diff --git a/aiserver.py b/aiserver.py index b2b47bbf..b9b579db 100644 --- a/aiserver.py +++ b/aiserver.py @@ -821,8 +821,10 @@ def loadmodelsettings(): if("nobreakmodel" in js): koboldai_vars.nobreakmodel = js["nobreakmodel"] if("sampler_order" in js): - koboldai_vars.sampler_order = js["sampler_order"] - koboldai_vars.default_preset['sampler_order'] = js["sampler_order"] + sampler_order = koboldai_vars.sampler_order + if(len(sampler_order) < 7): + sampler_order = [6] + sampler_order + koboldai_vars.sampler_order = sampler_order if("temp" in js): koboldai_vars.temp = js["temp"] koboldai_vars.default_preset['temp'] = js["temp"] @@ -965,7 +967,10 @@ def processsettings(js): if("andepth" in js): koboldai_vars.andepth = js["andepth"] if("sampler_order" in js): - koboldai_vars.sampler_order = js["sampler_order"] + sampler_order = koboldai_vars.sampler_order + if(len(sampler_order) < 7): + sampler_order = [6] + sampler_order + koboldai_vars.sampler_order = sampler_order if("temp" in js): koboldai_vars.temp = js["temp"] if("top_p" in js): @@ -1363,23 +1368,25 @@ def get_model_info(model, directory=""): def get_layer_count(model, directory=""): - if(model not in ["InferKit", "Colab", "OAI", "GooseAI" , "ReadOnly", "TPUMeshTransformerGPTJ"]): - if(koboldai_vars.model == "GPT2Custom"): - model_config = open(directory + "/config.json", "r") + if(model not in ["InferKit", "Colab", "API", "OAI", "GooseAI" , "ReadOnly", "TPUMeshTransformerGPTJ"]): + if(model == "GPT2Custom"): + with open(os.path.join(directory, "config.json"), "r") as f: + model_config = json.load(f) # Get the model_type from the config or assume a model type if it isn't present else: + if(directory): + model = directory from transformers import AutoConfig - if directory == "": - model_config = AutoConfig.from_pretrained(model, revision=koboldai_vars.revision, cache_dir="cache") + if(os.path.isdir(model.replace('/', '_'))): + model_config = AutoConfig.from_pretrained(model.replace('/', '_'), revision=koboldai_vars.revision, cache_dir="cache") + elif(os.path.isdir("models/{}".format(model.replace('/', '_')))): + model_config = AutoConfig.from_pretrained("models/{}".format(model.replace('/', '_')), revision=koboldai_vars.revision, cache_dir="cache") elif(os.path.isdir(directory)): model_config = AutoConfig.from_pretrained(directory, revision=koboldai_vars.revision, cache_dir="cache") elif(os.path.isdir(koboldai_vars.custmodpth.replace('/', '_'))): model_config = AutoConfig.from_pretrained(koboldai_vars.custmodpth.replace('/', '_'), revision=koboldai_vars.revision, cache_dir="cache") else: - model_config = AutoConfig.from_pretrained(koboldai_vars.custmodpth, revision=koboldai_vars.revision, cache_dir="cache") - - - + model_config = AutoConfig.from_pretrained(model, revision=koboldai_vars.revision, cache_dir="cache") return utils.num_layers(model_config) else: return None @@ -1623,8 +1630,6 @@ def patch_transformers(): dynamic_processor_wrap(TailFreeLogitsWarper, "tfs", "tfs", cond=lambda x: x < 1.0) dynamic_processor_wrap(TypicalLogitsWarper, "typical", "typical", cond=lambda x: x < 1.0) dynamic_processor_wrap(TemperatureLogitsWarper, "temperature", "temp", cond=lambda x: x != 1.0) - RepetitionPenaltyLogitsProcessor.__init__ = AdvancedRepetitionPenaltyLogitsProcessor.__init__ - RepetitionPenaltyLogitsProcessor.__call__ = AdvancedRepetitionPenaltyLogitsProcessor.__call__ class PhraseBiasLogitsProcessor(LogitsProcessor): def __init__(self): @@ -1768,9 +1773,13 @@ def patch_transformers(): self.__warper_list.append(TailFreeLogitsWarper(tfs=0.5, min_tokens_to_keep=1 + (beams > 1))) self.__warper_list.append(TypicalLogitsWarper(typical=0.5, min_tokens_to_keep=1 + (beams > 1))) self.__warper_list.append(TemperatureLogitsWarper(temperature=0.5)) + self.__warper_list.append(AdvancedRepetitionPenaltyLogitsProcessor()) def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, *args, **kwargs): - for k in koboldai_vars.sampler_order: + sampler_order = koboldai_vars.sampler_order[:] + if len(sampler_order) < 7: # Add repetition penalty at beginning if it's not present + sampler_order = [6] + sampler_order + for k in sampler_order: scores = self.__warper_list[k](input_ids, scores, *args, **kwargs) return scores @@ -2525,6 +2534,9 @@ def load_model(use_gpu=True, gpu_layers=None, disk_layers=None, initial_load=Fal koboldai_vars.compiling = False def tpumtjgenerate_settings_callback() -> dict: + sampler_order = vars.sampler_order[:] + if len(sampler_order) < 7: # Add repetition penalty at beginning if it's not present + sampler_order = [6] + sampler_order return { "sampler_order": koboldai_vars.sampler_order, "top_p": float(koboldai_vars.top_p), @@ -3685,12 +3697,16 @@ def get_message(msg): sendUSStatItems() elif(msg['cmd'] == 'samplers'): sampler_order = msg["data"] + sampler_order_min_length = 6 + sampler_order_max_length = 7 if(not isinstance(sampler_order, list)): raise ValueError(f"Sampler order must be a list, but got a {type(sampler_order)}") - if(len(sampler_order) != len(koboldai_vars.sampler_order)): - raise ValueError(f"Sampler order must be a list of length {len(koboldai_vars.sampler_order)}, but got a list of length {len(sampler_order)}") + if(not (sampler_order_min_length <= len(sampler_order) <= sampler_order_max_length)): + raise ValueError(f"Sampler order must be a list of length greater than or equal to {sampler_order_min_length} and less than or equal to {sampler_order_max_length}, but got a list of length {len(sampler_order)}") if(not all(isinstance(e, int) for e in sampler_order)): raise ValueError(f"Sampler order must be a list of ints, but got a list with at least one non-int element") + if(min(sampler_order) != 0 or max(sampler_order) != len(sampler_order) - 1 or len(set(sampler_order)) != len(sampler_order)): + raise ValueError(f"Sampler order list of length {len(sampler_order)} must be a permutation of the first {len(sampler_order)} nonnegative integers") koboldai_vars.sampler_order = sampler_order settingschanged() elif(msg['cmd'] == 'list_model'): @@ -3701,8 +3717,8 @@ def get_message(msg): changed = True if not utils.HAS_ACCELERATE: msg['disk_layers'] = "0" - if os.path.exists("settings/" + koboldai_vars.model.replace('/', '_') + ".breakmodel"): - with open("settings/" + koboldai_vars.model.replace('/', '_') + ".breakmodel", "r") as file: + if os.path.exists("settings/" + koboldai_vars.model_selected.replace('/', '_') + ".breakmodel"): + with open("settings/" + koboldai_vars.model_selected.replace('/', '_') + ".breakmodel", "r") as file: data = file.read().split('\n')[:2] if len(data) < 2: data.append("0") @@ -3710,14 +3726,15 @@ def get_message(msg): if gpu_layers == msg['gpu_layers'] and disk_layers == msg['disk_layers']: changed = False if changed: - if koboldai_vars.model in ["NeoCustom", "GPT2Custom"]: + if koboldai_vars.model_selected in ["NeoCustom", "GPT2Custom"]: filename = "settings/{}.breakmodel".format(os.path.basename(os.path.normpath(koboldai_vars.custmodpth))) else: - filename = "settings/{}.breakmodel".format(koboldai_vars.model.replace('/', '_')) + filename = "settings/{}.breakmodel".format(koboldai_vars.model_selected.replace('/', '_')) f = open(filename, "w") f.write(str(msg['gpu_layers']) + '\n' + str(msg['disk_layers'])) f.close() koboldai_vars.colaburl = msg['url'] + "/request" + vars.model = vars.model_selected load_model(use_gpu=msg['use_gpu'], gpu_layers=msg['gpu_layers'], disk_layers=msg['disk_layers'], online_model=msg['online_model']) elif(msg['cmd'] == 'show_model'): print("Model Name: {}".format(getmodelname())) @@ -3742,18 +3759,18 @@ def get_message(msg): elif msg['data'] in ('NeoCustom', 'GPT2Custom') and 'path_modelname' in msg: #Here the user entered custom text in the text box. This could be either a model name or a path. if check_if_dir_is_model(msg['path_modelname']): - koboldai_vars.model = msg['data'] + koboldai_vars.model_selected = msg['data'] koboldai_vars.custmodpth = msg['path_modelname'] get_model_info(msg['data'], directory=msg['path']) else: - koboldai_vars.model = msg['path_modelname'] + koboldai_vars.model_selected = msg['path_modelname'] try: - get_model_info(koboldai_vars.model) + get_model_info(koboldai_vars.model_selected) except: emit('from_server', {'cmd': 'errmsg', 'data': "The model entered doesn't exist."}, room="UI_1") elif msg['data'] in ('NeoCustom', 'GPT2Custom'): if check_if_dir_is_model(msg['path']): - koboldai_vars.model = msg['data'] + koboldai_vars.model_selected = msg['data'] koboldai_vars.custmodpth = msg['path'] get_model_info(msg['data'], directory=msg['path']) else: @@ -3762,12 +3779,12 @@ def get_message(msg): else: sendModelSelection(menu=msg['data'], folder=msg['path']) else: - koboldai_vars.model = msg['data'] + koboldai_vars.model_selected = msg['data'] if 'path' in msg: koboldai_vars.custmodpth = msg['path'] get_model_info(msg['data'], directory=msg['path']) else: - get_model_info(koboldai_vars.model) + get_model_info(koboldai_vars.model_selected) elif(msg['cmd'] == 'delete_model'): if "{}/models".format(os.getcwd()) in os.path.abspath(msg['data']) or "{}\\models".format(os.getcwd()) in os.path.abspath(msg['data']): if check_if_dir_is_model(msg['data']): @@ -3873,7 +3890,6 @@ def get_message(msg): emit( 'from_server', {'cmd': 'showfieldbudget', 'data': {"length": None, "max": None, "field": field}}, - broadcast=True ) return @@ -4612,7 +4628,7 @@ def _generate(txt, minimum, maximum, found_entries): gen_in, do_sample=True, max_length=int(2e9), - repetition_penalty=1.1, + repetition_penalty=1.0, bad_words_ids=koboldai_vars.badwordsids, use_cache=True, num_return_sequences=numseqs diff --git a/colab/TPU.ipynb b/colab/TPU.ipynb index 448880e7..22c35e08 100644 --- a/colab/TPU.ipynb +++ b/colab/TPU.ipynb @@ -66,7 +66,7 @@ "#@title <-- Select your model below and then click this to start KoboldAI\n", "#@markdown You can find a description of the models below along with instructions on how to start KoboldAI.\n", "\n", - "Model = \"Nerys 13B V2\" #@param [\"Nerys 13B V2\", \"Janeway 13B\", \"Shinen 13B\", \"Skein 6B\", \"Janeway 6B\", \"Adventure 6B\", \"Shinen 6B\", \"Lit 6B\", \"NeoX 20B\", \"OPT 13B\", \"Fairseq Dense 13B\", \"GPT-J-6B\"] {allow-input: true}\n", + "Model = \"Nerys 13B V2\" #@param [\"Nerys 13B V2\", \"Janeway 13B\", \"Shinen 13B\", \"Skein 20B\", \"Skein 6B\", \"Janeway 6B\", \"Adventure 6B\", \"Shinen 6B\", \"Lit 6B\", \"NeoX 20B\", \"OPT 13B\", \"Fairseq Dense 13B\", \"GPT-J-6B\"] {allow-input: true}\n", "Version = \"Official\" #@param [\"Official\", \"United\"] {allow-input: true}\n", "Provider = \"Cloudflare\" #@param [\"Localtunnel\", \"Cloudflare\"]\n", "\n", @@ -93,6 +93,10 @@ " Model = \"KoboldAI/fairseq-dense-13B-Shinen\"\n", " path = \"\"\n", " download = \"\"\n", + "elif Model == \"Skein 20B\":\n", + " Model = \"KoboldAI/GPT-NeoX-20B-Skein\"\n", + " path = \"\"\n", + " download = \"\"\n", "elif Model == \"NeoX 20B\":\n", " Model = \"EleutherAI/gpt-neox-20b\"\n", " path = \"\"\n", @@ -128,7 +132,7 @@ "elif Model == \"GPT-J-6B\":\n", " Model = \"EleutherAI/gpt-j-6B\"\n", " path = \"\"\n", - " download = \"\"\n", + " download = \"\"\n", "else:\n", " path = \"\"\n", " download = \"\"\n", @@ -225,4 +229,4 @@ }, "nbformat": 4, "nbformat_minor": 0 -} +} \ No newline at end of file diff --git a/gensettings.py b/gensettings.py index e88b5edf..0f3f239b 100644 --- a/gensettings.py +++ b/gensettings.py @@ -495,6 +495,17 @@ gensettingstf = [ "classname": "story", "name": "andepth" }, + { + "uitype": "toggle", + "unit": "bool", + "label": "Show Field Budget", + "id": "setshowbudget", + "min": 0, + "max": 1, + "step": 1, + "default": 0, + "tooltip": "Shows token usage when typing in relevant text boxes. May lag slower devices." + }, ] gensettingsik =[{ diff --git a/static/application.js b/static/application.js index 6a956408..bdf7047b 100644 --- a/static/application.js +++ b/static/application.js @@ -241,8 +241,27 @@ function addSetting(ob) { if(ob.id == "setadventure"){ setadventure($(this).prop('checked')); } + }); } + + if (ob.id === "setshowbudget") { + $("#setshowbudget").on("change", function () { + for (const el of document.getElementsByClassName("input-token-usage")) { + if (this.checked) { + el.classList.remove("hidden"); + } else { + el.classList.add("hidden"); + } + } + }); + + if (!$("#setshowbudget")[0].checked) { + for (const el of document.getElementsByClassName("input-token-usage")) { + el.classList.add("hidden"); + } + } + } } function refreshTitle() { @@ -1287,12 +1306,13 @@ function buildSamplerList(samplers) { "Tail-free Sampling", "Typical Sampling", "Temperature", + "Repetition Penalty", ] for(i=0; i\
\
\ -
"+samplers_lookup_table[samplers[i]]+"
\ +
"+(samplers[i] < samplers_lookup_table.length ? samplers_lookup_table[samplers[i]] : "Unknown sampler #" + samplers[i])+"
\
\
\ "); @@ -2165,6 +2185,9 @@ function interpolateRGB(color0, color1, t) { } function updateInputBudget(inputElement) { + let budgetElement = document.getElementById("setshowbudget"); + if (budgetElement && !budgetElement.checked) return; + let data = {"unencoded": inputElement.value, "field": inputElement.id}; if (inputElement.id === "anoteinput") { @@ -2182,7 +2205,6 @@ function registerTokenCounters() { let span = document.createElement("span"); span.classList.add("input-token-usage"); - span.innerText = "?/? Tokens"; el.appendChild(span); let inputElement = el.querySelector("input, textarea"); @@ -2446,10 +2468,6 @@ $(document).ready(function(){ } else if(msg.cmd == "updatechunk") { hideMessage(); game_text.attr('contenteditable', allowedit); - if (typeof submit_start !== 'undefined') { - $("#runtime")[0].innerHTML = `Generation time: ${Math.round((Date.now() - submit_start)/1000)} sec`; - delete submit_start; - } var index = msg.data.index; var html = msg.data.html; var existingChunk = game_text.children('#n' + index); @@ -2963,6 +2981,7 @@ $(document).ready(function(){ $("#showmodelnamecontainer").removeClass("hidden"); } else if(msg.cmd == 'hide_model_name') { $("#showmodelnamecontainer").addClass("hidden"); + $(window).off('beforeunload'); location.reload(); //console.log("Closing window"); } else if(msg.cmd == 'model_load_status') { diff --git a/static/custom.css b/static/custom.css index af238dc7..d4bfe872 100644 --- a/static/custom.css +++ b/static/custom.css @@ -473,7 +473,7 @@ body.connected #popupfooter, #popupfooter.always-available { } #samplerslist { - height: 300px; + height: 310px; overflow-y: scroll; overflow-wrap: anywhere; } diff --git a/static/favicon.js b/static/favicon.js index 180059ff..fb40ac84 100644 --- a/static/favicon.js +++ b/static/favicon.js @@ -2,6 +2,7 @@ var fav_icon2 = "data:image/x-icon;base64,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"; var fav_icon1 = "data:image/png;base64,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"; var fav_icon = "data:image/png;base64,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" +var submit_start; var favicon = { @@ -53,11 +54,16 @@ var favicon = { start_swap: function() { this.run = true; this.auto_swap(); + submit_start = Date.now(); }, stop_swap: function() { this.run = false; this.change(fav_icon); + if (typeof submit_start !== 'undefined') { + $("#runtime")[0].innerHTML = `Execution time: ${Math.round((Date.now() - submit_start)/1000)} sec`; + delete submit_start; + } }, docHead:document.getElementsByTagName("head")[0] diff --git a/tpu_mtj_backend.py b/tpu_mtj_backend.py index 55fcddcb..2cc008cb 100644 --- a/tpu_mtj_backend.py +++ b/tpu_mtj_backend.py @@ -176,7 +176,7 @@ def apply_repetition_penalty_dynamic(logits, tokens, repetition_penalty, generat logits[tokens] = penalty_logits return logits -def kobold_sample_dynamic(key, logits, sampler_order: Optional[np.ndarray] = None, top_p=0.9, temp=0.5, top_k=0, tfs=1.0, typical=1.0, top_a=0.0): +def kobold_sample_dynamic(key, logits, rpargs, sampler_order: Optional[np.ndarray] = None, top_p=0.9, temp=0.5, top_k=0, tfs=1.0, typical=1.0, top_a=0.0): ''' This gets called by generate_loop_fn to apply a series of 6 filters to the logits (top-k, then top-a, then top-p, then TFS, then typical, then temperature) @@ -312,6 +312,7 @@ def kobold_sample_dynamic(key, logits, sampler_order: Optional[np.ndarray] = Non if k == 3 and tfs < 1.0: logits = tail_free_filter(logits) if k == 4 and typical < 1.0: logits = typical_filter(logits) if k == 5 and temp != 1.0: logits = temp_filter(logits) + if k == 6 and rpargs[1] != 1.0: logits = apply_repetition_penalty_dynamic(logits, *rpargs) # Finally, pick one token using the softmax thingy again (it gives # an array whose elements sum to 1 so it can be used nicely as a # probability distribution) @@ -362,7 +363,7 @@ def apply_repetition_penalty_static(logits, tokens, repetition_penalty, generate # positions in the logits array return logits.at[tokens].set(penalty_logits) -def kobold_sample_static(key, logits, sampler_order: Optional[np.ndarray] = None, top_p=0.9, temp=0.5, top_k=0, tfs=1.0, typical=1.0, top_a=0.0): +def kobold_sample_static(key, logits, rpargs, sampler_order: Optional[np.ndarray] = None, top_p=0.9, temp=0.5, top_k=0, tfs=1.0, typical=1.0, top_a=0.0): ''' This gets called by generate_loop_fn to apply a series of 6 filters to the logits (top-k, then top-a, then top-p, then TFS, then typical, then temperature) @@ -497,6 +498,7 @@ def kobold_sample_static(key, logits, sampler_order: Optional[np.ndarray] = None logits = jax.lax.cond(jnp.logical_and(k == 3, tfs < 1.0), tail_free_filter, lambda x: x, logits) logits = jax.lax.cond(jnp.logical_and(k == 4, typical < 1.0), typical_filter, lambda x: x, logits) logits = jax.lax.cond(jnp.logical_and(k == 5, temp != 1.0), temp_filter, lambda x: x, logits) + logits = jax.lax.cond(jnp.logical_and(k == 6, rpargs[1] != 1.0), lambda x: apply_repetition_penalty_static(*x), lambda x: x[0], (logits, *rpargs)) # Finally, pick one token using the softmax thingy again (it gives # an array whose elements sum to 1 so it can be used nicely as a # probability distribution) @@ -513,17 +515,6 @@ def sample_func(data, key, numseqs_aux, badwords, repetition_penalty, generated_ # Get the pseudo-random number generator key that will # be used by kobold_sample_dynamic to randomly pick a token sample_key, new_key = jax.random.split(sample_key, num=2) - # Apply repetition penalty to all tokens that are - # currently inside the "generated" array - logits = apply_repetition_penalty_dynamic( - logits, - generated, - repetition_penalty, - generated_index, - gen_length, - rpslope, - rprange, - ) # Remove any tokens in the badwords list by setting # their logits to negative infinity which effectively # makes their probabilities of being chosen zero @@ -535,6 +526,14 @@ def sample_func(data, key, numseqs_aux, badwords, repetition_penalty, generated_ next_token = kobold_sample_dynamic( sample_key, logits, + ( + generated, + repetition_penalty, + generated_index, + gen_length, + rpslope, + rprange, + ), **sampler_options, ) # Remember what token was picked @@ -606,18 +605,6 @@ class PenalizingCausalTransformer(CausalTransformer): assert logits.shape == (1, config["n_vocab"]) # Flatten it into a 1D array to make it easier to use logits = logits[0] - # Apply repetition penalty to all tokens that are - # currently inside the "generated" array - if repetition_penalty is not None: - logits = apply_repetition_penalty_static( - logits, - generated, - repetition_penalty, - generated_index, - gen_length, - rpslope, - rprange, - ) # Remove any tokens in the badwords list by setting # their logits to negative infinity which effectively # makes their probabilities of being chosen zero @@ -629,6 +616,14 @@ class PenalizingCausalTransformer(CausalTransformer): next_token = kobold_sample_static( sample_key, logits, + ( + generated, + repetition_penalty, + generated_index, + gen_length, + rpslope, + rprange, + ), **sampler_options, ) # Remember what token was picked @@ -863,6 +858,9 @@ def infer_static( maps.thread_resources.env = thread_resources_env if sampler_order is None: sampler_order = utils.default_sampler_order.copy() + sampler_order = sampler_order[:] + if len(sampler_order) < 7: # Add repetition penalty at beginning if it's not present + sampler_order = [6] + sampler_order sampler_order = np.uint32(sampler_order) total_batch = 1 tokens = context diff --git a/utils.py b/utils.py index 375e9dfb..98b8006e 100644 --- a/utils.py +++ b/utils.py @@ -33,7 +33,7 @@ layers_module_names: Optional[List[str]] = None module_names: Optional[List[str]] = None named_buffers: Optional[List[tuple]] = None -default_sampler_order = [0, 1, 2, 3, 4, 5] +default_sampler_order = [6, 0, 1, 2, 3, 4, 5] #==================================================================# # Decorator to prevent a function's actions from being run until @@ -167,7 +167,7 @@ def decodenewlines(txt): # Returns number of layers given an HF model config #==================================================================# def num_layers(config): - return config.num_layers if hasattr(config, "num_layers") else config.n_layer if hasattr(config, "n_layer") else config.num_hidden_layers if hasattr(config, 'num_hidden_layers') else None + return config["n_layer"] if isinstance(config, dict) else config.num_layers if hasattr(config, "num_layers") else config.n_layer if hasattr(config, "n_layer") else config.num_hidden_layers if hasattr(config, 'num_hidden_layers') else None #==================================================================# # Downloads huggingface checkpoints using aria2c if possible @@ -177,6 +177,7 @@ class Send_to_socketio(object): def write(self, bar): time.sleep(0.01) try: + print(bar) emit('from_server', {'cmd': 'model_load_status', 'data': bar.replace(" ", " ")}, broadcast=True) except: pass diff --git a/warpers.py b/warpers.py index fb683f50..488a901e 100644 --- a/warpers.py +++ b/warpers.py @@ -28,10 +28,10 @@ SOFTWARE. ''' import torch -from transformers import LogitsWarper, LogitsProcessor +from transformers import LogitsWarper -class AdvancedRepetitionPenaltyLogitsProcessor(LogitsProcessor): +class AdvancedRepetitionPenaltyLogitsProcessor(LogitsWarper): def __init__(self, *args, **kwargs): pass