diff --git a/.github/workflows/build-wheels-cuda.yaml b/.github/workflows/build-wheels-cuda.yaml index 3d410148f..745b2e602 100644 --- a/.github/workflows/build-wheels-cuda.yaml +++ b/.github/workflows/build-wheels-cuda.yaml @@ -61,11 +61,9 @@ jobs: - name: Setup Mamba uses: conda-incubator/setup-miniconda@v3.1.0 with: - activate-environment: "build" + activate-environment: "llamacpp" python-version: ${{ matrix.pyver }} - miniforge-variant: Mambaforge miniforge-version: latest - use-mamba: true add-pip-as-python-dependency: true auto-activate-base: false diff --git a/CHANGELOG.md b/CHANGELOG.md index 6f6355893..affbd5db7 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,25 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.3.9] + +- feat: Update llama.cpp to ggerganov/llama.cpp@8733e0cf6eefc7c7752297cc22d0836706f4222c + +## [0.3.8] + +- feat: Update llama.cpp to ggerganov/llama.cpp@7841fc723e059d1fd9640e5c0ef19050fcc7c698 + +## [0.3.7] + +- feat: Update llama.cpp to ggerganov/llama.cpp@794fe23f29fb40104975c91fe19f23798f7c726e +- fix(ci): Fix the CUDA workflow by @oobabooga in #1894 +- fix: error showing time spent in llama perf context print, adds `no_perf` flag to `Llama` class by @shakalaca in #1898 + +## [0.3.6] + +- feat: Update llama.cpp to ggerganov/llama.cpp@f7cd13301c2a88f97073fd119072b4cc92c08df1 +- fix(server): streaming resource lock by @gjpower in #1879 + ## [0.3.5] - feat: Update llama.cpp to ggerganov/llama.cpp@26a8406ba9198eb6fdd8329fa717555b4f77f05f diff --git a/CMakeLists.txt b/CMakeLists.txt index 64a0304a1..505c024b2 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -62,6 +62,9 @@ if (LLAMA_BUILD) # Enable building of the common library set(LLAMA_BUILD_COMMON ON CACHE BOOL "Build llama.cpp common library" FORCE) + # Disable building curl support + set(LLAMA_CURL OFF CACHE BOOL "llama.cpp: enable curl" FORCE) + # Architecture detection and settings for Apple platforms if (APPLE) # Get the target architecture @@ -93,7 +96,15 @@ if (LLAMA_BUILD) set(GGML_METAL_EMBED_LIBRARY "ON" CACHE BOOL "ggml: embed metal library" FORCE) endif() + add_subdirectory(vendor/llama.cpp) + + if (WIN32) + if (TARGET llama) + set_target_properties(llama PROPERTIES WINDOWS_EXPORT_ALL_SYMBOLS ON) + endif() + endif() + llama_cpp_python_install_target(llama) llama_cpp_python_install_target(ggml) @@ -132,46 +143,46 @@ if (LLAMA_BUILD) ) endif() - if (LLAVA_BUILD) - if (LLAMA_CUBLAS OR LLAMA_CUDA) - add_compile_definitions(GGML_USE_CUBLAS) - add_compile_definitions(GGML_USE_CUDA) - endif() - - if (LLAMA_METAL) - add_compile_definitions(GGML_USE_METAL) - endif() - - # Building llava - add_subdirectory(vendor/llama.cpp/examples/llava) - set_target_properties(llava_shared PROPERTIES OUTPUT_NAME "llava") - - if (WIN32) - set_target_properties(llava_shared PROPERTIES CUDA_ARCHITECTURES OFF) - endif() - llama_cpp_python_install_target(llava_shared) - if (WIN32) - install( - FILES $ - DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib - ) - install( - FILES $ - DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib - ) - endif() - - # Fix for llava build: Add include directory for llama.h - # Move these commands after the add_subdirectory call - target_include_directories(llava PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/include) - target_include_directories(llava PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/ggml/include) - - if (BUILD_SHARED_LIBS) - target_include_directories(llava_shared PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/include) - target_include_directories(llava_shared PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/ggml/include) - endif() - - target_include_directories(llama-llava-cli PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/include) - target_include_directories(llama-minicpmv-cli PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/include) - endif() + # if (LLAVA_BUILD) + # if (LLAMA_CUBLAS OR LLAMA_CUDA) + # add_compile_definitions(GGML_USE_CUBLAS) + # add_compile_definitions(GGML_USE_CUDA) + # endif() + # + # if (LLAMA_METAL) + # add_compile_definitions(GGML_USE_METAL) + # endif() + # + # # Building llava + # add_subdirectory(vendor/llama.cpp/tools/mtmd) + # set_target_properties(llava_shared PROPERTIES OUTPUT_NAME "llava") + # + # if (WIN32) + # set_target_properties(llava_shared PROPERTIES CUDA_ARCHITECTURES OFF) + # endif() + # llama_cpp_python_install_target(llava_shared) + # if (WIN32) + # install( + # FILES $ + # DESTINATION ${CMAKE_CURRENT_SOURCE_DIR}/llama_cpp/lib + # ) + # install( + # FILES $ + # DESTINATION ${SKBUILD_PLATLIB_DIR}/llama_cpp/lib + # ) + # endif() + # + # # Fix for llava build: Add include directory for llama.h + # # Move these commands after the add_subdirectory call + # target_include_directories(llava PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/include) + # target_include_directories(llava PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/ggml/include) + # + # if (BUILD_SHARED_LIBS) + # target_include_directories(llava_shared PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/include) + # target_include_directories(llava_shared PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/ggml/include) + # endif() + # + # target_include_directories(llama-llava-cli PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/include) + # target_include_directories(llama-minicpmv-cli PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/vendor/llama.cpp/include) + # endif() endif() diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index d7d9be881..2c9c527cd 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.3.5" +__version__ = "0.3.9" diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py index 8fa2b447f..18d733481 100644 --- a/llama_cpp/_internals.py +++ b/llama_cpp/_internals.py @@ -9,6 +9,8 @@ Tuple, Optional, Sequence, + Callable, + Union, ) from dataclasses import dataclass, field from contextlib import ExitStack @@ -48,43 +50,55 @@ def __init__( raise ValueError(f"Model path does not exist: {path_model}") with suppress_stdout_stderr(disable=verbose): - model = llama_cpp.llama_load_model_from_file( + model = llama_cpp.llama_model_load_from_file( self.path_model.encode("utf-8"), self.params ) if model is None: raise ValueError(f"Failed to load model from file: {path_model}") + vocab = llama_cpp.llama_model_get_vocab(model) + + if vocab is None: + raise ValueError(f"Failed to get vocab from model: {path_model}") + self.model = model + self.vocab = vocab + self.sampler = None # LlamaModel doesn't use samplers, but some cleanup code expects this attribute def free_model(): if self.model is None: return - llama_cpp.llama_free_model(self.model) + llama_cpp.llama_model_free(self.model) self.model = None self._exit_stack.callback(free_model) def close(self): + if self.sampler is not None: + # NOTE: Must remove custom samplers before free or llama.cpp will try to free them + for i, _ in reversed(self.custom_samplers): + llama_cpp.llama_sampler_chain_remove(self.sampler, i) + self.custom_samplers.clear() self._exit_stack.close() def __del__(self): self.close() def vocab_type(self) -> int: - return llama_cpp.llama_vocab_type(self.model) + return llama_cpp.llama_vocab_type(self.vocab) def n_vocab(self) -> int: - return llama_cpp.llama_n_vocab(self.model) + return llama_cpp.llama_vocab_n_tokens(self.vocab) def n_ctx_train(self) -> int: - return llama_cpp.llama_n_ctx_train(self.model) + return llama_cpp.llama_model_n_ctx_train(self.model) def n_embd(self) -> int: - return llama_cpp.llama_n_embd(self.model) + return llama_cpp.llama_model_n_embd(self.model) def rope_freq_scale_train(self) -> float: - return llama_cpp.llama_rope_freq_scale_train(self.model) + return llama_cpp.llama_model_rope_freq_scale_train(self.model) def desc(self) -> str: buf = ctypes.create_string_buffer(1024) @@ -98,53 +112,53 @@ def n_params(self) -> int: return llama_cpp.llama_model_n_params(self.model) def get_tensor(self, name: str) -> ctypes.c_void_p: - return llama_cpp.llama_get_model_tensor(self.model, name.encode("utf-8")) + raise NotImplementedError("get_tensor is not implemented in llama.cpp") # Vocab def token_get_text(self, token: int) -> str: - return llama_cpp.llama_token_get_text(self.model, token).decode("utf-8") + return llama_cpp.llama_vocab_get_text(self.vocab, token).decode("utf-8") def token_get_score(self, token: int) -> float: - return llama_cpp.llama_token_get_score(self.model, token) + return llama_cpp.llama_vocab_get_score(self.vocab, token) def token_get_attr(self, token: int) -> int: - return llama_cpp.llama_token_get_attr(self.model, token) + return llama_cpp.llama_vocab_get_attr(self.vocab, token) # Special tokens def token_bos(self) -> int: - return llama_cpp.llama_token_bos(self.model) + return llama_cpp.llama_vocab_bos(self.vocab) def token_eos(self) -> int: - return llama_cpp.llama_token_eos(self.model) + return llama_cpp.llama_vocab_eos(self.vocab) def token_cls(self) -> int: - return llama_cpp.llama_token_cls(self.model) + return llama_cpp.llama_vocab_cls(self.vocab) def token_sep(self) -> int: - return llama_cpp.llama_token_sep(self.model) + return llama_cpp.llama_vocab_sep(self.vocab) def token_nl(self) -> int: - return llama_cpp.llama_token_nl(self.model) + return llama_cpp.llama_vocab_nl(self.vocab) def token_prefix(self) -> int: - return llama_cpp.llama_token_prefix(self.model) + return llama_cpp.llama_vocab_fim_pre(self.vocab) def token_middle(self) -> int: - return llama_cpp.llama_token_middle(self.model) + return llama_cpp.llama_vocab_fim_mid(self.vocab) def token_suffix(self) -> int: - return llama_cpp.llama_token_suffix(self.model) + return llama_cpp.llama_vocab_fim_suf(self.vocab) def token_eot(self) -> int: - return llama_cpp.llama_token_eot(self.model) + return llama_cpp.llama_vocab_eot(self.vocab) def add_bos_token(self) -> bool: - return llama_cpp.llama_add_bos_token(self.model) + return llama_cpp.llama_vocab_get_add_bos(self.vocab) def add_eos_token(self) -> bool: - return llama_cpp.llama_add_eos_token(self.model) + return llama_cpp.llama_vocab_get_add_eos(self.vocab) # Tokenization @@ -152,13 +166,13 @@ def tokenize(self, text: bytes, add_bos: bool, special: bool): n_ctx = self.n_ctx_train() tokens = (llama_cpp.llama_token * n_ctx)() n_tokens = llama_cpp.llama_tokenize( - self.model, text, len(text), tokens, n_ctx, add_bos, special + self.vocab, text, len(text), tokens, n_ctx, add_bos, special ) if n_tokens < 0: n_tokens = abs(n_tokens) tokens = (llama_cpp.llama_token * n_tokens)() n_tokens = llama_cpp.llama_tokenize( - self.model, text, len(text), tokens, n_tokens, add_bos, special + self.vocab, text, len(text), tokens, n_tokens, add_bos, special ) if n_tokens < 0: raise RuntimeError( @@ -168,7 +182,7 @@ def tokenize(self, text: bytes, add_bos: bool, special: bool): def token_to_piece(self, token: int, special: bool = False) -> bytes: buf = ctypes.create_string_buffer(32) - llama_cpp.llama_token_to_piece(self.model, token, buf, 32, 0, special) + llama_cpp.llama_token_to_piece(self.vocab, token, buf, 32, 0, special) return bytes(buf) def detokenize(self, tokens: List[int], special: bool = False) -> bytes: @@ -177,7 +191,7 @@ def detokenize(self, tokens: List[int], special: bool = False) -> bytes: buffer = (ctypes.c_char * size)() for token in tokens: n = llama_cpp.llama_token_to_piece( - self.model, llama_cpp.llama_token(token), buffer, size, 0, special + self.vocab, llama_cpp.llama_token(token), buffer, size, 0, special ) assert n <= size output += bytes(buffer[:n]) @@ -243,12 +257,14 @@ def __init__( self.verbose = verbose self._exit_stack = ExitStack() - ctx = llama_cpp.llama_new_context_with_model(self.model.model, self.params) + ctx = llama_cpp.llama_init_from_model(self.model.model, self.params) if ctx is None: raise ValueError("Failed to create llama_context") self.ctx = ctx + self.memory = llama_cpp.llama_get_memory(self.ctx) + self.sampler = None # LlamaContext doesn't manage samplers directly, but some cleanup code expects this attribute def free_ctx(): if self.ctx is None: @@ -271,22 +287,22 @@ def pooling_type(self) -> int: return llama_cpp.llama_pooling_type(self.ctx) def kv_cache_clear(self): - llama_cpp.llama_kv_cache_clear(self.ctx) + llama_cpp.llama_memory_clear(self.memory, True) def kv_cache_seq_rm(self, seq_id: int, p0: int, p1: int): - llama_cpp.llama_kv_cache_seq_rm(self.ctx, seq_id, p0, p1) + llama_cpp.llama_memory_seq_rm(self.memory, seq_id, p0, p1) def kv_cache_seq_cp(self, seq_id_src: int, seq_id_dst: int, p0: int, p1: int): - llama_cpp.llama_kv_cache_seq_cp(self.ctx, seq_id_src, seq_id_dst, p0, p1) + llama_cpp.llama_memory_seq_cp(self.memory, seq_id_src, seq_id_dst, p0, p1) def kv_cache_seq_keep(self, seq_id: int): - llama_cpp.llama_kv_cache_seq_keep(self.ctx, seq_id) + llama_cpp.llama_memory_seq_keep(self.memory, seq_id) def kv_cache_seq_shift(self, seq_id: int, p0: int, p1: int, shift: int): - llama_cpp.llama_kv_cache_seq_add(self.ctx, seq_id, p0, p1, shift) + llama_cpp.llama_memory_seq_add(self.memory, seq_id, p0, p1, shift) def get_state_size(self) -> int: - return llama_cpp.llama_get_state_size(self.ctx) + return llama_cpp.llama_state_get_size(self.ctx) # TODO: copy_state_data @@ -304,6 +320,14 @@ def decode(self, batch: LlamaBatch): if return_code != 0: raise RuntimeError(f"llama_decode returned {return_code}") + def encode(self, batch: LlamaBatch): + return_code = llama_cpp.llama_encode( + self.ctx, + batch.batch, + ) + if return_code != 0: + raise RuntimeError(f"llama_encode returned {return_code}") + def set_n_threads(self, n_threads: int, n_threads_batch: int): llama_cpp.llama_set_n_threads(self.ctx, n_threads, n_threads_batch) @@ -316,11 +340,16 @@ def get_logits_ith(self, i: int): def get_embeddings(self): return llama_cpp.llama_get_embeddings(self.ctx) - # Sampling functions + def get_embeddings_ith(self, i: int): + return llama_cpp.llama_get_embeddings_ith(self.ctx, i) + + def get_embeddings_seq(self, seq_id: int): + return llama_cpp.llama_get_embeddings_seq(self.ctx, seq_id) + + # Sampling functions - deprecated, use LlamaSampler instead def set_rng_seed(self, seed: int): - # TODO: Fix - llama_cpp.llama_set_rng_seed(self.ctx, seed) + raise NotImplementedError("set_rng_seed is deprecated, use LlamaSampler instead") def sample_repetition_penalties( self, @@ -331,55 +360,30 @@ def sample_repetition_penalties( penalty_freq: float, penalty_present: float, ): - llama_cpp.llama_sample_repetition_penalties( - self.ctx, - llama_cpp.byref(candidates.candidates), - last_tokens_data, - penalty_last_n, - penalty_repeat, - penalty_freq, - penalty_present, - ) + raise NotImplementedError("sample_repetition_penalties is deprecated, use LlamaSampler instead") def sample_softmax(self, candidates: "_LlamaTokenDataArray"): - llama_cpp.llama_sample_softmax( - self.ctx, - llama_cpp.byref(candidates.candidates), - ) + raise NotImplementedError("sample_softmax is deprecated, use LlamaSampler instead") def sample_top_k(self, candidates: "_LlamaTokenDataArray", k: int, min_keep: int): - llama_cpp.llama_sample_top_k( - self.ctx, llama_cpp.byref(candidates.candidates), k, min_keep - ) + raise NotImplementedError("sample_top_k is deprecated, use LlamaSampler instead") def sample_top_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int): - llama_cpp.llama_sample_top_p( - self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep - ) + raise NotImplementedError("sample_top_p is deprecated, use LlamaSampler instead") def sample_min_p(self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int): - llama_cpp.llama_sample_min_p( - self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep - ) + raise NotImplementedError("sample_min_p is deprecated, use LlamaSampler instead") def sample_typical( self, candidates: "_LlamaTokenDataArray", p: float, min_keep: int ): - llama_cpp.llama_sample_typical( - self.ctx, llama_cpp.byref(candidates.candidates), p, min_keep - ) + raise NotImplementedError("sample_typical is deprecated, use LlamaSampler instead") def sample_temp(self, candidates: "_LlamaTokenDataArray", temp: float): - llama_cpp.llama_sample_temp( - self.ctx, llama_cpp.byref(candidates.candidates), temp - ) + raise NotImplementedError("sample_temp is deprecated, use LlamaSampler instead") def sample_grammar(self, candidates: "_LlamaTokenDataArray", grammar: LlamaGrammar): - llama_cpp.llama_sample_grammar( - self.ctx, - llama_cpp.byref(candidates.candidates), - grammar.grammar, - ) + raise NotImplementedError("sample_grammar is deprecated, use LlamaSampler instead") def sample_token_mirostat( self, @@ -389,14 +393,7 @@ def sample_token_mirostat( m: int, mu: llama_cpp.CtypesPointerOrRef[ctypes.c_float], ) -> int: - return llama_cpp.llama_sample_token_mirostat( - self.ctx, - llama_cpp.byref(candidates.candidates), - tau, - eta, - m, - mu, - ) + raise NotImplementedError("sample_token_mirostat is deprecated, use LlamaSampler instead") def sample_token_mirostat_v2( self, @@ -405,29 +402,17 @@ def sample_token_mirostat_v2( eta: float, mu: llama_cpp.CtypesPointerOrRef[ctypes.c_float], ) -> int: - return llama_cpp.llama_sample_token_mirostat_v2( - self.ctx, - llama_cpp.byref(candidates.candidates), - tau, - eta, - mu, - ) + raise NotImplementedError("sample_token_mirostat_v2 is deprecated, use LlamaSampler instead") def sample_token_greedy(self, candidates: "_LlamaTokenDataArray") -> int: - return llama_cpp.llama_sample_token_greedy( - self.ctx, - llama_cpp.byref(candidates.candidates), - ) + raise NotImplementedError("sample_token_greedy is deprecated, use LlamaSampler instead") def sample_token(self, candidates: "_LlamaTokenDataArray") -> int: - return llama_cpp.llama_sample_token( - self.ctx, - llama_cpp.byref(candidates.candidates), - ) + raise NotImplementedError("sample_token is deprecated, use LlamaSampler instead") # Grammar def grammar_accept_token(self, grammar: LlamaGrammar, token: int): - llama_cpp.llama_grammar_accept_token(grammar.grammar, self.ctx, token) + raise NotImplementedError("grammar_accept_token is deprecated, use LlamaSampler instead") def reset_timings(self): llama_cpp.llama_perf_context_reset(self.ctx) @@ -458,6 +443,7 @@ def __init__( raise ValueError("Failed to create llama_batch") self.batch = batch + self.sampler = None # LlamaBatch doesn't use samplers, but some cleanup code expects this attribute def free_batch(): if self.batch is None: @@ -520,6 +506,7 @@ def __init__(self, *, n_vocab: int): ) self.default_candidates_data_id = np.arange(self.n_vocab, dtype=np.intc) # type: ignore self.default_candidates_data_p = np.zeros(self.n_vocab, dtype=np.single) + self.sampler = None # LlamaTokenDataArray doesn't use samplers, but some cleanup code expects this attribute def copy_logits(self, logits: npt.NDArray[np.single]): self.candidates_data.id[:] = self.default_candidates_data_id @@ -608,103 +595,16 @@ def sample( idx: int = 0, logits_array: Optional[npt.NDArray[np.single]] = None, ): - n_vocab = ctx_main.model.n_vocab() - id: int = 0 - - if logits_array is None: - logits = ctx_main.get_logits_ith(idx) - logits_array = np.array( - ctypes.cast(logits, ctypes.POINTER(ctypes.c_float * n_vocab)).contents, - dtype=np.single, - ) - - # apply logit_bias - for token, logit_bias in self.params.logit_bias.items(): - logits_array[token] += logit_bias - - token_data_array = LlamaTokenDataArray( - n_vocab=n_vocab - ) # TODO: Only create this once - token_data_array.copy_logits(logits_array) - - # apply penalties - if len(self.prev) > 0: - nl_token = ctx_main.model.token_nl() - nl_logit = logits_array[nl_token] - last_tokens = self.prev[-self.params.penalty_last_n :] - last_tokens_size = min(len(last_tokens), self.params.penalty_last_n) - if last_tokens_size > 0: - last_tokens_p = (llama_cpp.llama_token * len(last_tokens))(*last_tokens) - ctx_main.sample_repetition_penalties( - token_data_array, - last_tokens_p, - last_tokens_size, - self.params.penalty_repeat, - self.params.penalty_freq, - self.params.penalty_present, - ) - if not self.params.penalize_nl: - token_data_array.candidates_data.logit[nl_token] = nl_logit - - if self.grammar is not None: - ctx_main.sample_grammar(token_data_array, self.grammar) - - if self.params.temp < 0: - ctx_main.sample_softmax(token_data_array) - id = token_data_array.candidates_data.id[0] - elif self.params.temp == 0: - id = ctx_main.sample_token_greedy(token_data_array) - else: - if self.params.mirostat == 1: - mirostat_m = 100 - ctx_main.sample_temp(token_data_array, self.params.temp) - id = ctx_main.sample_token_mirostat( - token_data_array, - self.params.mirostat_tau, - self.params.mirostat_eta, - mirostat_m, - ctypes.pointer(self.mirostat_mu), - ) - elif self.params.mirostat == 2: - ctx_main.sample_temp(token_data_array, self.params.temp) - id = ctx_main.sample_token_mirostat_v2( - token_data_array, - self.params.mirostat_tau, - self.params.mirostat_eta, - ctypes.pointer(self.mirostat_mu), - ) - else: - min_keep = max(1, self.params.n_probs) - ctx_main.sample_top_k( - token_data_array, self.params.top_k, min_keep=min_keep - ) - ctx_main.sample_typical( - token_data_array, self.params.typical_p, min_keep=min_keep - ) - ctx_main.sample_top_p( - token_data_array, self.params.top_p, min_keep=min_keep - ) - ctx_main.sample_min_p( - token_data_array, self.params.min_p, min_keep=min_keep - ) - ctx_main.sample_temp(token_data_array, self.params.temp) - id = ctx_main.sample_token(token_data_array) - return id + # This method is deprecated in favor of using LlamaSampler directly + raise NotImplementedError("LlamaSamplingContext.sample is deprecated, use LlamaSampler instead") def accept(self, ctx_main: LlamaContext, id: int, apply_grammar: bool): - if apply_grammar and self.grammar is not None: - ctx_main.grammar_accept_token(self.grammar, id) self.prev.append(id) -from typing import List, Callable, Optional, Union -import ctypes -import llama_cpp - - class CustomSampler: def __init__( - self, apply_func: typing.Callable[[llama_cpp.llama_token_data_array], None] + self, apply_func: Callable[[llama_cpp.llama_token_data_array], None] ): self.apply_func = apply_func @@ -737,72 +637,117 @@ def get_sampler(self) -> llama_cpp.llama_sampler_p: class LlamaSampler: def __init__(self): - params = llama_cpp.llama_sampler_chain_params() + params = llama_cpp.llama_sampler_chain_default_params() self.sampler = llama_cpp.llama_sampler_chain_init(params) - self.samplers: List[llama_cpp.llama_sampler_p] = [] self.custom_samplers: List[Tuple[int, CustomSampler]] = [] + self._exit_stack = ExitStack() + + def free_sampler(): + if self.sampler is not None: + # NOTE: Must remove custom samplers before free or llama.cpp will try to free them + for i, _ in reversed(self.custom_samplers): + llama_cpp.llama_sampler_chain_remove(self.sampler, i) + llama_cpp.llama_sampler_free(self.sampler) + self.sampler = None + + self._exit_stack.callback(free_sampler) + + def close(self): + self._exit_stack.close() + + def __del__(self): + self.close() def add_greedy(self): sampler = llama_cpp.llama_sampler_init_greedy() - self._add_sampler(sampler) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) def add_dist(self, seed: int): sampler = llama_cpp.llama_sampler_init_dist(seed) - self._add_sampler(sampler) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) def add_softmax(self): sampler = llama_cpp.llama_sampler_init_softmax() - self._add_sampler(sampler) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) def add_top_k(self, k: int): sampler = llama_cpp.llama_sampler_init_top_k(k) - self._add_sampler(sampler) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) - def add_top_p(self, p: float, min_keep: int): + def add_top_p(self, p: float, min_keep: int = 1): sampler = llama_cpp.llama_sampler_init_top_p(p, min_keep) - self._add_sampler(sampler) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) - def add_min_p(self, p: float, min_keep: int): + def add_min_p(self, p: float, min_keep: int = 1): sampler = llama_cpp.llama_sampler_init_min_p(p, min_keep) - self._add_sampler(sampler) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) - def add_typical(self, p: float, min_keep: int): + def add_typical(self, p: float, min_keep: int = 1): sampler = llama_cpp.llama_sampler_init_typical(p, min_keep) - self._add_sampler(sampler) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) def add_temp(self, temp: float): sampler = llama_cpp.llama_sampler_init_temp(temp) - self._add_sampler(sampler) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) def add_temp_ext(self, t: float, delta: float, exponent: float): sampler = llama_cpp.llama_sampler_init_temp_ext(t, delta, exponent) - self._add_sampler(sampler) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) + + def add_xtc(self, p: float, t: float, min_keep: int, seed: int): + sampler = llama_cpp.llama_sampler_init_xtc(p, t, min_keep, seed) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) + + def add_top_n_sigma(self, n: float): + sampler = llama_cpp.llama_sampler_init_top_n_sigma(n) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) def add_mirostat(self, n_vocab: int, seed: int, tau: float, eta: float, m: int): sampler = llama_cpp.llama_sampler_init_mirostat(n_vocab, seed, tau, eta, m) - self._add_sampler(sampler) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) def add_mirostat_v2(self, seed: int, tau: float, eta: float): sampler = llama_cpp.llama_sampler_init_mirostat_v2(seed, tau, eta) - self._add_sampler(sampler) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) def add_grammar(self, model: LlamaModel, grammar: LlamaGrammar): sampler = llama_cpp.llama_sampler_init_grammar( - model.model, grammar._grammar.encode("utf-8"), grammar._root.encode("utf-8") + model.vocab, grammar._grammar.encode("utf-8"), grammar._root.encode("utf-8") ) - self._add_sampler(sampler) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) + + def add_grammar_lazy_patterns( + self, + model: LlamaModel, + grammar: LlamaGrammar, + trigger_patterns: List[str], + trigger_tokens: List[int] + ): + # Convert patterns to C array + pattern_ptrs = (ctypes.c_char_p * len(trigger_patterns))() + for i, pattern in enumerate(trigger_patterns): + pattern_ptrs[i] = pattern.encode("utf-8") + + # Convert tokens to C array + token_array = (llama_cpp.llama_token * len(trigger_tokens))(*trigger_tokens) + + sampler = llama_cpp.llama_sampler_init_grammar_lazy_patterns( + model.vocab, + grammar._grammar.encode("utf-8"), + grammar._root.encode("utf-8"), + pattern_ptrs, + len(trigger_patterns), + token_array, + len(trigger_tokens) + ) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) def add_penalties( self, - n_vocab: int, - special_eos_id: int, - linefeed_id: int, penalty_last_n: int, penalty_repeat: float, penalty_freq: float, penalty_present: float, - penalize_nl: bool, - ignore_eos: bool, ): sampler = llama_cpp.llama_sampler_init_penalties( penalty_last_n, @@ -810,49 +755,96 @@ def add_penalties( penalty_freq, penalty_present, ) - self._add_sampler(sampler) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) + + def add_dry( + self, + model: LlamaModel, + n_ctx_train: int, + dry_multiplier: float, + dry_base: float, + dry_allowed_length: int, + dry_penalty_last_n: int, + seq_breakers: List[str] + ): + # Convert seq_breakers to C array + breaker_ptrs = (ctypes.c_char_p * len(seq_breakers))() + for i, breaker in enumerate(seq_breakers): + breaker_ptrs[i] = breaker.encode("utf-8") + + sampler = llama_cpp.llama_sampler_init_dry( + model.vocab, + n_ctx_train, + dry_multiplier, + dry_base, + dry_allowed_length, + dry_penalty_last_n, + breaker_ptrs, + len(seq_breakers) + ) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) - def init_logit_bias( - self, n_vocab: int, n_logit_bias, logit_bias: llama_cpp.llama_logit_bias_p + def add_logit_bias( + self, + n_vocab: int, + logit_bias: Dict[int, float] ): + # Convert logit_bias dict to C array + bias_array = (llama_cpp.llama_logit_bias * len(logit_bias))() + for i, (token, bias) in enumerate(logit_bias.items()): + bias_array[i].token = token + bias_array[i].bias = bias + sampler = llama_cpp.llama_sampler_init_logit_bias( - n_vocab, n_logit_bias, logit_bias + n_vocab, + len(logit_bias), + bias_array ) - self._add_sampler(sampler) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) + + def add_infill(self, model: LlamaModel): + sampler = llama_cpp.llama_sampler_init_infill(model.vocab) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) def add_custom( self, apply_func: Callable[[llama_cpp.llama_token_data_array], None] ): custom_sampler = CustomSampler(apply_func) sampler = custom_sampler.get_sampler() - self._add_sampler(sampler) + llama_cpp.llama_sampler_chain_add(self.sampler, sampler) # NOTE: Must remove custom samplers before free or llama.cpp will try to free them self.custom_samplers.append( (llama_cpp.llama_sampler_chain_n(self.sampler) - 1, custom_sampler) ) - def _add_sampler(self, sampler: llama_cpp.llama_sampler_p): - assert self.sampler is not None - llama_cpp.llama_sampler_chain_add(self.sampler, sampler) - self.samplers.append(sampler) - def get_seed(self) -> int: - assert self.sampler is not None return llama_cpp.llama_sampler_get_seed(self.sampler) - def sample(self, ctx: LlamaContext, idx: int) -> int: - assert self.sampler is not None + def sample(self, ctx: LlamaContext, idx: int = -1) -> int: return llama_cpp.llama_sampler_sample(self.sampler, ctx.ctx, idx) - def close(self): - if self.sampler: - # NOTE: Must remove custom samplers before free or llama.cpp will try to free them - for i, _ in reversed(self.custom_samplers): - llama_cpp.llama_sampler_chain_remove(self.sampler, i) - llama_cpp.llama_sampler_free(self.sampler) - self.sampler = None - self.samplers.clear() - self.custom_samplers.clear() + def accept(self, token: int): + llama_cpp.llama_sampler_accept(self.sampler, token) - def __del__(self): - self.close() + def reset(self): + llama_cpp.llama_sampler_reset(self.sampler) + + def clone(self): + # NOTE: Custom samplers cannot be cloned due to Python callback limitations + if self.custom_samplers: + raise NotImplementedError("Cannot clone LlamaSampler that contains custom samplers") + + cloned_sampler = llama_cpp.llama_sampler_clone(self.sampler) + # Create a new wrapper around the cloned sampler + new_sampler = LlamaSampler.__new__(LlamaSampler) + new_sampler.sampler = cloned_sampler + new_sampler.custom_samplers = [] + new_sampler._exit_stack = ExitStack() + + def free_sampler(): + if new_sampler.sampler is not None: + llama_cpp.llama_sampler_free(new_sampler.sampler) + new_sampler.sampler = None + + new_sampler._exit_stack.callback(free_sampler) + return new_sampler diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py index 2fd7ff193..cdc05c7ad 100644 --- a/llama_cpp/llama.py +++ b/llama_cpp/llama.py @@ -66,7 +66,6 @@ def __init__( split_mode: int = llama_cpp.LLAMA_SPLIT_MODE_LAYER, main_gpu: int = 0, tensor_split: Optional[List[float]] = None, - rpc_servers: Optional[str] = None, vocab_only: bool = False, use_mmap: bool = True, use_mlock: bool = False, @@ -93,7 +92,10 @@ def __init__( embedding: bool = False, offload_kqv: bool = True, flash_attn: bool = False, + op_offloat: Optional[bool] = None, + swa_full: Optional[bool] = None, # Sampling Params + no_perf: bool = False, last_n_tokens_size: int = 64, # LoRA Params lora_base: Optional[str] = None, @@ -149,7 +151,6 @@ def __init__( split_mode: How to split the model across GPUs. See llama_cpp.LLAMA_SPLIT_* for options. main_gpu: main_gpu interpretation depends on split_mode: LLAMA_SPLIT_MODE_NONE: the GPU that is used for the entire model. LLAMA_SPLIT_MODE_ROW: the GPU that is used for small tensors and intermediate results. LLAMA_SPLIT_MODE_LAYER: ignored tensor_split: How split tensors should be distributed across GPUs. If None, the model is not split. - rpc_servers: Comma separated list of RPC servers to use for offloading vocab_only: Only load the vocabulary no weights. use_mmap: Use mmap if possible. use_mlock: Force the system to keep the model in RAM. @@ -173,6 +174,9 @@ def __init__( embedding: Embedding mode only. offload_kqv: Offload K, Q, V to GPU. flash_attn: Use flash attention. + op_offloat: offload host tensor operations to device + swa_full: use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055) + no_perf: Measure performance timings. last_n_tokens_size: Maximum number of tokens to keep in the last_n_tokens deque. lora_base: Optional path to base model, useful if using a quantized base model and you want to apply LoRA to an f16 model. lora_path: Path to a LoRA file to apply to the model. @@ -224,11 +228,6 @@ def __init__( ) # 0x7FFFFFFF is INT32 max, will be auto set to all layers self.model_params.split_mode = split_mode self.model_params.main_gpu = main_gpu - if rpc_servers is not None: - self.model_params.rpc_servers = rpc_servers.encode("utf-8") - self._rpc_servers = rpc_servers - else: - self._rpc_servers = None self.tensor_split = tensor_split self._c_tensor_split = None if self.tensor_split is not None: @@ -339,18 +338,24 @@ def __init__( yarn_beta_slow if yarn_beta_slow != 0.0 else 0 ) self.context_params.yarn_orig_ctx = yarn_orig_ctx if yarn_orig_ctx != 0 else 0 - self.context_params.logits_all = ( - logits_all if draft_model is None else True - ) # Must be set to True for speculative decoding + self._logits_all = logits_all if draft_model is None else True self.context_params.embeddings = embedding # TODO: Rename to embeddings self.context_params.offload_kqv = offload_kqv self.context_params.flash_attn = flash_attn + + if op_offloat is not None: + self.context_params.op_offloat = op_offloat + + if swa_full is not None: + self.context_params.swa_full = swa_full + # KV cache quantization if type_k is not None: self.context_params.type_k = type_k if type_v is not None: self.context_params.type_v = type_v # Sampling Params + self.context_params.no_perf = no_perf self.last_n_tokens_size = last_n_tokens_size self.cache: Optional[BaseLlamaCache] = None @@ -406,10 +411,10 @@ def __init__( ) ) - self._lora_adapter: Optional[llama_cpp.llama_lora_adapter_p] = None + self._lora_adapter: Optional[llama_cpp.llama_adapter_lora_p] = None if self.lora_path: - self._lora_adapter = llama_cpp.llama_lora_adapter_init( + self._lora_adapter = llama_cpp.llama_adapter_lora_init( self._model.model, self.lora_path.encode("utf-8"), ) @@ -421,12 +426,12 @@ def __init__( def free_lora_adapter(): if self._lora_adapter is None: return - llama_cpp.llama_lora_adapter_free(self._lora_adapter) + llama_cpp.llama_adapter_lora_free(self._lora_adapter) self._lora_adapter = None self._stack.callback(free_lora_adapter) - if llama_cpp.llama_lora_adapter_set( + if llama_cpp.llama_set_adapter_lora( self._ctx.ctx, self._lora_adapter, self.lora_scale ): raise RuntimeError( @@ -565,7 +570,7 @@ def eval_tokens(self) -> Deque[int]: def eval_logits(self) -> Deque[List[float]]: return deque( self.scores[: self.n_tokens, :].tolist(), - maxlen=self._n_ctx if self.context_params.logits_all else 1, + maxlen=self._n_ctx if self._logits_all else 1, ) def tokenize( @@ -638,13 +643,13 @@ def eval(self, tokens: Sequence[int]): n_past = self.n_tokens n_tokens = len(batch) self._batch.set_batch( - batch=batch, n_past=n_past, logits_all=self.context_params.logits_all + batch=batch, n_past=n_past, logits_all=self._logits_all ) self._ctx.decode(self._batch) # Save tokens self.input_ids[n_past : n_past + n_tokens] = batch # Save logits - if self.context_params.logits_all: + if self._logits_all: rows = n_tokens cols = self._n_vocab logits = np.ctypeslib.as_array( @@ -706,15 +711,15 @@ def apply_func(token_data_array: llama_cpp.llama_token_data_array_p): sampler.add_custom(apply_func) sampler.add_penalties( - n_vocab=self._n_vocab, - special_eos_id=self._token_eos, - linefeed_id=self._token_nl, + # n_vocab=self._n_vocab, + # special_eos_id=self._token_eos, + # linefeed_id=self._token_nl, penalty_last_n=self.last_n_tokens_size, penalty_repeat=repeat_penalty, penalty_freq=frequency_penalty, penalty_present=presence_penalty, - penalize_nl=penalize_nl, - ignore_eos=False, + # penalize_nl=penalize_nl, + # ignore_eos=False, ) if grammar is not None: @@ -1152,9 +1157,9 @@ def _create_completion( bos_token_id: int = self.token_bos() cls_token_id: int = self._model.token_cls() sep_token_id: int = self._model.token_sep() - prefix_token_id: int = self._model.token_prefix() - middle_token_id: int = self._model.token_middle() - suffix_token_id: int = self._model.token_suffix() + prefix_token_id: int = 0 # self._model.token_prefix() # TODO: Fix + middle_token_id: int = 0 # self._model.token_middle() # TODO: Fix + suffix_token_id: int = 0 # self._model.token_suffix() # TODO: Fix add_space_prefix: bool = ( self.metadata.get("tokenizer.ggml.add_space_prefix", "true") == "true" ) @@ -1285,7 +1290,7 @@ def logit_bias_processor( else: stop_sequences = [] - if logprobs is not None and self.context_params.logits_all is False: + if logprobs is not None and self._logits_all is False: raise ValueError( "logprobs is not supported for models created with logits_all=False" ) @@ -1332,7 +1337,7 @@ def logit_bias_processor( logits_processor=logits_processor, grammar=grammar, ): - if llama_cpp.llama_token_is_eog(self._model.model, token): + if llama_cpp.llama_token_is_eog(self._model.vocab, token): text = self.detokenize(completion_tokens, prev_tokens=prompt_tokens) finish_reason = "stop" break @@ -2088,11 +2093,14 @@ def __getstate__(self): yarn_beta_fast=self.context_params.yarn_beta_fast, yarn_beta_slow=self.context_params.yarn_beta_slow, yarn_orig_ctx=self.context_params.yarn_orig_ctx, - logits_all=self.context_params.logits_all, + logits_all=self._logits_all, embedding=self.context_params.embeddings, offload_kqv=self.context_params.offload_kqv, flash_attn=self.context_params.flash_attn, + op_offloat=self.context_params.op_offloat, + swa_full=self.context_params.swa_full, # Sampling Params + no_perf=self.context_params.no_perf, last_n_tokens_size=self.last_n_tokens_size, # LoRA Params lora_base=self.lora_base, diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 0481cdbcd..d13d60458 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -149,6 +149,10 @@ # define LLAMA_STATE_SEQ_VERSION 2 LLAMA_STATE_SEQ_VERSION = 2 +# struct llama_vocab; +llama_vocab_p = NewType("llama_vocab_p", int) +llama_vocab_p_ctypes = ctypes.c_void_p + # struct llama_model; llama_model_p = NewType("llama_model_p", int) llama_model_p_ctypes = ctypes.c_void_p @@ -157,9 +161,13 @@ llama_context_p = NewType("llama_context_p", int) llama_context_p_ctypes = ctypes.c_void_p -# # struct llama_sampler; -# llama_sampler_p = NewType("llama_sampler_p", int) -# llama_sampler_p_ctypes = ctypes.c_void_p +# typedef struct llama_memory_i * llama_memory_t; +llama_memory_t = NewType("llama_memory_t", int) +llama_memory_t_ctypes = ctypes.c_void_p + +# struct llama_kv_cache; (DEPRECATED) +llama_kv_cache_p = NewType("llama_kv_cache_p", int) +llama_kv_cache_p_ctypes = ctypes.c_void_p # typedef int32_t llama_pos; llama_pos = ctypes.c_int32 @@ -222,6 +230,14 @@ # LLAMA_VOCAB_PRE_TYPE_EXAONE = 25, # LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26, # LLAMA_VOCAB_PRE_TYPE_MINERVA = 27, +# LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28, +# LLAMA_VOCAB_PRE_TYPE_GPT4O = 29, +# LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30, +# LLAMA_VOCAB_PRE_TYPE_TRILLION = 31, +# LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32, +# LLAMA_VOCAB_PRE_TYPE_LLAMA4 = 33, +# LLAMA_VOCAB_PRE_TYPE_PIXTRAL = 34, +# LLAMA_VOCAB_PRE_TYPE_SEED_CODER = 35, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -239,7 +255,7 @@ LLAMA_VOCAB_PRE_TYPE_DBRX = 13 LLAMA_VOCAB_PRE_TYPE_SMAUG = 14 LLAMA_VOCAB_PRE_TYPE_PORO = 15 -LLAMA_VOCAV_PRE_TYPE_CHATGLM3 = 16 +LLAMA_VOCAB_PRE_TYPE_CHATGLM3 = 16 LLAMA_VOCAB_PRE_TYPE_CHATGLM4 = 17 LLAMA_VOCAB_PRE_TYPE_VIKING = 18 LLAMA_VOCAB_PRE_TYPE_JAIS = 19 @@ -251,6 +267,14 @@ LLAMA_VOCAB_PRE_TYPE_EXAONE = 25 LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26 LLAMA_VOCAB_PRE_TYPE_MINERVA = 27 +LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28 +LLAMA_VOCAB_PRE_TYPE_GPT4O = 29 +LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30 +LLAMA_VOCAB_PRE_TYPE_TRILLION = 31 +LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32 +LLAMA_VOCAB_PRE_TYPE_LLAMA4 = 33 +LLAMA_VOCAB_PRE_TYPE_PIXTRAL = 34 +LLAMA_VOCAB_PRE_TYPE_SEED_CODER = 35 # // note: these values should be synchronized with ggml_rope @@ -399,14 +423,14 @@ # LLAMA_ROPE_SCALING_TYPE_LINEAR = 1, # LLAMA_ROPE_SCALING_TYPE_YARN = 2, # LLAMA_ROPE_SCALING_TYPE_LONGROPE = 3, -# LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_YARN, +# LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_LONGROPE, # }; LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED = -1 LLAMA_ROPE_SCALING_TYPE_NONE = 0 LLAMA_ROPE_SCALING_TYPE_LINEAR = 1 LLAMA_ROPE_SCALING_TYPE_YARN = 2 LLAMA_ROPE_SCALING_TYPE_LONGROPE = 3 -LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_YARN +LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_LONGROPE # enum llama_pooling_type { # LLAMA_POOLING_TYPE_UNSPECIFIED = -1, @@ -436,7 +460,7 @@ # enum llama_split_mode { # LLAMA_SPLIT_MODE_NONE = 0, // single GPU # LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs -# LLAMA_SPLIT_MODE_ROW = 2, // split rows across GPUs +# LLAMA_SPLIT_MODE_ROW = 2, // split layers and KV across GPUs, use tensor parallelism if supported # }; LLAMA_SPLIT_MODE_NONE = 0 LLAMA_SPLIT_MODE_LAYER = 1 @@ -510,18 +534,21 @@ class llama_token_data_array(ctypes.Structure): ) -# // Input data for llama_decode +# // Input data for llama_encode/llama_decode # // A llama_batch object can contain input about one or many sequences # // The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens # // # // - token : the token ids of the input (used when embd is NULL) # // - embd : token embeddings (i.e. float vector of size n_embd) (used when token is NULL) # // - pos : the positions of the respective token in the sequence -# // (if set to NULL, the token position will be tracked automatically by llama_decode) +# // (if set to NULL, the token position will be tracked automatically by llama_encode/llama_decode) # // - seq_id : the sequence to which the respective token belongs # // (if set to NULL, the sequence ID will be assumed to be 0) # // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output -# // (if set to NULL, only the logits for last token will be returned) +# // (if set to NULL: +# // - if embeddings: all tokens are output +# // - if not: only the last token is output +# // ) # // # typedef struct llama_batch { # int32_t n_tokens; @@ -531,10 +558,10 @@ class llama_token_data_array(ctypes.Structure): # llama_pos * pos; # int32_t * n_seq_id; # llama_seq_id ** seq_id; -# int8_t * logits; // TODO: rename this to "output" +# int8_t * logits; // TODO: rename this to "output" # } llama_batch; class llama_batch(ctypes.Structure): - """Input data for llama_decode + """Input data for llama_encode/llama_decode A llama_batch object can contain input about one or many sequences @@ -622,25 +649,28 @@ class llama_model_kv_override(ctypes.Structure): value: Union[int, float, bool, bytes] +# struct llama_model_tensor_buft_override { +# const char * pattern; +# ggml_backend_buffer_type_t buft; +# }; + + # struct llama_model_params { # // NULL-terminated list of devices to use for offloading (if NULL, all available devices are used) # ggml_backend_dev_t * devices; +# // NULL-terminated list of buffer types to use for tensors that match a pattern +# const struct llama_model_tensor_buft_override * tensor_buft_overrides; + # int32_t n_gpu_layers; // number of layers to store in VRAM # enum llama_split_mode split_mode; // how to split the model across multiple GPUs -# // main_gpu interpretation depends on split_mode: -# // LLAMA_SPLIT_MODE_NONE: the GPU that is used for the entire model -# // LLAMA_SPLIT_MODE_ROW: the GPU that is used for small tensors and intermediate results -# // LLAMA_SPLIT_MODE_LAYER: ignored +# // the GPU that is used for the entire model when split_mode is LLAMA_SPLIT_MODE_NONE # int32_t main_gpu; # // 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. @@ -652,7 +682,6 @@ class llama_model_kv_override(ctypes.Structure): # // override key-value pairs of the model meta data # const struct llama_model_kv_override * kv_overrides; - # // Keep the booleans together to avoid misalignment during copy-by-value. # bool vocab_only; // only load the vocabulary, no weights # bool use_mmap; // use mmap if possible @@ -663,11 +692,12 @@ class llama_model_params(ctypes.Structure): """Parameters for llama_model Attributes: + devices (ctypes.Array[ggml_backend_dev_t]): NULL-terminated list of devices to use for offloading (if NULL, all available devices are used) + tensor_buft_overrides (ctypes.Array[llama_model_tensor_buft_override]): NULL-terminated list of buffer types to use for tensors that match a pattern n_gpu_layers (int): number of layers to store in VRAM 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 + main_gpu (int): the GPU that is used for the entire model when split_mode is LLAMA_SPLIT_MODE_NONE 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 @@ -677,11 +707,12 @@ class llama_model_params(ctypes.Structure): check_tensors (bool): validate model tensor data""" if TYPE_CHECKING: + devices: CtypesArray[ctypes.c_void_p] # NOTE: unused + tensor_buft_overrides: CtypesArray[llama_model_tensor_buft_override] # NOTE: unused n_gpu_layers: int 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] @@ -692,11 +723,11 @@ class llama_model_params(ctypes.Structure): _fields_ = [ ("devices", ctypes.c_void_p), # NOTE: unnused + ("tensor_buft_overrides", ctypes.c_void_p), # NOTE: unused ("n_gpu_layers", ctypes.c_int32), ("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)), @@ -708,7 +739,7 @@ class llama_model_params(ctypes.Structure): # // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations -# // https://github.com/ggerganov/llama.cpp/pull/7544 +# // https://github.com/ggml-org/llama.cpp/pull/7544 # struct llama_context_params { # uint32_t n_ctx; // text context, 0 = from model # uint32_t n_batch; // logical maximum batch size that can be submitted to llama_decode @@ -721,7 +752,7 @@ class llama_model_params(ctypes.Structure): # enum llama_pooling_type pooling_type; // whether to pool (sum) embedding results by sequence id # enum llama_attention_type attention_type; // attention type to use for embeddings -# // ref: https://github.com/ggerganov/llama.cpp/pull/2054 +# // ref: https://github.com/ggml-org/llama.cpp/pull/2054 # float rope_freq_base; // RoPE base frequency, 0 = from model # float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model # float yarn_ext_factor; // YaRN extrapolation mix factor, negative = from model @@ -729,7 +760,7 @@ class llama_model_params(ctypes.Structure): # float yarn_beta_fast; // YaRN low correction dim # float yarn_beta_slow; // YaRN high correction dim # uint32_t yarn_orig_ctx; // YaRN original context size -# float defrag_thold; // defragment the KV cache if holes/size > thold, < 0 disabled (default) +# float defrag_thold; // defragment the KV cache if holes/size > thold, <= 0 disabled (default) # ggml_backend_sched_eval_callback cb_eval; # void * cb_eval_user_data; @@ -737,19 +768,21 @@ class llama_model_params(ctypes.Structure): # enum ggml_type type_k; // data type for K cache [EXPERIMENTAL] # enum ggml_type type_v; // data type for V cache [EXPERIMENTAL] -# // Keep the booleans together to avoid misalignment during copy-by-value. -# bool logits_all; // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) -# bool embeddings; // if true, extract embeddings (together with logits) -# bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU -# bool flash_attn; // whether to use flash attention [EXPERIMENTAL] -# bool no_perf; // whether to measure performance timings - - # // Abort callback # // if it returns true, execution of llama_decode() will be aborted # // currently works only with CPU execution # ggml_abort_callback abort_callback; # void * abort_callback_data; + +# // Keep the booleans together and at the end of the struct to avoid misalignment during copy-by-value. +# bool embeddings; // if true, extract embeddings (together with logits) +# bool offload_kqv; // offload the KQV ops (including the KV cache) to GPU +# bool flash_attn; // use flash attention [EXPERIMENTAL] +# bool no_perf; // measure performance timings +# bool op_offload; // offload host tensor operations to device +# bool swa_full; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055) +# // NOTE: setting to false when n_seq_max > 1 can cause bad performance in some cases +# // ref: https://github.com/ggml-org/llama.cpp/pull/13845#issuecomment-2924800573 # }; class llama_context_params(ctypes.Structure): """Parameters for llama_context @@ -771,17 +804,19 @@ class llama_context_params(ctypes.Structure): yarn_beta_fast (float): YaRN low correction dim yarn_beta_slow (float): YaRN high correction dim yarn_orig_ctx (int): YaRN original context size - defrag_thold (float): defragment the KV cache if holes/size > thold, < 0 disabled (default) + defrag_thold (float): defragment the KV cache if holes/size > thold, <= 0 disabled (default) cb_eval (ggml_backend_sched_eval_callback): callback for scheduling eval cb_eval_user_data (ctypes.ctypes.c_void_p): user data for cb_eval type_k (int): data type for K cache type_v (int): data type for V cache - logits_all (bool): the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) + abort_callback (ggml_abort_callback): abort callback if it returns true, execution of llama_decode() will be aborted + abort_callback_data (ctypes.ctypes.c_void_p): data for abort_callback embeddings (bool): if true, extract embeddings (together with logits) offload_kqv (bool): whether to offload the KQV ops (including the KV cache) to GPU flash_attn (bool): whether to use flash attention - abort_callback (ggml_abort_callback): abort callback if it returns true, execution of llama_decode() will be aborted - abort_callback_data (ctypes.ctypes.c_void_p): data for abort_callback + no_perf (bool): whether to measure performance timings + op_offload (bool): offload host tensor operations to device + swa_full (bool): use full-size SWA cache """ if TYPE_CHECKING: @@ -806,12 +841,14 @@ class llama_context_params(ctypes.Structure): cb_eval_user_data: ctypes.c_void_p type_k: int type_v: int - logits_all: bool + abort_callback: Callable[[ctypes.c_void_p], bool] + abort_callback_data: ctypes.c_void_p embeddings: bool offload_kqv: bool flash_attn: bool - abort_callback: Callable[[ctypes.c_void_p], bool] - abort_callback_data: ctypes.c_void_p + no_perf: bool + op_offload: bool + swa_full: bool _fields_ = [ ("n_ctx", ctypes.c_uint32), @@ -835,12 +872,14 @@ class llama_context_params(ctypes.Structure): ("cb_eval_user_data", ctypes.c_void_p), ("type_k", ctypes.c_int), ("type_v", ctypes.c_int), - ("logits_all", ctypes.c_bool), + ("abort_callback", ggml_abort_callback), + ("abort_callback_data", ctypes.c_void_p), ("embeddings", ctypes.c_bool), ("offload_kqv", ctypes.c_bool), ("flash_attn", ctypes.c_bool), - ("abort_callback", ggml_abort_callback), - ("abort_callback_data", ctypes.c_void_p), + ("no_perf", ctypes.c_bool), + ("op_offload", ctypes.c_bool), + ("swa_full", ctypes.c_bool), ] @@ -862,17 +901,19 @@ class llama_context_params(ctypes.Structure): # // model quantization parameters # typedef struct llama_model_quantize_params { -# int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() -# enum llama_ftype ftype; // quantize to this llama_ftype -# enum ggml_type output_tensor_type; // output tensor type -# enum ggml_type token_embedding_type; // token embeddings tensor type -# bool allow_requantize; // allow quantizing non-f32/f16 tensors -# bool quantize_output_tensor; // quantize output.weight -# bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored -# bool pure; // quantize all tensors to the default type -# bool keep_split; // quantize to the same number of shards -# void * imatrix; // pointer to importance matrix data -# void * kv_overrides; // pointer to vector containing overrides +# int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() +# enum llama_ftype ftype; // quantize to this llama_ftype +# enum ggml_type output_tensor_type; // output tensor type +# enum ggml_type token_embedding_type; // token embeddings tensor type +# bool allow_requantize; // allow quantizing non-f32/f16 tensors +# bool quantize_output_tensor; // quantize output.weight +# bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored +# bool pure; // quantize all tensors to the default type +# bool keep_split; // quantize to the same number of shards +# void * imatrix; // pointer to importance matrix data +# void * kv_overrides; // pointer to vector containing overrides +# void * tensor_types; // pointer to vector containing tensor types +# void * prune_layers; // pointer to vector containing layer indices to prune # } llama_model_quantize_params; class llama_model_quantize_params(ctypes.Structure): """Parameters for llama_model_quantize @@ -889,6 +930,8 @@ class llama_model_quantize_params(ctypes.Structure): keep_split (bool): quantize to the same number of shards imatrix (ctypes.c_void_p): pointer to importance matrix data kv_overrides (ctypes.c_void_p): pointer to vector containing overrides + tensor_types (ctypes.c_void_p): pointer to vector containing tensor types + prune_layers (ctypes.c_void_p): pointer to vector containing layer indices to prune """ if TYPE_CHECKING: @@ -903,6 +946,8 @@ class llama_model_quantize_params(ctypes.Structure): keep_split: bool imatrix: ctypes.c_void_p kv_overrides: ctypes.c_void_p + tensor_types: ctypes.c_void_p + prune_layers: ctypes.c_void_p _fields_ = [ ("nthread", ctypes.c_int32), @@ -916,6 +961,8 @@ class llama_model_quantize_params(ctypes.Structure): ("keep_split", ctypes.c_bool), ("imatrix", ctypes.c_void_p), ("kv_overrides", ctypes.c_void_p), + ("tensor_types", ctypes.c_void_p), + ("prune_layers", ctypes.c_void_p), ] @@ -973,9 +1020,9 @@ class llama_chat_message(ctypes.Structure): # // lora adapter -# struct llama_lora_adapter; -llama_lora_adapter_p = ctypes.c_void_p -llama_lora_adapter_p_ctypes = ctypes.POINTER(ctypes.c_void_p) +# struct llama_adapter_lora; +llama_adapter_lora_p = ctypes.c_void_p +llama_adapter_lora_p_ctypes = ctypes.POINTER(ctypes.c_void_p) # // Helpers for getting default parameters @@ -1026,7 +1073,6 @@ def llama_model_quantize_default_params() -> llama_model_quantize_params: # // Initialize the llama + ggml backend # // If numa is true, use NUMA optimizations # // Call once at the start of the program -# LLAMA_API void llama_backend_init(bool numa); # LLAMA_API void llama_backend_init(void); @ctypes_function( "llama_backend_init", @@ -1035,7 +1081,6 @@ def llama_model_quantize_default_params() -> llama_model_quantize_params: ) def llama_backend_init(): """Initialize the llama + ggml backend - If numa is true, use NUMA optimizations Call once at the start of the program""" ... @@ -1057,6 +1102,18 @@ def llama_backend_init(): GGML_NUMA_STRATEGY_COUNT = 5 +# // Call once at the end of the program - currently only used for MPI +# LLAMA_API void llama_backend_free(void); +@ctypes_function( + "llama_backend_free", + [], + None, +) +def llama_backend_free(): + """Call once at the end of the program - currently only used for MPI""" + ... + + # //optional: # LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa); @ctypes_function( @@ -1070,41 +1127,88 @@ def llama_numa_init(numa: int, /): # // Optional: an auto threadpool gets created in ggml if not passed explicitly # LLAMA_API void llama_attach_threadpool( -# struct llama_context * ctx, -# ggml_threadpool_t threadpool, -# ggml_threadpool_t threadpool_batch); +# struct llama_context * ctx, +# ggml_threadpool_t threadpool, +# ggml_threadpool_t threadpool_batch); +# TODO: Add llama_attach_threadpool # LLAMA_API void llama_detach_threadpool(struct llama_context * ctx); +# TODO: Add llama_detach_threadpool -# // Call once at the end of the program - currently only used for MPI -# LLAMA_API void llama_backend_free(void); +# DEPRECATED(LLAMA_API struct llama_model * llama_load_model_from_file( +# const char * path_model, +# struct llama_model_params params), +# "use llama_model_load_from_file instead"); @ctypes_function( - "llama_backend_free", - [], - None, + "llama_load_model_from_file", + [ctypes.c_char_p, llama_model_params], + llama_model_p_ctypes, ) -def llama_backend_free(): - """Call once at the end of the program - currently only used for MPI""" +def llama_load_model_from_file( + path_model: bytes, params: llama_model_params, / +) -> Optional[llama_model_p]: ... -# LLAMA_API struct llama_model * llama_load_model_from_file( +# // Load the model from a file +# // If the file is split into multiple parts, the file name must follow this pattern: -%05d-of-%05d.gguf +# // If the split file name does not follow this pattern, use llama_model_load_from_splits +# LLAMA_API struct llama_model * llama_model_load_from_file( # const char * path_model, # struct llama_model_params params); @ctypes_function( - "llama_load_model_from_file", + "llama_model_load_from_file", [ctypes.c_char_p, llama_model_params], llama_model_p_ctypes, ) -def llama_load_model_from_file( +def llama_model_load_from_file( path_model: bytes, params: llama_model_params, / ) -> Optional[llama_model_p]: + """Load the model from a file + + If the file is split into multiple parts, the file name must follow this pattern: -%05d-of-%05d.gguf + + If the split file name does not follow this pattern, use llama_model_load_from_splits""" + ... + + +# // Load the model from multiple splits (support custom naming scheme) +# // The paths must be in the correct order +# LLAMA_API struct llama_model * llama_model_load_from_splits( +# const char ** paths, +# size_t n_paths, +# struct llama_model_params params); +@ctypes_function( + "llama_model_load_from_splits", + [ctypes.POINTER(ctypes.c_char_p), ctypes.c_size_t, llama_model_params], + llama_model_p_ctypes, +) +def llama_model_load_from_splits( + paths: List[bytes], n_paths: int, params: llama_model_params, / +) -> Optional[llama_model_p]: + """Load the model from multiple splits (support custom naming scheme) + + The paths must be in the correct order""" + ... + + +# LLAMA_API void llama_model_save_to_file( +# const struct llama_model * model, +# const char * path_model); +@ctypes_function( + "llama_model_save_to_file", + [llama_model_p_ctypes, ctypes.c_char_p], + None, +) +def llama_model_save_to_file(model: llama_model_p, path_model: bytes, /): + """Save the model to a file""" ... -# LLAMA_API void llama_free_model(struct llama_model * model); +# DEPRECATED(LLAMA_API void llama_free_model(struct llama_model * model), +# "use llama_model_free instead"); @ctypes_function( "llama_free_model", [llama_model_p_ctypes], @@ -1114,9 +1218,34 @@ def llama_free_model(model: llama_model_p, /): ... -# LLAMA_API struct llama_context * llama_new_context_with_model( +# LLAMA_API void llama_model_free(struct llama_model * model); +@ctypes_function( + "llama_model_free", + [llama_model_p_ctypes], + None, +) +def llama_model_free(model: llama_model_p, /): + ... + + +# LLAMA_API struct llama_context * llama_init_from_model( # struct llama_model * model, # struct llama_context_params params); +@ctypes_function( + "llama_init_from_model", + [llama_model_p_ctypes, llama_context_params], + llama_context_p_ctypes, +) +def llama_init_from_model( + model: llama_model_p, params: llama_context_params, / +) -> Optional[llama_context_p]: + ... + + +# DEPRECATED(LLAMA_API struct llama_context * llama_new_context_with_model( +# struct llama_model * model, +# struct llama_context_params params), +# "use llama_init_from_model instead"); @ctypes_function( "llama_new_context_with_model", [llama_model_p_ctypes, llama_context_params], @@ -1156,6 +1285,12 @@ def llama_max_devices() -> int: ... +# LLAMA_API size_t llama_max_parallel_sequences(void); +@ctypes_function("llama_max_parallel_sequences", [], ctypes.c_size_t) +def llama_max_parallel_sequences() -> int: + ... + + # LLAMA_API bool llama_supports_mmap (void); @ctypes_function("llama_supports_mmap", [], ctypes.c_bool) def llama_supports_mmap() -> bool: @@ -1204,65 +1339,147 @@ def llama_n_seq_max(ctx: llama_context_p, /) -> int: ... -# LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); -@ctypes_function("llama_n_vocab", [llama_model_p_ctypes], ctypes.c_int32) -def llama_n_vocab(model: llama_model_p, /) -> int: - ... - - -# LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model); +# DEPRECATED(LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model), "use llama_model_n_ctx_train instead"); @ctypes_function("llama_n_ctx_train", [llama_model_p_ctypes], ctypes.c_int32) def llama_n_ctx_train(model: llama_model_p, /) -> int: ... -# LLAMA_API int32_t llama_n_embd (const struct llama_model * model); +# DEPRECATED(LLAMA_API int32_t llama_n_embd (const struct llama_model * model), "use llama_model_n_embd instead"); @ctypes_function("llama_n_embd", [llama_model_p_ctypes], ctypes.c_int32) def llama_n_embd(model: llama_model_p, /) -> int: ... -# LLAMA_API int32_t llama_n_layer (const struct llama_model * model); +# DEPRECATED(LLAMA_API int32_t llama_n_layer (const struct llama_model * model), "use llama_model_n_layer instead"); @ctypes_function("llama_n_layer", [llama_model_p_ctypes], ctypes.c_int32) def llama_n_layer(model: llama_model_p, /) -> int: ... -# LLAMA_API int32_t llama_n_head (const struct llama_model * model); +# DEPRECATED(LLAMA_API int32_t llama_n_head (const struct llama_model * model), "use llama_model_n_head instead"); @ctypes_function("llama_n_head", [llama_model_p_ctypes], ctypes.c_int32) def llama_n_head(model: llama_model_p, /) -> int: ... -# LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx); +# DEPRECATED(LLAMA_API int32_t llama_n_vocab (const struct llama_vocab * vocab), "use llama_vocab_n_tokens instead"); +@ctypes_function("llama_n_vocab", [llama_vocab_p_ctypes], ctypes.c_int32) +def llama_n_vocab(model: llama_vocab_p, /) -> int: + ... + + +# LLAMA_API const struct llama_model * llama_get_model (const struct llama_context * ctx); @ctypes_function("llama_get_model", [llama_context_p_ctypes], llama_model_p_ctypes) def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: ... -# LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); +# LLAMA_API llama_memory_t llama_get_memory (const struct llama_context * ctx); +@ctypes_function("llama_get_memory", [llama_context_p_ctypes], llama_memory_t_ctypes) +def llama_get_memory(ctx: llama_context_p, /) -> Optional[llama_memory_t]: + """Get the memory for the context""" + ... + + +# LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); @ctypes_function("llama_pooling_type", [llama_context_p_ctypes], ctypes.c_int) def llama_pooling_type(ctx: llama_context_p, /) -> int: ... -# LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); -@ctypes_function("llama_vocab_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_vocab_type(model: llama_model_p, /) -> int: +# DEPRECATED(LLAMA_API struct llama_kv_cache * llama_get_kv_self(struct llama_context * ctx), "use llama_get_memory instead"); +@ctypes_function( + "llama_get_kv_self", + [llama_context_p_ctypes], + llama_kv_cache_p_ctypes, +) +def llama_get_kv_self(ctx: llama_context_p, /) -> Optional[llama_kv_cache_p]: + """Get the KV cache for self-attention (DEPRECATED)""" + ... + + +# LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model); +@ctypes_function("llama_model_get_vocab", [llama_model_p_ctypes], llama_vocab_p_ctypes) +def llama_model_get_vocab(model: llama_model_p, /) -> Optional[llama_vocab_p]: + ... + + +# LLAMA_API enum llama_rope_type llama_model_rope_type(const struct llama_model * model); +@ctypes_function("llama_model_rope_type", [llama_model_p_ctypes], ctypes.c_int) +def llama_model_rope_type(model: llama_model_p, /) -> int: + ... + + +# LLAMA_API int32_t llama_model_n_ctx_train(const struct llama_model * model); +@ctypes_function("llama_model_n_ctx_train", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_ctx_train(model: llama_model_p, /) -> int: + ... + + +# LLAMA_API int32_t llama_model_n_embd (const struct llama_model * model); +@ctypes_function("llama_model_n_embd", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_embd(model: llama_model_p, /) -> int: + ... + + +# LLAMA_API int32_t llama_model_n_layer (const struct llama_model * model); +@ctypes_function("llama_model_n_layer", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_layer(model: llama_model_p, /) -> int: + ... + + +# LLAMA_API int32_t llama_model_n_head (const struct llama_model * model); +@ctypes_function("llama_model_n_head", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_head(model: llama_model_p, /) -> int: + ... + + +# LLAMA_API int32_t llama_model_n_head_kv (const struct llama_model * model); +@ctypes_function("llama_model_n_head_kv", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_head_kv(model: llama_model_p, /) -> int: ... -# LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); -@ctypes_function("llama_rope_type", [llama_model_p_ctypes], ctypes.c_int) -def llama_rope_type(model: llama_model_p, /) -> int: +# LLAMA_API int32_t llama_model_n_swa (const struct llama_model * model); +@ctypes_function("llama_model_n_swa", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_swa(model: llama_model_p, /) -> int: ... # // Get the model's RoPE frequency scaling factor -# LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model); -@ctypes_function("llama_rope_freq_scale_train", [llama_model_p_ctypes], ctypes.c_float) -def llama_rope_freq_scale_train(model: llama_model_p, /) -> float: - """Get the model's RoPE frequency scaling factor""" +# LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model); +@ctypes_function("llama_model_rope_freq_scale_train", [llama_model_p_ctypes], ctypes.c_float) +def llama_model_rope_freq_scale_train(model: llama_model_p, /) -> float: + ... + + +# // Returns the number of classifier outputs (only valid for classifier models) +# // Undefined behavior for non-classifier models +# LLAMA_API uint32_t llama_model_n_cls_out(const struct llama_model * model); +@ctypes_function("llama_model_n_cls_out", [llama_model_p_ctypes], ctypes.c_uint32) +def llama_model_n_cls_out(model: llama_model_p, /) -> int: + """Returns the number of classifier outputs (only valid for classifier models)""" + ... + + +# // Returns label of classifier output by index ( Optional[bytes]: + """Returns label of classifier output by index. Returns None if no label provided""" + ... + + +# LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); +@ctypes_function("llama_vocab_type", [llama_vocab_p_ctypes], ctypes.c_int) +def llama_vocab_type(vocab: llama_vocab_p, /) -> int: + ... + + +# LLAMA_API int32_t llama_vocab_n_tokens(const struct llama_vocab * vocab); +@ctypes_function("llama_vocab_n_tokens", [llama_vocab_p_ctypes], ctypes.c_int32) +def llama_vocab_n_tokens(vocab: llama_vocab_p, /) -> int: ... @@ -1375,6 +1592,16 @@ def llama_model_size(model: llama_model_p, /) -> int: ... +# // Get the default chat template. Returns nullptr if not available +# // If name is NULL, returns the default chat template +# LLAMA_API const char * llama_model_chat_template(const struct llama_model * model, const char * name); +@ctypes_function("llama_model_chat_template", [llama_model_p_ctypes, ctypes.c_char_p], ctypes.c_char_p) +def llama_model_chat_template(model: llama_model_p, name: Optional[bytes], /) -> Optional[bytes]: + """Get the default chat template. Returns None if not available + If name is None, returns the default chat template""" + ... + + # // Returns the total number of parameters in the model # LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model); @ctypes_function("llama_model_n_params", [llama_model_p_ctypes], ctypes.c_uint64) @@ -1444,38 +1671,53 @@ def llama_model_quantize( ... +# // +# // Adapters +# // + # // Load a LoRA adapter from file -# // The loaded adapter will be associated to the given model, and will be free when the model is deleted -# LLAMA_API struct llama_lora_adapter * llama_lora_adapter_init( +# LLAMA_API struct llama_adapter_lora * llama_adapter_lora_init( # struct llama_model * model, # const char * path_lora); @ctypes_function( - "llama_lora_adapter_init", + "llama_adapter_lora_init", [llama_model_p_ctypes, ctypes.c_char_p], - llama_lora_adapter_p_ctypes, + llama_adapter_lora_p_ctypes, ) -def llama_lora_adapter_init( +def llama_adapter_lora_init( model: llama_model_p, path_lora: bytes, / -) -> Optional[llama_lora_adapter_p]: - """Load a LoRA adapter from file - The loaded adapter will be associated to the given model, and will be free when the model is deleted - """ +) -> Optional[llama_adapter_lora_p]: + ... + + +# // Manually free a LoRA adapter +# // Note: loaded adapters will be free when the associated model is deleted +# LLAMA_API void llama_adapter_lora_free(struct llama_adapter_lora * adapter); +@ctypes_function( + "llama_adapter_lora_free", + [llama_adapter_lora_p_ctypes], + None, +) +def llama_adapter_lora_free(adapter: llama_adapter_lora_p, /): ... +# // The following functions operate on a llama_context, hence the naming: llama_verb_... + + # // Add a loaded LoRA adapter to given context # // This will not modify model's weight -# LLAMA_API int32_t llama_lora_adapter_set( +# LLAMA_API int32_t llama_set_adapter_lora( # struct llama_context * ctx, -# struct llama_lora_adapter * adapter, +# struct llama_adapter_lora * adapter, # float scale); @ctypes_function( - "llama_lora_adapter_set", - [llama_context_p_ctypes, llama_lora_adapter_p_ctypes, ctypes.c_float], + "llama_set_adapter_lora", + [llama_context_p_ctypes, llama_adapter_lora_p_ctypes, ctypes.c_float], ctypes.c_int32, ) -def llama_lora_adapter_set( - ctx: llama_context_p, adapter: llama_lora_adapter_p, scale: float, / +def llama_set_adapter_lora( + ctx: llama_context_p, adapter: llama_adapter_lora_p, scale: float, / ) -> int: """Add a loaded LoRA adapter to given context This will not modify model's weight""" @@ -1484,64 +1726,49 @@ def llama_lora_adapter_set( # // Remove a specific LoRA adapter from given context # // Return -1 if the adapter is not present in the context -# LLAMA_API int32_t llama_lora_adapter_remove( +# LLAMA_API int32_t llama_rm_adapter_lora( # struct llama_context * ctx, -# struct llama_lora_adapter * adapter); +# struct llama_adapter_lora * adapter); @ctypes_function( - "llama_lora_adapter_remove", - [llama_context_p_ctypes, llama_lora_adapter_p_ctypes], + "llama_rm_adapter_lora", + [llama_context_p_ctypes, llama_adapter_lora_p_ctypes], ctypes.c_int32, ) -def llama_lora_adapter_remove( - ctx: llama_context_p, adapter: llama_lora_adapter_p, / +def llama_rm_adapter_lora( + ctx: llama_context_p, adapter: llama_adapter_lora_p, / ) -> int: - """Remove a LoRA adapter from given context + """Remove a specific LoRA adapter from given context Return -1 if the adapter is not present in the context""" ... # // Remove all LoRA adapters from given context -# LLAMA_API void llama_lora_adapter_clear( -# struct llama_context * ctx); +# LLAMA_API void llama_clear_adapter_lora(struct llama_context * ctx); @ctypes_function( - "llama_lora_adapter_clear", + "llama_clear_adapter_lora", [llama_context_p_ctypes], None, ) -def llama_lora_adapter_clear(ctx: llama_context_p, /): +def llama_clear_adapter_lora(ctx: llama_context_p, /): """Remove all LoRA adapters from given context""" ... -# // Manually free a LoRA adapter -# // Note: loaded adapters will be free when the associated model is deleted -# LLAMA_API void llama_lora_adapter_free(struct llama_lora_adapter * adapter); -@ctypes_function( - "llama_lora_adapter_free", - [llama_lora_adapter_p_ctypes], - None, -) -def llama_lora_adapter_free(adapter: llama_lora_adapter_p, /): - """Manually free a LoRA adapter - Note: loaded adapters will be free when the associated model is deleted""" - ... - - # // Apply a loaded control vector to a llama_context, or if data is NULL, clear # // the currently loaded vector. # // n_embd should be the size of a single layer's control, and data should point # // to an n_embd x n_layers buffer starting from layer 1. # // il_start and il_end are the layer range the vector should apply to (both inclusive) # // See llama_control_vector_load in common to load a control vector. -# LLAMA_API int32_t llama_control_vector_apply( -# struct llama_context * lctx, +# LLAMA_API int32_t llama_apply_adapter_cvec( +# struct llama_context * ctx, # const float * data, # size_t len, # int32_t n_embd, # int32_t il_start, # int32_t il_end); @ctypes_function( - "llama_control_vector_apply", + "llama_apply_adapter_cvec", [ llama_context_p_ctypes, ctypes.POINTER(ctypes.c_float), @@ -1552,8 +1779,8 @@ def llama_lora_adapter_free(adapter: llama_lora_adapter_p, /): ], ctypes.c_int32, ) -def llama_control_vector_apply( - lctx: llama_context_p, +def llama_apply_adapter_cvec( + ctx: llama_context_p, data: CtypesPointerOrRef[ctypes.c_float], len: int, n_embd: int, @@ -1571,148 +1798,22 @@ def llama_control_vector_apply( # // -# // KV cache +# // Memory # // - -# // Information associated with an individual cell in the KV cache view. -# struct llama_kv_cache_view_cell { -# // The position for this cell. Takes KV cache shifts into account. -# // May be negative if the cell is not populated. -# llama_pos pos; -# }; -class llama_kv_cache_view_cell(ctypes.Structure): - """Information associated with an individual cell in the KV cache view. - - Attributes: - pos (llama_pos): The position for this cell. Takes KV cache shifts into account. - May be negative if the cell is not populated.""" - - if TYPE_CHECKING: - pos: llama_pos - - _fields_ = [("pos", llama_pos)] - - -# // An updateable view of the KV cache. -# struct llama_kv_cache_view { -# // Number of KV cache cells. This will be the same as the context size. -# int32_t n_cells; - -# // Maximum number of sequences that can exist in a cell. It's not an error -# // if there are more sequences in a cell than this value, however they will -# // not be visible in the view cells_sequences. -# int32_t n_seq_max; - -# // Number of tokens in the cache. For example, if there are two populated -# // cells, the first with 1 sequence id in it and the second with 2 sequence -# // ids then you'll have 3 tokens. -# int32_t token_count; - -# // Number of populated cache cells. -# int32_t used_cells; - -# // Maximum contiguous empty slots in the cache. -# int32_t max_contiguous; - -# // Index to the start of the max_contiguous slot range. Can be negative -# // when cache is full. -# int32_t max_contiguous_idx; - -# // Information for an individual cell. -# struct llama_kv_cache_view_cell * cells; - - -# // The sequences for each cell. There will be n_seq_max items per cell. -# llama_seq_id * cells_sequences; -# }; -class llama_kv_cache_view(ctypes.Structure): - if TYPE_CHECKING: - n_cells: int - n_max_seq: int - token_count: int - used_cells: int - max_contiguous: int - max_contiguous_idx: int - cells: CtypesArray[llama_kv_cache_view_cell] - cells_sequences: CtypesArray[llama_seq_id] - - _fields_ = [ - ("n_cells", ctypes.c_int32), - ("n_max_seq", ctypes.c_int32), - ("token_count", ctypes.c_int32), - ("used_cells", ctypes.c_int32), - ("max_contiguous", ctypes.c_int32), - ("max_contiguous_idx", ctypes.c_int32), - ("cells", ctypes.POINTER(llama_kv_cache_view_cell)), - ("cells_sequences", ctypes.POINTER(llama_seq_id)), - ] - - -llama_kv_cache_view_p = ctypes.POINTER(llama_kv_cache_view) - - -# // Create an empty KV cache view. (use only for debugging purposes) -# LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_seq_max); -@ctypes_function( - "llama_kv_cache_view_init", - [llama_context_p_ctypes, ctypes.c_int32], - llama_kv_cache_view, -) -def llama_kv_cache_view_init( - ctx: llama_context_p, n_seq_max: Union[ctypes.c_int32, int], / -) -> llama_kv_cache_view: - """Create an empty KV cache view. (use only for debugging purposes)""" - ... - - -# // Free a KV cache view. (use only for debugging purposes) -# LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view); -@ctypes_function("llama_kv_cache_view_free", [llama_kv_cache_view_p], None) -def llama_kv_cache_view_free(view: "ctypes.pointer[llama_kv_cache_view]", /): # type: ignore - """Free a KV cache view. (use only for debugging purposes)""" - ... - - -# // Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes) -# LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view); -@ctypes_function( - "llama_kv_cache_view_update", [llama_context_p_ctypes, llama_kv_cache_view_p], None -) -def llama_kv_cache_view_update(ctx: llama_context_p, view: CtypesPointerOrRef[llama_kv_cache_view], /): # type: ignore - """Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes)""" - ... - - -# // Returns the number of tokens in the KV cache (slow, use only for debug) -# // If a KV cell has multiple sequences assigned to it, it will be counted multiple times -# LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx); -@ctypes_function( - "llama_get_kv_cache_token_count", [llama_context_p_ctypes], ctypes.c_int32 -) -def llama_get_kv_cache_token_count(ctx: llama_context_p, /) -> int: - """Returns the number of tokens in the KV cache (slow, use only for debug) - If a KV cell has multiple sequences assigned to it, it will be counted multiple times - """ - ... - - -# // Returns the number of used KV cells (i.e. have at least one sequence assigned to them) -# LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx); +# // Clear the memory contents +# // If data == true, the data buffers will also be cleared together with the metadata +# LLAMA_API void llama_memory_clear( +# llama_memory_t mem, +# bool data); @ctypes_function( - "llama_get_kv_cache_used_cells", [llama_context_p_ctypes], ctypes.c_int32 + "llama_memory_clear", + [llama_memory_t_ctypes, ctypes.c_bool], + None, ) -def llama_get_kv_cache_used_cells(ctx: llama_context_p, /) -> int: - """Returns the number of used KV cells (i.e. have at least one sequence assigned to them)""" - ... - - -# // Clear the KV cache - both cell info is erased and KV data is zeroed -# LLAMA_API void llama_kv_cache_clear( -# struct llama_context * ctx); -@ctypes_function("llama_kv_cache_clear", [llama_context_p_ctypes], None) -def llama_kv_cache_clear(ctx: llama_context_p, /): - """Clear the KV cache""" +def llama_memory_clear(mem: llama_memory_t, data: bool, /): + """Clear the memory contents + If data == true, the data buffers will also be cleared together with the metadata""" ... @@ -1721,23 +1822,23 @@ def llama_kv_cache_clear(ctx: llama_context_p, /): # // seq_id < 0 : match any sequence # // p0 < 0 : [0, p1] # // p1 < 0 : [p0, inf) -# LLAMA_API bool llama_kv_cache_seq_rm( -# struct llama_context * ctx, -# llama_seq_id seq_id, -# llama_pos p0, -# llama_pos p1); +# LLAMA_API bool llama_memory_seq_rm( +# llama_memory_t mem, +# llama_seq_id seq_id, +# llama_pos p0, +# llama_pos p1); @ctypes_function( - "llama_kv_cache_seq_rm", + "llama_memory_seq_rm", [ - llama_context_p_ctypes, + llama_memory_t_ctypes, llama_seq_id, llama_pos, llama_pos, ], ctypes.c_bool, ) -def llama_kv_cache_seq_rm( - ctx: llama_context_p, +def llama_memory_seq_rm( + mem: llama_memory_t, seq_id: Union[llama_seq_id, int], p0: Union[llama_pos, int], p1: Union[llama_pos, int], @@ -1754,19 +1855,18 @@ def llama_kv_cache_seq_rm( # // Copy all tokens that belong to the specified sequence to another sequence -# // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence # // p0 < 0 : [0, p1] # // p1 < 0 : [p0, inf) -# LLAMA_API void llama_kv_cache_seq_cp( -# struct llama_context * ctx, -# llama_seq_id seq_id_src, -# llama_seq_id seq_id_dst, -# llama_pos p0, -# llama_pos p1); +# LLAMA_API void llama_memory_seq_cp( +# llama_memory_t mem, +# llama_seq_id seq_id_src, +# llama_seq_id seq_id_dst, +# llama_pos p0, +# llama_pos p1); @ctypes_function( - "llama_kv_cache_seq_cp", + "llama_memory_seq_cp", [ - llama_context_p_ctypes, + llama_memory_t_ctypes, llama_seq_id, llama_seq_id, llama_pos, @@ -1774,8 +1874,8 @@ def llama_kv_cache_seq_rm( ], None, ) -def llama_kv_cache_seq_cp( - ctx: llama_context_p, +def llama_memory_seq_cp( + mem: llama_memory_t, seq_id_src: Union[llama_seq_id, int], seq_id_dst: Union[llama_seq_id, int], p0: Union[llama_pos, int], @@ -1783,40 +1883,36 @@ def llama_kv_cache_seq_cp( /, ): """Copy all tokens that belong to the specified sequence to another sequence - Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence p0 < 0 : [0, p1] p1 < 0 : [p0, inf)""" ... # // Removes all tokens that do not belong to the specified sequence -# LLAMA_API void llama_kv_cache_seq_keep( -# struct llama_context * ctx, -# llama_seq_id seq_id); +# LLAMA_API void llama_memory_seq_keep( +# llama_memory_t mem, +# llama_seq_id seq_id); @ctypes_function( - "llama_kv_cache_seq_keep", [llama_context_p_ctypes, llama_seq_id], None + "llama_memory_seq_keep", [llama_memory_t_ctypes, llama_seq_id], None ) -def llama_kv_cache_seq_keep(ctx: llama_context_p, seq_id: Union[llama_seq_id, int], /): +def llama_memory_seq_keep(mem: llama_memory_t, seq_id: Union[llama_seq_id, int], /): """Removes all tokens that do not belong to the specified sequence""" ... # // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) -# // If the KV cache is RoPEd, the KV data is updated accordingly: -# // - lazily on next llama_decode() -# // - explicitly with llama_kv_cache_update() # // p0 < 0 : [0, p1] # // p1 < 0 : [p0, inf) -# LLAMA_API void llama_kv_cache_seq_add( -# struct llama_context * ctx, -# llama_seq_id seq_id, -# llama_pos p0, -# llama_pos p1, -# llama_pos delta); +# LLAMA_API void llama_memory_seq_add( +# llama_memory_t mem, +# llama_seq_id seq_id, +# llama_pos p0, +# llama_pos p1, +# llama_pos delta); @ctypes_function( - "llama_kv_cache_seq_add", + "llama_memory_seq_add", [ - llama_context_p_ctypes, + llama_memory_t_ctypes, llama_seq_id, llama_pos, llama_pos, @@ -1824,8 +1920,8 @@ def llama_kv_cache_seq_keep(ctx: llama_context_p, seq_id: Union[llama_seq_id, in ], None, ) -def llama_kv_cache_seq_add( - ctx: llama_context_p, +def llama_memory_seq_add( + mem: llama_memory_t, seq_id: Union[llama_seq_id, int], p0: Union[llama_pos, int], p1: Union[llama_pos, int], @@ -1833,28 +1929,24 @@ def llama_kv_cache_seq_add( /, ): """Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) - If the KV cache is RoPEd, the KV data is updated accordingly: - - lazily on next llama_decode() - - explicitly with llama_kv_cache_update() p0 < 0 : [0, p1] p1 < 0 : [p0, inf)""" ... # // Integer division of the positions by factor of `d > 1` -# // If the KV cache is RoPEd, the KV data is updated accordingly # // p0 < 0 : [0, p1] # // p1 < 0 : [p0, inf) -# LLAMA_API void llama_kv_cache_seq_div( -# struct llama_context * ctx, -# llama_seq_id seq_id, -# llama_pos p0, -# llama_pos p1, -# int d); +# LLAMA_API void llama_memory_seq_div( +# llama_memory_t mem, +# llama_seq_id seq_id, +# llama_pos p0, +# llama_pos p1, +# int d); @ctypes_function( - "llama_kv_cache_seq_div", + "llama_memory_seq_div", [ - llama_context_p_ctypes, + llama_memory_t_ctypes, llama_seq_id, llama_pos, llama_pos, @@ -1862,8 +1954,8 @@ def llama_kv_cache_seq_add( ], None, ) -def llama_kv_cache_seq_div( - ctx: llama_context_p, +def llama_memory_seq_div( + mem: llama_memory_t, seq_id: Union[llama_seq_id, int], p0: Union[llama_pos, int], p1: Union[llama_pos, int], @@ -1871,39 +1963,304 @@ def llama_kv_cache_seq_div( /, ): """Integer division of the positions by factor of `d > 1` - If the KV cache is RoPEd, the KV data is updated accordingly p0 < 0 : [0, p1] p1 < 0 : [p0, inf)""" ... -# // Defragment the KV cache -# // This will be applied: -# // - lazily on next llama_decode() -# // - explicitly with llama_kv_cache_update() -# LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx); -@ctypes_function("llama_kv_cache_defrag", [llama_context_p_ctypes], None) -def llama_kv_cache_defrag(ctx: llama_context_p, /): - """Defragment the KV cache - This will be applied: - - lazily on next llama_decode() - - explicitly with llama_kv_cache_update()""" - ... +# // Returns the smallest position present in the memory for the specified sequence +# // This is typically non-zero only for SWA caches +# // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the memory +# // Return -1 if the sequence is empty +# LLAMA_API llama_pos llama_memory_seq_pos_min( +# llama_memory_t mem, +# llama_seq_id seq_id); +@ctypes_function( + "llama_memory_seq_pos_min", [llama_memory_t_ctypes, llama_seq_id], llama_pos +) +def llama_memory_seq_pos_min( + mem: llama_memory_t, seq_id: Union[llama_seq_id, int], / +) -> int: + """Returns the smallest position present in the memory for the specified sequence + This is typically non-zero only for SWA caches + Return -1 if the sequence is empty""" + ... -# // Apply the KV cache updates (such as K-shifts, defragmentation, etc.) -# LLAMA_API void llama_kv_cache_update(struct llama_context * ctx); -@ctypes_function("llama_kv_cache_update", [llama_context_p_ctypes], None) -def llama_kv_cache_update(ctx: llama_context_p, /): - """Apply the KV cache updates (such as K-shifts, defragmentation, etc.)""" +# // Returns the largest position present in the memory for the specified sequence +# // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the memory +# // Return -1 if the sequence is empty +# LLAMA_API llama_pos llama_memory_seq_pos_max( +# llama_memory_t mem, +# llama_seq_id seq_id); +@ctypes_function( + "llama_memory_seq_pos_max", [llama_memory_t_ctypes, llama_seq_id], llama_pos +) +def llama_memory_seq_pos_max( + mem: llama_memory_t, seq_id: Union[llama_seq_id, int], / +) -> int: + """Returns the largest position present in the memory for the specified sequence + Return -1 if the sequence is empty""" + ... + + +# // Check if the memory supports shifting +# LLAMA_API bool llama_memory_can_shift(llama_memory_t mem); +@ctypes_function("llama_memory_can_shift", [llama_memory_t_ctypes], ctypes.c_bool) +def llama_memory_can_shift(mem: llama_memory_t, /) -> bool: + """Check if the memory supports shifting""" + ... + + +# // +# // KV cache for self-attention (TODO: deprecate in favor of llama_memory) +# // + +# // Returns the number of tokens in the KV cache (slow, use only for debug) +# // If a KV cell has multiple sequences assigned to it, it will be counted multiple times +# DEPRECATED(LLAMA_API int32_t llama_kv_self_n_tokens(const struct llama_context * ctx), +# "Use llama_kv_self_seq_pos_max() and llama_kv_self_seq_pos_min() instead (https://github.com/ggml-org/llama.cpp/issues/13793)"); +@ctypes_function( + "llama_kv_self_n_tokens", [llama_context_p_ctypes], ctypes.c_int32 +) +def llama_kv_self_n_tokens(ctx: llama_context_p, /) -> int: + """Returns the number of tokens in the KV cache (slow, use only for debug) (DEPRECATED)""" + ... + + +# // Returns the number of used KV cells (i.e. have at least one sequence assigned to them) +# DEPRECATED(LLAMA_API int32_t llama_kv_self_used_cells(const struct llama_context * ctx), +# "Use llama_kv_self_seq_pos_max() and llama_kv_self_seq_pos_min() instead (https://github.com/ggml-org/llama.cpp/issues/13793)"); +@ctypes_function( + "llama_kv_self_used_cells", [llama_context_p_ctypes], ctypes.c_int32 +) +def llama_kv_self_used_cells(ctx: llama_context_p, /) -> int: + """Returns the number of used KV cells (DEPRECATED)""" + ... + + +# // Clear the KV cache - both cell info is erased and KV data is zeroed +# DEPRECATED(LLAMA_API void llama_kv_self_clear( +# struct llama_context * ctx), +# "Use llama_memory_clear() instead"); +@ctypes_function( + "llama_kv_self_clear", [llama_context_p_ctypes], None +) +def llama_kv_self_clear(ctx: llama_context_p, /): + """Clear the KV cache (DEPRECATED)""" + ... + + +# // Removes all tokens that belong to the specified sequence and have positions in [p0, p1) +# // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails +# // seq_id < 0 : match any sequence +# // p0 < 0 : [0, p1] +# // p1 < 0 : [p0, inf) +# DEPRECATED(LLAMA_API bool llama_kv_self_seq_rm( +# struct llama_context * ctx, +# llama_seq_id seq_id, +# llama_pos p0, +# llama_pos p1), +# "Use llama_memory_seq_rm() instead"); +@ctypes_function( + "llama_kv_self_seq_rm", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_pos, + llama_pos, + ], + ctypes.c_bool, +) +def llama_kv_self_seq_rm( + ctx: llama_context_p, + seq_id: Union[llama_seq_id, int], + p0: Union[llama_pos, int], + p1: Union[llama_pos, int], + /, +) -> bool: + """Remove tokens from KV cache (DEPRECATED)""" + ... + + +# // Copy all tokens that belong to the specified sequence to another sequence +# // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence +# // p0 < 0 : [0, p1] +# // p1 < 0 : [p0, inf) +# DEPRECATED(LLAMA_API void llama_kv_self_seq_cp( +# struct llama_context * ctx, +# llama_seq_id seq_id_src, +# llama_seq_id seq_id_dst, +# llama_pos p0, +# llama_pos p1), +# "Use llama_memory_seq_cp() instead"); +@ctypes_function( + "llama_kv_self_seq_cp", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_seq_id, + llama_pos, + llama_pos, + ], + None, +) +def llama_kv_self_seq_cp( + ctx: llama_context_p, + seq_id_src: Union[llama_seq_id, int], + seq_id_dst: Union[llama_seq_id, int], + p0: Union[llama_pos, int], + p1: Union[llama_pos, int], + /, +): + """Copy tokens in KV cache (DEPRECATED)""" + ... + + +# // Removes all tokens that do not belong to the specified sequence +# DEPRECATED(LLAMA_API void llama_kv_self_seq_keep( +# struct llama_context * ctx, +# llama_seq_id seq_id), +# "Use llama_memory_seq_keep() instead"); +@ctypes_function( + "llama_kv_self_seq_keep", [llama_context_p_ctypes, llama_seq_id], None +) +def llama_kv_self_seq_keep(ctx: llama_context_p, seq_id: Union[llama_seq_id, int], /): + """Keep only specified sequence in KV cache (DEPRECATED)""" + ... + + +# // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) +# // If the KV cache is RoPEd, the KV data is updated accordingly: +# // - lazily on next llama_decode() +# // p0 < 0 : [0, p1] +# // p1 < 0 : [p0, inf) +# DEPRECATED(LLAMA_API void llama_kv_self_seq_add( +# struct llama_context * ctx, +# llama_seq_id seq_id, +# llama_pos p0, +# llama_pos p1, +# llama_pos delta), +# "Use llama_memory_seq_add() instead"); +@ctypes_function( + "llama_kv_self_seq_add", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_pos, + llama_pos, + llama_pos, + ], + None, +) +def llama_kv_self_seq_add( + ctx: llama_context_p, + seq_id: Union[llama_seq_id, int], + p0: Union[llama_pos, int], + p1: Union[llama_pos, int], + delta: Union[llama_pos, int], + /, +): + """Add delta to sequence positions in KV cache (DEPRECATED)""" + ... + + +# // Integer division of the positions by factor of `d > 1` +# // If the KV cache is RoPEd, the KV data is updated accordingly: +# // - lazily on next llama_decode() +# // p0 < 0 : [0, p1] +# // p1 < 0 : [p0, inf) +# DEPRECATED(void llama_kv_self_seq_div( +# struct llama_context * ctx, +# llama_seq_id seq_id, +# llama_pos p0, +# llama_pos p1, +# int d), +# "Use llama_memory_seq_div() instead"); +@ctypes_function( + "llama_kv_self_seq_div", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_pos, + llama_pos, + ctypes.c_int, + ], + None, +) +def llama_kv_self_seq_div( + ctx: llama_context_p, + seq_id: Union[llama_seq_id, int], + p0: Union[llama_pos, int], + p1: Union[llama_pos, int], + d: Union[ctypes.c_int, int], + /, +): + """Divide sequence positions in KV cache (DEPRECATED)""" + ... + + +# // Returns the smallest position present in the KV cache for the specified sequence +# // This is typically non-zero only for SWA caches +# // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache +# // Return -1 if the sequence is empty +# DEPRECATED(LLAMA_API llama_pos llama_kv_self_seq_pos_min( +# struct llama_context * ctx, +# llama_seq_id seq_id), +# "Use llama_memory_seq_pos_min() instead"); +@ctypes_function( + "llama_kv_self_seq_pos_min", [llama_context_p_ctypes, llama_seq_id], llama_pos +) +def llama_kv_self_seq_pos_min( + ctx: llama_context_p, seq_id: Union[llama_seq_id, int], / +) -> int: + """Returns the smallest position in KV cache for sequence (DEPRECATED)""" + ... + + +# // Returns the largest position present in the KV cache for the specified sequence +# // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache +# // Return -1 if the sequence is empty +# DEPRECATED(LLAMA_API llama_pos llama_kv_self_seq_pos_max( +# struct llama_context * ctx, +# llama_seq_id seq_id), +# "Use llama_memory_seq_pos_max() instead"); +@ctypes_function( + "llama_kv_self_seq_pos_max", [llama_context_p_ctypes, llama_seq_id], llama_pos +) +def llama_kv_self_seq_pos_max( + ctx: llama_context_p, seq_id: Union[llama_seq_id, int], / +) -> int: + """Returns the largest position in KV cache for sequence (DEPRECATED)""" + ... + + +# // Defragment the KV cache +# // This will be applied: +# // - lazily on next llama_decode() +# DEPRECATED(LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx), +# "simply remove this call, the context will automatically decide when to do a defragmentation based on 'defrag_thold'"); +@ctypes_function("llama_kv_self_defrag", [llama_context_p_ctypes], None) +def llama_kv_self_defrag(ctx: llama_context_p, /): + """Defragment the KV cache (DEPRECATED)""" ... # // Check if the context supports KV cache shifting -# LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx); -@ctypes_function("llama_kv_cache_can_shift", [llama_context_p_ctypes], ctypes.c_bool) -def llama_kv_cache_can_shift(ctx: llama_context_p, /) -> bool: - """Check if the context supports KV cache shifting""" +# DEPRECATED(LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx), +# "use llama_memory_can_shift() instead"); +@ctypes_function("llama_kv_self_can_shift", [llama_context_p_ctypes], ctypes.c_bool) +def llama_kv_self_can_shift(ctx: llama_context_p, /) -> bool: + """Check if the context supports KV cache shifting (DEPRECATED)""" + ... + + +# // Apply the KV cache updates (such as K-shifts, defragmentation, etc.) +# DEPRECATED(LLAMA_API void llama_kv_self_update(struct llama_context * ctx), +# "simply remove this call, updates are applied lazily on the next llama_decode()"); +@ctypes_function("llama_kv_self_update", [llama_context_p_ctypes], None) +def llama_kv_self_update(ctx: llama_context_p, /): + """Apply the KV cache updates (DEPRECATED)""" ... @@ -1911,14 +2268,13 @@ def llama_kv_cache_can_shift(ctx: llama_context_p, /) -> bool: # // State / sessions # // - # // Returns the *actual* size in bytes of the state -# // (logits, embedding and kv_cache) +# // (logits, embedding and memory) # // Only use when saving the state, not when restoring it, otherwise the size may be too small. # LLAMA_API size_t llama_state_get_size(struct llama_context * ctx); @ctypes_function("llama_state_get_size", [llama_context_p_ctypes], ctypes.c_size_t) def llama_state_get_size(ctx: llama_context_p, /) -> int: - """Returns the *actual* size in bytes of the state (rng, logits, embedding and kv_cache) - will often be smaller after compacting tokens""" + """Returns the *actual* size in bytes of the state (logits, embedding and memory)""" ... @@ -1926,8 +2282,7 @@ def llama_state_get_size(ctx: llama_context_p, /) -> int: # "use llama_state_get_size instead"); @ctypes_function("llama_get_state_size", [llama_context_p_ctypes], ctypes.c_size_t) def llama_get_state_size(ctx: llama_context_p, /) -> int: - """Returns the maximum size in bytes of the state (rng, logits, embedding - and kv_cache) - will often be smaller after compacting tokens""" + """Returns the size in bytes of the state (DEPRECATED)""" ... @@ -1974,9 +2329,7 @@ def llama_state_get_data( def llama_copy_state_data( ctx: llama_context_p, dst: CtypesArray[ctypes.c_uint8], / ) -> int: - """Copies the state to the specified destination address. - Destination needs to have allocated enough memory. - Returns the number of bytes copied""" + """Copies the state to the specified destination address (DEPRECATED)""" ... @@ -2014,7 +2367,7 @@ def llama_state_set_data( def llama_set_state_data( ctx: llama_context_p, src: CtypesArray[ctypes.c_uint8], / ) -> int: - """Set the state reading from the specified address""" + """Set the state reading from the specified address (DEPRECATED)""" ... @@ -2063,7 +2416,7 @@ def llama_state_load_file( ctypes.c_size_t, ctypes.POINTER(ctypes.c_size_t), ], - ctypes.c_size_t, + ctypes.c_bool, ) def llama_load_session_file( ctx: llama_context_p, @@ -2072,7 +2425,7 @@ def llama_load_session_file( n_token_capacity: Union[ctypes.c_size_t, int], n_token_count_out: CtypesPointerOrRef[ctypes.c_size_t], /, -) -> int: +) -> bool: ... @@ -2115,7 +2468,7 @@ def llama_state_save_file( llama_token_p, ctypes.c_size_t, ], - ctypes.c_size_t, + ctypes.c_bool, ) def llama_save_session_file( ctx: llama_context_p, @@ -2123,11 +2476,11 @@ def llama_save_session_file( tokens: CtypesArray[llama_token], n_token_count: Union[ctypes.c_size_t, int], /, -) -> int: +) -> bool: ... -# // Get the exact size needed to copy the KV cache of a single sequence +# // Get the exact size needed to copy the state of a single sequence # LLAMA_API size_t llama_state_seq_get_size( # struct llama_context * ctx, # llama_seq_id seq_id); @@ -2137,11 +2490,11 @@ def llama_save_session_file( ctypes.c_size_t, ) def llama_state_seq_get_size(ctx: llama_context_p, seq_id: llama_seq_id, /) -> int: - """Get the exact size needed to copy the KV cache of a single sequence""" + """Get the exact size needed to copy the state of a single sequence""" ... -# // Copy the KV cache of a single sequence into the specified buffer +# // Copy the state of a single sequence into the specified buffer # LLAMA_API size_t llama_state_seq_get_data( # struct llama_context * ctx, # uint8_t * dst, @@ -2164,7 +2517,7 @@ def llama_state_seq_get_data( seq_id: llama_seq_id, /, ) -> int: - """Copy the KV cache of a single sequence into the specified buffer""" + """Copy the state of a single sequence into the specified buffer""" ... @@ -2194,7 +2547,7 @@ def llama_state_seq_set_data( dest_seq_id: llama_seq_id, /, ) -> int: - """Copy the sequence data (originally copied with `llama_state_seq_get_data`) into the specified sequence""" + """Copy the sequence data into the specified sequence""" ... @@ -2261,7 +2614,6 @@ def llama_state_seq_load_file( # // Decoding # // - # // Return batch for single sequence of tokens # // The sequence ID will be fixed to 0 # // The position of the tokens will be tracked automatically by llama_decode @@ -2284,7 +2636,7 @@ def llama_batch_get_one( n_tokens: Union[ctypes.c_int, int], /, ) -> llama_batch: - """Return batch for single sequence of tokens starting at pos_0 + """Return batch for single sequence of tokens NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it """ @@ -2329,35 +2681,46 @@ def llama_batch_free(batch: llama_batch, /): ... -# // Processes a batch of tokens with the ecoder part of the encoder-decoder model. -# // Stores the encoder output internally for later use by the decoder cross-attention layers. +# // Process a batch of tokens. +# // In contrast to llama_decode() - this call does not use KV cache. +# // For encode-decoder contexts, processes the batch using the encoder. +# // Can store the encoder output internally for later use by the decoder's cross-attention layers. # // 0 - success -# // < 0 - error +# // < 0 - error. the memory state is restored to the state before this call # LLAMA_API int32_t llama_encode( # struct llama_context * ctx, # struct llama_batch batch); @ctypes_function("llama_encode", [llama_context_p_ctypes, llama_batch], ctypes.c_int32) def llama_encode(ctx: llama_context_p, batch: llama_batch, /) -> int: - """Processes a batch of tokens with the ecoder part of the encoder-decoder model. - Stores the encoder output internally for later use by the decoder cross-attention layers. + """Process a batch of tokens using the encoder. 0 - success < 0 - error""" ... +# // Process a batch of tokens. +# // Requires the context to have a memory. +# // For encode-decoder contexts, processes the batch using the decoder. # // Positive return values does not mean a fatal error, but rather a warning. -# // 0 - success -# // 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context) -# // < 0 - error +# // Upon fatal-error or abort, the ubatches that managed to be been processed will remain in the memory state of the context +# // To handle this correctly, query the memory state using llama_memory_seq_pos_min() and llama_memory_seq_pos_max() +# // Upon other return values, the memory state is restored to the state before this call +# // 0 - success +# // 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context) +# // 2 - aborted (processed ubatches will remain in the context's memory) +# // -1 - invalid input batch +# // < -1 - fatal error (processed ubatches will remain in the context's memory) # LLAMA_API int32_t llama_decode( # struct llama_context * ctx, # struct llama_batch batch); @ctypes_function("llama_decode", [llama_context_p_ctypes, llama_batch], ctypes.c_int32) def llama_decode(ctx: llama_context_p, batch: llama_batch, /) -> int: - """Positive return values does not mean a fatal error, but rather a warning. + """Process a batch of tokens. 0 - success 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context) - < 0 - error""" + 2 - aborted (processed ubatches will remain in the context's memory) + -1 - invalid input batch + < -1 - fatal error (processed ubatches will remain in the context's memory)""" ... @@ -2403,13 +2766,12 @@ def llama_n_threads_batch(ctx: llama_context_p, /) -> int: ... -# // Set whether the model is in embeddings mode or not -# // If true, embeddings will be returned but logits will not +# // Set whether the context outputs embeddings or not +# // TODO: rename to avoid confusion with llama_get_embeddings() # LLAMA_API void llama_set_embeddings(struct llama_context * ctx, bool embeddings); @ctypes_function("llama_set_embeddings", [llama_context_p_ctypes, ctypes.c_bool], None) def llama_set_embeddings(ctx: llama_context_p, embeddings: bool, /): - """Set whether the model is in embeddings model or not - If true, embeddings will be returned but logits will not""" + """Set whether the context outputs embeddings or not""" ... @@ -2423,6 +2785,16 @@ def llama_set_causal_attn(ctx: llama_context_p, causal_attn: bool, /): ... +# // Set whether the model is in warmup mode or not +# // If true, all model tensors are activated during llama_decode() to load and cache their weights. +# LLAMA_API void llama_set_warmup(struct llama_context * ctx, bool warmup); +@ctypes_function("llama_set_warmup", [llama_context_p_ctypes, ctypes.c_bool], None) +def llama_set_warmup(ctx: llama_context_p, warmup: bool, /): + """Set whether the model is in warmup mode or not + If true, all model tensors are activated during llama_decode() to load and cache their weights.""" + ... + + # // Set abort callback # LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data); @ctypes_function( @@ -2528,7 +2900,7 @@ def llama_get_embeddings_ith( # // Get the embeddings for a sequence id # // Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE -# // when pooling_type == LLAMA_POOLING_TYPE_RANK, returns float[1] with the rank of the sequence +# // when pooling_type == LLAMA_POOLING_TYPE_RANK, returns float[n_cls_out] with the rank(s) of the sequence # // otherwise: float[n_embd] (1-dimensional) # LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id); @ctypes_function( @@ -2549,182 +2921,415 @@ def llama_get_embeddings_seq( # // Vocab # // - -# LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token); +# LLAMA_API const char * llama_vocab_get_text(const struct llama_vocab * vocab, llama_token token); @ctypes_function( - "llama_token_get_text", [llama_model_p_ctypes, llama_token], ctypes.c_char_p + "llama_vocab_get_text", [llama_vocab_p_ctypes, llama_token], ctypes.c_char_p ) -def llama_token_get_text( - model: llama_model_p, token: Union[llama_token, int], / +def llama_vocab_get_text( + vocab: llama_vocab_p, token: Union[llama_token, int], / ) -> bytes: ... -# LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token); +# LLAMA_API float llama_vocab_get_score(const struct llama_vocab * vocab, llama_token token); @ctypes_function( - "llama_token_get_score", [llama_model_p_ctypes, llama_token], ctypes.c_float + "llama_vocab_get_score", [llama_vocab_p_ctypes, llama_token], ctypes.c_float ) -def llama_token_get_score( - model: llama_model_p, token: Union[llama_token, int], / +def llama_vocab_get_score( + vocab: llama_vocab_p, token: Union[llama_token, int], / ) -> float: ... -# LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_model * model, llama_token token); +# LLAMA_API enum llama_token_attr llama_vocab_get_attr(const struct llama_vocab * vocab, llama_token token); @ctypes_function( - "llama_token_get_attr", [llama_model_p_ctypes, llama_token], ctypes.c_int + "llama_vocab_get_attr", [llama_vocab_p_ctypes, llama_token], ctypes.c_int ) -def llama_token_get_attr( - model: llama_model_p, token: Union[llama_token, int], / +def llama_vocab_get_attr( + vocab: llama_vocab_p, token: Union[llama_token, int], / ) -> int: ... # // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.) -# LLAMA_API bool llama_token_is_eog(const struct llama_model * model, llama_token token); +# LLAMA_API bool llama_vocab_is_eog(const struct llama_vocab * vocab, llama_token token); @ctypes_function( - "llama_token_is_eog", [llama_model_p_ctypes, llama_token], ctypes.c_bool + "llama_vocab_is_eog", [llama_vocab_p_ctypes, llama_token], ctypes.c_bool ) -def llama_token_is_eog(model: llama_model_p, token: Union[llama_token, int], /) -> bool: +def llama_vocab_is_eog(vocab: llama_vocab_p, token: Union[llama_token, int], /) -> bool: """Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.)""" ... # // Identify if Token Id is a control token or a render-able token -# LLAMA_API bool llama_token_is_control(const struct llama_model * model, llama_token token); +# LLAMA_API bool llama_vocab_is_control(const struct llama_vocab * vocab, llama_token token); @ctypes_function( - "llama_token_is_control", [llama_model_p_ctypes, llama_token], ctypes.c_bool + "llama_vocab_is_control", [llama_vocab_p_ctypes, llama_token], ctypes.c_bool ) -def llama_token_is_control( - model: llama_model_p, token: Union[llama_token, int], / +def llama_vocab_is_control( + vocab: llama_vocab_p, token: Union[llama_token, int], / ) -> bool: """Identify if Token Id is a control token or a render-able token""" ... # // Special tokens - - -# LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence -@ctypes_function("llama_token_bos", [llama_model_p_ctypes], llama_token) -def llama_token_bos(model: llama_model_p, /) -> int: +# LLAMA_API llama_token llama_vocab_bos(const struct llama_vocab * vocab); // beginning-of-sentence +@ctypes_function("llama_vocab_bos", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_bos(vocab: llama_vocab_p, /) -> llama_token: """beginning-of-sentence""" ... -# LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence -@ctypes_function("llama_token_eos", [llama_model_p_ctypes], llama_token) -def llama_token_eos(model: llama_model_p, /) -> int: +# LLAMA_API llama_token llama_vocab_eos(const struct llama_vocab * vocab); // end-of-sentence +@ctypes_function("llama_vocab_eos", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_eos(vocab: llama_vocab_p, /) -> llama_token: """end-of-sentence""" ... -# LLAMA_API llama_token llama_token_eot(const struct llama_model * model); // end-of-turn -@ctypes_function("llama_token_eot", [llama_model_p_ctypes], llama_token) -def llama_token_eot(model: llama_model_p, /) -> int: +# LLAMA_API llama_token llama_vocab_eot(const struct llama_vocab * vocab); // end-of-turn +@ctypes_function("llama_vocab_eot", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_eot(vocab: llama_vocab_p, /) -> llama_token: """end-of-turn""" ... -# LLAMA_API llama_token llama_token_cls(const struct llama_model * model); // classification -@ctypes_function("llama_token_cls", [llama_model_p_ctypes], llama_token) -def llama_token_cls(model: llama_model_p, /) -> int: - """classification""" +# LLAMA_API llama_token llama_vocab_sep(const struct llama_vocab * vocab); // sentence separator +@ctypes_function("llama_vocab_sep", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_sep(vocab: llama_vocab_p, /) -> llama_token: + """sentence separator""" ... -# LLAMA_API llama_token llama_token_sep(const struct llama_model * model); // sentence separator -@ctypes_function("llama_token_sep", [llama_model_p_ctypes], llama_token) -def llama_token_sep(model: llama_model_p, /) -> int: - """sentence separator""" +# LLAMA_API llama_token llama_vocab_nl (const struct llama_vocab * vocab); // next-line +@ctypes_function("llama_vocab_nl", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_nl(vocab: llama_vocab_p, /) -> llama_token: + """next-line""" ... -# LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line -@ctypes_function("llama_token_nl", [llama_model_p_ctypes], llama_token) -def llama_token_nl(model: llama_model_p, /) -> int: - """next-line""" +# LLAMA_API llama_token llama_vocab_pad(const struct llama_vocab * vocab); // padding +@ctypes_function("llama_vocab_pad", [llama_vocab_p_ctypes], llama_token) +def llama_vocab_pad(vocab: llama_vocab_p, /) -> llama_token: + """padding""" + ... + + +# LLAMA_API bool llama_vocab_get_add_bos(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_get_add_bos", + [llama_vocab_p_ctypes], + ctypes.c_bool, +) +def llama_vocab_get_add_bos(vocab: llama_vocab_p, /) -> bool: + ... + + +# LLAMA_API bool llama_vocab_get_add_eos(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_get_add_eos", + [llama_vocab_p_ctypes], + ctypes.c_bool, +) +def llama_vocab_get_add_eos(vocab: llama_vocab_p, /) -> bool: + ... + + +# LLAMA_API bool llama_vocab_get_add_sep(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_get_add_sep", + [llama_vocab_p_ctypes], + ctypes.c_bool, +) +def llama_vocab_get_add_sep(vocab: llama_vocab_p, /) -> bool: + ... + + +# LLAMA_API llama_token llama_vocab_fim_pre(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_pre", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_pre(vocab: llama_vocab_p, /) -> llama_token: + ... + + +# LLAMA_API llama_token llama_vocab_fim_suf(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_suf", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_suf(vocab: llama_vocab_p, /) -> llama_token: + ... + + +# LLAMA_API llama_token llama_vocab_fim_mid(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_mid", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_mid(vocab: llama_vocab_p, /) -> llama_token: + ... + + +# LLAMA_API llama_token llama_vocab_fim_pad(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_pad", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_pad(vocab: llama_vocab_p, /) -> llama_token: ... -# LLAMA_API bool llama_add_bos_token(const struct llama_model * model); -@ctypes_function("llama_add_bos_token", [llama_model_p_ctypes], ctypes.c_bool) -def llama_add_bos_token(model: llama_model_p, /) -> bool: +# LLAMA_API llama_token llama_vocab_fim_rep(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_rep", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_rep(vocab: llama_vocab_p, /) -> llama_token: ... -# LLAMA_API bool llama_add_eos_token(const struct llama_model * model); -@ctypes_function("llama_add_eos_token", [llama_model_p_ctypes], ctypes.c_bool) -def llama_add_eos_token(model: llama_model_p, /) -> bool: +# LLAMA_API llama_token llama_vocab_fim_sep(const struct llama_vocab * vocab); +@ctypes_function( + "llama_vocab_fim_sep", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_fim_sep(vocab: llama_vocab_p, /) -> llama_token: ... -# // Codellama infill tokens -# DEPRECATED(LLAMA_API llama_token llama_token_prefix(const struct llama_model * model), "use llama_token_fim_pre instead"); -@ctypes_function("llama_token_prefix", [llama_model_p_ctypes], llama_token) -def llama_token_prefix(model: llama_model_p) -> int: - """codellama infill tokens""" +# DEPRECATED functions +# DEPRECATED(LLAMA_API const char * llama_token_get_text(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_text instead"); +@ctypes_function( + "llama_token_get_text", + [llama_vocab_p_ctypes, llama_token], + ctypes.c_char_p, +) +def llama_token_get_text( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> bytes: ... -# DEPRECATED(LLAMA_API llama_token llama_token_middle(const struct llama_model * model), "use llama_token_fim_mid instead"); -@ctypes_function("llama_token_middle", [llama_model_p_ctypes], llama_token) -def llama_token_middle(model: llama_model_p, /) -> int: +# DEPRECATED(LLAMA_API float llama_token_get_score(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_score instead"); +@ctypes_function( + "llama_token_get_score", + [llama_vocab_p_ctypes, llama_token], + ctypes.c_float, +) +def llama_token_get_score( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> float: ... +# DEPRECATED(LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_attr instead"); +@ctypes_function( + "llama_token_get_attr", + [llama_vocab_p_ctypes, llama_token], + ctypes.c_int, +) +def llama_token_get_attr( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> int: + ... + +# DEPRECATED(LLAMA_API bool llama_token_is_eog(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_is_eog instead"); +@ctypes_function( + "llama_token_is_eog", + [llama_vocab_p_ctypes, llama_token], + ctypes.c_bool, +) +def llama_token_is_eog( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> bool: + ... -# DEPRECATED(LLAMA_API llama_token llama_token_suffix(const struct llama_model * model), "use llama_token_fim_suf instead"); -@ctypes_function("llama_token_suffix", [llama_model_p_ctypes], llama_token) -def llama_token_suffix(model: llama_model_p, /) -> int: +# DEPRECATED(LLAMA_API bool llama_token_is_control(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_is_control instead"); +@ctypes_function( + "llama_token_is_control", + [llama_vocab_p_ctypes, llama_token], + ctypes.c_bool, +) +def llama_token_is_control( + vocab: llama_vocab_p, token: Union[llama_token, int], / +) -> bool: ... +# DEPRECATED(LLAMA_API llama_token llama_token_bos(const struct llama_vocab * vocab), "use llama_vocab_bos instead"); +@ctypes_function( + "llama_token_bos", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_bos(vocab: llama_vocab_p, /) -> int: + ... -# LLAMA_API llama_token llama_token_fim_pre(const struct llama_model * model); -@ctypes_function("llama_token_fim_pre", [llama_model_p_ctypes], llama_token) -def llama_token_fim_pre(model: llama_model_p, /) -> int: +# DEPRECATED(LLAMA_API llama_token llama_token_eos(const struct llama_vocab * vocab), "use llama_vocab_eos instead"); +@ctypes_function( + "llama_token_eos", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_eos(vocab: llama_vocab_p, /) -> int: ... -# LLAMA_API llama_token llama_token_fim_suf(const struct llama_model * model); -@ctypes_function("llama_token_fim_suf", [llama_model_p_ctypes], llama_token) -def llama_token_fim_suf(model: llama_model_p, /) -> int: +# DEPRECATED(LLAMA_API llama_token llama_token_eot(const struct llama_vocab * vocab), "use llama_vocab_eot instead"); +@ctypes_function( + "llama_token_eot", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_eot(vocab: llama_vocab_p, /) -> int: ... -# LLAMA_API llama_token llama_token_fim_mid(const struct llama_model * model); -@ctypes_function("llama_token_fim_mid", [llama_model_p_ctypes], llama_token) -def llama_token_fim_mid(model: llama_model_p, /) -> int: +# DEPRECATED(LLAMA_API llama_token llama_token_cls(const struct llama_vocab * vocab), "use llama_vocab_cls instead"); +@ctypes_function( + "llama_token_cls", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_cls(vocab: llama_vocab_p, /) -> int: ... -# LLAMA_API llama_token llama_token_fim_pad(const struct llama_model * model); -@ctypes_function("llama_token_fim_pad", [llama_model_p_ctypes], llama_token) -def llama_token_fim_pad(model: llama_model_p, /) -> int: +# DEPRECATED(LLAMA_API llama_token llama_token_sep(const struct llama_vocab * vocab), "use llama_vocab_sep instead"); +@ctypes_function( + "llama_token_sep", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_sep(vocab: llama_vocab_p, /) -> int: ... -# LLAMA_API llama_token llama_token_fim_rep(const struct llama_model * model); -@ctypes_function("llama_token_fim_rep", [llama_model_p_ctypes], llama_token) -def llama_token_fim_rep(model: llama_model_p, /) -> int: + +# DEPRECATED(LLAMA_API llama_token llama_token_nl (const struct llama_vocab * vocab), "use llama_vocab_nl instead"); +@ctypes_function( + "llama_token_nl", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_nl(vocab: llama_vocab_p, /) -> int: ... -# LLAMA_API llama_token llama_token_fim_sep(const struct llama_model * model); -@ctypes_function("llama_token_fim_sep", [llama_model_p_ctypes], llama_token) -def llama_token_fim_sep(model: llama_model_p, /) -> int: + +# DEPRECATED(LLAMA_API llama_token llama_token_pad(const struct llama_vocab * vocab), "use llama_vocab_pad instead"); +@ctypes_function( + "llama_token_pad", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_pad(vocab: llama_vocab_p, /) -> int: ... + +# DEPRECATED(LLAMA_API bool llama_add_bos_token(const struct llama_vocab * vocab), "use llama_vocab_get_add_bos instead"); +@ctypes_function( + "llama_add_bos_token", + [llama_vocab_p_ctypes], + ctypes.c_bool, +) +def llama_add_bos_token(vocab: llama_vocab_p, /) -> bool: + ... + +# DEPRECATED(LLAMA_API bool llama_add_eos_token(const struct llama_vocab * vocab), "use llama_vocab_get_add_eos instead"); +@ctypes_function( + "llama_add_eos_token", + [llama_vocab_p_ctypes], + ctypes.c_bool, +) +def llama_add_eos_token(vocab: llama_vocab_p, /) -> bool: + ... + + +# DEPRECATED(LLAMA_API llama_token llama_token_fim_pre(const struct llama_vocab * vocab), "use llama_vocab_fim_pre instead"); +@ctypes_function( + "llama_token_fim_pre", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_pre(vocab: llama_vocab_p, /) -> llama_token: + ... + +# DEPRECATED(LLAMA_API llama_token llama_token_fim_suf(const struct llama_vocab * vocab), "use llama_vocab_fim_suf instead"); +@ctypes_function( + "llama_token_fim_suf", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_suf(vocab: llama_vocab_p, /) -> llama_token: + ... + +# DEPRECATED(LLAMA_API llama_token llama_token_fim_mid(const struct llama_vocab * vocab), "use llama_vocab_fim_mid instead"); +@ctypes_function( + "llama_token_fim_mid", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_mid(vocab: llama_vocab_p, /) -> llama_token: + ... + +# DEPRECATED(LLAMA_API llama_token llama_token_fim_pad(const struct llama_vocab * vocab), "use llama_vocab_fim_pad instead"); +@ctypes_function( + "llama_token_fim_pad", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_pad(vocab: llama_vocab_p, /) -> llama_token: + ... + +# DEPRECATED(LLAMA_API llama_token llama_token_fim_rep(const struct llama_vocab * vocab), "use llama_vocab_fim_rep instead"); +@ctypes_function( + "llama_token_fim_rep", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_rep(vocab: llama_vocab_p, /) -> llama_token: + ... + +# DEPRECATED(LLAMA_API llama_token llama_token_fim_sep(const struct llama_vocab * vocab), "use llama_vocab_fim_sep instead"); +@ctypes_function( + "llama_token_fim_sep", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_token_fim_sep(vocab: llama_vocab_p, /) -> llama_token: + ... + +# // CLS is equivalent to BOS +# DEPRECATED(LLAMA_API llama_token llama_vocab_cls(const struct llama_vocab * vocab), // classification +# "use llama_vocab_bos instead"); +@ctypes_function( + "llama_vocab_cls", + [llama_vocab_p_ctypes], + llama_token, +) +def llama_vocab_cls(vocab: llama_vocab_p, /) -> llama_token: + ... + + # // # // Tokenization # // # // The API is thread-safe. # // - # /// @details Convert the provided text into tokens. # /// @param tokens The tokens pointer must be large enough to hold the resulting tokens. # /// @return Returns the number of tokens on success, no more than n_tokens_max # /// @return Returns a negative number on failure - the number of tokens that would have been returned +# /// @return Returns INT32_MIN on overflow (e.g., tokenization result size exceeds int32_t limit) # /// @param add_special Allow to add BOS and EOS tokens if model is configured to do so. # /// @param parse_special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated # /// as plaintext. Does not insert a leading space. # LLAMA_API int32_t llama_tokenize( -# const struct llama_model * model, +# const struct llama_vocab * vocab, # const char * text, # int32_t text_len, # llama_token * tokens, @@ -2734,7 +3339,7 @@ def llama_token_fim_sep(model: llama_model_p, /) -> int: @ctypes_function( "llama_tokenize", [ - llama_model_p_ctypes, + llama_vocab_p_ctypes, ctypes.c_char_p, ctypes.c_int32, llama_token_p, @@ -2745,7 +3350,7 @@ def llama_token_fim_sep(model: llama_model_p, /) -> int: ctypes.c_int32, ) def llama_tokenize( - model: llama_model_p, + vocab: llama_vocab_p, text: bytes, text_len: Union[ctypes.c_int, int], tokens: CtypesArray[llama_token], @@ -2757,12 +3362,12 @@ def llama_tokenize( """Convert the provided text into tokens. Args: - model: The model to use for tokenization. + vocab: The vocabulary to use for tokenization. text: The text to tokenize. text_len: The length of the text. tokens: The tokens pointer must be large enough to hold the resulting tokens. n_max_tokens: The maximum number of tokens to return. - add_special: Allow adding special tokenns if the model is configured to do so. + add_special: Allow adding special tokens if the model is configured to do so. parse_special: Allow parsing special tokens. Returns: @@ -2778,7 +3383,7 @@ def llama_tokenize( # // User can skip up to 'lstrip' leading spaces before copying (useful when encoding/decoding multiple tokens with 'add_space_prefix') # // @param special If true, special tokens are rendered in the output. # LLAMA_API int32_t llama_token_to_piece( -# const struct llama_model * model, +# const struct llama_vocab * vocab, # llama_token token, # char * buf, # int32_t length, @@ -2787,7 +3392,7 @@ def llama_tokenize( @ctypes_function( "llama_token_to_piece", [ - llama_model_p_ctypes, + llama_vocab_p_ctypes, llama_token, ctypes.c_char_p, ctypes.c_int32, @@ -2797,7 +3402,7 @@ def llama_tokenize( ctypes.c_int32, ) def llama_token_to_piece( - model: llama_model_p, + vocab: llama_vocab_p, token: Union[llama_token, int], buf: Union[ctypes.c_char_p, bytes, CtypesArray[ctypes.c_char]], length: Union[ctypes.c_int, int], @@ -2811,7 +3416,7 @@ def llama_token_to_piece( User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens. Args: - model: The model to use for tokenization. + vocab: The vocabulary to use for tokenization. token: The token to convert. buf: The buffer to write the token to. length: The length of the buffer. @@ -2820,23 +3425,6 @@ def llama_token_to_piece( ... -# # // check if token0 is contained as a prefix in token1 -# # LLAMA_API bool llama_token_is_prefix( -# # const struct llama_model * model, -# # llama_token token0, -# # llama_token token1); -# @ctypes_function( -# "llama_token_is_prefix", -# [llama_model_p_ctypes, llama_token, llama_token], -# ctypes.c_bool, -# ) -# def llama_token_is_prefix( -# model: llama_model_p, token0: Union[llama_token, int], token1: Union[llama_token, int], / -# ) -> bool: -# """Check if token0 is contained as a prefix in token1""" -# ... - - # /// @details Convert the provided tokens into text (inverse of llama_tokenize()). # /// @param text The char pointer must be large enough to hold the resulting text. # /// @return Returns the number of chars/bytes on success, no more than text_len_max. @@ -2844,7 +3432,7 @@ def llama_token_to_piece( # /// @param remove_special Allow to remove BOS and EOS tokens if model is configured to do so. # /// @param unparse_special If true, special tokens are rendered in the output. # LLAMA_API int32_t llama_detokenize( -# const struct llama_model * model, +# const struct llama_vocab * vocab, # const llama_token * tokens, # int32_t n_tokens, # char * text, @@ -2854,7 +3442,7 @@ def llama_token_to_piece( @ctypes_function( "llama_detokenize", [ - llama_model_p_ctypes, + llama_vocab_p_ctypes, ctypes.POINTER(llama_token), ctypes.c_int32, ctypes.c_char_p, @@ -2865,7 +3453,7 @@ def llama_token_to_piece( ctypes.c_int32, ) def llama_detokenize( - model: llama_model_p, + vocab: llama_vocab_p, tokens: CtypesArray[llama_token], n_tokens: Union[ctypes.c_int, int], text: bytes, @@ -2877,7 +3465,7 @@ def llama_detokenize( """Convert the provided tokens into text (inverse of llama_tokenize()). Args: - model: The model to use for tokenization. + vocab: The vocabulary to use for tokenization. tokens: The tokens to convert. n_tokens: The number of tokens. text: The buffer to write the text to. @@ -2891,11 +3479,10 @@ def llama_detokenize( # // Chat templates # // - # /// Apply chat template. Inspired by hf apply_chat_template() on python. # /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model" -# /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template -# /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead. +# /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template +# /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model's default chat template will be used instead. # /// @param chat Pointer to a list of multiple llama_chat_message # /// @param n_msg Number of llama_chat_message in this chat # /// @param add_ass Whether to end the prompt with the token(s) that indicate the start of an assistant message. @@ -2903,7 +3490,6 @@ def llama_detokenize( # /// @param length The size of the allocated buffer # /// @return The total number of bytes of the formatted prompt. If is it larger than the size of buffer, you may need to re-alloc it and then re-apply the template. # LLAMA_API int32_t llama_chat_apply_template( -# const struct llama_model * model, # const char * tmpl, # const struct llama_chat_message * chat, # size_t n_msg, @@ -2913,20 +3499,37 @@ def llama_detokenize( @ctypes_function( "llama_chat_apply_template", [ - ctypes.c_void_p, - ctypes.c_char_p, - ctypes.POINTER(llama_chat_message), - ctypes.c_size_t, + ctypes.c_char_p, # tmpl + ctypes.POINTER(llama_chat_message), # chat + ctypes.c_size_t, # n_msg + ctypes.c_bool, # add_ass (added) + ctypes.c_char_p, # buf + ctypes.c_int32, # length ], ctypes.c_int32, ) def llama_chat_apply_template( - model: llama_model_p, tmpl: bytes, chat: CtypesArray[llama_chat_message], n_msg: int, + add_ass: bool, # Added parameter + buf: bytes, + length: int, /, ) -> int: + """Apply chat template. + + Args: + tmpl: Template to use. If None, uses model's default + chat: Array of chat messages + n_msg: Number of messages + add_ass: Whether to end prompt with assistant token + buf: Output buffer + length: Buffer length + + Returns: + Number of bytes written, or needed if buffer too small + """ ... @@ -2961,42 +3564,6 @@ def llama_chat_builtin_templates( # // # // Sampling API # // -# // Sample usage: -# // -# // // prepare the sampling chain at the start -# // auto sparams = llama_sampler_chain_default_params(); -# // -# // llama_sampler * smpl = llama_sampler_chain_init(sparams); -# // -# // llama_sampler_chain_add(smpl, llama_sampler_init_top_k(50)); -# // llama_sampler_chain_add(smpl, llama_sampler_init_top_p(0.9, 1)); -# // llama_sampler_chain_add(smpl, llama_sampler_init_temp (0.8)); -# // -# // // typically, the chain should end with a sampler such as "greedy", "dist" or "mirostat" -# // // this sampler will be responsible to select the actual token -# // llama_sampler_chain_add(smpl, llama_sampler_init_dist(seed)); -# // -# // ... -# // -# // // decoding loop: -# // while (...) { -# // ... -# // -# // llama_decode(ctx, batch); -# // -# // // sample from the logits of the last token in the batch -# // const llama_token id = llama_sampler_sample(smpl, ctx, -1); -# // -# // // accepting the token updates the internal state of certain samplers (e.g. grammar, repetition, etc.) -# // llama_sampler_accept(smpl, id); -# // ... -# // } -# // -# // llama_sampler_free(smpl); -# // -# // TODO: In the future, llama_sampler will be utilized to offload the sampling to the backends (e.g. GPU). -# // TODO: in the future, the entire sampling API that uses llama_model should start using llama_vocab -# // # typedef void * llama_sampler_context_t; llama_sampler_context_t = ctypes.c_void_p @@ -3010,7 +3577,7 @@ def llama_chat_builtin_templates( # void (*reset) ( struct llama_sampler * smpl); // can be NULL # struct llama_sampler * (*clone) (const struct llama_sampler * smpl); // can be NULL if ctx is NULL # void (*free) ( struct llama_sampler * smpl); // can be NULL if ctx is NULL -# + # // TODO: API for internal libllama usage for appending the sampling to an existing ggml_cgraph # //void (*apply_ggml) (struct llama_sampler * smpl, ...); # }; @@ -3019,8 +3586,8 @@ class llama_sampler_i(ctypes.Structure): # struct llama_sampler { -# struct llama_sampler_i * iface; -# llama_sampler_context_t ctx; +# const struct llama_sampler_i * iface; +# llama_sampler_context_t ctx; # }; class llama_sampler(ctypes.Structure): _fields_ = [ @@ -3054,6 +3621,18 @@ class llama_sampler(ctypes.Structure): # // mirror of llama_sampler_i: +# LLAMA_API struct llama_sampler * llama_sampler_init (const struct llama_sampler_i * iface, llama_sampler_context_t ctx); +@ctypes_function( + "llama_sampler_init", + [ctypes.POINTER(llama_sampler_i), llama_sampler_context_t], + llama_sampler_p_ctypes, +) +def llama_sampler_init( + iface: ctypes.POINTER(llama_sampler_i), ctx: llama_sampler_context_t, / +) -> llama_sampler_p: + ... + + # LLAMA_API const char * llama_sampler_name (const struct llama_sampler * smpl); @ctypes_function( "llama_sampler_name", @@ -3119,7 +3698,7 @@ def llama_sampler_free(smpl: llama_sampler_p, /): # // llama_sampler_chain # // a type of llama_sampler that can chain multiple samplers one after another -# + # LLAMA_API struct llama_sampler * llama_sampler_chain_init(struct llama_sampler_chain_params params); @ctypes_function( "llama_sampler_chain_init", @@ -3177,7 +3756,7 @@ def llama_sampler_chain_remove( # // available samplers: -# + # LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void); @ctypes_function("llama_sampler_init_greedy", [], llama_sampler_p_ctypes) def llama_sampler_init_greedy() -> llama_sampler_p: @@ -3193,13 +3772,14 @@ def llama_sampler_init_dist(seed: int) -> llama_sampler_p: # /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits. # /// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first. # DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax (void), -# "will be removed in the future (see https://github.com/ggerganov/llama.cpp/pull/9896#discussion_r1800920915)"); +# "will be removed in the future (see https://github.com/ggml-org/llama.cpp/pull/9896#discussion_r1800920915)"); @ctypes_function("llama_sampler_init_softmax", [], llama_sampler_p_ctypes) def llama_sampler_init_softmax() -> llama_sampler_p: ... # /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 +# /// Setting k <= 0 makes this a noop # LLAMA_API struct llama_sampler * llama_sampler_init_top_k (int32_t k); @ctypes_function("llama_sampler_init_top_k", [ctypes.c_int32], llama_sampler_p_ctypes) def llama_sampler_init_top_k(k: int) -> llama_sampler_p: @@ -3217,7 +3797,7 @@ def llama_sampler_init_top_p(p: float, min_keep: int) -> llama_sampler_p: ... -# /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841 +# /// @details Minimum P sampling as described in https://github.com/ggml-org/llama.cpp/pull/3841 # LLAMA_API struct llama_sampler * llama_sampler_init_min_p (float p, size_t min_keep); @ctypes_function( "llama_sampler_init_min_p", @@ -3239,6 +3819,7 @@ def llama_sampler_init_typical(p: float, min_keep: int) -> llama_sampler_p: ... +# /// #details Updates the logits l_i` = l_i/t. When t <= 0.0f, the maximum logit is kept at it's original value, the rest are set to -inf # LLAMA_API struct llama_sampler * llama_sampler_init_temp (float t); @ctypes_function("llama_sampler_init_temp", [ctypes.c_float], llama_sampler_p_ctypes) def llama_sampler_init_temp(t: float) -> llama_sampler_p: @@ -3271,12 +3852,18 @@ def llama_sampler_init_xtc( ... +# /// @details Top n sigma sampling as described in academic paper "Top-nσ: Not All Logits Are You Need" https://arxiv.org/pdf/2411.07641 +# LLAMA_API struct llama_sampler * llama_sampler_init_top_n_sigma(float n); +@ctypes_function( + "llama_sampler_init_top_n_sigma", + [ctypes.c_float], + llama_sampler_p_ctypes, +) +def llama_sampler_init_top_n_sigma(n: float, /) -> llama_sampler_p: + ... + + # /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. -# /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. -# /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. -# /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. -# /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm. -# /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. # LLAMA_API struct llama_sampler * llama_sampler_init_mirostat( # int32_t n_vocab, # uint32_t seed, @@ -3295,10 +3882,6 @@ def llama_sampler_init_mirostat( # /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. -# /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. -# /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. -# /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. -# /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. # LLAMA_API struct llama_sampler * llama_sampler_init_mirostat_v2( # uint32_t seed, # float tau, @@ -3314,17 +3897,88 @@ def llama_sampler_init_mirostat_v2( ... +# /// @details Intializes a GBNF grammar, see grammars/README.md for details. # LLAMA_API struct llama_sampler * llama_sampler_init_grammar( -# const struct llama_model * model, +# const struct llama_vocab * vocab, # const char * grammar_str, # const char * grammar_root); @ctypes_function( "llama_sampler_init_grammar", - [llama_model_p_ctypes, ctypes.c_char_p, ctypes.c_char_p], + [llama_vocab_p_ctypes, ctypes.c_char_p, ctypes.c_char_p], llama_sampler_p_ctypes, ) def llama_sampler_init_grammar( - model: llama_model_p, grammar_str: bytes, grammar_root: bytes, / + vocab: llama_vocab_p, grammar_str: bytes, grammar_root: bytes, / +) -> llama_sampler_p: + ... + + +# DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy( +# const struct llama_vocab * vocab, +# const char * grammar_str, +# const char * grammar_root, +# const char ** trigger_words, +# size_t num_trigger_words, +# const llama_token * trigger_tokens, +# size_t num_trigger_tokens), +# "use llama_sampler_init_grammar_lazy_patterns instead"); +@ctypes_function( + "llama_sampler_init_grammar_lazy", + [ + llama_vocab_p_ctypes, + ctypes.c_char_p, + ctypes.c_char_p, + ctypes.POINTER(ctypes.c_char_p), + ctypes.c_size_t, + ctypes.POINTER(llama_token), + ctypes.c_size_t, + ], + llama_sampler_p_ctypes, +) +def llama_sampler_init_grammar_lazy( + vocab: llama_vocab_p, + grammar_str: bytes, + grammar_root: bytes, + trigger_words: CtypesArray[bytes], + num_trigger_words: int, + trigger_tokens: CtypesArray[llama_token], + num_trigger_tokens: int, + /, +) -> llama_sampler_p: + ... + + +# /// @details Lazy grammar sampler, introduced in https://github.com/ggml-org/llama.cpp/pull/9639 +# LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy_patterns( +# const struct llama_vocab * vocab, +# const char * grammar_str, +# const char * grammar_root, +# const char ** trigger_patterns, +# size_t num_trigger_patterns, +# const llama_token * trigger_tokens, +# size_t num_trigger_tokens); +@ctypes_function( + "llama_sampler_init_grammar_lazy_patterns", + [ + llama_vocab_p_ctypes, + ctypes.c_char_p, + ctypes.c_char_p, + ctypes.POINTER(ctypes.c_char_p), + ctypes.c_size_t, + ctypes.POINTER(llama_token), + ctypes.c_size_t, + ], + llama_sampler_p_ctypes, +) +def llama_sampler_init_grammar_lazy_patterns( + vocab: llama_vocab_p, + grammar_str: bytes, + grammar_root: bytes, + trigger_patterns: CtypesArray[bytes], + num_trigger_patterns: int, + trigger_tokens: CtypesArray[llama_token], + num_trigger_tokens: int, + /, ) -> llama_sampler_p: ... @@ -3352,7 +4006,8 @@ def llama_sampler_init_penalties( # /// @details DRY sampler, designed by p-e-w, as described in: https://github.com/oobabooga/text-generation-webui/pull/5677, porting Koboldcpp implementation authored by pi6am: https://github.com/LostRuins/koboldcpp/pull/982 # LLAMA_API struct llama_sampler * llama_sampler_init_dry( -# const struct llama_model * model, +# const struct llama_vocab * vocab, +# int32_t n_ctx_train, # float dry_multiplier, # float dry_base, # int32_t dry_allowed_length, @@ -3362,7 +4017,8 @@ def llama_sampler_init_penalties( @ctypes_function( "llama_sampler_init_dry", [ - llama_model_p_ctypes, + llama_vocab_p_ctypes, + ctypes.c_int32, ctypes.c_float, ctypes.c_float, ctypes.c_int32, @@ -3373,12 +4029,13 @@ def llama_sampler_init_penalties( llama_sampler_p_ctypes, ) def llama_sampler_init_dry( - model: llama_model_p, + vocab: llama_vocab_p, + n_ctx_train: int, dry_multiplier: float, dry_base: float, dry_allowed_length: int, dry_penalty_last_n: int, - seq_breakers: CtypesArray[bytes], + seq_breakers, num_breakers: int, /, ) -> llama_sampler_p: @@ -3401,35 +4058,13 @@ def llama_sampler_init_logit_bias( # // this sampler is meant to be used for fill-in-the-middle infilling -# // it's supposed to be used after top_k + top_p sampling -# // -# // 1. if the sum of the EOG probs times the number of candidates is higher than the sum of the other probs -> pick EOG -# // 2. combine probs of tokens that have the same prefix -# // -# // example: -# // -# // - before: -# // "hel": 0.5 -# // "hell": 0.2 -# // "hello": 0.1 -# // "dummy": 0.1 -# // -# // - after: -# // "hel": 0.8 -# // "dummy": 0.1 -# // -# // 3. discard non-EOG tokens with low prob -# // 4. if no tokens are left -> pick EOT -# // -# LLAMA_API struct llama_sampler * llama_sampler_init_infill(const struct llama_model * model); +# LLAMA_API struct llama_sampler * llama_sampler_init_infill(const struct llama_vocab * vocab); @ctypes_function( "llama_sampler_init_infill", - [llama_model_p_ctypes], + [llama_vocab_p_ctypes], llama_sampler_p_ctypes, ) -def llama_sampler_init_infill(model: llama_model_p, /) -> llama_sampler_p: - """This sampler is meant to be used for fill-in-the-middle infilling. - """ +def llama_sampler_init_infill(vocab: llama_vocab_p, /) -> llama_sampler_p: ... @@ -3445,15 +4080,6 @@ def llama_sampler_get_seed(smpl: llama_sampler_p, /) -> int: # /// @details Sample and accept a token from the idx-th output of the last evaluation -# // -# // Shorthand for: -# // const auto * logits = llama_get_logits_ith(ctx, idx); -# // llama_token_data_array cur_p = { ... init from logits ... }; -# // llama_sampler_apply(smpl, &cur_p); -# // auto token = cur_p.data[cur_p.selected].id; -# // llama_sampler_accept(smpl, token); -# // return token; -# // Returns the sampled token # LLAMA_API llama_token llama_sampler_sample(struct llama_sampler * smpl, struct llama_context * ctx, int32_t idx); @ctypes_function( "llama_sampler_sample", @@ -3470,10 +4096,7 @@ def llama_sampler_sample( # // Model split # // - # /// @details Build a split GGUF final path for this chunk. -# /// llama_split_path(split_path, sizeof(split_path), "/models/ggml-model-q4_0", 2, 4) => split_path = "/models/ggml-model-q4_0-00002-of-00004.gguf" -# // Returns the split_path length. # LLAMA_API int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count); @ctypes_function( "llama_split_path", @@ -3493,8 +4116,6 @@ def llama_split_path( # /// @details Extract the path prefix from the split_path if and only if the split_no and split_count match. -# /// llama_split_prefix(split_prefix, 64, "/models/ggml-model-q4_0-00002-of-00004.gguf", 2, 4) => split_prefix = "/models/ggml-model-q4_0" -# // Returns the split_prefix length. # LLAMA_API int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count); @ctypes_function( "llama_split_prefix", @@ -3542,16 +4163,13 @@ def llama_log_set( # // # // Performance utils # // -# // NOTE: Used by llama.cpp examples, avoid using in third-party apps. Instead, do your own performance measurements. -# // - # struct llama_perf_context_data { # double t_start_ms; # double t_load_ms; # double t_p_eval_ms; # double t_eval_ms; -# + # int32_t n_p_eval; # int32_t n_eval; # }; @@ -3568,7 +4186,7 @@ class llama_perf_context_data(ctypes.Structure): # struct llama_perf_sampler_data { # double t_sample_ms; -# + # int32_t n_sample; # }; class llama_perf_sampler_data(ctypes.Structure): @@ -3639,3 +4257,83 @@ def llama_perf_sampler_reset(chain: llama_sampler_p, /): ... +# // +# // training +# // + +# // function that returns whether or not a given tensor contains trainable parameters +# typedef bool (*llama_opt_param_filter)(const struct ggml_tensor * tensor, void * userdata); +llama_opt_param_filter = ctypes.CFUNCTYPE(ctypes.c_bool, ctypes.c_void_p, ctypes.c_void_p) + +# // always returns true +# LLAMA_API bool llama_opt_param_filter_all(const struct ggml_tensor * tensor, void * userdata); +@ctypes_function( + "llama_opt_param_filter_all", + [ctypes.c_void_p, ctypes.c_void_p], + ctypes.c_bool, +) +def llama_opt_param_filter_all(tensor: ctypes.c_void_p, userdata: ctypes.c_void_p, /) -> bool: + ... + + +# struct llama_opt_params { +# uint32_t n_ctx_train; // assumed context size post training, use context size specified in llama_context if 0 + +# llama_opt_param_filter param_filter; // callback for determining which tensors contain trainable parameters +# void * param_filter_ud; // userdata for determining which tensors contain trainable parameters + +# ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters +# void * get_opt_pars_ud; // userdata for calculating optimizer parameters +# }; +class llama_opt_params(ctypes.Structure): + _fields_ = [ + ("n_ctx_train", ctypes.c_uint32), + ("param_filter", llama_opt_param_filter), + ("param_filter_ud", ctypes.c_void_p), + ("get_opt_pars", ctypes.c_void_p), # ggml_opt_get_optimizer_params - not implemented here + ("get_opt_pars_ud", ctypes.c_void_p), + ] + + +# LLAMA_API void llama_opt_init(struct llama_context * lctx, struct llama_model * model, struct llama_opt_params lopt_params); +@ctypes_function( + "llama_opt_init", + [llama_context_p_ctypes, llama_model_p_ctypes, llama_opt_params], + None, +) +def llama_opt_init(lctx: llama_context_p, model: llama_model_p, lopt_params: llama_opt_params, /): + ... + + +# LLAMA_API void llama_opt_epoch( +# struct llama_context * lctx, +# ggml_opt_dataset_t dataset, +# ggml_opt_result_t result_train, +# ggml_opt_result_t result_eval, +# int64_t idata_split, +# ggml_opt_epoch_callback callback_train, +# ggml_opt_epoch_callback callback_eval); +@ctypes_function( + "llama_opt_epoch", + [ + llama_context_p_ctypes, + ctypes.c_void_p, # ggml_opt_dataset_t + ctypes.c_void_p, # ggml_opt_result_t + ctypes.c_void_p, # ggml_opt_result_t + ctypes.c_int64, + ctypes.c_void_p, # ggml_opt_epoch_callback + ctypes.c_void_p, # ggml_opt_epoch_callback + ], + None, +) +def llama_opt_epoch( + lctx: llama_context_p, + dataset: ctypes.c_void_p, + result_train: ctypes.c_void_p, + result_eval: ctypes.c_void_p, + idata_split: int, + callback_train: ctypes.c_void_p, + callback_eval: ctypes.c_void_p, + /, +): + ... diff --git a/llama_cpp/server/app.py b/llama_cpp/server/app.py index b6db453b8..5120f2416 100644 --- a/llama_cpp/server/app.py +++ b/llama_cpp/server/app.py @@ -7,7 +7,7 @@ from anyio import Lock from functools import partial -from typing import Iterator, List, Optional, Union, Dict +from typing import List, Optional, Union, Dict import llama_cpp @@ -155,34 +155,71 @@ def create_app( return app +def prepare_request_resources( + body: CreateCompletionRequest | CreateChatCompletionRequest, + llama_proxy: LlamaProxy, + body_model: str | None, + kwargs, +) -> llama_cpp.Llama: + if llama_proxy is None: + raise HTTPException( + status_code=status.HTTP_503_SERVICE_UNAVAILABLE, + detail="Service is not available", + ) + llama = llama_proxy(body_model) + if body.logit_bias is not None: + kwargs["logit_bias"] = ( + _logit_bias_tokens_to_input_ids(llama, body.logit_bias) + if body.logit_bias_type == "tokens" + else body.logit_bias + ) + + 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) + return llama + + async def get_event_publisher( request: Request, inner_send_chan: MemoryObjectSendStream[typing.Any], - iterator: Iterator[typing.Any], - on_complete: typing.Optional[typing.Callable[[], typing.Awaitable[None]]] = None, + body: CreateCompletionRequest | CreateChatCompletionRequest, + body_model: str | None, + llama_call, + kwargs, ): server_settings = next(get_server_settings()) interrupt_requests = ( server_settings.interrupt_requests if server_settings else False ) - async with inner_send_chan: - try: - async for chunk in iterate_in_threadpool(iterator): - await inner_send_chan.send(dict(data=json.dumps(chunk))) - if await request.is_disconnected(): - raise anyio.get_cancelled_exc_class()() - if interrupt_requests and llama_outer_lock.locked(): - await inner_send_chan.send(dict(data="[DONE]")) - raise anyio.get_cancelled_exc_class()() - await inner_send_chan.send(dict(data="[DONE]")) - except anyio.get_cancelled_exc_class() as e: - print("disconnected") - with anyio.move_on_after(1, shield=True): - print(f"Disconnected from client (via refresh/close) {request.client}") - raise e - finally: - if on_complete: - await on_complete() + async with contextlib.asynccontextmanager(get_llama_proxy)() as llama_proxy: + llama = prepare_request_resources(body, llama_proxy, body_model, kwargs) + async with inner_send_chan: + try: + iterator = await run_in_threadpool(llama_call, llama, **kwargs) + async for chunk in iterate_in_threadpool(iterator): + await inner_send_chan.send(dict(data=json.dumps(chunk))) + if await request.is_disconnected(): + raise anyio.get_cancelled_exc_class()() + if interrupt_requests and llama_outer_lock.locked(): + await inner_send_chan.send(dict(data="[DONE]")) + raise anyio.get_cancelled_exc_class()() + await inner_send_chan.send(dict(data="[DONE]")) + except anyio.get_cancelled_exc_class() as e: + print("disconnected") + with anyio.move_on_after(1, shield=True): + print( + f"Disconnected from client (via refresh/close) {request.client}" + ) + raise e def _logit_bias_tokens_to_input_ids( @@ -267,18 +304,11 @@ async def create_completion( request: Request, body: CreateCompletionRequest, ) -> llama_cpp.Completion: - exit_stack = contextlib.AsyncExitStack() - llama_proxy = await exit_stack.enter_async_context(contextlib.asynccontextmanager(get_llama_proxy)()) - if llama_proxy is None: - raise HTTPException( - status_code=status.HTTP_503_SERVICE_UNAVAILABLE, - detail="Service is not available", - ) if isinstance(body.prompt, list): assert len(body.prompt) <= 1 body.prompt = body.prompt[0] if len(body.prompt) > 0 else "" - llama = llama_proxy( + body_model = ( body.model if request.url.path != "/v1/engines/copilot-codex/completions" else "copilot-codex" @@ -293,44 +323,8 @@ async def create_completion( } kwargs = body.model_dump(exclude=exclude) - if body.logit_bias is not None: - kwargs["logit_bias"] = ( - _logit_bias_tokens_to_input_ids(llama, body.logit_bias) - if body.logit_bias_type == "tokens" - else body.logit_bias - ) - - 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) - - try: - iterator_or_completion: Union[ - llama_cpp.CreateCompletionResponse, - Iterator[llama_cpp.CreateCompletionStreamResponse], - ] = await run_in_threadpool(llama, **kwargs) - except Exception as err: - await exit_stack.aclose() - raise err - - if isinstance(iterator_or_completion, Iterator): - # EAFP: It's easier to ask for forgiveness than permission - first_response = await run_in_threadpool(next, iterator_or_completion) - - # If no exception was raised from first_response, we can assume that - # the iterator is valid and we can use it to stream the response. - def iterator() -> Iterator[llama_cpp.CreateCompletionStreamResponse]: - yield first_response - yield from iterator_or_completion - + # handle streaming request + if kwargs.get("stream", False): send_chan, recv_chan = anyio.create_memory_object_stream(10) return EventSourceResponse( recv_chan, @@ -338,15 +332,29 @@ def iterator() -> Iterator[llama_cpp.CreateCompletionStreamResponse]: get_event_publisher, request=request, inner_send_chan=send_chan, - iterator=iterator(), - on_complete=exit_stack.aclose, + body=body, + body_model=body_model, + llama_call=llama_cpp.Llama.__call__, + kwargs=kwargs, ), sep="\n", ping_message_factory=_ping_message_factory, ) - else: - await exit_stack.aclose() - return iterator_or_completion + + # handle regular request + async with contextlib.asynccontextmanager(get_llama_proxy)() as llama_proxy: + llama = prepare_request_resources(body, llama_proxy, body_model, kwargs) + + if await request.is_disconnected(): + print( + f"Disconnected from client (via refresh/close) before llm invoked {request.client}" + ) + raise HTTPException( + status_code=status.HTTP_400_BAD_REQUEST, + detail="Client closed request", + ) + + return await run_in_threadpool(llama, **kwargs) @router.post( @@ -474,13 +482,8 @@ async def create_chat_completion( # where the dependency is cleaned up before a StreamingResponse # is complete. # https://github.com/tiangolo/fastapi/issues/11143 - exit_stack = contextlib.AsyncExitStack() - llama_proxy = await exit_stack.enter_async_context(contextlib.asynccontextmanager(get_llama_proxy)()) - if llama_proxy is None: - raise HTTPException( - status_code=status.HTTP_503_SERVICE_UNAVAILABLE, - detail="Service is not available", - ) + + body_model = body.model exclude = { "n", "logit_bias_type", @@ -488,44 +491,9 @@ async def create_chat_completion( "min_tokens", } kwargs = body.model_dump(exclude=exclude) - llama = llama_proxy(body.model) - if body.logit_bias is not None: - kwargs["logit_bias"] = ( - _logit_bias_tokens_to_input_ids(llama, body.logit_bias) - if body.logit_bias_type == "tokens" - else body.logit_bias - ) - - 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) - - try: - iterator_or_completion: Union[ - llama_cpp.ChatCompletion, Iterator[llama_cpp.ChatCompletionChunk] - ] = await run_in_threadpool(llama.create_chat_completion, **kwargs) - except Exception as err: - await exit_stack.aclose() - raise err - - if isinstance(iterator_or_completion, Iterator): - # EAFP: It's easier to ask for forgiveness than permission - first_response = await run_in_threadpool(next, iterator_or_completion) - - # If no exception was raised from first_response, we can assume that - # the iterator is valid and we can use it to stream the response. - def iterator() -> Iterator[llama_cpp.ChatCompletionChunk]: - yield first_response - yield from iterator_or_completion + # handle streaming request + if kwargs.get("stream", False): send_chan, recv_chan = anyio.create_memory_object_stream(10) return EventSourceResponse( recv_chan, @@ -533,15 +501,29 @@ def iterator() -> Iterator[llama_cpp.ChatCompletionChunk]: get_event_publisher, request=request, inner_send_chan=send_chan, - iterator=iterator(), - on_complete=exit_stack.aclose, + body=body, + body_model=body_model, + llama_call=llama_cpp.Llama.create_chat_completion, + kwargs=kwargs, ), sep="\n", ping_message_factory=_ping_message_factory, ) - else: - await exit_stack.aclose() - return iterator_or_completion + + # handle regular request + async with contextlib.asynccontextmanager(get_llama_proxy)() as llama_proxy: + llama = prepare_request_resources(body, llama_proxy, body_model, kwargs) + + if await request.is_disconnected(): + print( + f"Disconnected from client (via refresh/close) before llm invoked {request.client}" + ) + raise HTTPException( + status_code=status.HTTP_400_BAD_REQUEST, + detail="Client closed request", + ) + + return await run_in_threadpool(llama.create_chat_completion, **kwargs) @router.get( diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 716bd6dec..8846aace4 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 716bd6dec3e044e5c325386b5b0483392b24cefe +Subproject commit 8846aace4934ad29651ea61b8c7e3f6b0556e3d2 pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy