llama.cpp/examples/low_level_api/low_level_api_chat_cpp.py

"""
This is an example implementation of main.cpp from llama.cpp
Quirks:
 * Its not exactly alike since this port is designed around programmatic I/O
 * Input is always echoed if on, so it should be turned off when using "input()"
 * The first antiprompt should be the userprompt like "\nUser:", 
   because its added when n_predict is reached (aka generation ended prematurely)
 * n_predict can be set to -1 for unlimited length responses (or just a really high value)
 * Instruction mode adds its own antiprompt.
   You should also still be feeding the model with a "primer" prompt that 
   shows it the expected format.
"""
import ctypes
import sys
from time import time
from os import cpu_count, path

import llama_cpp
from common import GptParams, gpt_params_parse, gpt_random_prompt
import util

# A LLaMA interactive session
class LLaMAInteract:
	def __init__(self, params: GptParams) -> None:
		# input args
		self.params = params

		if (self.params.perplexity):
			raise NotImplementedError("""************
please use the 'perplexity' tool for perplexity calculations
************""")

		if (self.params.embedding):
			raise NotImplementedError("""************
please use the 'embedding' tool for embedding calculations
************""")

		if (self.params.n_ctx > 2048):
			print(f"""warning: model does not support \
context sizes greater than 2048 tokens ({self.params.n_ctx} \
specified) expect poor results""", file=sys.stderr)

		if (self.params.seed <= 0):
			self.params.seed = int(time())

		print(f"seed = {self.params.seed}", file=sys.stderr)

		if (self.params.random_prompt):
			self.params.prompt = gpt_random_prompt(self.params.seed)

		# runtime args
		self.input_consumed = 0
		self.n_past = 0
		self.n_session_consumed = 0
		self.first_antiprompt = []
		self.remaining_tokens = self.params.n_predict
		self.output_echo = self.params.input_echo
		self.multibyte_fix = []

		# model load
		self.lparams = llama_cpp.llama_context_default_params()
		self.lparams.n_ctx = self.params.n_ctx
		self.lparams.n_parts = self.params.n_parts
		self.lparams.seed = self.params.seed
		self.lparams.memory_f16 = self.params.memory_f16
		self.lparams.use_mlock = self.params.use_mlock
		self.lparams.use_mmap = self.params.use_mmap

		self.ctx = llama_cpp.llama_init_from_file(self.params.model.encode("utf8"), self.lparams)
		if (not self.ctx):
			raise RuntimeError(f"error: failed to load model '{self.params.model}'")

		if (self.params.ignore_eos):
			self.params.logit_bias[llama_cpp.llama_token_eos()] = -float("inf")

		if (len(self.params.lora_adapter) > 0):
			if (llama_cpp.llama_apply_lora_from_file(
				self.ctx, 
				self.params.lora_adapter, 
				self.params.lora_base if len(self.params.lora_base) > 0 else None,
				self.params.n_threads
			) != 0):
				print("error: failed to apply lora adapter")
				return

		print(file=sys.stderr)
		print(f"system_info: n_threads = {self.params.n_threads} / {cpu_count()} \
| {llama_cpp.llama_print_system_info().decode('utf8')}", file=sys.stderr)

		# determine the required inference memory per token:
		if (self.params.mem_test):
			tmp = [0, 1, 2, 3]
			llama_cpp.llama_eval(self.ctx, (llama_cpp.c_int * len(tmp))(*tmp), len(tmp), 0, self.n_threads)
			llama_cpp.llama_print_timings(self.ctx)
			self.exit()
			return

		# create internal context
		self.n_ctx = llama_cpp.llama_n_ctx(self.ctx)

		# Add a space in front of the first character to match OG llama tokenizer behavior
		self.params.prompt = " " + self.params.prompt

		# Load prompt file
		if (self.params.file):
			with open(self.params.file) as f:
				self.params.prompt = f.read()

		self.session_tokens: list[llama_cpp.llama_token] = []
		if (len(self.params.path_session) > 0):
			print(f"attempting to load saved session from '{self.params.path_session}'", file=sys.stderr)

			if (path.exists(self.params.path_session)):
				_session_tokens = (llama_cpp.llama_token * (self.params.n_ctx))()
				_n_token_count_out = llama_cpp.c_int()
				if (llama_cpp.llama_load_session_file(
					self.ctx, 
					self.params.path_session.encode("utf8"),
					_session_tokens,
					self.params.n_ctx,
					ctypes.byref(_n_token_count_out)
				) != 0):
					print(f"error: failed to load session file '{self.params.path_session}'", file=sys.stderr)
					return
				self.session_tokens = _session_tokens[:_n_token_count_out]
				print(f"loaded a session with prompt size of {_n_token_count_out} tokens", file=sys.stderr)
			else:
				print(f"session file does not exist, will create", file=sys.stderr)

		# tokenize the prompt
		self.embd = []
		self.embd_inp = self._tokenize(self.params.prompt)

		if (len(self.embd_inp) > self.n_ctx - 4):
			raise RuntimeError(f"error: prompt is too long ({len(self.embd_inp)} tokens, max {self.params.n_ctx - 4})")

		# debug message about similarity of saved session, if applicable
		n_matching_session_tokens = 0
		if len(self.session_tokens) > 0:
			for id in self.session_tokens:
				if n_matching_session_tokens >= len(self.embd_inp) or id != self.embd_inp[n_matching_session_tokens]:
					break
				n_matching_session_tokens += 1
			
			if n_matching_session_tokens >= len(self.embd_inp):
				print(f"session file has exact match for prompt!")
			elif n_matching_session_tokens < (len(self.embd_inp) / 2):
				print(f"warning: session file has low similarity to prompt ({n_matching_session_tokens} / {len(self.embd_inp)} tokens); will mostly be reevaluated")
			else:
				print(f"session file matches {n_matching_session_tokens} / {len(self.embd_inp)} tokens of prompt")

		# number of tokens to keep when resetting context
		if (self.params.n_keep < 0 or self.params.n_keep > len(self.embd_inp) or self.params.instruct):
			self.params.n_keep = len(self.embd_inp)

		self.inp_prefix = self._tokenize(self.params.instruct_inp_prefix)
		self.inp_suffix = self._tokenize(self.params.instruct_inp_suffix, False)

		# in instruct mode, we inject a prefix and a suffix to each input by the user
		self.antiecho = None
		if (self.params.instruct):
			self.params.interactive_start = True
			_ptn = self._tokenize(self.params.instruct_inp_prefix.strip(), False)
			self.first_antiprompt.append(_ptn)
			self.antiecho = util.IterSearch(_ptn)

		# enable interactive mode if reverse prompt or interactive start is specified
		if (len(self.params.antiprompt) != 0 or self.params.interactive_start):
			self.params.interactive = True

		# determine newline token
		self.llama_token_newline = self._tokenize("\n", False)
		self.llama_token_eot = self._tokenize(" [end of text]\n", False)

		if (self.params.verbose_prompt):
			print(f"""
prompt: '{self.params.prompt}'
number of tokens in prompt = {len(self.embd_inp)}""", file=sys.stderr)

			for i in range(len(self.embd_inp)):
				print(f"{self.embd_inp[i]} -> '{llama_cpp.llama_token_to_str(self.ctx, self.embd_inp[i])}'", file=sys.stderr)

			if (self.params.n_keep > 0):
				print("static prompt based on n_keep: '")
				for i in range(self.params.n_keep):
					print(llama_cpp.llama_token_to_str(self.ctx, self.embd_inp[i]), file=sys.stderr)
				print("'", file=sys.stderr)
			print(file=sys.stderr)

		if (self.params.interactive):
			print("interactive mode on.", file=sys.stderr)

			if (len(self.params.antiprompt) > 0):
				for antiprompt in self.params.antiprompt:
					print(f"Reverse prompt: '{antiprompt}'", file=sys.stderr)

			if len(self.params.input_prefix) > 0:
				print(f"Input prefix: '{self.params.input_prefix}'", file=sys.stderr)

		print(f"""sampling: repeat_last_n = {self.params.repeat_last_n},\
repeat_penalty = {self.params.repeat_penalty},\
presence_penalty = {self.params.presence_penalty},\
frequency_penalty = {self.params.frequency_penalty},\
top_k = {self.params.top_k},\
tfs_z = {self.params.tfs_z},\
top_p = {self.params.top_p},\
typical_p = {self.params.typical_p},\
temp = {self.params.temp},\
mirostat = {self.params.mirostat},\
mirostat_lr = {self.params.mirostat_eta},\
mirostat_ent = {self.params.mirostat_tau},\

generate: n_ctx = {self.n_ctx},\
n_batch = {self.params.n_batch},\
n_predict = {self.params.n_predict},\
n_keep = {self.params.n_keep}

""", file=sys.stderr)

		# determine antiprompt tokens
		for i in self.params.antiprompt:
			self.first_antiprompt.append(self._tokenize(i, False))

		self.last_n_tokens = [0]*self.n_ctx #TODO: deque doesnt support slices

		if (params.interactive):
			print("""== Running in interactive mode. ==
 - Press Ctrl+C to interject at any time.
 - Press Return to return control to LLaMa.
 - If you want to submit another line, end your input in '\\'.

""", file=sys.stderr)
		self.set_color(util.CONSOLE_COLOR_PROMPT)

		self.need_to_save_session = len(self.params.path_session) > 0 and n_matching_session_tokens < (len(self.embd_inp) * 3 / 4)


	# tokenize a prompt
	def _tokenize(self, prompt, bos=True):
		_arr = (llama_cpp.llama_token * ((len(prompt) + 1) * 4))()
		_n = llama_cpp.llama_tokenize(self.ctx, prompt.encode("utf8", errors="ignore"), _arr, len(_arr), bos)
		return _arr[:_n]

	def set_color(self, c):
		if (self.params.use_color):
			print(c, end="")

	def use_antiprompt(self):
		return len(self.first_antiprompt) > 0

	# generate tokens
	def generate(self):
		while self.remaining_tokens > 0 or self.params.interactive or self.params.n_predict == -1:
			# predict
			if len(self.embd) > 0:
				# infinite text generation via context swapping
				# if we run out of context:
				# - take the n_keep first tokens from the original prompt (via n_past)
				# - take half of the last (n_ctx - n_keep) tokens and recompute the logits in a batch
				if (self.n_past + len(self.embd) > self.n_ctx):
					n_left = self.n_past - self.params.n_keep
					self.n_past = self.params.n_keep

					# insert n_left/2 tokens at the start of embd from last_n_tokens
					_insert = self.last_n_tokens[
						self.n_ctx - int(n_left/2) - len(self.embd):-len(self.embd)
					]
					self.embd = _insert + self.embd
					self.params.path_session = ""

				# try to reuse a matching prefix from the loaded session instead of re-eval (via n_past)
				if self.n_session_consumed < len(self.session_tokens):
					for i in range(len(self.embd)):
						if self.embd[i] != self.session_tokens[self.n_session_consumed]:
							self.session_tokens = self.session_tokens[:self.n_session_consumed]
							break
						
						self.n_past += 1
						self.n_session_consumed += 1
						
						if self.n_session_consumed >= len(self.session_tokens):
							i += 1
							break
					
					if i > 0:
						self.embd = self.embd[i:]

				# evaluate tokens in batches
				# embd is typically prepared beforehand to fit within a batch, but not always
				#TODO BUG: The batching code causes nonsensical generation
				"""for i in range(0, len(self.embd), self.params.n_batch):
					n_eval = self.params.n_batch
					_arr = (llama_cpp.llama_token * n_eval)(*self.embd[i:i + n_eval])
					if llama_cpp.llama_eval(self.ctx, _arr, n_eval, self.n_past, self.params.n_threads) != 0:
						print(f"failed to eval")
						return
					
					self.n_past += n_eval"""

				if (llama_cpp.llama_eval(
					self.ctx, (llama_cpp.llama_token * len(self.embd))(*self.embd), len(self.embd), self.n_past, self.params.n_threads
				) != 0):
					raise Exception("Failed to llama_eval!")

				if len(self.embd) > 0 and not len(self.params.path_session) > 0:
					self.session_tokens.extend(self.embd)
					self.n_session_consumed = len(self.session_tokens)

			self.n_past += len(self.embd)
			self.embd = []
			if len(self.embd_inp) <= self.input_consumed: #&& !is_interacting
				# out of user input, sample next token
				top_k = llama_cpp.llama_n_vocab(self.ctx) if self.params.top_k <= 0 else self.params.top_k
				repeat_last_n = self.n_ctx if self.params.repeat_last_n < 0 else self.params.repeat_last_n

				# optionally save the session on first sample (for faster prompt loading next time)
				if len(self.params.path_session) > 0 and self.need_to_save_session:
					self.need_to_save_session = False
					llama_cpp.llama_save_session_file(
						self.ctx,
						self.params.path_session.encode("utf8"),
						self.session_tokens,
						len(self.session_tokens)
					)

				id = 0

				logits = llama_cpp.llama_get_logits(self.ctx)
				n_vocab = llama_cpp.llama_n_vocab(self.ctx)

				# Apply params.logit_bias map
				for key, value in self.params.logit_bias.items():
					logits[key] += value

				_arr = (llama_cpp.llama_token_data * n_vocab)(*[
					llama_cpp.llama_token_data(token_id, logits[token_id], 0.0)
					for token_id in range(n_vocab)
				])
				candidates_p = llama_cpp.ctypes.pointer(llama_cpp.llama_token_data_array(_arr, len(_arr), False))

				# Apply penalties
				nl_logit = logits[llama_cpp.llama_token_nl()]
				last_n_repeat = min(len(self.last_n_tokens), repeat_last_n, self.n_ctx)

				_arr = (llama_cpp.llama_token * last_n_repeat)(*self.last_n_tokens[len(self.last_n_tokens) - last_n_repeat:])
				llama_cpp.llama_sample_repetition_penalty(self.ctx, candidates_p,
					_arr,
					last_n_repeat, llama_cpp.c_float(self.params.repeat_penalty))
				llama_cpp.llama_sample_frequency_and_presence_penalties(self.ctx, candidates_p,
					_arr,
					last_n_repeat, llama_cpp.c_float(self.params.frequency_penalty), llama_cpp.c_float(self.params.presence_penalty))

				if not self.params.penalize_nl:
					logits[llama_cpp.llama_token_nl()] = nl_logit
				
				if self.params.temp <= 0:
					# Greedy sampling
					id = llama_cpp.llama_sample_token_greedy(self.ctx, candidates_p)
				else:
					if self.params.mirostat == 1:
						mirostat_mu = 2.0 * self.params.mirostat_tau
						mirostat_m = 100
						llama_cpp.llama_sample_temperature(self.ctx, candidates_p, llama_cpp.c_float(self.params.temp))
						id = llama_cpp.llama_sample_token_mirostat(self.ctx, candidates_p, llama_cpp.c_float(self.params.mirostat_tau), llama_cpp.c_float(self.params.mirostat_eta), llama_cpp.c_int(mirostat_m), llama_cpp.c_float(mirostat_mu))
					elif self.params.mirostat == 2:
						mirostat_mu = 2.0 * self.params.mirostat_tau
						llama_cpp.llama_sample_temperature(self.ctx, candidates_p, llama_cpp.c_float(self.params.temp))
						id = llama_cpp.llama_sample_token_mirostat_v2(self.ctx, candidates_p, llama_cpp.c_float(self.params.mirostat_tau), llama_cpp.c_float(self.params.mirostat_eta), llama_cpp.c_float(mirostat_mu))
					else:
						# Temperature sampling
						llama_cpp.llama_sample_top_k(self.ctx, candidates_p, top_k)
						llama_cpp.llama_sample_tail_free(self.ctx, candidates_p, llama_cpp.c_float(self.params.tfs_z))
						llama_cpp.llama_sample_typical(self.ctx, candidates_p, llama_cpp.c_float(self.params.typical_p))
						llama_cpp.llama_sample_top_p(self.ctx, candidates_p, llama_cpp.c_float(self.params.top_p))
						llama_cpp.llama_sample_temperature(self.ctx, candidates_p, llama_cpp.c_float(self.params.temp))
						id = llama_cpp.llama_sample_token(self.ctx, candidates_p)
				# print("`{}`".format(candidates_p.size))

				self.last_n_tokens.pop(0)
				self.last_n_tokens.append(id)

				# replace end of text token with newline token when in interactive mode
				if (id == llama_cpp.llama_token_eos() and self.params.interactive and not self.params.instruct):
					id = self.llama_token_newline[0]
					if (self.use_antiprompt()):
						# tokenize and inject first reverse prompt
						self.embd_inp += self.first_antiprompt[0]

				# add it to the context
				self.embd.append(id)

				# echo this to console
				self.output_echo = True

				# decrement remaining sampling budget
				self.remaining_tokens -= 1
			else:
				# output to console if input echo is on
				self.output_echo = self.params.input_echo

				# some user input remains from prompt or interaction, forward it to processing
				while len(self.embd_inp) > self.input_consumed:
					self.embd.append(self.embd_inp[self.input_consumed])
					self.last_n_tokens.pop(0)
					self.last_n_tokens.append(self.embd_inp[self.input_consumed])
					self.input_consumed += 1
					if len(self.embd) >= self.params.n_batch:
						break

			# display tokens
			if self.output_echo:
				for id in self.embd:
					if self.antiecho != None:
						for r in self.antiecho(id):
							yield r
					else:
						yield id

			# reset color to default if we there is no pending user input
			if (self.params.input_echo and len(self.embd_inp) == self.input_consumed):
				self.set_color(util.CONSOLE_COLOR_DEFAULT)

			if (self.params.interactive and len(self.embd_inp) <= self.input_consumed):
				# if antiprompt is present, stop
				if (self.use_antiprompt()):
					if True in [
						i == self.last_n_tokens[-len(i):] 
						for i in self.first_antiprompt
					]:
						break

				# if we are using instruction mode, and we have processed the initial prompt
				if (self.params.interactive_start):
					break

			# end of text token
			if len(self.embd) > 0 and self.embd[-1] == llama_cpp.llama_token_eos():
				if (not self.params.instruct):
					for i in self.llama_token_eot:
						yield i
					break

			# respect n_predict even if antiprompt is present
			if (self.params.interactive and self.remaining_tokens <= 0 and self.params.n_predict != -1):
				# If we arent in instruction mode, fix the current generation by appending the antiprompt.
				# Makes it so if chat ends prematurely you dont append the AI's text etc.
				if not self.params.instruct:
					self.embd_inp += self.first_antiprompt[0]
				self.n_remain = self.params.n_predict
				break

		self.params.interactive_start = False

	def __enter__(self):
		return self

	def __exit__(self, type, value, tb):
		self.exit()

	def exit(self):
		llama_cpp.llama_free(self.ctx)
		self.set_color(util.CONSOLE_COLOR_DEFAULT)

	# return past text
	def past(self):
		for id in self.last_n_tokens[-self.n_past:]:
			yield llama_cpp.llama_token_to_str(self.ctx, id).decode("utf8", errors="ignore")

	# write input
	def input(self, prompt: str):
		if (self.params.instruct and self.last_n_tokens[-len(self.inp_prefix):] != self.inp_prefix):
			self.embd_inp += self.inp_prefix
		self.embd_inp += self._tokenize(prompt)
		if (self.params.instruct):
			self.embd_inp += self.inp_suffix

	# write output
	def output(self):
		self.remaining_tokens = self.params.n_predict
		for id in self.generate():
			cur_char = llama_cpp.llama_token_to_str(self.ctx, id)

			# Add remainder of missing bytes
			if None in self.multibyte_fix:
				self.multibyte_fix[self.multibyte_fix.index(None)] = cur_char

			# Return completed utf char
			if len(self.multibyte_fix) > 0 and not None in self.multibyte_fix:
				yield (b"".join(self.multibyte_fix)).decode("utf8")
				self.multibyte_fix = []
				continue

			# Contains multi-byte UTF8
			for num, pattern in [(2, 192), (3, 224), (4, 240)]:
				# Bitwise AND check
				if pattern & int.from_bytes(cur_char) == pattern:
					self.multibyte_fix = [cur_char] + ([None] * (num-1))

			# Stop incomplete bytes from passing
			if len(self.multibyte_fix) > 0:
				continue

			yield cur_char.decode("utf8")

	# read user input
	def read_input(self):
		out = ""
		while (t := input()).endswith("\\"):
			out += t[:-1] + "\n"
		return out + t + "\n"

	# interactive mode
	def interact(self):
		for i in self.output():
			print(i,end="",flush=True)
		self.params.input_echo = False

		while self.params.interactive:
			self.set_color(util.CONSOLE_COLOR_USER_INPUT)
			if (self.params.instruct):
				print('\n> ', end="")
				self.input(self.read_input())
			else:
				print(self.params.input_prefix, end="")
				self.input(f"{self.params.input_prefix}{self.read_input()}{self.params.input_suffix}")
				print(self.params.input_suffix,end="")
			self.set_color(util.CONSOLE_COLOR_DEFAULT)

			try:
				for i in self.output():
					print(i,end="",flush=True)
			except KeyboardInterrupt:
				self.set_color(util.CONSOLE_COLOR_DEFAULT)
				if not self.params.instruct:
					print(self.params.fix_prefix,end="")
					self.input(self.params.fix_prefix)

if __name__ == "__main__":
	from datetime import datetime

	USER_NAME="User"
	AI_NAME="ChatLLaMa"

	time_now = datetime.now()
	prompt = f"""Text transcript of a never ending dialog, where {USER_NAME} interacts with an AI assistant named {AI_NAME}.
{AI_NAME} is helpful, kind, honest, friendly, good at writing and never fails to answer {USER_NAME}’s requests immediately and with details and precision.
There are no annotations like (30 seconds passed...) or (to himself), just what {USER_NAME} and {AI_NAME} say aloud to each other.
The dialog lasts for years, the entirety of it is shared below. It's 10000 pages long.
The transcript only includes text, it does not include markup like HTML and Markdown.

{USER_NAME}: Hello, {AI_NAME}!
{AI_NAME}: Hello {USER_NAME}! How may I help you today?
{USER_NAME}: What time is it?
{AI_NAME}: It is {time_now.strftime("%H:%M")}.
{USER_NAME}: What year is it?
{AI_NAME}: We are in {time_now.strftime("%Y")}.
{USER_NAME}: What is a cat?
{AI_NAME}: A cat is a domestic species of small carnivorous mammal. It is the only domesticated species in the family Felidae.
{USER_NAME}: Name a color.
{AI_NAME}: Blue
{USER_NAME}:"""
	params = gpt_params_parse()

	with LLaMAInteract(params) as m:
		m.interact()