diff --git a/aiserver.py b/aiserver.py index 977db3d2..80db8893 100644 --- a/aiserver.py +++ b/aiserver.py @@ -89,7 +89,6 @@ class vars: submission = "" # Same as above, but after applying input formatting lastctx = "" # The last context submitted to the generator model = "" # Model ID string chosen at startup - model_orig = "" # Original model string before being changed by auto model type detection model_type = "" # Model Type (Automatically taken from the model config) noai = False # Runs the script without starting up the transformers pipeline aibusy = False # Stops submissions while the AI is working @@ -124,6 +123,7 @@ class vars: lua_running = False # Whether or not Lua is running (i.e. wasn't stopped due to an error) lua_edited = set() # Set of chunk numbers that were edited from a Lua generation modifier lua_deleted = set() # Set of chunk numbers that were deleted from a Lua generation modifier + generated_tkns = 0 # If using a backend that supports Lua generation modifiers, how many tokens have already been generated, otherwise 0 spfilename = "" # Filename of soft prompt to load, or an empty string if not using a soft prompt userscripts = [] # List of userscripts to load last_userscripts = [] # List of previous userscript filenames from the previous time userscripts were send via usstatitems @@ -158,6 +158,7 @@ class vars: saveow = False # Whether or not overwrite confirm has been displayed genseqs = [] # Temporary storage for generated sequences recentback = False # Whether Back button was recently used without Submitting or Retrying after + recentrng = None # If a new random game was recently generated without Submitting after, this is the topic used (as a string), otherwise this is None useprompt = False # Whether to send the full prompt with every submit action breakmodel = False # For GPU users, whether to use both system RAM and VRAM to conserve VRAM while offering speedup compared to CPU-only bmsupported = False # Whether the breakmodel option is supported (GPT-Neo/GPT-J only, currently) @@ -194,7 +195,7 @@ def getModelSelection(): while(vars.model == ''): modelsel = input("Model #> ") if(modelsel.isnumeric() and int(modelsel) > 0 and int(modelsel) <= len(modellist)): - vars.model = vars.model_orig = modellist[int(modelsel)-1][1] + vars.model = modellist[int(modelsel)-1][1] else: print("{0}Please enter a valid selection.{1}".format(colors.RED, colors.END)) @@ -375,7 +376,7 @@ parser.add_argument("--override_rename", action='store_true', help="Renaming sto parser.add_argument("--configname", help="Force a fixed configuration name to aid with config management.") args = parser.parse_args() -vars.model = vars.model_orig = args.model; +vars.model = args.model; if args.remote: vars.remote = True; @@ -801,7 +802,10 @@ if(not vars.model in ["InferKit", "Colab", "OAI", "ReadOnly", "TPUMeshTransforme scores: torch.FloatTensor, **kwargs, ) -> bool: - if(vars.lua_koboldbridge.generated_cols >= vars.genamt): + vars.generated_tkns += 1 + if(vars.lua_koboldbridge.generated_cols and vars.generated_tkns != vars.lua_koboldbridge.generated_cols): + raise RuntimeError(f"Inconsistency detected between KoboldAI Python and Lua backends ({vars.generated_tkns} != {vars.lua_koboldbridge.generated_cols})") + if(vars.generated_tkns >= vars.genamt): self.regeneration_required = False self.halt = False return True @@ -813,7 +817,7 @@ if(not vars.model in ["InferKit", "Colab", "OAI", "ReadOnly", "TPUMeshTransforme vars.lua_koboldbridge.regeneration_required = False for i in range(vars.numseqs): - vars.lua_koboldbridge.generated[i+1][vars.lua_koboldbridge.generated_cols] = input_ids[i, -1].item() + vars.lua_koboldbridge.generated[i+1][vars.generated_tkns] = int(input_ids[i, -1].item()) if(not vars.dynamicscan): return self.regeneration_required or self.halt @@ -1150,7 +1154,7 @@ def lua_compute_context(submission, entries, folders): i += 1 winfo, mem, anotetxt, _ = calcsubmitbudgetheader(submission, allowed_entries=allowed_entries, allowed_folders=allowed_folders, force_use_txt=True) txt, _, _ = calcsubmitbudget(len(actions), winfo, mem, anotetxt, actions) - return txt + return tokenizer.decode(txt) #==================================================================# # Get property of a world info entry given its UID and property name @@ -1452,6 +1456,8 @@ def lua_is_custommodel(): #==================================================================# def execute_inmod(): vars.lua_logname = ... + vars.lua_edited = set() + vars.lua_deleted = set() try: tpool.execute(vars.lua_koboldbridge.execute_inmod) except lupa.LuaError as e: @@ -1465,8 +1471,6 @@ def execute_inmod(): set_aibusy(0) def execute_genmod(): - vars.lua_edited = set() - vars.lua_deleted = set() vars.lua_koboldbridge.execute_genmod() def execute_outmod(): @@ -1606,6 +1610,13 @@ def get_message(msg): if(msg['cmd'] == 'submit'): if(vars.mode == "play"): vars.lua_koboldbridge.feedback = None + if(vars.chatmode): + if(type(msg['chatname']) is not str): + raise ValueError("Chatname must be a string") + vars.chatname = msg['chatname'] + settingschanged() + emit('from_server', {'cmd': 'setchatname', 'data': vars.chatname}, broadcast=True) + vars.recentrng = None actionsubmit(msg['data'], actionmode=msg['actionmode']) elif(vars.mode == "edit"): editsubmit(msg['data']) @@ -1613,6 +1624,12 @@ def get_message(msg): memsubmit(msg['data']) # Retry Action elif(msg['cmd'] == 'retry'): + if(vars.chatmode): + if(type(msg['chatname']) is not str): + raise ValueError("Chatname must be a string") + vars.chatname = msg['chatname'] + settingschanged() + emit('from_server', {'cmd': 'setchatname', 'data': vars.chatname}, broadcast=True) actionretry(msg['data']) # Back/Undo Action elif(msg['cmd'] == 'back'): @@ -2056,7 +2073,7 @@ def settingschanged(): #==================================================================# # Take input text from SocketIO and decide what to do with it #==================================================================# -def actionsubmit(data, actionmode=0, force_submit=False): +def actionsubmit(data, actionmode=0, force_submit=False, force_prompt_gen=False, disable_recentrng=False): # Ignore new submissions if the AI is currently busy if(vars.aibusy): return @@ -2064,6 +2081,9 @@ def actionsubmit(data, actionmode=0, force_submit=False): while(True): set_aibusy(1) + if(disable_recentrng): + vars.recentrng = None + vars.recentback = False vars.recentedit = False vars.actionmode = actionmode @@ -2093,7 +2113,7 @@ def actionsubmit(data, actionmode=0, force_submit=False): assert False # Start the game vars.gamestarted = True - if(not vars.noai and vars.lua_koboldbridge.generating and not vars.nopromptgen): + if(not vars.noai and vars.lua_koboldbridge.generating and (not vars.nopromptgen or force_prompt_gen)): # Save this first action as the prompt vars.prompt = data # Clear the startup text from game screen @@ -2102,6 +2122,7 @@ def actionsubmit(data, actionmode=0, force_submit=False): if(vars.lua_koboldbridge.restart_sequence is not None and len(vars.genseqs) == 0): data = "" force_submit = True + disable_recentrng = True continue emit('from_server', {'cmd': 'scrolldown', 'data': ''}, broadcast=True) break @@ -2122,6 +2143,7 @@ def actionsubmit(data, actionmode=0, force_submit=False): refresh_story() data = "" force_submit = True + disable_recentrng = True continue else: if(vars.lua_koboldbridge.restart_sequence is not None and vars.lua_koboldbridge.restart_sequence > 0): @@ -2129,6 +2151,7 @@ def actionsubmit(data, actionmode=0, force_submit=False): refresh_story() data = "" force_submit = True + disable_recentrng = True continue genselect(genout) refresh_story() @@ -2157,6 +2180,7 @@ def actionsubmit(data, actionmode=0, force_submit=False): if(vars.lua_koboldbridge.restart_sequence is not None and len(vars.genseqs) == 0): data = "" force_submit = True + disable_recentrng = True continue emit('from_server', {'cmd': 'scrolldown', 'data': ''}, broadcast=True) break @@ -2174,12 +2198,14 @@ def actionsubmit(data, actionmode=0, force_submit=False): if(vars.lua_koboldbridge.restart_sequence is not None): data = "" force_submit = True + disable_recentrng = True continue else: if(vars.lua_koboldbridge.restart_sequence is not None and vars.lua_koboldbridge.restart_sequence > 0): genresult(genout[vars.lua_koboldbridge.restart_sequence-1]["generated_text"]) data = "" force_submit = True + disable_recentrng = True continue genselect(genout) set_aibusy(0) @@ -2195,6 +2221,9 @@ def actionretry(data): return if(vars.aibusy): return + if(vars.recentrng is not None): + randomGameRequest(vars.recentrng) + return # Remove last action if possible and resubmit if(vars.gamestarted if vars.useprompt else len(vars.actions) > 0): if(not vars.recentback and len(vars.actions) != 0 and len(vars.genseqs) == 0): # Don't pop if we're in the "Select sequence to keep" menu or if there are no non-prompt actions @@ -2246,38 +2275,53 @@ def calcsubmitbudgetheader(txt, **kwargs): return winfo, mem, anotetxt, found_entries -def calcsubmitbudget(actionlen, winfo, mem, anotetxt, actions): +def calcsubmitbudget(actionlen, winfo, mem, anotetxt, actions, submission=None, budget_deduction=0): forceanote = False # In case we don't have enough actions to hit A.N. depth anoteadded = False # In case our budget runs out before we hit A.N. depth anotetkns = [] # Placeholder for Author's Note tokens lnanote = 0 # Placeholder for Author's Note length - # Calculate token budget - prompttkns = tokenizer.encode(vars.comregex_ai.sub('', vars.prompt), max_length=1+int(vars.max_length), truncation=True) - lnprompt = len(prompttkns) - - memtokens = tokenizer.encode(mem, max_length=1+int(vars.max_length), truncation=True) - lnmem = len(memtokens) - - witokens = tokenizer.encode(winfo, max_length=1+int(vars.max_length), truncation=True) - lnwi = len(witokens) - - if(anotetxt != ""): - anotetkns = tokenizer.encode(anotetxt, max_length=1+int(vars.max_length), truncation=True) - lnanote = len(anotetkns) - lnsp = vars.sp.shape[0] if vars.sp is not None else 0 - + + # Calculate token budget + prompttkns = tokenizer.encode(vars.comregex_ai.sub('', vars.prompt), max_length=int(2e9), truncation=True) + lnprompt = len(prompttkns) + + memtokens = tokenizer.encode(mem, max_length=int(2e9), truncation=True) + lnmem = len(memtokens) + if(lnmem > vars.max_length - lnsp - vars.genamt - budget_deduction): + raise OverflowError("The memory in your story is too long. Please either write a shorter memory text or increase the Max Tokens setting. If you are using a soft prompt, additionally consider using a smaller soft prompt.") + + witokens = tokenizer.encode(winfo, max_length=int(2e9), truncation=True) + lnwi = len(witokens) + if(lnmem + lnwi > vars.max_length - lnsp - vars.genamt - budget_deduction): + raise OverflowError("The current active world info keys take up too many tokens. Please either write shorter world info, decrease World Info Depth or increase the Max Tokens setting. If you are using a soft prompt, additionally consider using a smaller soft prompt.") + + if(anotetxt != ""): + anotetkns = tokenizer.encode(anotetxt, max_length=int(2e9), truncation=True) + lnanote = len(anotetkns) + if(lnmem + lnwi + lnanote > vars.max_length - lnsp - vars.genamt - budget_deduction): + raise OverflowError("The author's note in your story is too long. Please either write a shorter author's note or increase the Max Tokens setting. If you are using a soft prompt, additionally consider using a smaller soft prompt.") + if(vars.useprompt): - budget = vars.max_length - lnsp - lnprompt - lnmem - lnanote - lnwi - vars.genamt + budget = vars.max_length - lnsp - lnprompt - lnmem - lnanote - lnwi - vars.genamt - budget_deduction else: - budget = vars.max_length - lnsp - lnmem - lnanote - lnwi - vars.genamt + budget = vars.max_length - lnsp - lnmem - lnanote - lnwi - vars.genamt - budget_deduction + + lnsubmission = len(tokenizer.encode(vars.comregex_ai.sub('', submission), max_length=int(2e9), truncation=True)) if submission is not None else 0 + maybe_lnprompt = lnprompt if vars.useprompt and actionlen > 0 else 0 + + if(lnmem + lnwi + lnanote + maybe_lnprompt + lnsubmission > vars.max_length - lnsp - vars.genamt - budget_deduction): + raise OverflowError("Your submission is too long. Please either write a shorter submission or increase the Max Tokens setting. If you are using a soft prompt, additionally consider using a smaller soft prompt. If you are using the Always Add Prompt setting, turning it off may help.") + + assert budget >= 0 if(actionlen == 0): # First/Prompt action - subtxt = vars.memory + winfo + anotetxt + vars.comregex_ai.sub('', vars.prompt) - lnsub = lnsp + lnmem + lnwi + lnprompt + lnanote - return subtxt, lnsub+1, lnsub+vars.genamt + tokens = memtokens + witokens + anotetkns + prompttkns + assert len(tokens) <= vars.max_length - lnsp - vars.genamt - budget_deduction + ln = len(tokens) + lnsp + return tokens, ln+1, ln+vars.genamt else: tokens = [] @@ -2290,9 +2334,10 @@ def calcsubmitbudget(actionlen, winfo, mem, anotetxt, actions): for key in reversed(actions): chunk = vars.comregex_ai.sub('', actions[key]) + assert budget >= 0 if(budget <= 0): break - acttkns = tokenizer.encode(chunk, max_length=int(vars.max_length), truncation=True) + acttkns = tokenizer.encode(chunk, max_length=int(2e9), truncation=True) tknlen = len(acttkns) if(tknlen < budget): tokens = acttkns + tokens @@ -2317,7 +2362,7 @@ def calcsubmitbudget(actionlen, winfo, mem, anotetxt, actions): prompttkns = prompttkns[-budget:] else: prompttkns = [] - + # Did we get to add the A.N.? If not, do it here if(anotetxt != ""): if((not anoteadded) or forceanote): @@ -2327,10 +2372,11 @@ def calcsubmitbudget(actionlen, winfo, mem, anotetxt, actions): else: # Prepend Memory, WI, and Prompt before action tokens tokens = memtokens + witokens + prompttkns + tokens - + # Send completed bundle to generator + assert len(tokens) <= vars.max_length - lnsp - vars.genamt - budget_deduction ln = len(tokens) + lnsp - return tokenizer.decode(tokens), ln+1, ln+vars.genamt + return tokens, ln+1, ln+vars.genamt #==================================================================# # Take submitted text and build the text to be given to generator @@ -2345,23 +2391,23 @@ def calcsubmit(txt): # For all transformers models if(vars.model != "InferKit"): - subtxt, min, max = calcsubmitbudget(actionlen, winfo, mem, anotetxt, vars.actions) + subtxt, min, max = calcsubmitbudget(actionlen, winfo, mem, anotetxt, vars.actions, submission=txt) if(actionlen == 0): if(not vars.model in ["Colab", "OAI", "TPUMeshTransformerGPTJ"]): generate(subtxt, min, max, found_entries=found_entries) elif(vars.model == "Colab"): - sendtocolab(subtxt, min, max) + sendtocolab(tokenizer.decode(subtxt), min, max) elif(vars.model == "OAI"): - oairequest(subtxt, min, max) + oairequest(tokenizer.decode(subtxt), min, max) elif(vars.model == "TPUMeshTransformerGPTJ"): tpumtjgenerate(subtxt, min, max, found_entries=found_entries) else: if(not vars.model in ["Colab", "OAI", "TPUMeshTransformerGPTJ"]): generate(subtxt, min, max, found_entries=found_entries) elif(vars.model == "Colab"): - sendtocolab(subtxt, min, max) + sendtocolab(tokenizer.decode(subtxt), min, max) elif(vars.model == "OAI"): - oairequest(subtxt, min, max) + oairequest(tokenizer.decode(subtxt), min, max) elif(vars.model == "TPUMeshTransformerGPTJ"): tpumtjgenerate(subtxt, min, max, found_entries=found_entries) @@ -2426,13 +2472,14 @@ def calcsubmit(txt): #==================================================================# def _generate(txt, minimum, maximum, found_entries): - gen_in = tokenizer.encode(txt, return_tensors="pt", max_length=int(vars.max_length), truncation=True).long() + gen_in = torch.tensor(txt, dtype=torch.long)[None] if(vars.sp is not None): soft_tokens = torch.arange( model.config.vocab_size, model.config.vocab_size + vars.sp.shape[0], ) gen_in = torch.cat((soft_tokens[None], gen_in), dim=-1) + assert gen_in.shape[-1] + vars.genamt <= vars.max_length if(vars.hascuda and vars.usegpu): gen_in = gen_in.to(vars.gpu_device) @@ -2464,11 +2511,14 @@ def _generate(txt, minimum, maximum, found_entries): num_return_sequences=numseqs ) already_generated += len(genout[0]) - len(gen_in[0]) + assert already_generated <= vars.genamt if(model.kai_scanner.halt or not model.kai_scanner.regeneration_required): break assert genout.ndim >= 2 assert genout.shape[0] == vars.numseqs - if(already_generated != vars.lua_koboldbridge.generated_cols): + if(vars.lua_koboldbridge.generated_cols and vars.generated_tkns != vars.lua_koboldbridge.generated_cols): + raise RuntimeError("Inconsistency detected between KoboldAI Python and Lua backends") + if(already_generated != vars.generated_tkns): raise RuntimeError("WI scanning error") for r in range(vars.numseqs): for c in range(already_generated): @@ -2479,8 +2529,8 @@ def _generate(txt, minimum, maximum, found_entries): txt = tokenizer.decode(genout[i, -already_generated:]) winfo, mem, anotetxt, _found_entries = calcsubmitbudgetheader(txt, force_use_txt=True) found_entries[i].update(_found_entries) - txt, _, _ = calcsubmitbudget(len(vars._actions), winfo, mem, anotetxt, vars._actions) - encoded.append(tokenizer.encode(txt, return_tensors="pt", max_length=int(vars.max_length), truncation=True)[0].long().to(genout.device)) + txt, _, _ = calcsubmitbudget(len(vars._actions), winfo, mem, anotetxt, vars._actions, submission=txt) + encoded.append(torch.tensor(txt, dtype=torch.long, device=genout.device)) max_length = len(max(encoded, key=len)) encoded = torch.stack(tuple(torch.nn.functional.pad(e, (max_length - len(e), 0), value=model.config.pad_token_id or model.config.eos_token_id) for e in encoded)) genout = torch.cat( @@ -2497,6 +2547,7 @@ def _generate(txt, minimum, maximum, found_entries): device=genout.device, ) genout = torch.cat((soft_tokens.tile(vars.numseqs, 1), genout), dim=-1) + assert genout.shape[-1] + vars.genamt - already_generated <= vars.max_length diff = genout.shape[-1] - gen_in.shape[-1] minimum += diff maximum += diff @@ -2508,14 +2559,16 @@ def _generate(txt, minimum, maximum, found_entries): def generate(txt, minimum, maximum, found_entries=None): + vars.generated_tkns = 0 + if(found_entries is None): found_entries = set() found_entries = tuple(found_entries.copy() for _ in range(vars.numseqs)) - print("{0}Min:{1}, Max:{2}, Txt:{3}{4}".format(colors.YELLOW, minimum, maximum, txt, colors.END)) + print("{0}Min:{1}, Max:{2}, Txt:{3}{4}".format(colors.YELLOW, minimum, maximum, tokenizer.decode(txt), colors.END)) # Store context in memory to use it for comparison with generated content - vars.lastctx = txt + vars.lastctx = tokenizer.decode(txt) # Clear CUDA cache if using GPU if(vars.hascuda and (vars.usegpu or vars.breakmodel)): @@ -2541,7 +2594,7 @@ def generate(txt, minimum, maximum, found_entries=None): return for i in range(vars.numseqs): - vars.lua_koboldbridge.generated[i+1][vars.lua_koboldbridge.generated_cols] = genout[i, -1].item() + vars.lua_koboldbridge.generated[i+1][vars.generated_tkns] = int(genout[i, -1].item()) vars.lua_koboldbridge.outputs[i+1] = tokenizer.decode(genout[i, -already_generated:]) execute_outmod() @@ -2624,7 +2677,7 @@ def selectsequence(n): vars.genseqs = [] if(vars.lua_koboldbridge.restart_sequence is not None): - actionsubmit("", actionmode=vars.actionmode, force_submit=True) + actionsubmit("", actionmode=vars.actionmode, force_submit=True, disable_recentrng=True) #==================================================================# # Send transformers-style request to ngrok/colab host @@ -2712,7 +2765,7 @@ def tpumtjgenerate(txt, minimum, maximum, found_entries=None): found_entries = set() found_entries = tuple(found_entries.copy() for _ in range(vars.numseqs)) - print("{0}Min:{1}, Max:{2}, Txt:{3}{4}".format(colors.YELLOW, minimum, maximum, txt, colors.END)) + print("{0}Min:{1}, Max:{2}, Txt:{3}{4}".format(colors.YELLOW, minimum, maximum, tokenizer.decode(txt), colors.END)) # Submit input text to generator try: @@ -2742,7 +2795,7 @@ def tpumtjgenerate(txt, minimum, maximum, found_entries=None): genout = tpool.execute( tpu_mtj_backend.infer, - txt, + np.uint32(txt), gen_len = maximum-minimum+1, temp=vars.temp, top_p=vars.top_p, @@ -2809,8 +2862,8 @@ def getnewcontent(txt): return txt # Tokenize the last context and the generated content - ctxtokens = tokenizer.encode(vars.lastctx, max_length=1+int(vars.max_length), truncation=True) - txttokens = tokenizer.encode(txt, max_length=1+int(vars.max_length), truncation=True) + ctxtokens = tokenizer.encode(vars.lastctx, max_length=int(2e9), truncation=True) + txttokens = tokenizer.encode(txt, max_length=int(2e9), truncation=True) dif = (len(txttokens) - len(ctxtokens)) * -1 # Remove the context from the returned text @@ -4126,10 +4179,11 @@ def newGameRequest(): setStartState() def randomGameRequest(topic): + vars.recentrng = topic newGameRequest() vars.memory = "You generate the following " + topic + " story concept :" vars.lua_koboldbridge.feedback = None - actionsubmit("", force_submit=True) + actionsubmit("", force_submit=True, force_prompt_gen=True) vars.memory = "" #==================================================================# diff --git a/bridge.lua b/bridge.lua index 39acc222..f9ce25fb 100644 --- a/bridge.lua +++ b/bridge.lua @@ -888,6 +888,8 @@ return function(_python, _bridged) ---@field rmspch boolean ---@field adsnsp boolean ---@field singleline boolean + ---@field chatmode boolean + ---@field chatname string local KoboldSettings = setmetatable({ _name = "KoboldSettings", }, metawrapper) @@ -1038,7 +1040,7 @@ return function(_python, _bridged) ---@param t KoboldLib ---@return string function KoboldLib_getters.model(t) - return bridged.vars.model_orig + return bridged.vars.model end ---@param t KoboldLib @@ -1526,7 +1528,7 @@ return function(_python, _bridged) end local old_loadfile = loadfile - local old_package_loaded = package.loaded + local package_loaded = {} ---@type table