
+    # offsets to the sin_cos_cache for each lora_id loaded.
+    # This value is dynamically modified.
+    offsets_by_lora_id: Dict[int, int] = field(default_factory=dict)
+
+
+def get_lora_id():
+    global _GLOBAL_LORA_ID
+    _GLOBAL_LORA_ID += 1
+    return _GLOBAL_LORA_ID
+
+
+class LoRAModel(AdapterModel):
+    """A LoRA fine-tuned model."""
+
+    def __init__(
+        self,
+        lora_model_id: int,
+        rank: int,
+        loras: Dict[str, LoRALayerWeights],
+        scaling_factor: Optional[float] = None,
+    ) -> None:
+        """
+        Args:
+            lora_model_id: The integer id for the lora model.
+            rank: lora rank.
+            loras: module name -> weights for lora-replaced layers.
+            scaling_factor: Scaling factor to support long context lora model.
+                None if the lora is not tuned for long context support.
+        """
+        self.id = lora_model_id
+        # Scaling factor for long context lora model. None if it is not
+        # fine tuned for the long context.
+        self.scaling_factor = scaling_factor
+        assert (
+            lora_model_id
+            > 0), f"a valid lora id should be greater than 0, got {self.id}"
+        self.rank = rank
+        self.loras: Dict[str, LoRALayerWeights] = loras
+
+    def clone(self, lora_model_id: int) -> "LoRAModel":
+        """Return a copy of the object with different ids.
+
+        Will share the underlying tensors."""
+        return self.__class__(
+            lora_model_id,
+            rank=self.rank,
+            loras=self.loras.copy(),
+        )
+
+    @property
+    def extra_vocab_size(self) -> int:
+        return max(lora.extra_vocab_size
+                   for lora in self.loras.values()) if self.loras else 0
+
+    def get_lora(self, module_name: str) -> Optional[LoRALayerWeights]:
+        """Get LoRA for a given module by name"""
+        return self.loras.get(module_name, None)
+
+    # (yard1): TODO see if we can derive target_embedding_padding automatically
+    @classmethod
+    def from_lora_tensors(
+        cls,
+        lora_model_id: int,
+        tensors: Dict[str, torch.Tensor],
+        peft_helper: PEFTHelper,
+        device: str = "cuda",
+        dtype: Optional[torch.dtype] = None,
+        embeddings: Optional[Dict[str, torch.Tensor]] = None,
+        target_embedding_padding: Optional[int] = None,
+        embedding_modules: Optional[Dict[str, str]] = None,
+        embedding_padding_modules: Optional[List[str]] = None,
+        weights_mapper: Optional[WeightsMapper] = None,
+    ) -> "LoRAModel":
+        """Create a LoRAModel from a dictionary of tensors."""
+        pin_memory = str(device) == "cpu" and is_pin_memory_available()
+        loras: Dict[str, LoRALayerWeights] = {}
+        for tensor_name, tensor in tensors.items():
+            module_name, is_lora_a, is_bias = parse_fine_tuned_lora_name(
+                tensor_name, weights_mapper)
+            if module_name not in loras:
+                lora_embeddings_tensor = None
+                if embeddings:
+                    assert embedding_modules is not None
+                    embeddings_module = next(
+                        (k for k in embedding_modules if k in module_name),
+                        None)
+                    if embeddings_module:
+                        lora_embeddings_tensor = embeddings[
+                            embedding_modules[embeddings_module]].to(
+                                device=device, dtype=dtype)
+                        if pin_memory:
+                            lora_embeddings_tensor = (
+                                lora_embeddings_tensor.pin_memory())
+                loras[module_name] = LoRALayerWeights.from_config(
+                    module_name, peft_helper, lora_embeddings_tensor)
+
+            if is_bias:
+                loras[module_name].bias = tensor.to(device=device,
+                                                    dtype=dtype).t()
+                bias = tensor.to(device=device, dtype=dtype).t()
+                if pin_memory:
+                    bias = bias.pin_memory()
+                loras[module_name].bias = bias
+            elif is_lora_a:
+                loras[module_name].lora_a = tensor.to(device=device,
+                                                      dtype=dtype).t()
+                if pin_memory:
+                    loras[module_name].lora_a = loras[
+                        module_name].lora_a.pin_memory()
+            else:
+                loras[module_name].lora_b = tensor.to(device=device,
+                                                      dtype=dtype).t()
+                assert embedding_padding_modules is not None
+                if any(name in module_name
+                       for name in embedding_padding_modules
+                       ) and target_embedding_padding is not None:
+                    lora_b = loras[module_name].lora_b
+                    assert target_embedding_padding >= lora_b.shape[1]
+                    addition = target_embedding_padding - lora_b.shape[1]
+                    loras[module_name].lora_b = torch.nn.functional.pad(
+                        lora_b, (0, addition))
+                if pin_memory:
+                    loras[module_name].lora_b = loras[
+                        module_name].lora_b.pin_memory()
+
+        for lora in loras.values():
+            lora.optimize()
+
+        return cls(lora_model_id,
+                   peft_helper.r,
+                   loras,
+                   scaling_factor=peft_helper.vllm_long_context_scaling_factor)
+
+    @classmethod
+    def from_local_checkpoint(
+        cls,
+        lora_dir: str,
+        expected_lora_modules: List[str],
+        peft_helper: PEFTHelper,
+        *,
+        lora_model_id: Optional[int] = None,
+        device: str = "cuda",
+        dtype: Optional[torch.dtype] = None,
+        target_embedding_padding: Optional[int] = None,
+        embedding_modules: Optional[Dict[str, str]] = None,
+        embedding_padding_modules: Optional[List[str]] = None,
+        weights_mapper: Optional[WeightsMapper] = None,
+    ) -> "LoRAModel":
+        """Create a LoRAModel from a local checkpoint.
+        
+        Args:
+            lora_dir: The local path that has lora data.
+            expected_lora_modules: Name of modules that are expected to be
+                replaced by lora.
+            peft_helper: Loaded lora configuration information.
+            lora_model_id: Lora model id. If not given, automatically set by
+                a global counter.
+            device: Device where the lora model is loaded.
+            dtype: dtype of the lora model weights.
+
+        Returns:
+            Loaded LoRA Model.
+        """
+        lora_tensor_path = os.path.join(lora_dir, "adapter_model.safetensors")
+        lora_bin_file_path = os.path.join(lora_dir, "adapter_model.bin")
+        new_embeddings_tensor_path = os.path.join(
+            lora_dir, "new_embeddings.safetensors")
+        new_embeddings_bin_file_path = os.path.join(lora_dir,
+                                                    "new_embeddings.bin")
+
+        unexpected_modules: List[Union[list[str], str]]
+        if os.path.isfile(lora_tensor_path):
+            tensors: Dict[str, torch.Tensor] = {}
+            # Find unexpected modules.
+            # Use safetensor key as a source of truth to find expected modules.
+            # in peft if you have target_modules A, B, C and C does not exist
+            # in the model it won’t error and model will be trained with A, B
+            # loraified. C won’t exist in the safetensor but it will exist in
+            # the target_modules of the adapter_config.json.
+            unexpected_modules = []
+            with safetensors.safe_open(lora_tensor_path,
+                                       framework="pt") as f:  # type: ignore
+                for lora_module in f.keys():  # noqa
+                    module_name, _, _ = parse_fine_tuned_lora_name(
+                        lora_module, weights_mapper)
+                    part_name = module_name.split(".")[-1]
+                    if part_name not in expected_lora_modules:
+                        unexpected_modules.append(module_name)
+                if unexpected_modules:
+                    raise ValueError(
+                        f"While loading {lora_dir}, expected"
+                        f" target modules in {expected_lora_modules}"
+                        f" but received {unexpected_modules}."
+                        f" Please verify that the loaded LoRA module is correct"
+                    )
+                # Load tensors if there are only expected modules.
+                for module in f.keys():  # noqa
+                    tensors[module] = f.get_tensor(module)
+        elif os.path.isfile(lora_bin_file_path):
+            # When a bin file is provided, we rely on config to find unexpected
+            # modules.
+            unexpected_modules = []
+            target_modules = peft_helper.target_modules
+            if not isinstance(target_modules, list):
+                target_modules = [target_modules]
+            for module in target_modules:
+                # Compatible with more modules,
+                # such as:layers.11.self_attn.k_proj
+                part_name = module.split(".")[-1]
+                if part_name not in expected_lora_modules:
+                    unexpected_modules.append(module)
+            # loaded lora's target modules must be a subset of
+            # expected_lora_modules. It is not reliable. See
+            # https://github.com/vllm-project/vllm/pull/5909. But there's no
+            # other better mechanism.
+            if unexpected_modules and not is_regex_target_modules(
+                    peft_helper.target_modules, expected_lora_modules):
+                raise ValueError(
+                    f"While loading {lora_dir}, expected"
+                    f" target modules in {expected_lora_modules}"
+                    f" but received {unexpected_modules}."
+                    f" Please verify that the loaded LoRA module is correct")
+            tensors = torch.load(lora_bin_file_path, map_location=device)
+        else:
+            raise ValueError(f"{lora_dir} doesn't contain tensors")
+
+        embeddings = None
+        if os.path.isfile(new_embeddings_tensor_path):
+            embeddings = safetensors.torch.load_file(
+                new_embeddings_tensor_path)
+        elif os.path.isfile(new_embeddings_bin_file_path):
+            embeddings = torch.load(new_embeddings_bin_file_path,
+                                    map_location=device,
+                                    weights_only=True)
+
+        return cls.from_lora_tensors(
+            lora_model_id=get_lora_id()
+            if lora_model_id is None else lora_model_id,
+            tensors=tensors,
+            peft_helper=peft_helper,
+            device=device,
+            dtype=dtype,
+            embeddings=embeddings,
+            target_embedding_padding=target_embedding_padding,
+            embedding_modules=embedding_modules,
+            embedding_padding_modules=embedding_padding_modules,
+            weights_mapper=weights_mapper)
+
+
+class LoRAModelManager(AdapterModelManager):
+    """A manager that manages multiple LoRA-fine-tuned models."""
+
+    def __init__(
+        self,
+        model: SupportsLoRA,
+        max_num_seqs: int,
+        max_num_batched_tokens: int,
+        vocab_size: int,
+        lora_config: LoRAConfig,
+        device: torch.device,
+    ):
+        """Create a LoRAModelManager and adapter for a given model.
+
+        Args:
+            model: the model to be adapted.
+            max_num_seqs: the maximum number of sequences model can run in a
+                single batch.
+            max_num_batched_tokens: the maximum number of tokens model can run
+                in a single batch.
+            vocab_size: the vocab size of the model.
+            lora_config: the LoRA configuration.
+        """
+        self.lora_config = lora_config
+        self.device = device
+        self.max_num_seqs = max_num_seqs
+        assert self.capacity >= self.lora_slots
+        self.max_num_batched_tokens = math.ceil(max_num_batched_tokens / 8) * 8
+        self.lora_index_to_id: List[Optional[int]] = [None] * self.lora_slots
+        self.vocab_size = vocab_size
+        self.long_lora_context: Optional[LongContextLoRAContext] = None
+        self.punica_wrapper = get_punica_wrapper(max_num_batched_tokens,
+                                                 max_batches=self.max_num_seqs,
+                                                 device=self.device)
+        # Scaling factor -> offset to the sin_cos_cache to it.
+        # Used for long context lora.
+        self.scaling_factor_to_offset: Dict[float, int] = {}
+        super().__init__(model)
+        if hasattr(self.model, "supported_lora_modules"):
+            self.supported_lora_modules = copy.deepcopy(
+                self.model.supported_lora_modules)
+            if lora_config.long_lora_scaling_factors:
+                # We need to replace rotary emb layer to do batch computation
+                # for long lora.
+                self.supported_lora_modules.append("rotary_emb")
+            self.packed_modules_mapping = copy.deepcopy(
+                self.model.packed_modules_mapping)
+        # Used to indicate whether the model is a multimodal model
+        self.supports_mm: bool = (
+            supports_multimodal(self.model)
+            # In case the model only supports LoRA for
+            # text modules (e.g. ChatGLM)
+            and hasattr(self.model, "get_mm_mapping"))
+        self.packed_modules: Dict[str, List[str]] = {}
+        self.modules: Dict[str, BaseLayerWithLoRA] = {}
+        # Dict instead of a Set for compatibility with LRUCache.
+        self._last_mapping: Optional[LoRAMapping] = None
+        self._create_lora_modules()
+        self.model.lora_manager = self
+        self.adapter_type = 'LoRa'
+
+    @property
+    def capacity(self) -> int:
+        return self.lora_config.max_cpu_loras
+
+    @property
+    def lora_slots(self) -> int:
+        return self.lora_config.max_loras
+
+    @property
+    def adapter_slots(self) -> int:
+        return self.lora_slots
+
+    def activate_adapter(
+        self,
+        lora_id: int,
+    ) -> bool:
+        """Move LoRA into a GPU buffer to be used in the forward pass."""
+        if lora_id in self._active_adapters:
+            return False
+        first_free_slot = next(
+            ((i, lora_id) for i, lora_id in enumerate(self.lora_index_to_id)
+             if lora_id is None), None)
+        if first_free_slot is None:
+            raise ValueError("No free lora slots")
+        index, _ = first_free_slot
+        self._active_adapters[lora_id] = None
+        lora_model = self._registered_adapters[lora_id]
+        logger.debug("Activating LoRA. int id: %d, slot index: %d",
+                     lora_model.id, index)
+        self.lora_index_to_id[index] = lora_model.id
+        for module_name, module in self.modules.items():
+            module_lora = lora_model.get_lora(module_name)
+            if module_lora:
+                module_lora.optimize()
+                # Bias is not explicitly enabled with the flag enable_lora_bias.
+                bias = module_lora.bias
+                if ((torch.is_tensor(bias) or
+                     (isinstance(bias, Sequence) and any(b is not None
+                                                         for b in bias)))
+                        and not self.lora_config.bias_enabled):
+                    module_lora.bias = None
+                    raise ValueError(
+                        f"Adapter bias cannot be used for {module_name}"
+                        " without --enable-lora-bias.")
+                module.set_lora(index, module_lora.lora_a, module_lora.lora_b,
+                                module_lora.embeddings_tensor,
+                                module_lora.bias)
+            else:
+                module.reset_lora(index)
+        return True
+
+    def _deactivate_adapter(self, lora_id: int):
+        try:
+            index = self.lora_index_to_id.index(lora_id)
+            self.lora_index_to_id[index] = None
+        except ValueError:
+            pass
+
+    def _set_long_lora_context(self, lora: LoRAModel):
+        if self.long_lora_context is None:
+            return
+
+        if lora.scaling_factor is None:
+            return
+
+        if (lora.scaling_factor not in self.scaling_factor_to_offset):
+            raise ValueError(f"Long LoRA scaling factor {lora.scaling_factor}"
+                             " has not been initialized.")
+
+        offsets = self.scaling_factor_to_offset.get(lora.scaling_factor)
+        if offsets:
+            self.long_lora_context.offsets_by_lora_id[lora.id] = offsets
+
+    def _add_adapter(self, lora: LoRAModel):
+        self._create_merged_loras_inplace(lora)
+        self._registered_adapters[lora.id] = lora
+        self._set_long_lora_context(lora)
+
+    def pin_adapter(self, lora_id: int) -> bool:
+        """Pin a LoRAModel in the manager cache."""
+        raise NotImplementedError(
+            "Pinning is not supported in LoRAModelManager."
+            "Use LRUCacheLoRAModelManager for pinning")  # type: ignore
+
+    def _set_adapter_mapping(self, mapping: LoRAMapping) -> None:
+        # update lora states
+        self.punica_wrapper.update_metadata(
+            mapping,
+            self.lora_index_to_id,
+            self.lora_slots + 1,
+            self.vocab_size,
+            self.lora_config.lora_extra_vocab_size,
+            self.long_lora_context,
+        )
+
+    def remove_all_adapters(self):
+        """Remove all LoRAModels from the manager."""
+        self._registered_adapters.clear()
+        self.lora_index_to_id = [None] * self.lora_slots
+        self._active_adapters.clear()
+
+    def _create_lora_modules(self):
+        for module_name, module in self.model.named_modules(
+                remove_duplicate=False):
+            if isinstance(module, PPMissingLayer):
+                continue
+            if not self._match_target_modules(module_name):
+                continue
+            # A temporary approach for multimodal models to support LoRA
+            # TODO: Remove this restriction
+            if self._filter_unsupported_mm_module(module_name):
+                logger.warning(
+                    "Regarding multimodal models, vLLM currently only supports "
+                    "adding LoRA to language model, %s will be ignored.",
+                    module_name,
+                )
+                continue
+            parts = module_name.split(".")[-1]
+            packed_moduled_lst = self.packed_modules_mapping.get(parts, [])
+            new_module = replace_submodule(
+                self.model, module_name,
+                from_layer(module, self.lora_slots, self.lora_config,
+                           packed_moduled_lst, self.model.config))
+
+            # LinearScalingRotaryEmbeddingWithLora is used to handle
+            # long context lora. Register relevant metadata.
+            if isinstance(new_module, LinearScalingRotaryEmbeddingWithLora):
+                self.long_lora_context = LongContextLoRAContext(
+                    new_module.scaling_factors, new_module.rotary_dim)
+                self.scaling_factor_to_offset = \
+                    new_module.scaling_factor_to_offset
+            # (yard1): TODO make this more robust
+            if "lm_head" in module_name:
+                logits_processor_module = self.model.get_submodule(
+                    "logits_processor")
+                new_module = replace_submodule(
+                    self.model, "logits_processor",
+                    from_layer_logits_processor(logits_processor_module,
+                                                module, self.lora_slots,
+                                                self.lora_config,
+                                                self.model.config))
+
+            # In some models, especially multimodal ones, layers with the same
+            # name may have different types, such as nn.Linear and
+            # ReplicatedLinear. The nn.Linear layers cannot be replaced with
+            # LoRA layers, leading to assertion error. The following check
+            # aims to prevent this error
+            if self.supports_mm and not isinstance(new_module,
+                                                   BaseLayerWithLoRA):
+                continue
+            self.register_module(module_name, new_module)
+            self._register_packed_modules(module_name)
+            # All lora layers share the same punica_wrapper based on reference.
+            new_module.set_mapping(self.punica_wrapper)
+
+    def register_module(self, module_name: str, module: "BaseLayerWithLoRA"):
+        assert isinstance(module, BaseLayerWithLoRA)
+        self.modules[module_name] = module
+
+    def create_dummy_lora(
+            self,
+            lora_id: int,
+            rank: int,
+            scaling_factor: Optional[float],
+            embedding_modules: Optional[Dict[str, str]] = None) -> LoRAModel:
+        """Create zero-initialized LoRAModel for warmup."""
+        model = LoRAModel(lora_id, rank, {}, scaling_factor)
+        for module_name, module in self.model.named_modules():
+            bias_enabled = self.lora_config.bias_enabled
+            if (not self._match_target_modules(module_name)
+                    or not isinstance(module, BaseLayerWithLoRA)
+                    or isinstance(module, LinearScalingRotaryEmbeddingWithLora)
+                    or self._filter_unsupported_mm_module(module_name)):
+                continue
+            parts = module_name.split(".")
+            if module_name not in self.packed_modules:
+                assert embedding_modules is not None
+                if parts[-1] in embedding_modules:
+                    input_dim = (module.base_layer.org_vocab_size +
+                                 self.lora_config.lora_extra_vocab_size if
+                                 hasattr(module.base_layer, "org_vocab_size")
+                                 else module.base_layer.weight.shape[1])
+                    output_dim = module.base_layer.embedding_dim if hasattr(
+                        module.base_layer,
+                        "embedding_dim") else module.base_layer.weight.shape[0]
+                    embeddings_tensor_dim = (module.base_layer.embedding_dim if
+                                             hasattr(module.base_layer,
+                                                     "embedding_dim") else
+                                             module.base_layer.weight.shape[1])
+                    lora = LoRALayerWeights.create_dummy_lora_weights(
+                        module_name,
+                        input_dim,
+                        output_dim,
+                        rank,
+                        module.lora_a_stacked[0].dtype,
+                        "cpu",
+                        embeddings_tensor_dim=embeddings_tensor_dim,
+                        bias_enabled=bias_enabled)
+                else:
+                    lora = LoRALayerWeights.create_dummy_lora_weights(
+                        module_name,
+                        module.lora_a_stacked[0].shape[-1],
+                        module.lora_b_stacked[0].shape[-2],
+                        rank,
+                        module.lora_a_stacked[0].dtype,
+                        "cpu",
+                        bias_enabled=bias_enabled,
+                    )
+                lora.optimize()
+            else:
+                parts = module_name.split(".")
+                replacements = self.packed_modules_mapping[parts[-1]]
+                subloras: List[Optional[LoRALayerWeights]] = []
+                for i, r in enumerate(replacements):
+                    lora = LoRALayerWeights.create_dummy_lora_weights(
+                        module_name + "." + r,
+                        module.lora_a_stacked[i].shape[-1],
+                        module.lora_b_stacked[i].shape[-2],
+                        rank,
+                        module.lora_a_stacked[i].dtype,
+                        "cpu",
+                        bias_enabled=bias_enabled,
+                    )
+                    lora.optimize()
+                    subloras.append(lora)
+                lora = PackedLoRALayerWeights.pack(subloras)
+            model.loras[module_name] = lora
+        return model
+
+    def _match_target_modules(self, module_name: str):
+        return any(
+            re.match(
+                r".*\.{target_module}$".format(target_module=target_module),
+                module_name) or target_module == module_name
+            for target_module in self.supported_lora_modules)
+
+    def _filter_unsupported_mm_module(self, module_name: str) -> bool:
+        """
+        Regarding multimodal models, vLLM currently only supports adding LoRA to
+        language model. LoRA for other modules, such as the vision tower, will 
+        be filtered out.
+        """
+        if self.supports_mm:
+            module_mapping: MultiModelKeys = self.model.get_mm_mapping()
+            prefix_lst = module_mapping.connector + module_mapping.tower_model
+            return any(
+                [module_name.startswith(prefix) for prefix in prefix_lst])
+        return False
+
+    def _register_packed_modules(self, module_full_name: str) -> None:
+        parts = module_full_name.split(".")
+        module_name = parts[-1]
+        replacements = self.packed_modules_mapping.get(module_name, [])
+        # When replacements is less than or equal to 1, it indicates that this
+        # module is not a packed module.
+        if len(replacements) <= 1:
+            return
+        prefix = ".".join(parts[:-1])
+        self.packed_modules[module_full_name] = [
+            prefix + "." + r if prefix else r for r in replacements
+        ]
+
+    def _create_merged_loras_inplace(self, lora_model: LoRAModel) -> None:
+        for module_name, new_module_names in self.packed_modules.items():
+            replacement_loras: List[Optional[LoRALayerWeights]] = []
+            has_replacement = False
+            for r in new_module_names:
+                lora = lora_model.get_lora(r)
+                replacement_loras.append(lora)
+                if lora:
+                    has_replacement = True
+            if not has_replacement:
+                continue
+            for i in range(len(replacement_loras)):
+                if replacement_loras[i]:
+                    continue
+                replacement_loras[i] = None
+            lora_model.loras[module_name] = PackedLoRALayerWeights.pack(
+                replacement_loras)
+
+    def deactivate_adapter(self, adapter_id: int) -> bool:
+        return deactivate_adapter(adapter_id, self._active_adapters,
+                                  self._deactivate_adapter)
+
+    def add_adapter(self, adapter: LoRAModel) -> bool:
+        logger.debug(
+            "Adding lora. Model id: %d, "
+            "int id: %d, "
+            "scaling factor: %s", adapter.id, adapter.id,
+            adapter.scaling_factor)
+        return add_adapter(adapter, self._registered_adapters, self.capacity,
+                           self._add_adapter)
+
+    def set_adapter_mapping(self, mapping: LoRAMapping) -> None:
+        self._last_mapping = set_adapter_mapping(mapping, self._last_mapping,
+                                                 self._set_adapter_mapping)
+
+    def remove_adapter(self, adapter_id: int) -> bool:
+        return remove_adapter(adapter_id, self._registered_adapters,
+                              self.deactivate_adapter)
+
+    def list_adapters(self) -> Dict[int, Any]:
+        return list_adapters(self._registered_adapters)
+
+    def get_adapter(self, adapter_id: int) -> Optional[Any]:
+        return get_adapter(adapter_id, self._registered_adapters)
+
+
+class LoRALRUCache(AdapterLRUCache[LoRAModel]):
+
+    def __init__(self, capacity: int, deactivate_lora_fn: Callable[[int],
+                                                                   bool]):
+        super().__init__(capacity, deactivate_lora_fn)
+
+
+class LRUCacheLoRAModelManager(LoRAModelManager):
+    """A model manager that manages multiple LoRAs with LRU cache."""
+
+    def __init__(self, model: nn.Module, max_num_seqs: int,
+                 max_num_batched_tokens: int, vocab_size: int,
+                 lora_config: LoRAConfig, device: torch.device):
+        super().__init__(model, max_num_seqs, max_num_batched_tokens,
+                         vocab_size, lora_config, device)
+        self._registered_adapters: LoRALRUCache = LoRALRUCache(
+            self.capacity, self.deactivate_adapter)
+        self._active_adapters: LoRALRUCache = LoRALRUCache(
+            self.lora_slots, self._deactivate_adapter)
+
+    def list_adapters(self) -> Dict[int, LoRAModel]:
+        """List all registered LoRAModels."""
+        return dict(self._registered_adapters.cache)
+
+    def add_adapter(self, lora: LoRAModel) -> bool:
+        """Add a LoRAModel to the manager."""
+        logger.debug(
+            "Adding lora. Model id: %d, "
+            "int id: %d, "
+            "scaling factor: %s", lora.id, lora.id, lora.scaling_factor)
+        if lora.id not in self._registered_adapters:
+            self._add_adapter(lora)
+            was_added = True
+        else:
+            # We always touch to update the LRU cache order
+            self._registered_adapters.touch(lora.id)
+            was_added = False
+        return was_added
+
+    def activate_adapter(
+        self,
+        lora_id: int,
+    ) -> bool:
+        if lora_id not in self._active_adapters and len(
+                self._active_adapters) >= self.lora_slots:
+            self._active_adapters.remove_oldest()
+        result = super().activate_adapter(lora_id)
+        # We always touch to update the LRU cache order
+        self._active_adapters.touch(lora_id)
+        return result
+
+    def remove_oldest_adapter(self) -> bool:
+        if len(self._registered_adapters) > 0:
+            self._registered_adapters.remove_oldest()
+            return True
+        return False
+
+    def pin_adapter(self, lora_id: int) -> bool:
+        """Pin a LoRAModel in the manager cache."""
+        self._pin_lora_in_cpu_cache(lora_id)
+        self._pin_lora_in_gpu_cache(lora_id)
+        return True
+
+    def _pin_lora_in_cpu_cache(self, lora_id: int):
+        try:
+            self._registered_adapters.pin(lora_id)
+        except ValueError as err:
+            raise ValueError("Pinning failed. "
+                             f"LoRA {lora_id} is not registered.") from err
+
+    def _pin_lora_in_gpu_cache(self, lora_id: int):
+        if lora_id not in self._active_adapters:
+            # move lora to gpu if not already active
+            self.activate_adapter(lora_id)
+
+        self._active_adapters.pin(lora_id)
+
+
+def create_lora_manager(
+        model: nn.Module,
+        max_num_seqs: int,
+        max_num_batched_tokens: int,
+        vocab_size: int,
+        lora_config: LoRAConfig,
+        device: torch.device,
+        lora_manager_cls: Type[LoRAModelManager] = LoRAModelManager,
+        **kwargs) -> LoRAModelManager:
+    """Create a LoRA adapter for a given model."""
+    if not hasattr(model, "supported_lora_modules"):
+        raise ValueError(f"Model {type(model)} is not supported for LoRA.")
+    lora_manager = lora_manager_cls(
+        model=model,
+        max_num_seqs=max_num_seqs,
+        max_num_batched_tokens=max_num_batched_tokens,
+        vocab_size=vocab_size,
+        lora_config=lora_config,
+        device=device,
+        **kwargs)
+    return lora_manager

.venv/lib/python3.11/site-packages/vllm/lora/ops/torch_ops/__init__.py

@@ -0,0 +1,15 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from vllm.lora.ops.torch_ops.lora_ops import bgmv_expand  # noqa: F401
+from vllm.lora.ops.torch_ops.lora_ops import (bgmv_expand_slice, bgmv_shrink,
+                                              sgmv_expand, sgmv_expand_slice,
+                                              sgmv_shrink)
+
+__all__ = [
+    "bgmv_expand",
+    "bgmv_expand_slice",
+    "bgmv_shrink",
+    "sgmv_expand",
+    "sgmv_expand_slice",
+    "sgmv_shrink",
+]

.venv/lib/python3.11/site-packages/vllm/lora/ops/torch_ops/lora_ops.py

@@ -0,0 +1,115 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+
+
+def sgmv_expand(inputs: torch.Tensor,
+                lora_b_weights: torch.Tensor,
+                output_tensor: torch.Tensor,
+                b_seq_start_loc: torch.Tensor,
+                seq_len_tensor: torch.Tensor,
+                lora_indices_tensor: torch.Tensor,
+                batches: int,
+                max_seq_length: int,
+                token_nums: int,
+                add_inputs: bool = False):
+    exploded_indices = torch.repeat_interleave(lora_indices_tensor,
+                                               seq_len_tensor)
+
+    bgmv_expand(inputs, lora_b_weights, output_tensor, exploded_indices,
+                add_inputs)
+
+
+def bgmv_expand(inputs: torch.Tensor,
+                lora_b_weights: torch.Tensor,
+                output_tensor: torch.Tensor,
+                lora_indices_tensor: torch.Tensor,
+                add_inputs: bool = True):
+    selected_loras = lora_b_weights[lora_indices_tensor].to(
+        dtype=output_tensor.dtype)
+    if len(selected_loras.shape) == 4:
+        selected_loras = selected_loras.squeeze(dim=1)
+    inputs = inputs.to(dtype=output_tensor.dtype)
+    outputs = torch.einsum("bi, boi -> bo", inputs, selected_loras)
+
+    limit = output_tensor.shape[0]
+    if outputs.shape[0] == 1 and output_tensor.shape[0] != 1:
+        limit = 1
+
+    if add_inputs:
+        output_tensor[:, :outputs.shape[1]] += outputs[:limit, :]
+    else:
+        output_tensor[:, :outputs.shape[1]] = outputs[:limit, :]
+
+
+def sgmv_shrink(
+    inputs: torch.Tensor,
+    lora_a_weights: torch.Tensor,
+    output_tensor: torch.Tensor,
+    b_seq_start_loc: torch.Tensor,
+    seq_len_tensor: torch.Tensor,
+    lora_indices_tensor: torch.Tensor,
+    batches: int,
+    max_seq_length: int,
+    token_nums: int,
+    scaling: float,
+):
+    exploded_indices = torch.repeat_interleave(lora_indices_tensor,
+                                               seq_len_tensor)
+
+    bgmv_shrink(inputs, lora_a_weights, output_tensor, exploded_indices,
+                scaling)
+
+
+def bgmv_shrink(inputs: torch.Tensor,
+                lora_b_weights: torch.Tensor,
+                output_tensor: torch.Tensor,
+                lora_indices_tensor: torch.Tensor,
+                scaling: float = 1.0):
+    selected_loras = lora_b_weights[lora_indices_tensor].to(
+        dtype=output_tensor.dtype)
+    if len(selected_loras.shape) == 4:
+        selected_loras = selected_loras.squeeze(dim=1)
+    inputs = inputs.to(dtype=output_tensor.dtype)
+    outputs = torch.einsum("bi, boi -> bo", inputs, selected_loras)
+
+    output_tensor[:, :outputs.shape[1]] = scaling * outputs[:]
+
+
+def sgmv_expand_slice(inputs: torch.Tensor,
+                      lora_b_weights: torch.Tensor,
+                      output_tensor: torch.Tensor,
+                      b_seq_start_loc: torch.Tensor,
+                      seq_len_tensor: torch.Tensor,
+                      lora_indices_tensor: torch.Tensor,
+                      batches: int,
+                      max_seq_length: int,
+                      token_nums: int,
+                      slice_offset: int,
+                      slice_size: int,
+                      add_inputs: bool = False):
+    exploded_indices = torch.repeat_interleave(lora_indices_tensor,
+                                               seq_len_tensor)
+
+    bgmv_expand_slice(inputs, lora_b_weights, output_tensor, exploded_indices,
+                      slice_offset, slice_size, add_inputs)
+
+
+def bgmv_expand_slice(inputs: torch.Tensor,
+                      lora_b_weights: torch.Tensor,
+                      output_tensor: torch.Tensor,
+                      lora_indices_tensor: torch.Tensor,
+                      slice_offset: int,
+                      slice_size: int,
+                      add_inputs: bool = True):
+    selected_loras = lora_b_weights[lora_indices_tensor].to(
+        dtype=output_tensor.dtype)
+    inputs = inputs.to(dtype=output_tensor.dtype)
+    if len(selected_loras.shape) == 4:
+        selected_loras = selected_loras.squeeze(dim=1)
+    outputs = torch.einsum("bi, boi -> bo", inputs, selected_loras)
+
+    if add_inputs:
+        output_tensor[:, slice_offset:slice_offset + slice_size] += outputs[:]
+    else:
+        output_tensor[:, slice_offset:slice_offset + slice_size] = outputs[:]

.venv/lib/python3.11/site-packages/vllm/lora/ops/triton_ops/__init__.py

@@ -0,0 +1,15 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from vllm.lora.ops.triton_ops.bgmv_expand import bgmv_expand
+from vllm.lora.ops.triton_ops.bgmv_expand_slice import bgmv_expand_slice
+from vllm.lora.ops.triton_ops.bgmv_shrink import bgmv_shrink
+from vllm.lora.ops.triton_ops.sgmv_expand import sgmv_expand
+from vllm.lora.ops.triton_ops.sgmv_shrink import sgmv_shrink  # noqa: F401
+
+__all__ = [
+    "bgmv_expand",
+    "bgmv_expand_slice",
+    "bgmv_shrink",
+    "sgmv_expand",
+    "sgmv_shrink",
+]