Merge https://github.com/abetlen/llama-cpp-python

* implement min_tokens

* set default to 0

* pass min_tokens

* fix

* remove copy

* implement MinTokensLogitsProcessor

* format

* fix condition

.github/dependabot.yml

@@ -8,8 +8,12 @@ updates:
   - package-ecosystem: "pip" # See documentation for possible values
     directory: "/" # Location of package manifests
     schedule:
-      interval: "weekly"
+      interval: "daily"
   - package-ecosystem: "github-actions"
     directory: "/"
     schedule:
-      interval: "weekly"    
+      interval: "daily"
+  - package-ecosystem: "docker"
+    directory: "/"
+    schedule:
+      interval: "daily"   

.github/workflows/build-and-release.yaml

@@ -29,7 +29,7 @@ jobs:
           python -m pip install -e .[all]
 
       - name: Build wheels
-        uses: pypa/cibuildwheel@v2.17.0
+        uses: pypa/cibuildwheel@v2.18.0
         env:
           # disable repair
           CIBW_REPAIR_WHEEL_COMMAND: ""
@@ -56,7 +56,7 @@ jobs:
           platforms: linux/arm64
 
       - name: Build wheels
-        uses: pypa/cibuildwheel@v2.17.0
+        uses: pypa/cibuildwheel@v2.18.0
         env:
           CIBW_SKIP: "*musllinux* pp*"
           CIBW_REPAIR_WHEEL_COMMAND: ""

CHANGELOG.md

@@ -7,9 +7,61 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.2.75]
+
+- feat: Update llama.cpp to ggerganov/llama.cpp@13ad16af1231ab2d245d35df3295bcfa23de1305
+- fix: segfault for models without eos / bos tokens by @abetlen in d99a6ba607a4885fb00e63e967964aa41bdbbbcb
+- feat: add MinTokensLogitProcessor and min_tokens argument to server by @twaka in #1333
+- misc: Remove unnecessary metadata lookups by @CISC in #1448
+
+## [0.2.74]
+
+- feat: Update llama.cpp to ggerganov/llama.cpp@b228aba91ac2cd9eb90e9d423ba1d0d20e0117e2
+- fix: Enable CUDA backend for llava by @abetlen in 7f59856fa6f3e23f07e12fc15aeb9359dc6c3bb4
+- docs: Fix typo in README.md by @yupbank in #1444
+
+## [0.2.73]
+
+- feat: Update llama.cpp to ggerganov/llama.cpp@25c6e82e7a1ad25a42b0894e87d9b5c557409516
+- fix: Clear kv cache at beginning of image chat formats to avoid bug when image is evaluated first by @abetlen in ac55d0a175115d1e719672ce1cb1bec776c738b1
+
+## [0.2.72]
+
+- fix(security): Remote Code Execution by Server-Side Template Injection in Model Metadata by @retr0reg in b454f40a9a1787b2b5659cd2cb00819d983185df
+- fix(security): Update remaining jinja chat templates to use immutable sandbox by @CISC in #1441
+
+## [0.2.71]
+
+- feat: Update llama.cpp to ggerganov/llama.cpp@911b3900dded9a1cfe0f0e41b82c7a29baf3a217
+- fix: Make leading bos_token optional for image chat formats, fix nanollava system message by @abetlen in 77122638b4153e31d9f277b3d905c2900b536632
+- fix: free last image embed in llava chat handler by @abetlen in 3757328b703b2cd32dcbd5853271e3a8c8599fe7
+
+## [0.2.70]
+
+- feat: Update llama.cpp to ggerganov/llama.cpp@c0e6fbf8c380718102bd25fcb8d2e55f8f9480d1
+- feat: fill-in-middle support by @CISC in #1386
+- fix: adding missing args in create_completion for functionary chat handler by @skalade in #1430
+- docs: update README.md @eltociear in #1432
+- fix: chat_format log where auto-detected format prints None by @balvisio in #1434
+- feat(server): Add support for setting root_path by @abetlen in 0318702cdc860999ee70f277425edbbfe0e60419
+- feat(ci): Add docker checks and check deps more frequently by @Smartappli in #1426
+- fix: detokenization case where first token does not start with a leading space by @noamgat in #1375
+- feat: Implement streaming for Functionary v2 + Bug fixes by @jeffrey-fong in #1419
+- fix: Use memmove to copy str_value kv_override by @abetlen in 9f7a85571ae80d3b6ddbd3e1bae407b9f1e3448a
+- feat(server): Remove temperature bounds checks for server by @abetlen in 0a454bebe67d12a446981eb16028c168ca5faa81
+- fix(server): Propagate flash_attn to model load by @dthuerck in #1424
+
+## [0.2.69]
+
+- feat: Update llama.cpp to ggerganov/llama.cpp@6ecf3189e00a1e8e737a78b6d10e1d7006e050a2
+- feat: Add llama-3-vision-alpha chat format by @abetlen in 31b1d95a6c19f5b615a3286069f181a415f872e8
+- fix: Change default verbose value of verbose in image chat format handlers to True to match Llama by @abetlen in 4f01c452b6c738dc56eacac3758119b12c57ea94
+- fix: Suppress all logs when verbose=False, use hardcoded fileno's to work in colab notebooks by @abetlen in f116175a5a7c84569c88cad231855c1e6e59ff6e
+- fix: UTF-8 handling with grammars by @jsoma in #1415
+
 ## [0.2.68]
 
-- feat: Update llama.cpp to ggerganov/llama.cpp@
+- feat: Update llama.cpp to ggerganov/llama.cpp@77e15bec6217a39be59b9cc83d6b9afb6b0d8167
 - feat: Add option to enable flash_attn to Lllama params and ModelSettings by @abetlen in 22d77eefd2edaf0148f53374d0cac74d0e25d06e
 - fix(ci): Fix build-and-release.yaml by @Smartappli in #1413
 

CMakeLists.txt

@@ -51,8 +51,9 @@ if (LLAMA_BUILD)
     )
 
     if (LLAVA_BUILD)
-        if (LLAMA_CUBLAS)
+        if (LLAMA_CUBLAS OR LLAMA_CUDA)
             add_compile_definitions(GGML_USE_CUBLAS)
+            add_compile_definitions(GGML_USE_CUDA)
         endif()
 
         if (LLAMA_METAL)

Makefile

@@ -16,7 +16,7 @@ build.debug:
 	CMAKE_ARGS="-DCMAKE_BUILD_TYPE=Debug" python3 -m pip install --verbose --config-settings=cmake.verbose=true --config-settings=logging.level=INFO --config-settings=install.strip=false  --editable .
 
 build.cuda:
-	CMAKE_ARGS="-DLLAMA_CUBLAS=on" python3 -m pip install --verbose -e .
+	CMAKE_ARGS="-DLLAMA_CUDA=on" python3 -m pip install --verbose -e .
 
 build.opencl:
 	CMAKE_ARGS="-DLLAMA_CLBLAST=on" python3 -m pip install --verbose -e .

README.md

@@ -516,7 +516,7 @@ chat_handler = Llava15ChatHandler(clip_model_path="path/to/llava/mmproj.bin")
 llm = Llama(
   model_path="./path/to/llava/llama-model.gguf",
   chat_handler=chat_handler,
-  n_ctx=2048, # n_ctx should be increased to accomodate the image embedding
+  n_ctx=2048, # n_ctx should be increased to accommodate the image embedding
 )
 llm.create_chat_completion(
     messages = [
@@ -547,10 +547,10 @@ llm = Llama.from_pretrained(
   repo_id="vikhyatk/moondream2",
   filename="*text-model*",
   chat_handler=chat_handler,
-  n_ctx=2048, # n_ctx should be increased to accomodate the image embedding
+  n_ctx=2048, # n_ctx should be increased to accommodate the image embedding
 )
 
-respoonse = llm.create_chat_completion(
+response = llm.create_chat_completion(
     messages = [
         {
             "role": "user",

examples/ray/README.md

@@ -0,0 +1,19 @@
+This is an example of doing LLM inference with [Ray](https://docs.ray.io/en/latest/index.html) and [Ray Serve](https://docs.ray.io/en/latest/serve/index.html).
+
+First, install the requirements:
+
+```bash
+$ pip install -r requirements.txt
+```
+
+Deploy a GGUF model to Ray Serve with the following command:
+
+```bash
+$ serve run llm:llm_builder model_path='../models/mistral-7b-instruct-v0.2.Q4_K_M.gguf'
+```
+
+This will start an API endpoint at `http://localhost:8000/`. You can query the model like this:
+
+```bash
+$ curl -k -d '{"prompt": "tell me a joke", "max_tokens": 128}' -X POST http://localhost:8000
+```

examples/ray/llm.py

@@ -0,0 +1,20 @@
+from starlette.requests import Request
+from typing import Dict
+from ray import serve
+from ray.serve import Application
+from llama_cpp import Llama
+
+@serve.deployment
+class LlamaDeployment:
+    def __init__(self, model_path: str):
+        self._llm = Llama(model_path=model_path)
+
+    async def __call__(self, http_request: Request) -> Dict:
+        input_json = await http_request.json()
+        prompt = input_json["prompt"]
+        max_tokens = input_json.get("max_tokens", 64)
+        return self._llm(prompt, max_tokens=max_tokens)
+
+
+def llm_builder(args: Dict[str, str]) -> Application:
+    return LlamaDeployment.bind(args["model_path"])

examples/ray/requirements.txt

@@ -0,0 +1,3 @@
+ray[serve]
+--extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu
+llama-cpp-python

llama_cpp/__init__.py

@@ -1,4 +1,4 @@
 from .llama_cpp import *
 from .llama import *
 
-__version__ = "0.2.68"
+__version__ = "0.2.75"

llama_cpp/_internals.py

@@ -203,7 +203,7 @@ class _LlamaModel:
         # NOTE: Llama1 models automatically added a space at the start of the prompt
         # this line removes a leading space if the first token is a beginning of sentence token
         return (
-            output[1:] if len(tokens) > 0 and tokens[0] == self.token_bos() else output
+            output[1:] if len(tokens) > 0 and tokens[0] == self.token_bos() and output[0:1] == b' ' else output
         )
 
     # Extra

llama_cpp/llama.py

@@ -262,7 +262,12 @@ class Llama:
                         raise ValueError(f"Value for {k} is too long: {v}")
                     v_bytes = v_bytes.ljust(128, b"\0")
                     self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_TYPE_STR
-                    self._kv_overrides_array[i].value.str_value[:128] = v_bytes
+                    # copy min(v_bytes, 128) to str_value
+                    ctypes.memmove(
+                        self._kv_overrides_array[i].value.str_value,
+                        v_bytes,
+                        min(len(v_bytes), 128),
+                    )
                 else:
                     raise ValueError(f"Unknown value type for {k}: {v}")
 
@@ -373,6 +378,7 @@ class Llama:
 
         self.chat_format = chat_format
         self.chat_handler = chat_handler
+        self._chat_handlers: Dict[str, llama_chat_format.LlamaChatCompletionHandler] = {}
 
         self.draft_model = draft_model
 
@@ -404,10 +410,33 @@ class Llama:
         if self.verbose:
             print(f"Model metadata: {self.metadata}", file=sys.stderr)
 
+        eos_token_id = self.token_eos()
+        bos_token_id = self.token_bos()
+
+        eos_token = self._model.token_get_text(eos_token_id) if eos_token_id != -1 else ""
+        bos_token = self._model.token_get_text(bos_token_id) if bos_token_id != -1 else ""
+
+        # Unfortunately the llama.cpp API does not return metadata arrays, so we can't get template names from tokenizer.chat_templates
+        template_choices = dict((name[10:], template) for name, template in self.metadata.items() if name.startswith("tokenizer.chat_template."))
+
+        if "tokenizer.chat_template" in self.metadata:
+            template_choices["chat_template.default"] = self.metadata["tokenizer.chat_template"]
+
+        if self.verbose and template_choices:
+            print(f"Available chat formats from metadata: {', '.join(template_choices.keys())}", file=sys.stderr)
+
+        for name, template in template_choices.items():
+            self._chat_handlers[name] = llama_chat_format.Jinja2ChatFormatter(
+                template=template,
+                eos_token=eos_token,
+                bos_token=bos_token,
+                stop_token_ids=[eos_token_id],
+            ).to_chat_handler()
+
         if (
             self.chat_format is None
             and self.chat_handler is None
-            and "tokenizer.chat_template" in self.metadata
+            and "chat_template.default" in template_choices
         ):
             chat_format = llama_chat_format.guess_chat_format_from_gguf_metadata(
                 self.metadata
@@ -418,35 +447,17 @@ class Llama:
                 if self.verbose:
                     print(f"Guessed chat format: {chat_format}", file=sys.stderr)
             else:
-                template = self.metadata["tokenizer.chat_template"]
-                try:
-                    eos_token_id = int(self.metadata["tokenizer.ggml.eos_token_id"])
-                except:
-                    eos_token_id = self.token_eos()
-                try:
-                    bos_token_id = int(self.metadata["tokenizer.ggml.bos_token_id"])
-                except:
-                    bos_token_id = self.token_bos()
-
-                eos_token = self._model.token_get_text(eos_token_id)
-                bos_token = self._model.token_get_text(bos_token_id)
-
                 if self.verbose:
-                    print(f"Using gguf chat template: {template}", file=sys.stderr)
+                    print(f"Using gguf chat template: {template_choices['chat_template.default']}", file=sys.stderr)
                     print(f"Using chat eos_token: {eos_token}", file=sys.stderr)
                     print(f"Using chat bos_token: {bos_token}", file=sys.stderr)
 
-                self.chat_handler = llama_chat_format.Jinja2ChatFormatter(
+                self.chat_format = "chat_template.default"
-                    template=template,
-                    eos_token=eos_token,
-                    bos_token=bos_token,
-                    stop_token_ids=[eos_token_id],
-                ).to_chat_handler()
 
         if self.chat_format is None and self.chat_handler is None:
             self.chat_format = "llama-2"
             if self.verbose:
-                print(f"Using fallback chat format: {chat_format}", file=sys.stderr)
+                print(f"Using fallback chat format: {self.chat_format}", file=sys.stderr)
 
     @property
     def ctx(self) -> llama_cpp.llama_context_p:
@@ -950,19 +961,54 @@ class Llama:
 
         completion_id: str = f"cmpl-{str(uuid.uuid4())}"
         created: int = int(time.time())
+        prefix_token_id: int = self._model.token_prefix()
+        middle_token_id: int = self._model.token_middle()
+        suffix_token_id: int = self._model.token_suffix()
         # If prompt is empty, initialize completion with BOS token to avoid
         # detokenization including a space at the beginning of the completion
         completion_tokens: List[int] = [] if len(prompt) > 0 else [self.token_bos()]
         # Add blank space to start of prompt to match OG llama tokenizer
         prompt_tokens: List[int] = (
             (
-                self.tokenize(prompt.encode("utf-8"), special=True)
+                [prefix_token_id]
+                if prefix_token_id >= 0 and suffix is not None
+                else []
+            )
+            +
+            (
+                (
+                    self.tokenize(prompt.encode("utf-8"), add_bos=(prefix_token_id < 0 or suffix is None), special=(prefix_token_id < 0 or suffix is None))
                     if prompt != ""
+                    else (
+                        []
+                        if prefix_token_id >= 0 and suffix is not None
                         else [self.token_bos()]
                     )
+                )
                 if isinstance(prompt, str)
                 else prompt
             )
+            +
+            (
+                (
+                    [suffix_token_id]
+                    +
+                    (
+                        self.tokenize(suffix.encode("utf-8"), add_bos=False, special=False)
+                        if suffix
+                        else []
+                    )
+                )
+                if suffix_token_id >= 0 and suffix is not None
+                else []
+            )
+            +
+            (
+                [middle_token_id]
+                if middle_token_id >= 0 and suffix is not None
+                else []
+            )
+        )
         text: bytes = b""
         returned_tokens: int = 0
         stop = (
@@ -1341,7 +1387,7 @@ class Llama:
         if echo:
             text_str = prompt + text_str
 
-        if suffix is not None:
+        if suffix_token_id < 0 and suffix is not None:
             text_str = text_str + suffix
 
         logprobs_or_none: Optional[CompletionLogprobs] = None
@@ -1679,7 +1725,7 @@ class Llama:
         Returns:
             Generated chat completion or a stream of chat completion chunks.
         """
-        handler = self.chat_handler or llama_chat_format.get_chat_completion_handler(
+        handler = self.chat_handler or self._chat_handlers.get(self.chat_format) or llama_chat_format.get_chat_completion_handler(
             self.chat_format
         )
         return handler(
@@ -2038,3 +2084,19 @@ class StoppingCriteriaList(List[StoppingCriteria]):
         self, input_ids: npt.NDArray[np.intc], logits: npt.NDArray[np.single]
     ) -> bool:
         return any([stopping_criteria(input_ids, logits) for stopping_criteria in self])
+
+
+class MinTokensLogitsProcessor(LogitsProcessor):
+    def __init__(self, min_tokens: int, token_eos: int):
+        self.min_tokens = min_tokens
+        self.token_eos = token_eos
+        self.prompt_tokens = None
+
+    def __call__(
+        self, input_ids: npt.NDArray[np.intc], scores: npt.NDArray[np.single]
+    ) -> npt.NDArray[np.single]:
+        if self.prompt_tokens is None:
+            self.prompt_tokens = len(input_ids)
+        if len(input_ids) - self.prompt_tokens < self.min_tokens:
+            scores[self.token_eos] = -np.inf
+        return scores

llama_cpp/llama_chat_format.py

@@ -11,6 +11,7 @@ from contextlib import ExitStack
 from typing import Any, Dict, Iterator, List, Literal, Optional, Tuple, Union, Protocol, cast
 
 import jinja2
+from jinja2.sandbox import ImmutableSandboxedEnvironment
 
 import numpy as np
 import numpy.typing as npt
@@ -191,7 +192,7 @@ class Jinja2ChatFormatter(ChatFormatter):
         self.add_generation_prompt = add_generation_prompt
         self.stop_token_ids = set(stop_token_ids) if stop_token_ids is not None else None
 
-        self._environment = jinja2.Environment(
+        self._environment = ImmutableSandboxedEnvironment(
             loader=jinja2.BaseLoader(),
             trim_blocks=True,
             lstrip_blocks=True,
@@ -684,8 +685,7 @@ def hf_tokenizer_config_to_chat_formatter(
     assert isinstance(tokenizer_config["eos_token"], str)
     eos_token = tokenizer_config["eos_token"]
 
-    env = jinja2.Environment(
+    env = ImmutableSandboxedEnvironment(
-        loader=jinja2.BaseLoader(),
         trim_blocks=True,
         lstrip_blocks=True,
     ).from_string(chat_template)
@@ -1894,6 +1894,8 @@ def functionary_v1_v2_chat_handler(
         function_call = (
             tool_choice if isinstance(tool_choice, str) else tool_choice["function"]
         )
+    elif function_call is not None:
+        pass
     else:
         function_call = "auto"
 
@@ -1930,11 +1932,10 @@ def functionary_v1_v2_chat_handler(
             logits_processor=logits_processor,
             grammar=grammar,
         )
+        if stream is False:
             completion_or_completion_chunks["choices"][0]["text"] = completion_or_completion_chunks["choices"][0]["text"].lstrip()
         return _convert_completion_to_chat(completion_or_completion_chunks, stream=stream)  # type: ignore
 
-    assert stream is False  # TODO: support stream mode
-
     def get_grammar(function_call):
         function_body = None
         for function in functions or []:
@@ -1968,7 +1969,7 @@ def functionary_v1_v2_chat_handler(
 
         return grammar
 
-    def create_completion(stop):
+    def create_completion(prompt, stop, grammar):
         completion = cast(llama_types.Completion, llama.create_completion(
             prompt=prompt,
             temperature=temperature,
@@ -1976,7 +1977,7 @@ def functionary_v1_v2_chat_handler(
             top_k=top_k,
             min_p=min_p,
             typical_p=typical_p,
-            stream=False,
+            stream=stream,
             stop=stop,
             max_tokens=max_tokens,
             presence_penalty=presence_penalty,
@@ -1997,6 +1998,315 @@ def functionary_v1_v2_chat_handler(
     function_calls, function_bodies = [], []
     completion_tokens = 0
     
+    def generate_streaming(tools, functions, function_call, prompt):
+        assert version == "v2", "Streaming for v1 is not supported"
+        
+        chunk_id, chunk_created = None, None
+        
+        # If tool_choice/function_call is provided
+        if isinstance(function_call, dict):
+            prompt += f"{function_call['name']}\n{CONTENT_TOKEN}"
+            grammar = get_grammar(function_call["name"])
+            stops = [STOP_TOKEN, FROM_TOKEN]
+            tool_id = "".join([random.choice(string.ascii_letters + string.digits) for _ in range(24)])
+            completion = create_completion(prompt=prompt, stop=stops, grammar=grammar)
+            completion_text = ""
+            first = True
+            for chunk in completion:
+                # Yield the tool/function name first
+                if first:
+                    if tools is not None:
+                        func_call_dict = {
+                            "tool_calls": [
+                                {
+                                    "index": 0,
+                                    "id": "call_" + tool_id,
+                                    "type": "function",
+                                    "function": {"name": function_call["name"], "arguments": ""},
+                                }
+                            ]
+                        }
+                    else:
+                        func_call_dict = {"function_call": {"name": function_call["name"], "arguments": ""}}
+                    yield llama_types.CreateChatCompletionStreamResponse(
+                        id="chat" + chunk["id"],
+                        object="chat.completion.chunk",
+                        created=chunk["created"],
+                        model=chunk["model"],
+                        choices=[
+                            {"index": 0, "logprobs": None, "delta": {"role": None, "content": None, **func_call_dict}}
+                        ],
+                    )
+                    first = False
+                if tools is not None:
+                    func_call_dict = {
+                        "tool_calls": [
+                            {
+                                "index": 0,
+                                "id": "call_" + tool_id,
+                                "type": "function",
+                                "function": {
+                                    "name": None,
+                                    "arguments": chunk["choices"][0]["text"].rstrip(),
+                                },
+                            }
+                        ]
+                    }
+                else:
+                    func_call_dict = {"function_call": {"name": None, "arguments": chunk["choices"][0]["text"].rstrip()}}
+                if len(chunk["choices"][0]["text"].rstrip()) > 0:
+                    yield llama_types.CreateChatCompletionStreamResponse(
+                        id="chat" + chunk["id"],
+                        object="chat.completion.chunk",
+                        created=chunk["created"],
+                        model=chunk["model"],
+                        choices=[
+                            {
+                                "index": 0,
+                                "logprobs": chunk["choices"][0]["logprobs"],
+                                "delta": {
+                                    "role": None,
+                                    "content": None,
+                                    **func_call_dict,
+                                },
+                            }
+                        ],
+                    )
+            # Yield tool_call/function_call stop message
+            yield llama_types.CreateChatCompletionStreamResponse(
+                id="chat" + chunk["id"],
+                object="chat.completion.chunk",
+                created=chunk["created"],
+                model=chunk["model"],
+                choices=[
+                    {
+                        "index": 0,
+                        "finish_reason": "tool_calls" if tools is not None else "function_call",
+                        "logprobs": None,
+                        "delta": {
+                            "role": None, "content": None, "function_call": None, "tool_calls": None
+                        },
+                    }
+                ],
+            )
+        # If "auto" or no tool_choice/function_call
+        elif isinstance(function_call, str) and function_call == "auto":
+            tool_index = 0
+            while True:
+                # Generate function name first
+                grammar = None
+                stops = CONTENT_TOKEN
+                completion = create_completion(prompt=prompt, stop=stops, grammar=grammar)
+                completion_text = ""
+                for chunk in completion:
+                    completion_text += chunk["choices"][0]["text"]
+                if chunk_id is None:
+                    chunk_id = chunk["id"]
+                if chunk_created is None:
+                    chunk_created = chunk["created"]
+                function_name = completion_text.strip()
+                if function_name == "all":
+                    prompt += "all\n<|content|>"
+                    # Yield the first empty message for content
+                    yield llama_types.CreateChatCompletionStreamResponse(
+                        id="chat" + chunk_id,
+                        model=chunk["model"],
+                        created=chunk_created,
+                        object="chat.completion.chunk",
+                        choices=[
+                            {
+                                "index": 0,
+                                "delta": {"role": "assistant", "content": ""},
+                                "logprobs": None,
+                                "finish_reason": None,
+                            }
+                        ],
+                    )
+                else:
+                    prompt += f"{function_name}\n<|content|>"
+                    grammar = get_grammar(function_name)
+                    tool_id = "".join([random.choice(string.ascii_letters + string.digits) for _ in range(24)])
+                    if tools is not None:
+                        func_call_dict = {
+                            "tool_calls": [
+                                {
+                                    "index": tool_index,
+                                    "id": "call_" + tool_id,
+                                    "type": "function",
+                                    "function": {"name": function_name, "arguments": ""},
+                                }
+                            ]
+                        }
+                    else:
+                        func_call_dict = {"function_call": {"name": function_name, "arguments": ""}}
+                    # Stream function name
+                    yield llama_types.CreateChatCompletionStreamResponse(
+                        id="chat" + chunk_id,
+                        object="chat.completion.chunk",
+                        created=chunk_created,
+                        model=chunk["model"],
+                        choices=[
+                            {
+                                "index": 0,
+                                "logprobs": chunk["choices"][0]["logprobs"],
+                                "delta": {
+                                    "role": "assistant",
+                                    "content": None,
+                                    **func_call_dict,
+                                },
+                            }
+                        ],
+                    )
+                # Generate content
+                stops = [RECIPIENT_TOKEN, STOP_TOKEN]
+                completion = create_completion(prompt=prompt, stop=stops, grammar=grammar)
+                if function_name == "all":
+                    completion_text = ""
+                    stop_sequence, buffer, is_end = "\n<|from|>assistant\n<|recipient|>", [], False
+                    for i, chunk in enumerate(completion):
+                        completion_text += chunk["choices"][0]["text"]
+                        if is_end:
+                            buffer.append(chunk["choices"][0]["text"].strip(" "))
+                            if stop_sequence.startswith("".join(buffer)):
+                                continue
+                            else:
+                                buffer.pop()
+                                while len(buffer) > 0:
+                                    yield llama_types.CreateChatCompletionStreamResponse(
+                                        id="chat" + chunk_id,
+                                        object="chat.completion.chunk",
+                                        created=chunk_created,
+                                        model=chunk["model"],
+                                        choices=[
+                                            {
+                                                "index": 0,
+                                                "logprobs": chunk["choices"][0]["logprobs"],
+                                                "delta": {
+                                                    "role": "assistant", "content": buffer.pop(0)
+                                                },
+                                            }
+                                        ],
+                                    )
+                                is_end = False
+                        elif chunk["choices"][0]["text"] == "\n":
+                            is_end = True
+                            buffer.append(chunk["choices"][0]["text"].strip(" "))
+                            continue
+
+                        if len(buffer) == 0 and len(chunk["choices"][0]["text"]) > 0:
+                            yield llama_types.CreateChatCompletionStreamResponse(
+                                id="chat" + chunk_id,
+                                object="chat.completion.chunk",
+                                created=chunk_created,
+                                model=chunk["model"],
+                                choices=[
+                                    {
+                                        "index": 0,
+                                        "logprobs": chunk["choices"][0]["logprobs"],
+                                        "delta": {
+                                            "role": "assistant",
+                                            "content": chunk["choices"][0]["text"] if i > 0 else chunk["choices"][0]["text"].lstrip()
+                                        },
+                                    }
+                                ],
+                            )
+                    # Check whether the model wants to generate another turn
+                    if "<|from|> assistant" in completion_text or "<|from|>assistant" in completion_text:
+                        if completion_text.endswith("\n<|from|>assistant\n"):
+                            cleaned_completion_text = completion_text[:-len("\n<|from|>assistant\n")].strip()
+                        elif completion_text.endswith("\n<|from|> assistant\n"):
+                            cleaned_completion_text = completion_text[:-len("\n<|from|> assistant\n")].strip()
+                        else:
+                            cleaned_completion_text = completion_text.strip()
+                        prompt += f"{cleaned_completion_text}\n<|from|>assistant\n<|recipient|>"
+                    else:
+                        # Yield stop message
+                        yield llama_types.CreateChatCompletionStreamResponse(
+                            id="chat" + chunk_id,
+                            model=chunk["model"],
+                            created=chunk_created,
+                            object="chat.completion.chunk",
+                            choices=[
+                                {
+                                    "index": 0,
+                                    "delta": {},
+                                    "logprobs": None,
+                                    "finish_reason": "stop",
+                                }
+                            ],
+                        )
+                        break
+                else:
+                    # Check whether the model wants to generate another turn
+                    completion_text = ""
+                    for chunk in completion:
+                        completion_text += chunk["choices"][0]["text"]
+                        if len(chunk["choices"][0]["text"].rstrip()) > 0:
+                            if tools is not None:
+                                func_call_dict = {
+                                    "tool_calls": [
+                                        {
+                                            "index": tool_index,
+                                            "id": "call_" + tool_id,
+                                            "type": "function",
+                                            "function": {
+                                                "name": None,
+                                                "arguments": chunk["choices"][0]["text"].rstrip(),
+                                            },
+                                        }
+                                    ]
+                                }
+                            else:
+                                func_call_dict = {"function_call": {"name": None, "arguments": chunk["choices"][0]["text"].rstrip()}}
+                            yield llama_types.CreateChatCompletionStreamResponse(
+                                id="chat" + chunk_id,
+                                object="chat.completion.chunk",
+                                created=chunk_created,
+                                model=chunk["model"],
+                                choices=[
+                                    {
+                                        "index": 0,
+                                        "logprobs": chunk["choices"][0]["logprobs"],
+                                        "delta": {
+                                            "role": None,
+                                            "content": None,
+                                            **func_call_dict,
+                                        },
+                                    }
+                                ],
+                            )
+                    prompt += completion_text.strip()
+                    grammar = None
+                    completion = create_completion(prompt=prompt, stop=stops, grammar=grammar)
+                    completion_text += "".join([chunk["choices"][0]["text"] for chunk in completion])
+                    if ("<|from|> assistant" in completion_text or "<|from|>assistant" in completion_text) and tools is not None:
+                        prompt += "\n<|from|>assistant\n<|recipient|>"
+                        tool_index += 1
+                    else:
+                        # Yield tool_call/function_call stop message
+                        yield llama_types.CreateChatCompletionStreamResponse(
+                            id="chat" + chunk_id,
+                            object="chat.completion.chunk",
+                            created=chunk_created,
+                            model=chunk["model"],
+                            choices=[
+                                {
+                                    "index": 0,
+                                    "finish_reason": "tool_calls" if tools is not None else "function_call",
+                                    "logprobs": None,
+                                    "delta": {
+                                        "role": None, "content": None, "function_call": None, "tool_calls": None
+                                    },
+                                }
+                            ],
+                        )
+                        break
+        
+    if stream is not False:
+        return generate_streaming(
+            tools=tools, functions=functions, function_call=function_call, prompt=prompt
+        )
+    else:
         if version == "v1":
             # If no or "auto" tool_choice/function_call
             if isinstance(function_call, str) and function_call == "auto":
@@ -2012,7 +2322,7 @@ def functionary_v1_v2_chat_handler(
                 prompt = prompt
                 stops = ["\n", END_ASSISTANT_TOKEN]
 
-        completion = create_completion(stop=stops)
+            completion = create_completion(prompt=prompt, stop=stops, grammar=grammar)
             completion_text = completion["choices"][0]["text"]
             completion_tokens += completion["usage"]["completion_tokens"]
             
@@ -2039,7 +2349,7 @@ def functionary_v1_v2_chat_handler(
                     completion_text.split(START_FUNCTION_CALL_TOKEN)[-1][:-1].strip()
                 )
                 grammar = get_grammar(function_calls[-1])
-            completion = create_completion(stop=END_FUNCTION_CALL_TOKEN)
+                completion = create_completion(prompt=prompt, stop=END_FUNCTION_CALL_TOKEN, grammar=grammar)
                 completion_tokens += completion["usage"]["completion_tokens"]
                 function_bodies.append(completion["choices"][0]["text"].strip())
             # If the prompt involves a function call, just append generated parameters to function_bodies
@@ -2053,7 +2363,7 @@ def functionary_v1_v2_chat_handler(
                 function_calls.append(function_call)
                 grammar = get_grammar(function_call)
                 stops = [STOP_TOKEN, FROM_TOKEN]
-            completion = create_completion(stop=stops)
+                completion = create_completion(prompt=prompt, stop=stops, grammar=grammar)
                 completion_text = completion["choices"][0]["text"]
                 completion_tokens += completion["usage"]["completion_tokens"]
                 function_bodies.append(completion_text.strip())
@@ -2063,7 +2373,7 @@ def functionary_v1_v2_chat_handler(
                     # Generate function name first
                     grammar = None
                     stops = CONTENT_TOKEN
-                completion = create_completion(stop=stops)
+                    completion = create_completion(prompt=prompt, stop=stops, grammar=grammar)
                     completion_text = completion["choices"][0]["text"]
                     completion_tokens += completion["usage"]["completion_tokens"]
                     function_name = completion_text.strip()
@@ -2076,7 +2386,7 @@ def functionary_v1_v2_chat_handler(
                         grammar = get_grammar(function_call)
                     # Generate content
                     stops = [RECIPIENT_TOKEN, STOP_TOKEN]
-                completion = create_completion(stop=stops)
+                    completion = create_completion(prompt=prompt, stop=stops, grammar=grammar)
                     completion_text = completion["choices"][0]["text"]
                     completion_tokens += completion["usage"]["completion_tokens"]
                     if function_name == "all":
@@ -2103,7 +2413,7 @@ def functionary_v1_v2_chat_handler(
                         # Check whether the model wants to generate another turn
                         prompt += completion_text.strip()
                         grammar = None
-                    completion = create_completion(stop=stops)
+                        completion = create_completion(prompt=prompt, stop=stops, grammar=grammar)
                         completion_tokens += completion["usage"]["completion_tokens"]
                         if "<|from|> assistant" in completion["choices"][0]["text"] or "<|from|>assistant" in completion["choices"][0]["text"]:
                             prompt += "\n<|from|>assistant\n<|recipient|>"
@@ -2165,7 +2475,7 @@ def functionary_v1_v2_chat_handler(
 
 
 class Llava15ChatHandler:
-    DEFAULT_SYSTEM_MESSAGE =  "A chat between a curious human and an artificial intelligence assistant.  The assistant gives helpful, detailed, and polite answers to the human's questions."
+    DEFAULT_SYSTEM_MESSAGE: Optional[str] =  "A chat between a curious human and an artificial intelligence assistant.  The assistant gives helpful, detailed, and polite answers to the human's questions."
 
     CHAT_FORMAT = (
         "{% for message in messages %}"
@@ -2288,18 +2598,31 @@ class Llava15ChatHandler:
         assert self.clip_ctx is not None
 
         system_prompt = _get_system_message(messages)
-        if system_prompt == "":
+        if system_prompt == "" and self.DEFAULT_SYSTEM_MESSAGE is not None:
             messages = [llama_types.ChatCompletionRequestSystemMessage(role="system", content=self.DEFAULT_SYSTEM_MESSAGE)] + messages
 
         image_urls = self.get_image_urls(messages)
-        template = jinja2.Template(self.CHAT_FORMAT)
+        template = ImmutableSandboxedEnvironment(
-        text = template.render(messages=messages, add_generation_prompt=True)
+            trim_blocks=True,
+            lstrip_blocks=True,
+        ).from_string(self.CHAT_FORMAT)
+        text = template.render(
+            messages=messages,
+            add_generation_prompt=True,
+            eos_token=llama.detokenize([llama.token_eos()]),
+            bos_token=llama.detokenize([llama.token_bos()]),
+        )
         split_text = self.split_text_on_image_urls(text, image_urls)
 
         def embed_image_bytes(image_bytes: bytes):
             if self._last_image_embed is not None and self._last_image_hash is not None and hash(image_bytes) == self._last_image_hash:
                 return self._last_image_embed
             with suppress_stdout_stderr(disable=self.verbose):
+                # Free the previous image embed
+                if self._last_image_embed is not None:
+                    self._llava_cpp.llava_image_embed_free(self._last_image_embed)
+                    self._last_image_embed = None
+                    self._last_image_hash = None
                 embed = (
                     self._llava_cpp.llava_image_embed_make_with_bytes(
                         self.clip_ctx,
@@ -2314,9 +2637,10 @@ class Llava15ChatHandler:
 
         # Evaluate prompt
         llama.reset()
-        for i, (type_, value) in enumerate(split_text):
+        llama._ctx.kv_cache_clear()
+        for type_, value in split_text:
             if type_ == "text":
-                tokens = llama.tokenize(value.encode("utf8"), add_bos=i == 0)
+                tokens = llama.tokenize(value.encode("utf8"), add_bos=False, special=True)
                 if llama.n_tokens + len(tokens) > llama.n_ctx():
                     raise ValueError("Prompt exceeds n_ctx") # TODO: Fix
                 llama.eval(tokens)
@@ -2334,6 +2658,8 @@ class Llava15ChatHandler:
                         llama.n_batch,
                         n_past_p,
                     )
+                # Required to avoid issues with hf tokenizer
+                llama.input_ids[llama.n_tokens : n_past.value] = -1
                 llama.n_tokens = n_past.value
 
         # Get prompt tokens to avoid a cache miss
@@ -2723,6 +3049,7 @@ class NanoLlavaChatHandler(Llava15ChatHandler):
     # Answer the question<|im_end|><|im_start|>user
     # <image>
     # What is the picture about?<|im_end|><|im_start|>assistant
+    DEFAULT_SYSTEM_MESSAGE = "Answer the question"
 
     CHAT_FORMAT = (
         "{% for message in messages %}"
@@ -2771,6 +3098,66 @@ class NanoLlavaChatHandler(Llava15ChatHandler):
         "{% endif %}"
     )
 
+class Llama3VisionAlpha(Llava15ChatHandler):
+    # question = "<image>" + q
+
+    # prompt = f"<|start_header_id|>user<|end_header_id|>\n\n{question}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+    DEFAULT_SYSTEM_MESSAGE = None
+
+    CHAT_FORMAT = (
+        "{% for message in messages %}"
+
+        "<|start_header_id|>"
+
+        "{% if message.role == 'user' %}"
+
+        "user<|end_header_id|>\n\n"
+
+        "{% if message.content is iterable %}"
+
+        # <image>
+        "{% for content in message.content %}"
+        "{% if content.type == 'image_url' %}"
+        "{% if content.image_url is string %}"
+        "{{ content.image_url }}"
+        "{% endif %}"
+        "{% if content.image_url is mapping %}"
+        "{{ content.image_url.url }}"
+        "{% endif %}"
+        "{% endif %}"
+        "{% endfor %}"
+
+        # Question:
+        "{% for content in message.content %}"
+        "{% if content.type == 'text' %}"
+        "{{ content.text }}"
+        "{% endif %}"
+        "{% endfor %}"
+
+        "{% endif %}"
+
+        # Question:
+        "{% if message.content is string %}"
+        "{{ message.content }}"
+        "{% endif %}"
+
+        "{% endif %}"
+
+        # Answer:
+        "{% if message.role == 'assistant' %}"
+        "assistant<|end_header_id|>\n\n"
+        "{{ message.content }}"
+        "{% endif %}"
+
+        "<|eot_id|>"
+
+        "{% endfor %}"
+
+        # Generation prompt
+        "{% if add_generation_prompt %}"
+        "<|start_header_id|>assistant<|end_header_id|>\n\n"
+        "{% endif %}"
+    )
 
 @register_chat_completion_handler("chatml-function-calling")
 def chatml_function_calling(
@@ -2858,8 +3245,7 @@ def chatml_function_calling(
         "{% endfor %}"
         "{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}"
     )
-    template_renderer = jinja2.Environment(
+    template_renderer = ImmutableSandboxedEnvironment(
-        loader=jinja2.BaseLoader(),
         autoescape=jinja2.select_autoescape(["html", "xml"]),
         undefined=jinja2.StrictUndefined,
     ).from_string(function_calling_template)

llama_cpp/llama_cpp.py

@@ -294,6 +294,11 @@ LLAMA_VOCAB_TYPE_WPM = 3
 #     LLAMA_VOCAB_PRE_TYPE_MPT            = 5,
 #     LLAMA_VOCAB_PRE_TYPE_STARCODER      = 6,
 #     LLAMA_VOCAB_PRE_TYPE_GPT2           = 7,
+#     LLAMA_VOCAB_PRE_TYPE_REFACT         = 8,
+#     LLAMA_VOCAB_PRE_TYPE_COMMAND_R      = 9,
+#     LLAMA_VOCAB_PRE_TYPE_QWEN2          = 10,
+#     LLAMA_VOCAB_PRE_TYPE_OLMO           = 11,
+#     LLAMA_VOCAB_PRE_TYPE_DBRX           = 12,
 # };
 LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0
 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1
@@ -303,6 +308,11 @@ LLAMA_VOCAB_PRE_TYPE_FALCON = 4
 LLAMA_VOCAB_PRE_TYPE_MPT = 5
 LLAMA_VOCAB_PRE_TYPE_STARCODER = 6
 LLAMA_VOCAB_PRE_TYPE_GPT2 = 7
+LLAMA_VOCAB_PRE_TYPE_REFACT = 8
+LLAMA_VOCAB_PRE_TYPE_COMMAND_R = 9
+LLAMA_VOCAB_PRE_TYPE_QWEN2 = 10
+LLAMA_VOCAB_PRE_TYPE_OLMO = 11
+LLAMA_VOCAB_PRE_TYPE_DBRX = 12
 
 
 # // note: these values should be synchronized with ggml_rope
@@ -371,6 +381,7 @@ LLAMA_TOKEN_TYPE_BYTE = 6
 #     LLAMA_FTYPE_MOSTLY_IQ2_M         = 29, // except 1d tensors
 #     LLAMA_FTYPE_MOSTLY_IQ4_XS        = 30, // except 1d tensors
 #     LLAMA_FTYPE_MOSTLY_IQ1_M         = 31, // except 1d tensors
+#     LLAMA_FTYPE_MOSTLY_BF16          = 32, // except 1d tensors
 
 #     LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
 # };
@@ -403,6 +414,8 @@ LLAMA_FTYPE_MOSTLY_IQ3_M = 27
 LLAMA_FTYPE_MOSTLY_IQ2_S = 28
 LLAMA_FTYPE_MOSTLY_IQ2_M = 29
 LLAMA_FTYPE_MOSTLY_IQ4_XS = 30
+LLAMA_FTYPE_MOSTLY_IQ1_M = 31
+LLAMA_FTYPE_MOSTLY_BF16 = 32
 LLAMA_FTYPE_GUESSED = 1024
 
 # enum llama_rope_scaling_type {
@@ -494,7 +507,7 @@ class llama_token_data_array(ctypes.Structure):
 
 llama_token_data_array_p = ctypes.POINTER(llama_token_data_array)
 
-# typedef bool (*llama_progress_callback)(float progress, void *ctx);
+# typedef bool (*llama_progress_callback)(float progress, void * user_data);
 llama_progress_callback = ctypes.CFUNCTYPE(
     ctypes.c_bool, ctypes.c_float, ctypes.c_void_p
 )
@@ -635,6 +648,9 @@ class llama_model_kv_override(ctypes.Structure):
 #     // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
 #     const float * tensor_split;
 
+#     // comma separated list of RPC servers to use for offloading
+#     const char * rpc_servers;
+
 #     // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
 #     // If the provided progress_callback returns true, model loading continues.
 #     // If it returns false, model loading is immediately aborted.
@@ -661,6 +677,7 @@ class llama_model_params(ctypes.Structure):
         split_mode (int): how to split the model across multiple GPUs
         main_gpu (int): the GPU that is used for the entire model. main_gpu interpretation depends on split_mode: LLAMA_SPLIT_NONE: the GPU that is used for the entire model LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results LLAMA_SPLIT_LAYER: ignored
         tensor_split (ctypes.Array[ctypes.ctypes.c_float]): proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
+        rpc_servers (ctypes.c_char_p): comma separated list of RPC servers to use for offloading
         progress_callback (llama_progress_callback): called with a progress value between 0.0 and 1.0. Pass NULL to disable. If the provided progress_callback returns true, model loading continues. If it returns false, model loading is immediately aborted.
         progress_callback_user_data (ctypes.ctypes.c_void_p): context pointer passed to the progress callback
         kv_overrides (ctypes.Array[llama_model_kv_override]): override key-value pairs of the model meta data
@@ -674,6 +691,7 @@ class llama_model_params(ctypes.Structure):
         split_mode: int
         main_gpu: int
         tensor_split: CtypesArray[ctypes.c_float]
+        rpc_servers: ctypes.c_char_p
         progress_callback: Callable[[float, ctypes.c_void_p], bool]
         progress_callback_user_data: ctypes.c_void_p
         kv_overrides: CtypesArray[llama_model_kv_override]
@@ -687,6 +705,7 @@ class llama_model_params(ctypes.Structure):
         ("split_mode", ctypes.c_int),
         ("main_gpu", ctypes.c_int32),
         ("tensor_split", ctypes.POINTER(ctypes.c_float)),
+        ("rpc_servers", ctypes.c_char_p),
         ("progress_callback", llama_progress_callback),
         ("progress_callback_user_data", ctypes.c_void_p),
         ("kv_overrides", ctypes.POINTER(llama_model_kv_override)),

llama_cpp/server/app.py

@@ -132,6 +132,7 @@ def create_app(
         middleware=middleware,
         title="🦙 llama.cpp Python API",
         version=llama_cpp.__version__,
+        root_path=server_settings.root_path,
     )
     app.add_middleware(
         CORSMiddleware,
@@ -274,6 +275,7 @@ async def create_completion(
         "best_of",
         "logit_bias_type",
         "user",
+        "min_tokens",
     }
     kwargs = body.model_dump(exclude=exclude)
 
@@ -287,6 +289,15 @@ async def create_completion(
     if body.grammar is not None:
         kwargs["grammar"] = llama_cpp.LlamaGrammar.from_string(body.grammar)
 
+    if body.min_tokens > 0:
+        _min_tokens_logits_processor = llama_cpp.LogitsProcessorList(
+            [llama_cpp.MinTokensLogitsProcessor(body.min_tokens, llama.token_eos())]
+        )
+        if "logits_processor" not in kwargs:
+            kwargs["logits_processor"] = _min_tokens_logits_processor
+        else:
+            kwargs["logits_processor"].extend(_min_tokens_logits_processor)
+
     iterator_or_completion: Union[
         llama_cpp.CreateCompletionResponse,
         Iterator[llama_cpp.CreateCompletionStreamResponse],
@@ -444,6 +455,7 @@ async def create_chat_completion(
         "n",
         "logit_bias_type",
         "user",
+        "min_tokens",
     }
     kwargs = body.model_dump(exclude=exclude)
     llama = llama_proxy(body.model)
@@ -457,6 +469,15 @@ async def create_chat_completion(
     if body.grammar is not None:
         kwargs["grammar"] = llama_cpp.LlamaGrammar.from_string(body.grammar)
 
+    if body.min_tokens > 0:
+        _min_tokens_logits_processor = llama_cpp.LogitsProcessorList(
+            [llama_cpp.MinTokensLogitsProcessor(body.min_tokens, llama.token_eos())]
+        )
+        if "logits_processor" not in kwargs:
+            kwargs["logits_processor"] = _min_tokens_logits_processor
+        else:
+            kwargs["logits_processor"].extend(_min_tokens_logits_processor)
+
     iterator_or_completion: Union[
         llama_cpp.ChatCompletion, Iterator[llama_cpp.ChatCompletionChunk]
     ] = await run_in_threadpool(llama.create_chat_completion, **kwargs)

llama_cpp/server/model.py

@@ -140,6 +140,20 @@ class LlamaProxy:
                 chat_handler = llama_cpp.llama_chat_format.NanoLlavaChatHandler(
                     clip_model_path=settings.clip_model_path, verbose=settings.verbose
                 )
+        elif settings.chat_format == "llama-3-vision-alpha":
+            assert settings.clip_model_path is not None, "clip model not found"
+            if settings.hf_model_repo_id is not None:
+                chat_handler = (
+                    llama_cpp.llama_chat_format.Llama3VisionAlpha.from_pretrained(
+                        repo_id=settings.hf_model_repo_id,
+                        filename=settings.clip_model_path,
+                        verbose=settings.verbose,
+                    )
+                )
+            else:
+                chat_handler = llama_cpp.llama_chat_format.Llama3VisionAlpha(
+                    clip_model_path=settings.clip_model_path, verbose=settings.verbose
+                )
         elif settings.chat_format == "hf-autotokenizer":
             assert (
                 settings.hf_pretrained_model_name_or_path is not None
@@ -228,6 +242,7 @@ class LlamaProxy:
             logits_all=settings.logits_all,
             embedding=settings.embedding,
             offload_kqv=settings.offload_kqv,
+            flash_attn=settings.flash_attn,
             # Sampling Params
             last_n_tokens_size=settings.last_n_tokens_size,
             # LoRA Params

llama_cpp/server/settings.py

@@ -215,6 +215,10 @@ class ServerSettings(BaseSettings):
         default=False,
         description="Disable EventSource pings (may be needed for some clients).",
     )
+    root_path: str = Field(
+        default="",
+        description="The root path for the server. Useful when running behind a reverse proxy.",
+    )
 
 
 class Settings(ServerSettings, ModelSettings):

llama_cpp/server/types.py

@@ -16,10 +16,14 @@ max_tokens_field = Field(
     default=16, ge=1, description="The maximum number of tokens to generate."
 )
 
+min_tokens_field = Field(
+    default=0,
+    ge=0,
+    description="The minimum number of tokens to generate. It may return fewer tokens if another condition is met (e.g. max_tokens, stop).",
+)
+
 temperature_field = Field(
     default=0.8,
-    ge=0.0,
-    le=2.0,
     description="Adjust the randomness of the generated text.\n\n"
     + "Temperature is a hyperparameter that controls the randomness of the generated text. It affects the probability distribution of the model's output tokens. A higher temperature (e.g., 1.5) makes the output more random and creative, while a lower temperature (e.g., 0.5) makes the output more focused, deterministic, and conservative. The default value is 0.8, which provides a balance between randomness and determinism. At the extreme, a temperature of 0 will always pick the most likely next token, leading to identical outputs in each run.",
 )
@@ -113,6 +117,7 @@ class CreateCompletionRequest(BaseModel):
     max_tokens: Optional[int] = Field(
         default=16, ge=0, description="The maximum number of tokens to generate."
     )
+    min_tokens: int = min_tokens_field
     temperature: float = temperature_field
     top_p: float = top_p_field
     min_p: float = min_p_field
@@ -208,6 +213,7 @@ class CreateChatCompletionRequest(BaseModel):
         default=None,
         description="The maximum number of tokens to generate. Defaults to inf",
     )
+    min_tokens: int = min_tokens_field
     logprobs: Optional[bool] = Field(
         default=False,
         description="Whether to output the logprobs or not. Default is True"

scripts/releases-to-pep-503.sh

@@ -44,7 +44,9 @@ releases=$(echo $releases | tr ' ' '\n' | grep -E $pattern)
 # For each release, get all assets
 for release in $releases; do
     assets=$(curl -s https://api.github.com/repos/abetlen/llama-cpp-python/releases/tags/$release | jq -r .assets)
-    echo "    <h2>$release</h2>" >> index.html
+    # Get release version from release ie v0.1.0-cu121 -> v0.1.0
+    release_version=$(echo $release | grep -oE "^[v]?[0-9]+\.[0-9]+\.[0-9]+")
+    echo "    <h2>$release_version</h2>" >> index.html
     for asset in $(echo $assets | jq -r .[].browser_download_url); do
         if [[ $asset == *".whl" ]]; then
             echo "    <a href=\"$asset\">$asset</a>" >> index.html

vendor/llama.cpp

@@ -1 +1 @@
-Subproject commit f364eb6fb5d46118a76fa045f487318de4c24961
+Subproject commit 05834841dcb4f922983ea976539c70472272df9a