diff --git a/misc/llama/__init__.py b/misc/llama/__init__.py
new file mode 100644
index 0000000..2990ac4
--- /dev/null
+++ b/misc/llama/__init__.py
@@ -0,0 +1 @@
+from .cffi import *
\ No newline at end of file
diff --git a/misc/llama/cffi.py b/misc/llama/cffi.py
new file mode 100644
index 0000000..6d73d50
--- /dev/null
+++ b/misc/llama/cffi.py
@@ -0,0 +1,2 @@
+from .llama_cli_cffi import *
+from .llama_cli_options import *
diff --git a/misc/llama/ctypes.py b/misc/llama/ctypes.py
new file mode 100644
index 0000000..8fc25e3
--- /dev/null
+++ b/misc/llama/ctypes.py
@@ -0,0 +1,2 @@
+from .llama_cli_ctypes import *
+from .llama_cli_options import *
\ No newline at end of file
diff --git a/misc/llama/llama_cli_cffi.py b/misc/llama/llama_cli_cffi.py
new file mode 100644
index 0000000..fdaee55
--- /dev/null
+++ b/misc/llama/llama_cli_cffi.py
@@ -0,0 +1,125 @@
+__all__ = ['llama_generate', 'LlamaOptions']
+
+import json
+import ctypes
+from queue import Queue
+from copy import deepcopy
+from typing import Iterator
+from threading import Thread
+from functools import partial
+
+from huggingface_hub import hf_hub_download
+
+from .llama_cli_options import LlamaOptions, convert_options_to_bytes
+from ._llama_cli import lib, ffi
+
+
+FPRINTF_FUNC = ctypes.CFUNCTYPE(ctypes.c_int, ctypes.c_void_p, ctypes.c_char_p, ctypes.c_char_p)
+FFLUSH_FUNC = ctypes.CFUNCTYPE(ctypes.c_int, ctypes.c_void_p)
+
+
+def fprintf_func(file_obj, fmt, arg, queue=None, callback=None, metadata=None):
+    content = arg.decode('utf-8')
+
+    if metadata:
+        eos = metadata.get('eos')
+        eot = metadata.get('eot')
+
+        if eos is not None and eos in content:
+            content = content[:content.index(eos)]
+        elif eot is not None and eot in content:
+            content = content[:content.index(eot)]
+
+    if queue is not None:
+        if content:
+            queue.put(content)
+        else:
+            queue.put(None)
+    elif callback is not None:
+        callback(content)
+
+    size = len(content)
+    return size
+
+
+def fflush_func(file_obj):
+    return 0
+
+
+def _llama_cli_main(argc, argv, queue=None, callback=None, metadata=None):
+    r = lib.llama_cli_main(argc, argv)
+    assert r == 0
+    
+    if queue is not None:
+        queue.put(None)
+    elif callback is not None:
+        callback(None)
+
+
+def llama_generate(options: LlamaOptions, callback=None) -> Iterator[str] | None:
+    # check hf_repo, hf_file
+    if options.hf_repo and options.hf_file:
+        options.model = hf_hub_download(repo_id=options.hf_repo, filename=options.hf_file)
+        options.hf_repo = None
+        options.hf_file = None
+    elif options.model:
+        pass
+    else:
+        raise ValueError(f'hf_repo = {options.hf_repo}, hf_file = {options.hf_file}')
+
+    assert options.model
+
+    if callback:
+        queue = None
+    else:
+        queue = Queue()
+
+    # get bos, eos, and eot from metedata
+    metadata_options = deepcopy(options)
+    metadata_options.log_disable = True
+    metadata_argv: list[bytes] = [b'llama-cli'] + convert_options_to_bytes(metadata_options)
+    metadata_argv = [ffi.new('char[]', n) for n in metadata_argv]
+    metadata_argc = len(metadata_argv)
+
+    c_metadata: 'const char*' = lib.llama_get_metadata_as_json(metadata_argc, metadata_argv)
+    metadata: bytes = ffi.string(c_metadata)
+    lib.llama_free_metadata_as_json(c_metadata)
+    metadata: str = metadata.decode('utf-8')
+    metadata: dict = json.loads(metadata)
+    print(f'{metadata = }')
+
+    # intercept token generation
+    fprintf = FPRINTF_FUNC(partial(fprintf_func, queue=queue, metadata=metadata))
+    fflush = FFLUSH_FUNC(fflush_func)
+
+    fprintf_address = ctypes.cast(fprintf, ctypes.c_void_p).value
+    fflush_address = ctypes.cast(fflush, ctypes.c_void_p).value
+
+    cffi_fprintf_callback = ffi.cast('int (*func)(FILE*, const char* format, ...)', fprintf_address)
+    cffi_fflush_callback = ffi.cast('int (*func)(FILE*)', fflush_address)
+
+    lib.llama_set_fprintf(cffi_fprintf_callback)
+    lib.llama_set_fflush(cffi_fflush_callback)
+
+    argv: list[bytes] = [b'llama-cli'] + convert_options_to_bytes(options)
+    argv = [ffi.new('char[]', n) for n in argv]
+    argc = len(argv)
+
+    if callback:
+        _llama_cli_main(argc, argv, queue, callback, metadata)
+        yield ''
+    else:
+        t = Thread(target=_llama_cli_main, args=(argc, argv, queue, callback, metadata))
+        t.start()
+
+        while True:
+            chunk = queue.get()
+            queue.task_done()
+
+            if chunk is None:
+                break
+
+            yield chunk
+
+        queue.join()
+        t.join()
diff --git a/misc/llama/llama_cli_ctypes.py b/misc/llama/llama_cli_ctypes.py
new file mode 100644
index 0000000..07c4616
--- /dev/null
+++ b/misc/llama/llama_cli_ctypes.py
@@ -0,0 +1,132 @@
+__all__ = ['llama_generate', 'LlamaOptions']
+
+import os
+import json
+from ctypes import *
+from queue import Queue
+from copy import deepcopy
+from typing import Iterator
+from threading import Thread
+from functools import partial
+
+from huggingface_hub import hf_hub_download
+
+from .llama_cli_options import LlamaOptions, convert_options_to_bytes
+
+current_module_path = os.path.abspath(__file__)
+current_module_dir = os.path.dirname(current_module_path)
+libllama_cli_path = os.path.join(current_module_dir, 'libllama-cli.so')
+
+lib = CDLL(libllama_cli_path)
+
+lib.llama_cli_main.argtypes = [c_int, POINTER(c_char_p)]
+lib.llama_cli_main.restype = c_int
+
+lib.llama_get_metadata_as_json.argtypes = [c_int, POINTER(c_char_p)]
+lib.llama_get_metadata_as_json.restype = c_void_p
+
+lib.llama_free_metadata_as_json.argtypes = [c_void_p]
+lib.llama_free_metadata_as_json.restype = None
+
+FPRINTF_FUNC = CFUNCTYPE(c_int, c_void_p, c_char_p, c_char_p)
+FFLUSH_FUNC = CFUNCTYPE(c_int, c_void_p)
+
+
+def fprintf_func(file_obj, fmt, arg, queue=None, callback=None, metadata=None):
+    content = arg.decode('utf-8')
+
+    if metadata:
+        eos = metadata.get('eos')
+        eot = metadata.get('eot')
+
+        if eos is not None and eos in content:
+            content = content[:content.index(eos)]
+        elif eot is not None and eot in content:
+            content = content[:content.index(eot)]
+
+    if queue is not None:
+        if content:
+            queue.put(content)
+        else:
+            queue.put(None)
+    elif callback is not None:
+        callback(content)
+
+    size = len(content)
+    return size
+
+
+def fflush_func(file_obj):
+    return 0
+
+
+def _llama_cli_main(argc, argv, queue=None, callback=None, metadata=None):
+    r = lib.llama_cli_main(argc, argv)
+    assert r == 0
+
+    if queue is not None:
+        queue.put(None)
+    elif callback is not None:
+        callback(None)
+
+
+def llama_generate(options: LlamaOptions, callback=None, metadata=None) -> Iterator[str] | None:
+    # check hf_repo, hf_file
+    if options.hf_repo and options.hf_file:
+        options.model = hf_hub_download(repo_id=options.hf_repo, filename=options.hf_file)
+        options.hf_repo = None
+        options.hf_file = None
+    elif options.model:
+        pass
+    else:
+        raise ValueError(f'hf_repo = {options.hf_repo}, hf_file = {options.hf_file}')
+
+    assert options.model
+
+    if callback:
+        queue = None
+    else:
+        queue = Queue()
+
+    # get bos, eos, and eot from metedata
+    metadata_options = deepcopy(options)
+    metadata_options.log_disable = True
+    metadata_argv: list[bytes] = [b'llama-cli'] + convert_options_to_bytes(metadata_options)
+    metadata_argc = len(metadata_argv)
+    metadata_argv = (c_char_p * metadata_argc)(*metadata_argv)
+
+    c_metadata: 'const char*' = lib.llama_get_metadata_as_json(metadata_argc, metadata_argv)
+    metadata: str = string_at(c_metadata)
+    lib.llama_free_metadata_as_json(c_metadata)
+    metadata: dict = json.loads(metadata)
+    print(f'{metadata = }')
+
+    # intercept token generation
+    fprintf = FPRINTF_FUNC(partial(fprintf_func, queue=queue, callback=callback, metadata=metadata))
+    fflush = FFLUSH_FUNC(fflush_func)
+
+    lib.llama_set_fprintf(fprintf)
+    lib.llama_set_fflush(fflush)
+
+    argv: list[bytes] = [b'llama-cli'] + convert_options_to_bytes(options)
+    argc = len(argv)
+    argv = (c_char_p * argc)(*argv)
+
+    if callback:
+        _llama_cli_main(argc, argv, queue, callback, metadata)
+        yield ''
+    else:
+        t = Thread(target=_llama_cli_main, args=(argc, argv, queue, callback, metadata))
+        t.start()
+
+        while True:
+            chunk = queue.get()
+            queue.task_done()
+
+            if chunk is None:
+                break
+
+            yield chunk
+
+        queue.join()
+        t.join()
diff --git a/misc/llama/llama_cli_options.py b/misc/llama/llama_cli_options.py
new file mode 100644
index 0000000..8743a26
--- /dev/null
+++ b/misc/llama/llama_cli_options.py
@@ -0,0 +1,214 @@
+__all__ = ['LlamaOptions', 'convert_options_to_bytes']
+
+import attr
+
+
+@attr.s
+class LlamaOptions:
+    help = attr.ib(default=None, metadata={"description": "print usage and exit", "long_name": "--help", "alias": "-h"})
+    version = attr.ib(default=None, metadata={"description": "show version and build info", "long_name": "--version"})
+    verbose = attr.ib(default=None, metadata={"description": "print verbose information", "long_name": "--verbose", "alias": "-v"})
+    verbosity = attr.ib(default=None, metadata={"description": "set specific verbosity level", "long_name": "--verbosity", "default": 0})
+    verbose_prompt = attr.ib(default=None, metadata={"description": "print a verbose prompt before generation", "long_name": "--verbose-prompt", "default": False})
+    no_display_prompt = attr.ib(default=None, metadata={"description": "don't print prompt at generation", "long_name": "--no-display-prompt", "default": False})
+    color = attr.ib(default=None, metadata={"description": "colorise output to distinguish prompt and user input from generations", "long_name": "--color", "alias": "-co", "default": False})
+    seed = attr.ib(default=None, metadata={"description": "RNG seed", "long_name": "--seed", "alias": "-s", "default": -1})
+    threads = attr.ib(default=None, metadata={"description": "number of threads to use during generation", "long_name": "--threads", "alias": "-t", "default": 16})
+    threads_batch = attr.ib(default=None, metadata={"description": "number of threads to use during batch and prompt processing", "long_name": "--threads-batch", "alias": "-tb"})
+    threads_draft = attr.ib(default=None, metadata={"description": "number of threads to use during generation", "long_name": "--threads-draft", "alias": "-td"})
+    threads_batch_draft = attr.ib(default=None, metadata={"description": "number of threads to use during batch and prompt processing", "long_name": "--threads-batch-draft", "alias": "-tbd"})
+    draft = attr.ib(default=None, metadata={"description": "number of tokens to draft for speculative decoding", "long_name": "--draft", "default": 5})
+    p_split = attr.ib(default=None, metadata={"description": "speculative decoding split probability", "long_name": "--p-split", "alias": "-ps", "default": 0.1})
+    lookup_cache_static = attr.ib(default=None, metadata={"description": "path to static lookup cache to use for lookup decoding", "long_name": "--lookup-cache-static", "alias": "-lcs"})
+    lookup_cache_dynamic = attr.ib(default=None, metadata={"description": "path to dynamic lookup cache to use for lookup decoding", "long_name": "--lookup-cache-dynamic", "alias": "-lcd"})
+    ctx_size = attr.ib(default=None, metadata={"description": "size of the prompt context", "long_name": "--ctx-size", "alias": "-c", "default": 0})
+    predict = attr.ib(default=None, metadata={"description": "number of tokens to predict", "long_name": "--predict", "alias": "-n", "default": -1})
+    batch_size = attr.ib(default=None, metadata={"description": "logical maximum batch size", "long_name": "--batch-size", "alias": "-b", "default": 2048})
+    ubatch_size = attr.ib(default=None, metadata={"description": "physical maximum batch size", "long_name": "--ubatch-size", "alias": "-ub", "default": 512})
+    keep = attr.ib(default=None, metadata={"description": "number of tokens to keep from the initial prompt", "long_name": "--keep", "default": 0})
+    chunks = attr.ib(default=None, metadata={"description": "max number of chunks to process", "long_name": "--chunks", "default": -1})
+    flash_attn = attr.ib(default=None, metadata={"description": "enable Flash Attention", "long_name": "--flash-attn", "alias": "-fa", "default": False})
+    prompt = attr.ib(default=None, metadata={"description": "prompt to start generation with", "long_name": "--prompt", "alias": "-p", "default": ''})
+    file = attr.ib(default=None, metadata={"description": "a file containing the prompt", "long_name": "--file", "alias": "-f"})
+    in_file = attr.ib(default=None, metadata={"description": "an input file", "long_name": "--in-file"})
+    binary_file = attr.ib(default=None, metadata={"description": "binary file containing the prompt", "long_name": "--binary-file", "alias": "-bf"})
+    escape = attr.ib(default=None, metadata={"description": "process escapes sequences", "long_name": "--escape", "alias": "-e", "default": True})
+    no_escape = attr.ib(default=None, metadata={"description": "do not process escape sequences", "long_name": "--no-escape"})
+    print_token_count = attr.ib(default=None, metadata={"description": "print token count every N tokens", "long_name": "--print-token-count", "alias": "-ptc", "default": -1})
+    prompt_cache = attr.ib(default=None, metadata={"description": "file to cache prompt state for faster startup", "long_name": "--prompt-cache"})
+    prompt_cache_all = attr.ib(default=None, metadata={"description": "saves user input and generations to cache as well", "long_name": "--prompt-cache-all", "default": False})
+    prompt_cache_ro = attr.ib(default=None, metadata={"description": "uses the prompt cache but does not update it", "long_name": "--prompt-cache-ro", "default": False})
+    reverse_prompt = attr.ib(default=None, metadata={"description": "halt generation at PROMPT, return control in interactive mode", "long_name": "--reverse-prompt", "alias": "-r"})
+    special = attr.ib(default=None, metadata={"description": "special tokens output enabled", "long_name": "--special", "alias": "-sp", "default": False})
+    conversation = attr.ib(default=None, metadata={"description": "run in conversation mode", "long_name": "--conversation", "alias": "-cnv", "default": False})
+    interactive = attr.ib(default=None, metadata={"description": "run in interactive mode", "long_name": "--interactive", "alias": "-i", "default": False})
+    interactive_first = attr.ib(default=None, metadata={"description": "run in interactive mode and wait for input right away", "long_name": "--interactive-first", "alias": "-if", "default": False})
+    multiline_input = attr.ib(default=None, metadata={"description": "allows you to write or paste multiple lines without ending each in '\\'", "long_name": "--multiline-input"})
+    in_prefix_bos = attr.ib(default=None, metadata={"description": "prefix BOS to user inputs", "long_name": "--in-prefix-bos"})
+    in_prefix = attr.ib(default=None, metadata={"description": "string to prefix user inputs with", "long_name": "--in-prefix", "default": ''})
+    in_suffix = attr.ib(default=None, metadata={"description": "string to suffix after user inputs with", "long_name": "--in-suffix", "default": ''})
+    spm_infill = attr.ib(default=None, metadata={"description": "use Suffix/Prefix/Middle pattern for infill", "long_name": "--spm-infill", "default": False})
+
+    samplers = attr.ib(default=None, metadata={"description": "samplers that will be used for generation in the order", "long_name": "--samplers"})
+    sampling_seq = attr.ib(default=None, metadata={"description": "simplified sequence for samplers that will be used", "long_name": "--sampling-seq", "default": "kfypmt"})
+    ignore_eos = attr.ib(default=None, metadata={"description": "ignore end of stream token and continue generating", "long_name": "--ignore-eos"})
+    penalize_nl = attr.ib(default=None, metadata={"description": "penalize newline tokens", "long_name": "--penalize-nl", "default": False})
+    temp = attr.ib(default=None, metadata={"description": "temperature", "long_name": "--temp", "default": 0.8})
+    top_k = attr.ib(default=None, metadata={"description": "top-k sampling", "long_name": "--top-k", "default": 40})
+    top_p = attr.ib(default=None, metadata={"description": "top-p sampling", "long_name": "--top-p", "default": 0.9})
+    min_p = attr.ib(default=None, metadata={"description": "min-p sampling", "long_name": "--min-p", "default": 0.1})
+    tfs = attr.ib(default=None, metadata={"description": "tail free sampling, parameter z", "long_name": "--tfs", "default": 1.0})
+    typical = attr.ib(default=None, metadata={"description": "locally typical sampling, parameter p", "long_name": "--typical", "default": 1.0})
+    repeat_last_n = attr.ib(default=None, metadata={"description": "last n tokens to consider for penalize", "long_name": "--repeat-last-n", "default": 64})
+    repeat_penalty = attr.ib(default=None, metadata={"description": "penalize repeat sequence of tokens", "long_name": "--repeat-penalty", "default": 1.0})
+    presence_penalty = attr.ib(default=None, metadata={"description": "repeat alpha presence penalty", "long_name": "--presence-penalty", "default": 0.0})
+    frequency_penalty = attr.ib(default=None, metadata={"description": "repeat alpha frequency penalty", "long_name": "--frequency-penalty", "default": 0.0})
+    dynatemp_range = attr.ib(default=None, metadata={"description": "dynamic temperature range", "long_name": "--dynatemp-range", "default": 0.0})
+    dynatemp_exp = attr.ib(default=None, metadata={"description": "dynamic temperature exponent", "long_name": "--dynatemp-exp", "default": 1.0})
+    mirostat = attr.ib(default=None, metadata={"description": "use Mirostat sampling", "long_name": "--mirostat", "default": 0})
+    mirostat_lr = attr.ib(default=None, metadata={"description": "Mirostat learning rate, parameter eta", "long_name": "--mirostat-lr", "default": 0.1})
+    mirostat_ent = attr.ib(default=None, metadata={"description": "Mirostat target entropy, parameter tau", "long_name": "--mirostat-ent", "default": 5.0})
+    logit_bias = attr.ib(default=None, metadata={"description": "modifies the likelihood of token appearing in the completion", "long_name": "--logit-bias", "alias": "-l"})
+    cfg_negative_prompt = attr.ib(default=None, metadata={"description": "negative prompt to use for guidance", "long_name": "--cfg-negative-prompt", "default": ''})
+    cfg_negative_prompt_file = attr.ib(default=None, metadata={"description": "negative prompt file to use for guidance", "long_name": "--cfg-negative-prompt-file"})
+    cfg_scale = attr.ib(default=None, metadata={"description": "strength of guidance", "long_name": "--cfg-scale", "default": 1.0})
+    chat_template = attr.ib(default=None, metadata={"description": "set custom jinja chat template", "long_name": "--chat-template"})
+
+    grammar = attr.ib(default=None, metadata={"description": "BNF-like grammar to constrain generations", "long_name": "--grammar", "default": ''})
+    grammar_file = attr.ib(default=None, metadata={"description": "file to read grammar from", "long_name": "--grammar-file"})
+    json_schema = attr.ib(default=None, metadata={"description": "JSON schema to constrain generations", "long_name": "--json-schema", "alias": "-j"})
+
+    pooling = attr.ib(default=None, metadata={"description": "pooling type for embeddings", "long_name": "--pooling"})
+    attention = attr.ib(default=None, metadata={"description": "attention type for embeddings", "long_name": "--attention"})
+
+    rope_scaling = attr.ib(default=None, metadata={"description": "RoPE frequency scaling method", "long_name": "--rope-scaling"})
+    rope_scale = attr.ib(default=None, metadata={"description": "RoPE context scaling factor", "long_name": "--rope-scale"})
+    rope_freq_base = attr.ib(default=None, metadata={"description": "RoPE base frequency", "long_name": "--rope-freq-base"})
+    rope_freq_scale = attr.ib(default=None, metadata={"description": "RoPE frequency scaling factor", "long_name": "--rope-freq-scale"})
+    yarn_orig_ctx = attr.ib(default=None, metadata={"description": "YaRN: original context size of model", "long_name": "--yarn-orig-ctx"})
+    yarn_ext_factor = attr.ib(default=None, metadata={"description": "YaRN: extrapolation mix factor", "long_name": "--yarn-ext-factor"})
+    yarn_attn_factor = attr.ib(default=None, metadata={"description": "YaRN: scale sqrt(t) or attention magnitude", "long_name": "--yarn-attn-factor", "default": 1.0})
+    yarn_beta_slow = attr.ib(default=None, metadata={"description": "YaRN: high correction dim or alpha", "long_name": "--yarn-beta-slow", "default": 1.0})
+    yarn_beta_fast = attr.ib(default=None, metadata={"description": "YaRN: low correction dim or beta", "long_name": "--yarn-beta-fast", "default": 32.0})
+    grp_attn_n = attr.ib(default=None, metadata={"description": "group-attention factor", "long_name": "--grp-attn-n", "alias": "-gan", "default": 1})
+    grp_attn_w = attr.ib(default=None, metadata={"description": "group-attention width", "long_name": "--grp-attn-w", "alias": "-gaw", "default": 512.0})
+    dump_kv_cache = attr.ib(default=None, metadata={"description": "verbose print of the KV cache", "long_name": "--dump-kv-cache", "alias": "-dkvc"})
+    no_kv_offload = attr.ib(default=None, metadata={"description": "disable KV offload", "long_name": "--no-kv-offload", "alias": "-nkvo"})
+    cache_type_k = attr.ib(default=None, metadata={"description": "KV cache data type for K", "long_name": "--cache-type-k", "alias": "-ctk", "default": 'f16'})
+    cache_type_v = attr.ib(default=None, metadata={"description": "KV cache data type for V", "long_name": "--cache-type-v", "alias": "-ctv", "default": 'f16'})
+
+    all_logits = attr.ib(default=None, metadata={"description": "return logits for all tokens in the batch", "long_name": "--all-logits", "default": False})
+    hellaswag = attr.ib(default=None, metadata={"description": "compute HellaSwag score over random tasks from datafile supplied with -f", "long_name": "--hellaswag"})
+    hellaswag_tasks = attr.ib(default=None, metadata={"description": "number of tasks to use when computing the HellaSwag score", "long_name": "--hellaswag-tasks", "default": 400})
+    winogrande = attr.ib(default=None, metadata={"description": "compute Winogrande score over random tasks from datafile supplied with -f", "long_name": "--winogrande"})
+    winogrande_tasks = attr.ib(default=None, metadata={"description": "number of tasks to use when computing the Winogrande score", "long_name": "--winogrande-tasks", "default": 0})
+    multiple_choice = attr.ib(default=None, metadata={"description": "compute multiple choice score over random tasks from datafile supplied with -f", "long_name": "--multiple-choice"})
+    multiple_choice_tasks = attr.ib(default=None, metadata={"description": "number of tasks to use when computing the multiple choice score", "long_name": "--multiple-choice-tasks", "default": 0})
+    kl_divergence = attr.ib(default=None, metadata={"description": "computes KL-divergence to logits provided via --kl-divergence-base", "long_name": "--kl-divergence"})
+    ppl_stride = attr.ib(default=None, metadata={"description": "stride for perplexity calculation", "long_name": "--ppl-stride", "default": 0})
+    ppl_output_type = attr.ib(default=None, metadata={"description": "output type for perplexity calculation", "long_name": "--ppl-output-type", "default": 0})
+
+    defrag_thold = attr.ib(default=None, metadata={"description": "KV cache defragmentation threshold", "long_name": "--defrag-thold", "alias": "-dt", "default": -1.0})
+    parallel = attr.ib(default=None, metadata={"description": "number of parallel sequences to decode", "long_name": "--parallel", "alias": "-np", "default": 1})
+    sequences = attr.ib(default=None, metadata={"description": "number of sequences to decode", "long_name": "--sequences", "alias": "-ns", "default": 1})
+    cont_batching = attr.ib(default=None, metadata={"description": "enable continuous batching", "long_name": "--cont-batching", "alias": "-cb", "default": True})
+
+    mmproj = attr.ib(default=None, metadata={"description": "path to a multimodal projector file for LLaVA", "long_name": "--mmproj"})
+    image = attr.ib(default=None, metadata={"description": "path to an image file", "long_name": "--image"})
+
+    rpc = attr.ib(default=None, metadata={"description": "comma separated list of RPC servers", "long_name": "--rpc"})
+    mlock = attr.ib(default=None, metadata={"description": "force system to keep model in RAM rather than swapping or compressing", "long_name": "--mlock"})
+    no_mmap = attr.ib(default=None, metadata={"description": "do not memory-map model", "long_name": "--no-mmap"})
+    numa = attr.ib(default=None, metadata={"description": "attempt optimizations that help on some NUMA systems", "long_name": "--numa"})
+
+    check_tensors = attr.ib(default=None, metadata={"description": "check model tensor data for invalid values", "long_name": "--check-tensors", "default": False})
+    override_kv = attr.ib(default=None, metadata={"description": "advanced option to override model metadata by key", "long_name": "--override-kv"})
+    lora = attr.ib(default=None, metadata={"description": "apply LoRA adapter", "long_name": "--lora"})
+    lora_scaled = attr.ib(default=None, metadata={"description": "apply LoRA adapter with user defined scaling", "long_name": "--lora-scaled"})
+    lora_base = attr.ib(default=None, metadata={"description": "optional model to use as a base for the layers modified by the LoRA adapter", "long_name": "--lora-base"})
+    control_vector = attr.ib(default=None, metadata={"description": "add a control vector", "long_name": "--control-vector"})
+    control_vector_scaled = attr.ib(default=None, metadata={"description": "add a control vector with user defined scaling", "long_name": "--control-vector-scaled"})
+    control_vector_layer_range = attr.ib(default=None, metadata={"description": "layer range to apply the control vector(s) to", "long_name": "--control-vector-layer-range"})
+    model = attr.ib(default=None, metadata={"description": "model path", "long_name": "--model", "alias": "-m"})
+    model_draft = attr.ib(default=None, metadata={"description": "draft model for speculative decoding", "long_name": "--model-draft", "alias": "-md"})
+    model_url = attr.ib(default=None, metadata={"description": "model download url", "long_name": "--model-url", "alias": "-mu"})
+    hf_repo = attr.ib(default=None, metadata={"description": "Hugging Face model repository", "long_name": "--hf-repo", "alias": "-hfr"})
+    hf_file = attr.ib(default=None, metadata={"description": "Hugging Face model file", "long_name": "--hf-file", "alias": "-hff"})
+
+    context_file = attr.ib(default=None, metadata={"description": "file to load context from", "long_name": "--context-file"})
+    chunk_size = attr.ib(default=None, metadata={"description": "minimum length of embedded text chunks", "long_name": "--chunk-size", "default": 64})
+    chunk_separator = attr.ib(default=None, metadata={"description": "separator between chunks", "long_name": "--chunk-separator", "default": '\n'})
+
+    junk = attr.ib(default=None, metadata={"description": "number of times to repeat the junk text", "long_name": "--junk", "default": 250})
+    pos = attr.ib(default=None, metadata={"description": "position of the passkey in the junk text", "long_name": "--pos", "default": -1})
+
+    output = attr.ib(default=None, metadata={"description": "output file", "long_name": "--output", "alias": "-o", "default": 'imatrix.dat'})
+    output_frequency = attr.ib(default=None, metadata={"description": "output the imatrix every N iterations", "long_name": "--output-frequency", "default": 10})
+    save_frequency = attr.ib(default=None, metadata={"description": "save an imatrix copy every N iterations", "long_name": "--save-frequency", "default": 0})
+    process_output = attr.ib(default=None, metadata={"description": "collect data for the output tensor", "long_name": "--process-output", "default": False})
+    no_ppl = attr.ib(default=None, metadata={"description": "do not compute perplexity", "long_name": "--no-ppl", "default": True})
+    chunk = attr.ib(default=None, metadata={"description": "start processing the input from chunk N", "long_name": "--chunk", "default": 0})
+
+    pps = attr.ib(default=None, metadata={"description": "is the prompt shared across parallel sequences", "long_name": "-pps"})
+    npp = attr.ib(default=None, metadata={"description": "number of prompt tokens", "long_name": "-npp"})
+    ntg = attr.ib(default=None, metadata={"description": "number of text generation tokens", "long_name": "-ntg"})
+    npl = attr.ib(default=None, metadata={"description": "number of parallel prompts", "long_name": "-npl"})
+
+    embd_normalize = attr.ib(default=None, metadata={"description": "normalisation for embeddings", "long_name": "--embd-normalize", "default": 2})
+    embd_output_format = attr.ib(default=None, metadata={"description": "output format for embeddings", "long_name": "--embd-output-format"})
+    embd_separator = attr.ib(default=None, metadata={"description": "separator of embeddings", "long_name": "--embd-separator", "default": '\n'})
+
+    host = attr.ib(default=None, metadata={"description": "ip address to listen", "long_name": "--host", "default": '127.0.0.1'})
+    port = attr.ib(default=None, metadata={"description": "port to listen", "long_name": "--port", "default": 8080})
+    path = attr.ib(default=None, metadata={"description": "path to serve static files from", "long_name": "--path", "default": ''})
+    embedding = attr.ib(default=None, metadata={"description": "enable embedding endpoint", "long_name": "--embedding(s)", "default": False})
+    api_key = attr.ib(default=None, metadata={"description": "API key to use for authentication", "long_name": "--api-key"})
+    api_key_file = attr.ib(default=None, metadata={"description": "path to file containing API keys", "long_name": "--api-key-file"})
+    ssl_key_file = attr.ib(default=None, metadata={"description": "path to file a PEM-encoded SSL private key", "long_name": "--ssl-key-file"})
+    ssl_cert_file = attr.ib(default=None, metadata={"description": "path to file a PEM-encoded SSL certificate", "long_name": "--ssl-cert-file"})
+    timeout = attr.ib(default=None, metadata={"description": "server read/write timeout in seconds", "long_name": "--timeout", "default": 600})
+    threads_http = attr.ib(default=None, metadata={"description": "number of threads used to process HTTP requests", "long_name": "--threads-http", "default": -1})
+    system_prompt_file = attr.ib(default=None, metadata={"description": "set a file to load a system prompt", "long_name": "--system-prompt-file"})
+    log_format = attr.ib(default=None, metadata={"description": "log output format", "long_name": "--log-format", "default": 'json'})
+    metrics = attr.ib(default=None, metadata={"description": "enable prometheus compatible metrics endpoint", "long_name": "--metrics", "default": False})
+    no_slots = attr.ib(default=None, metadata={"description": "disables slots monitoring endpoint", "long_name": "--no-slots", "default": True})
+    slot_save_path = attr.ib(default=None, metadata={"description": "path to save slot kv cache", "long_name": "--slot-save-path"})
+    slot_prompt_similarity = attr.ib(default=None, metadata={"description": "how much the prompt of a request must match the prompt of a slot in order to use that slot", "long_name": "--slot-prompt-similarity", "alias": "-sps", "default": 0.50})
+
+    simple_io = attr.ib(default=None, metadata={"description": "use basic IO for better compatibility in subprocesses and limited consoles", "long_name": "--simple-io"})
+    logdir = attr.ib(default=None, metadata={"description": "path under which to save YAML logs", "long_name": "--logdir", "alias": "-ld"})
+    log_test = attr.ib(default=None, metadata={"description": "Run simple logging test", "long_name": "--log-test"})
+    log_disable = attr.ib(default=None, metadata={"description": "Disable trace logs", "long_name": "--log-disable"})
+    log_enable = attr.ib(default=None, metadata={"description": "Enable trace logs", "long_name": "--log-enable"})
+    log_file = attr.ib(default=None, metadata={"description": "Specify a log filename", "long_name": "--log-file"})
+    log_new = attr.ib(default=None, metadata={"description": "Create a separate new log file on start", "long_name": "--log-new"})
+    log_append = attr.ib(default=None, metadata={"description": "Don't truncate the old log file", "long_name": "--log-append"})
+
+    output = attr.ib(default=None, metadata={"description": "output file", "long_name": "--output", "alias": "-o", "default": 'control_vector.gguf'})
+    positive_file = attr.ib(default=None, metadata={"description": "positive prompts file", "long_name": "--positive-file", "default": 'examples/cvector-generator/positive.txt'})
+    negative_file = attr.ib(default=None, metadata={"description": "negative prompts file", "long_name": "--negative-file", "default": 'examples/cvector-generator/negative.txt'})
+    pca_batch = attr.ib(default=None, metadata={"description": "batch size used for PCA", "long_name": "--pca-batch", "default": 100})
+    pca_iter = attr.ib(default=None, metadata={"description": "number of iterations used for PCA", "long_name": "--pca-iter", "default": 1000})
+    method = attr.ib(default=None, metadata={"description": "dimensionality reduction method to be used", "long_name": "--method", "default": 'pca'})
+
+
+def convert_options_to_bytes(options: LlamaOptions) -> list[bytes]:
+    result = []
+
+    # Iterate over all attributes of the options class
+    for field in attr.fields(LlamaOptions):
+        value = getattr(options, field.name)
+        if value is not None:
+            long_name = field.metadata["long_name"]
+            alias = field.metadata.get("alias")
+
+            # Handle boolean options
+            if isinstance(value, bool):
+                if value:
+                    result.append(long_name.encode())
+            # Handle other options
+            else:
+                result.append(long_name.encode())
+                result.append(str(value).encode())
+
+    return result
diff --git a/scripts/build.py b/scripts/build.py
index d4b407b..57512ca 100644
--- a/scripts/build.py
+++ b/scripts/build.py
@@ -38,10 +38,10 @@
 
 
 def build(*args, **kwargs):
-    subprocess.run(['rm', '-rf', 'llama.cpp'], check=True)
-    subprocess.run(['git', 'clone', 'https://github.com/ggerganov/llama.cpp.git'], check=True)
-    subprocess.run(['patch', 'llama.cpp/examples/main/main.cpp', 'main_shared_library_1.patch'], check=True)
-    subprocess.run(['patch', 'llama.cpp/Makefile', 'makefile_static_library_0.patch'], check=True)
+    # subprocess.run(['rm', '-rf', 'llama.cpp'], check=True)
+    # subprocess.run(['git', 'clone', 'https://github.com/ggerganov/llama.cpp.git'], check=True)
+    # subprocess.run(['patch', 'llama.cpp/examples/main/main.cpp', 'main_shared_library_1.patch'], check=True)
+    # subprocess.run(['patch', 'llama.cpp/Makefile', 'makefile_static_library_0.patch'], check=True)
 
     # cffi
     env = os.environ.copy()
diff --git a/scripts/clean.py b/scripts/clean.py
index 3a18465..d5a7736 100644
--- a/scripts/clean.py
+++ b/scripts/clean.py
@@ -7,5 +7,5 @@ def clean():
     subprocess.run(['rm', '-fv'] + files, check=True)
     subprocess.run(['rm', '-fr', 'build'], check=True)
     subprocess.run(['rm', '-fr', 'dist'], check=True)
-    subprocess.run(['rm', '-fr', 'llama.cpp'], check=True)
+    # subprocess.run(['rm', '-fr', 'llama.cpp'], check=True)
     subprocess.run(['rm', '-fr', 'wheelhouse'], check=True)