router.post("/v1/chat/completions", tags=["Completions"])

router.post("/v1/completions", tags=["Completions"])

router.post( "/extras/tokenize/count", summary="Tokenize Count", dependencies=[Depends(authenticat

router.post( "/v1/chat/completions", summary="Chat", dependencies=[Depends(authenticate)], res

router.post( "/v1/completions", summary="Completion", dependencies=[Depends(authenticate)], re

router.post( "/v1/embeddings", summary="Embedding", dependencies=[Depends(authenticate)], tags

router.get("/dashboard/billing/credit_grants", tags=["MISC"])

router.post( "/extras/detokenize", summary="Detokenize", dependencies=[Depends(authenticate)],

router.post("/disable-state-cache", tags=["State Cache"])

router.post("/v1/embeddings", tags=["Embeddings"])

router.post("/enable-state-cache", tags=["State Cache"])

app.post("/exit", tags=["Root"])

router.post("/file-to-text", tags=["File Process"])

router.get("/albatross/profile", tags=["Albatross"])

router.get( "/v1/models", summary="Models", dependencies=[Depends(authenticate)], tags=[openai

ggml_base_function("ggml_backend_buffer_get_alignment", [ctypes.c_void_p], ctypes.c_size_t)

ggml_base_function("ggml_backend_buffer_get_alloc_size", [ ctypes.c_void_p, ggml_tensor_p, ], ctypes.c

ggml_base_function("ggml_backend_buffer_get_base", [ctypes.c_void_p], None)

ggml_base_function("ggml_backend_buffer_get_max_size", [ctypes.c_void_p], ctypes.c_size_t)

ggml_base_function("ggml_backend_buffer_get_size", [ctypes.c_void_p], ctypes.c_size_t)

ggml_base_function("ggml_backend_buffer_get_type", [ctypes.c_void_p], ctypes.c_void_p)

ggml_base_function("ggml_backend_buffer_get_usage", [ctypes.c_void_p], ctypes.c_int32)

ggml_base_function("ggml_backend_buft_get_alignment", [ctypes.c_void_p], ctypes.c_size_t)

ggml_base_function("ggml_backend_buft_get_alloc_size", [ ctypes.c_void_p, ggml_tensor_p, ], ctypes.c_s

ggml_base_function("ggml_backend_buft_get_device", [ctypes.c_void_p], ctypes.c_void_p)

ggml_base_function("ggml_backend_buft_get_max_size", [ctypes.c_void_p], ctypes.c_size_t)

ggml_function("ggml_backend_dev_get", [ctypes.c_size_t], ctypes.c_void_p)

ggml_function("ggml_backend_reg_get", [ctypes.c_size_t], ctypes.c_void_p)

ggml_base_function( "ggml_log_get", [ctypes.POINTER(ggml_log_callback), ctypes.POINTER(ctypes.c_void_p

ctypes_function( "llama_adapter_get_alora_invocation_tokens", [llama_adapter_lora_p_ctypes], ctype

ctypes_function( "llama_adapter_get_alora_n_invocation_tokens", [llama_adapter_lora_p_ctypes], cty

ctypes_function( "llama_batch_get_one", [ llama_token_p, ctypes.c_int32, ], ll

ctypes_function( "llama_get_embeddings", [llama_context_p_ctypes], ctypes.POINTER(ctypes.c_float) )

ctypes_function( "llama_get_embeddings_ith", [llama_context_p_ctypes, ctypes.c_int32], ctypes.POIN

ctypes_function( "llama_get_embeddings_seq", [llama_context_p_ctypes, llama_seq_id], ctypes.POINTE

ctypes_function( "llama_get_logits", [llama_context_p_ctypes], ctypes.POINTER(ctypes.c_float) )

ctypes_function( "llama_get_logits_ith", [llama_context_p_ctypes, ctypes.c_int32], ctypes.POINTER(

ctypes_function("llama_get_memory", [llama_context_p_ctypes], llama_memory_i_p_ctypes)

ctypes_function("llama_get_model", [llama_context_p_ctypes], llama_model_p_ctypes)

ctypes_function( "llama_get_sampled_candidates_count_ith", [llama_context_p_ctypes, ctypes.c_int32],

ctypes_function( "llama_get_sampled_candidates_ith", [llama_context_p_ctypes, ctypes.c_int32], cty

ctypes_function( "llama_get_sampled_logits_count_ith", [llama_context_p_ctypes, ctypes.c_int32], c

ctypes_function( "llama_get_sampled_logits_ith", [llama_context_p_ctypes, ctypes.c_int32], ctypes.

ctypes_function( "llama_get_sampled_probs_count_ith", [llama_context_p_ctypes, ctypes.c_int32], ct

ctypes_function( "llama_get_sampled_probs_ith", [llama_context_p_ctypes, ctypes.c_int32], ctypes.P

ctypes_function( "llama_get_sampled_token_ith", [llama_context_p_ctypes, ctypes.c_int32], llama_to

ctypes_function("llama_get_state_size", [llama_context_p_ctypes], ctypes.c_size_t)

ctypes_function( "llama_log_get", [ctypes.POINTER(ggml_log_callback), ctypes.POINTER(ctypes.c_void_p)]

ctypes_function("llama_model_get_vocab", [llama_model_p_ctypes], llama_vocab_p_ctypes)

ctypes_function( "llama_sampler_chain_get", [llama_sampler_p_ctypes, ctypes.c_int32], llama_sample

ctypes_function( "llama_sampler_get_seed", [llama_sampler_p_ctypes], ctypes.c_uint32, )

ctypes_function( "llama_state_get_data", [ llama_context_p_ctypes, ctypes.POINTER(ctyp

ctypes_function("llama_state_get_size", [llama_context_p_ctypes], ctypes.c_size_t)

ctypes_function( "llama_state_seq_get_data", [ llama_context_p_ctypes, ctypes.POINTER(

ctypes_function( "llama_state_seq_get_data_ext", [ llama_context_p_ctypes, ctypes.POIN

ctypes_function( "llama_state_seq_get_size", [llama_context_p_ctypes, llama_seq_id], llama_seq_id,

ctypes_function( "llama_state_seq_get_size_ext", [ llama_context_p_ctypes, llama_seq_i

ctypes_function( "llama_token_get_attr", [llama_vocab_p_ctypes, llama_token], ctypes.c_int, )

ctypes_function( "llama_token_get_score", [llama_vocab_p_ctypes, llama_token], ctypes.c_float, )

ctypes_function( "llama_token_get_text", [llama_vocab_p_ctypes, llama_token], ctypes.c_char_p, )

ctypes_function( "llama_vocab_get_add_bos", [llama_vocab_p_ctypes], ctypes.c_bool, )

ctypes_function( "llama_vocab_get_add_eos", [llama_vocab_p_ctypes], ctypes.c_bool, )

ctypes_function( "llama_vocab_get_add_sep", [llama_vocab_p_ctypes], ctypes.c_bool, )

ctypes_function( "llama_vocab_get_attr", [llama_vocab_p_ctypes, llama_token], ctypes.c_int )

ctypes_function( "llama_vocab_get_score", [llama_vocab_p_ctypes, llama_token], ctypes.c_float )

ctypes_function( "llama_vocab_get_text", [llama_vocab_p_ctypes, llama_token], ctypes.c_char_p )

router.post("/midi-to-text", tags=["MIDI"])

router.post("/midi-to-wav", tags=["MIDI"])

router.get("/v1/models/{model_id}", tags=["MISC"])

router.get("/v1/models", tags=["MISC"])

ctypes_function_mtmd("mtmd_bitmap_get_data", [mtmd_bitmap_p_ctypes], c_char_p)

ctypes_function_mtmd("mtmd_bitmap_get_id", [mtmd_bitmap_p_ctypes], c_char_p)

ctypes_function_mtmd("mtmd_bitmap_get_n_bytes", [mtmd_bitmap_p_ctypes], c_size_t)

ctypes_function_mtmd("mtmd_bitmap_get_nx", [mtmd_bitmap_p_ctypes], c_uint32)

ctypes_function_mtmd("mtmd_bitmap_get_ny", [mtmd_bitmap_p_ctypes], c_uint32)

ctypes_function_mtmd( "mtmd_decode_use_non_causal", [mtmd_context_p_ctypes, mtmd_input_chunk_p_ctypes], c_

ctypes_function_mtmd( "mtmd_encode_chunk", [ mtmd_context_p_ctypes, mtmd_input_chunk_p_cty

ctypes_function_mtmd( "mtmd_get_audio_sample_rate", [mtmd_context_p_ctypes], c_int)

ctypes_function_mtmd( "mtmd_get_output_embd", [mtmd_context_p_ctypes], POINTER(c_float))

ctypes_function_mtmd( "mtmd_helper_decode_image_chunk", [ mtmd_context_p_ctypes, llama_cpp

ctypes_function_mtmd( "mtmd_helper_eval_chunk_single", [ mtmd_context_p_ctypes, llama_cpp.

ctypes_function_mtmd( "mtmd_helper_eval_chunks", [ mtmd_context_p_ctypes, llama_cpp.llama_

ctypes_function_mtmd( "mtmd_helper_get_n_pos", [mtmd_input_chunk_p_ctypes], c_int32)

ctypes_function_mtmd( "mtmd_helper_get_n_tokens", [mtmd_input_chunk_p_ctypes], c_size_t)

ctypes_function_mtmd("mtmd_helper_image_get_decoder_pos", [ mtmd_image_tokens_p_ctypes

ctypes_function_mtmd( "mtmd_image_tokens_get_decoder_pos", [mtmd_image_tokens_p_ctypes, c_int32, c_size_t]

ctypes_function_mtmd( "mtmd_image_tokens_get_id", [mtmd_image_tokens_p_ctypes], c_char_p)

ctypes_function_mtmd( "mtmd_image_tokens_get_n_pos", [mtmd_image_tokens_p_ctypes], c_int32)

ctypes_function_mtmd( "mtmd_image_tokens_get_n_tokens", [mtmd_image_tokens_p_ctypes], c_size_t)

ctypes_function_mtmd( "mtmd_image_tokens_get_nx", [mtmd_image_tokens_p_ctypes], c_size_t)

ctypes_function_mtmd( "mtmd_image_tokens_get_ny", [mtmd_image_tokens_p_ctypes], c_size_t)

ctypes_function_mtmd("mtmd_input_chunk_copy", [mtmd_input_chunk_p_ctypes], mtmd_input_chunk_p_ctypes)

ctypes_function_mtmd("mtmd_input_chunk_free", [mtmd_input_chunk_p_ctypes], None)

ctypes_function_mtmd("mtmd_input_chunk_get_id", [mtmd_input_chunk_p_ctypes], c_char_p)

ctypes_function_mtmd("mtmd_input_chunk_get_n_pos", [mtmd_input_chunk_p_ctypes], c_int32)

ctypes_function_mtmd("mtmd_input_chunk_get_n_tokens", [mtmd_input_chunk_p_ctypes], c_size_t)

ctypes_function_mtmd("mtmd_input_chunk_get_tokens_image", [mtmd_input_chunk_p_ctypes], mtmd_image_tokens_p_cty

ctypes_function_mtmd( "mtmd_input_chunk_get_tokens_text", [mtmd_input_chunk_p_ctypes, POINTER(c_size_t

ctypes_function_mtmd("mtmd_input_chunk_get_type", [mtmd_input_chunk_p_ctypes], c_int32)

ctypes_function_mtmd("mtmd_input_chunks_free", [mtmd_input_chunks_p_ctypes], None)