Official as Default (TPU)

Release of 1.18

.gitignore

@@ -9,6 +9,8 @@ stories
 /.project
 *.bak
 miniconda3
+runtime
+bin
 *.settings
 __pycache__
 *.log
@@ -21,9 +23,14 @@ userscripts
 softprompts
 models
 !models/models go here.txt
+Uninstall
+.ipynb_checkpoints
 
 # Ignore PyCharm project files.
 .idea
 
 # Ignore compiled Python files.
 *.pyc
+
+# Don't ignore defaults
+!defaults/*

Jupyter.bat

@@ -7,23 +7,23 @@ IF %M%==2 GOTO subfolder
 IF %M%==3 GOTO drivemap_B
 
 :subfolder
-umamba.exe install --no-shortcuts -r miniconda3 -n base -c conda-forge jupyter
+umamba.exe install --no-shortcuts -r miniconda3 -n base -c conda-forge jupyterlab jupyterlab-git
 call miniconda3\condabin\activate
-jupyter notebook
+jupyter-lab
 cmd /k
 
 :drivemap
 subst K: miniconda3 >nul
-umamba.exe install --no-shortcuts -r K:\python\ -n base -c conda-forge jupyter
+umamba.exe install --no-shortcuts -r K:\python\ -n base -c conda-forge jupyterlab jupyterlab-git
 call K:\python\condabin\activate
-jupyter notebook
+jupyter-lab
 subst K: /D
 cmd /k
 
 :drivemap_B
 subst B: miniconda3 >nul
-umamba.exe install --no-shortcuts -r B:\python\ -n base -c conda-forge jupyter
+umamba.exe install --no-shortcuts -r B:\python\ -n base -c conda-forge jupyterlab jupyterlab-git
 call B:\python\condabin\activate
-jupyter notebook
+jupyter-lab
 subst B: /D
 cmd /k

Uninstall.bat

@@ -0,0 +1,32 @@
+@echo off 
+cd /D %~dp0
+TITLE KoboldAI Uninstall Helper
+SET /P M=<loader.settings
+IF %M%==3 subst /D B: >nul
+IF %M%==1 subst /D K: >nul
+
+IF "%1" == "FORCE" GOTO UNINSTALL
+
+IF EXIST "Uninstall\unins000.exe" (
+   start Uninstall\unins000.exe
+   exit
+) ELSE (
+   echo This will remove all KoboldAI folders that do not contain user data
+	pause
+	GOTO UNINSTALL
+)
+
+:UNINSTALL
+echo Uninstallation in progress, please wait...
+set DM=Y
+attrib -h .git >nul
+for /d %%D in (*) do if not "%%~nxD"=="stories" if not "%%~nxD"=="userscripts" if not "%%~nxD"=="settings" if not "%%~nxD"=="softprompts" if not "%%~nxD"=="models" if not "%%~nxD"=="Uninstall" rmdir /S /Q %%~nxD
+for %%i in (*) do if not "%%i"=="Uninstall.bat" del /q "%%i"
+set /P DM=Would you like to delete the models folder? (Y/n) :
+IF %DM%==Y rmdir models /s /q
+IF %DM%==y rmdir models /s /q
+set DM=N
+set /P DM=Would you like to delete all other user folders? (y/N) :
+IF %DM%==Y rmdir stories userscripts settings softprompts /s /q
+IF %DM%==y rmdir stories userscripts settings softprompts /s /q
+del Uninstall.bat

breakmodel.py

@@ -212,14 +212,17 @@ Copyright 2018 The Hugging Face team
 
 
 import torch
+from torch import nn
 import torch.cuda.comm
 import copy
 import gc
 import sys
 import itertools
 import bisect
+import random
+from typing import Optional
 
-from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.modeling_outputs import BaseModelOutputWithPast, BaseModelOutputWithPastAndCrossAttentions
 
 from transformers.utils import logging
 logger = logging.get_logger(__name__)
@@ -230,22 +233,40 @@ gpu_blocks = []
 primary_device = 0
 
 
-def move_hidden_layers(transformer):
+# Copied from transformers.models.bart.modeling_bart._expand_mask
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill(inverted_mask.bool(), torch.finfo(dtype).min)
+
+
+def move_hidden_layers(transformer, h=None):
+    if h is None:
+        h = transformer.h
+
     assert len(gpu_blocks) <= torch.cuda.device_count()
-    assert sum(gpu_blocks) <= len(transformer.h)
-    ram_blocks = len(transformer.h) - sum(gpu_blocks)
+    assert sum(gpu_blocks) <= len(h)
+    ram_blocks = len(h) - sum(gpu_blocks)
 
     transformer.extrastorage = {}
     torch.cuda.empty_cache()
     
     able_to_pin_layers = True
     for i in range(ram_blocks):
-        transformer.h[i].to("cpu")
-        transformer.extrastorage[i] = copy.deepcopy(transformer.h[i])
+        h[i].to("cpu")
+        transformer.extrastorage[i] = copy.deepcopy(h[i])
         smalltensor = torch.tensor(0).to(primary_device)
-        for param1 in transformer.h[i].parameters():
+        for param1 in h[i].parameters():
             param1.data = smalltensor
-        transformer.h[i].to(primary_device)
+        h[i].to(primary_device)
         for param in transformer.extrastorage[i].parameters():
             param.requires_grad = False
             param.data = param.data.detach()
@@ -259,34 +280,34 @@ def move_hidden_layers(transformer):
             torch.cuda.empty_cache()
 
     if ram_blocks:
-        for param1,param2 in zip(transformer.h[0].parameters(),transformer.extrastorage[0].parameters()):
+        for param1,param2 in zip(h[0].parameters(),transformer.extrastorage[0].parameters()):
             param1.data = param2.data.to(primary_device, non_blocking=False).detach()
 
-        for param1,param2 in zip(transformer.h[ram_blocks-1].parameters(),transformer.extrastorage[ram_blocks-1].parameters()):
+        for param1,param2 in zip(h[ram_blocks-1].parameters(),transformer.extrastorage[ram_blocks-1].parameters()):
             param1.data = param2.data.to(primary_device, non_blocking=False).detach()
 
     i = ram_blocks
     for j in range(len(gpu_blocks)):
         for _ in range(gpu_blocks[j]):
-            transformer.h[i].to(j)
+            h[i].to(j)
             i += 1
 
 
-def new_forward(
-        self,
-        input_ids=None,
-        past_key_values=None,
-        attention_mask=None,
-        token_type_ids=None,
-        position_ids=None,
-        head_mask=None,
-        inputs_embeds=None,
-        use_cache=None,
-        output_attentions=None,
-        output_hidden_states=None,
-        return_dict=None,
-        embs=None,
-    ):
+def new_forward_neo(
+    self,
+    input_ids=None,
+    past_key_values=None,
+    attention_mask=None,
+    token_type_ids=None,
+    position_ids=None,
+    head_mask=None,
+    inputs_embeds=None,
+    use_cache=None,
+    output_attentions=None,
+    output_hidden_states=None,
+    return_dict=None,
+    embs=None,
+):
     assert len(gpu_blocks) <= torch.cuda.device_count()
     assert sum(gpu_blocks) <= len(self.h)
     ram_blocks = len(self.h) - sum(gpu_blocks)
@@ -477,3 +498,365 @@ def new_forward(
         hidden_states=all_hidden_states,
         attentions=all_self_attentions,
     )
+
+
+def new_forward_xglm(
+    self,
+    input_ids=None,
+    attention_mask=None,
+    encoder_hidden_states=None,
+    encoder_attention_mask=None,
+    head_mask=None,
+    cross_attn_head_mask=None,
+    past_key_values=None,
+    inputs_embeds=None,
+    use_cache=None,
+    output_attentions=None,
+    output_hidden_states=None,
+    return_dict=None,
+):
+    assert len(gpu_blocks) <= torch.cuda.device_count()
+    assert sum(gpu_blocks) <= len(self.layers)
+    ram_blocks = len(self.layers) - sum(gpu_blocks)
+    cumulative_gpu_blocks = tuple(itertools.accumulate(gpu_blocks))
+
+
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+    use_cache = use_cache if use_cache is not None else self.config.use_cache
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+    # retrieve input_ids and inputs_embeds
+    if input_ids is not None and inputs_embeds is not None:
+        raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+    elif input_ids is not None:
+        input_shape = input_ids.size()
+        input_ids = input_ids.view(-1, input_shape[-1])
+    elif inputs_embeds is not None:
+        input_shape = inputs_embeds.size()[:-1]
+    else:
+        raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+    # past_key_values_length
+    past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+    if inputs_embeds is None:
+        if breakmodel:
+            input_ids = input_ids.to(primary_device)
+        inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale
+
+    attention_mask = self._prepare_decoder_attention_mask(
+        attention_mask, input_shape, inputs_embeds, past_key_values_length
+    )
+
+    # expand encoder attention mask
+    if encoder_hidden_states is not None and encoder_attention_mask is not None:
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        encoder_attention_mask = _expand_mask(encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])
+
+    # embed positions
+    if breakmodel:
+        inputs_embeds = inputs_embeds.to(primary_device)
+    positions = self.embed_positions(input_ids, inputs_embeds, past_key_values_length)
+    if breakmodel:
+        positions = positions.to(primary_device)
+
+    hidden_states = inputs_embeds + positions
+
+    hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+    # decoder layers
+    all_hidden_states = () if output_hidden_states else None
+    all_self_attns = () if output_attentions else None
+    all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
+    next_decoder_cache = () if use_cache else None
+
+    if breakmodel and ram_blocks:
+        copystream = torch.cuda.Stream(device=primary_device, priority=-1)
+
+    # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
+    for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]):
+        if attn_mask is not None:
+            assert attn_mask.size()[0] == (
+                len(self.layers)
+            ), f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for {head_mask.size()[0]}."
+    for idx, decoder_layer in enumerate(self.layers):
+        i = idx
+        if breakmodel:
+            if i in range(ram_blocks):
+                index1 = (i+1)%ram_blocks
+                for param1,param2 in zip(self.layers[index1].parameters(),self.layers[(i-1)%ram_blocks].parameters()):
+                    param1.data = param2.data
+                for param1,param2 in zip(self.layers[index1].parameters(),self.extrastorage[index1].parameters()):
+                    with torch.cuda.stream(copystream):
+                        torch.cuda.comm.broadcast(param2.data,out = [param1.data])
+
+        # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+        dropout_probability = random.uniform(0, 1)
+        if self.training and (dropout_probability < self.layerdrop):
+            continue
+
+        past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+        if self.gradient_checkpointing and self.training:
+
+            if use_cache:
+                logger.warning(
+                    "`use_cache = True` is incompatible with gradient checkpointing`. Setting `use_cache = False`..."
+                )
+                use_cache = False
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    # None for past_key_value
+                    return module(*inputs, output_attentions, use_cache)
+
+                return custom_forward
+
+            layer_outputs = torch.utils.checkpoint.checkpoint(
+                create_custom_forward(decoder_layer),
+                hidden_states,
+                attention_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                head_mask[idx] if head_mask is not None else None,
+                cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None,
+                None,
+            )
+        else:
+            if breakmodel:
+                device = primary_device if i < ram_blocks else bisect.bisect_right(cumulative_gpu_blocks, i - ram_blocks)
+            layer_outputs = decoder_layer(
+                hidden_states.to(device) if breakmodel and hidden_states is not None else hidden_states,
+                attention_mask=attention_mask.to(device) if breakmodel and attention_mask is not None else attention_mask,
+                encoder_hidden_states=encoder_hidden_states.to(device) if breakmodel and encoder_hidden_states is not None else encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask.to(device) if breakmodel and encoder_attention_mask is not None else encoder_attention_mask,
+                layer_head_mask=((head_mask[idx].to(device) if breakmodel and head_mask[idx] is not None else head_mask[idx]) if head_mask is not None else None),
+                cross_attn_layer_head_mask=(
+                    (cross_attn_head_mask[idx].to(device) if breakmodel and cross_attn_head_mask[idx] is not None else cross_attn_head_mask[idx]) if cross_attn_head_mask is not None else None
+                ),
+                past_key_value=tuple(v.to(device) for v in past_key_value if v is not None) if breakmodel and past_key_value is not None and i >= ram_blocks and len(past_key_value) and past_key_value[0].device.index != device else past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+            )
+        hidden_states = layer_outputs[0]
+
+        if use_cache:
+            next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)
+
+        if output_attentions:
+            all_self_attns += (layer_outputs[1],)
+
+            if encoder_hidden_states is not None:
+                all_cross_attentions += (layer_outputs[2],)
+        
+        if breakmodel:
+            if i in range(ram_blocks):
+                torch.cuda.synchronize()
+                torch.cuda.empty_cache()
+
+    if breakmodel:
+        if ram_blocks:
+            del copystream
+        torch.cuda.empty_cache()
+        hidden_states = hidden_states.to(primary_device)
+    hidden_states = self.layer_norm(hidden_states)
+    if breakmodel:
+        hidden_states = hidden_states.to(primary_device)
+
+    # add hidden states from the last decoder layer
+    if output_hidden_states:
+        all_hidden_states += (hidden_states,)
+
+    next_cache = next_decoder_cache if use_cache else None
+    if not return_dict:
+        return tuple(
+            v
+            for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
+            if v is not None
+        )
+    return BaseModelOutputWithPastAndCrossAttentions(
+        last_hidden_state=hidden_states,
+        past_key_values=next_cache,
+        hidden_states=all_hidden_states,
+        attentions=all_self_attns,
+        cross_attentions=all_cross_attentions,
+    )
+
+
+def new_forward_opt(
+    self,
+    input_ids=None,
+    attention_mask=None,
+    head_mask=None,
+    past_key_values=None,
+    inputs_embeds=None,
+    use_cache=None,
+    output_attentions=None,
+    output_hidden_states=None,
+    return_dict=None,
+):
+    assert len(gpu_blocks) <= torch.cuda.device_count()
+    assert sum(gpu_blocks) <= len(self.layers)
+    ram_blocks = len(self.layers) - sum(gpu_blocks)
+    cumulative_gpu_blocks = tuple(itertools.accumulate(gpu_blocks))
+
+
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+    use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+    # retrieve input_ids and inputs_embeds
+    if input_ids is not None and inputs_embeds is not None:
+        raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+    elif input_ids is not None:
+        input_shape = input_ids.size()
+        input_ids = input_ids.view(-1, input_shape[-1])
+    elif inputs_embeds is not None:
+        input_shape = inputs_embeds.size()[:-1]
+    else:
+        raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+    past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+    if inputs_embeds is None:
+        if breakmodel:
+            input_ids = input_ids.to(primary_device) 
+        inputs_embeds = self.embed_tokens(input_ids)
+
+    # embed positions
+    if breakmodel:
+        inputs_embeds = inputs_embeds.to(primary_device) 
+    if attention_mask is None:
+        attention_mask = torch.ones(inputs_embeds.shape[:2], dtype=torch.bool, device=inputs_embeds.device)
+
+    positions = self.embed_positions(attention_mask)[:, past_key_values_length:, :]
+    if breakmodel:
+        positions = positions.to(primary_device) 
+
+    attention_mask = self._prepare_decoder_attention_mask(
+        attention_mask, input_shape, inputs_embeds, past_key_values_length
+    )
+
+    if self.project_in is not None:
+        inputs_embeds = self.project_in(inputs_embeds)
+
+    hidden_states = inputs_embeds + positions
+
+    hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+    # decoder layers
+    all_hidden_states = () if output_hidden_states else None
+    all_self_attns = () if output_attentions else None
+    next_decoder_cache = () if use_cache else None
+
+    if breakmodel and ram_blocks:
+        copystream = torch.cuda.Stream(device=primary_device, priority=-1)
+
+    # check if head_mask has a correct number of layers specified if desired
+    for attn_mask, mask_name in zip([head_mask], ["head_mask"]):
+        if attn_mask is not None:
+            if attn_mask.size()[0] != (len(self.layers)):
+                raise ValueError(
+                    f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
+                    f" {head_mask.size()[0]}."
+                )
+
+    for idx, decoder_layer in enumerate(self.layers):
+        i = idx
+        if breakmodel:
+            if i in range(ram_blocks):
+                index1 = (i+1)%ram_blocks
+                for param1,param2 in zip(self.layers[index1].parameters(),self.layers[(i-1)%ram_blocks].parameters()):
+                    param1.data = param2.data
+                for param1,param2 in zip(self.layers[index1].parameters(),self.extrastorage[index1].parameters()):
+                    with torch.cuda.stream(copystream):
+                        torch.cuda.comm.broadcast(param2.data,out = [param1.data])
+
+        # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+        dropout_probability = random.uniform(0, 1)
+        if self.training and (dropout_probability < self.layerdrop):
+            continue
+
+        past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+        if self.gradient_checkpointing and self.training:
+
+            if use_cache:
+                logger.warning(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    # None for past_key_value
+                    return module(*inputs, output_attentions, None)
+
+                return custom_forward
+
+            layer_outputs = torch.utils.checkpoint.checkpoint(
+                create_custom_forward(decoder_layer),
+                hidden_states,
+                attention_mask,
+                head_mask[idx] if head_mask is not None else None,
+                None,
+            )
+        else:
+            if breakmodel:
+                device = primary_device if i < ram_blocks else bisect.bisect_right(cumulative_gpu_blocks, i - ram_blocks)
+            layer_outputs = decoder_layer(
+                hidden_states.to(device) if breakmodel and hidden_states is not None else hidden_states,
+                attention_mask=attention_mask.to(device) if breakmodel and attention_mask is not None else attention_mask,
+                layer_head_mask=((head_mask[idx].to(device) if breakmodel and head_mask[idx] is not None else head_mask[idx]) if head_mask is not None else None),
+                past_key_value=tuple(v.to(device) for v in past_key_value if v is not None) if breakmodel and past_key_value is not None and i >= ram_blocks and len(past_key_value) and past_key_value[0].device.index != device else past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+            )
+
+        hidden_states = layer_outputs[0]
+
+        if use_cache:
+            next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
+
+        if output_attentions:
+            all_self_attns += (layer_outputs[1],)
+        
+        if breakmodel:
+            if i in range(ram_blocks):
+                torch.cuda.synchronize()
+                torch.cuda.empty_cache()
+
+    if breakmodel:
+        if ram_blocks:
+            del copystream
+        torch.cuda.empty_cache()
+        hidden_states = hidden_states.to(primary_device)
+    if self.project_out is not None:
+        hidden_states = self.project_out(hidden_states)
+    if breakmodel:
+        hidden_states = hidden_states.to(primary_device)
+
+    # add hidden states from the last decoder layer
+    if output_hidden_states:
+        all_hidden_states += (hidden_states,)
+
+    next_cache = next_decoder_cache if use_cache else None
+    if not return_dict:
+        return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+    return BaseModelOutputWithPast(
+        last_hidden_state=hidden_states,
+        past_key_values=next_cache,
+        hidden_states=all_hidden_states,
+        attentions=all_self_attns,
+    )

bridge.lua

@@ -165,7 +165,7 @@ return function(_python, _bridged)
     ---@field num_outputs integer
     ---@field feedback string
     ---@field is_config_file_open boolean
-    local kobold = setmetatable({API_VERSION = 1.0}, metawrapper)
+    local kobold = setmetatable({API_VERSION = 1.1}, metawrapper)
     local KoboldLib_mt = setmetatable({}, metawrapper)
     local KoboldLib_getters = setmetatable({}, metawrapper)
     local KoboldLib_setters = setmetatable({}, metawrapper)
@@ -866,6 +866,7 @@ return function(_python, _bridged)
     ---@field settopp number
     ---@field settopk integer
     ---@field settfs number
+    ---@field settypical number
     ---@field setreppen number
     ---@field setreppenslope number
     ---@field setreppenrange number
@@ -882,6 +883,7 @@ return function(_python, _bridged)
     ---@field top_p number
     ---@field top_k integer
     ---@field tfs number
+    ---@field typical number
     ---@field reppen number
     ---@field reppenslope number
     ---@field reppenrange number
@@ -1048,11 +1050,34 @@ return function(_python, _bridged)
             return
         elseif not bridged.vars.gamestarted and v == "" then
             error("`KoboldLib.submission` must not be set to the empty string when the story is empty")
+            return
         end
         bridged.vars.submission = v
     end
 
 
+    --==========================================================================
+    -- Userscript API: Soft prompt
+    --==========================================================================
+
+    ---@param t KoboldLib
+    ---@return string?
+    function KoboldLib_getters.spfilename(t)
+        return bridged.get_spfilename()
+    end
+
+    ---@param t KoboldLib
+    ---@param v string?
+    function KoboldLib_setters.spfilename(t, v)
+        if v:find("/") or v:find("\\") then
+            error("Cannot set `KoboldLib.spfilename` to a string that contains slashes")
+        end
+        if bridged.set_spfilename(v) then
+            maybe_require_regeneration()
+        end
+    end
+
+
     --==========================================================================
     -- Userscript API: Model information
     --==========================================================================

colab/GPU.ipynb

@@ -7,7 +7,7 @@
       "private_outputs": true,
       "provenance": [],
       "collapsed_sections": [],
-      "authorship_tag": "ABX9TyOKIa/NDLlYI5j63GXPtkXv",
+      "authorship_tag": "ABX9TyPbwW79K9/RkYH9i9rkYFyj",
       "include_colab_link": true
     },
     "kernelspec": {
@@ -68,14 +68,20 @@
         "#@title <b><-- Click this to start KoboldAI</b>\n",
         "#@markdown You can find a description of the models below along with instructions on how to start KoboldAI.\n",
         "\n",
-        "Model = \"KoboldAI/GPT-Neo-2.7B-Janeway\" #@param [\"KoboldAI/GPT-Neo-2.7B-Janeway\", \"KoboldAI/GPT-Neo-2.7B-AID\", \"KoboldAI/GPT-Neo-2.7B-Picard\", \"KoboldAI/GPT-Neo-2.7B-Horni-LN\", \"KoboldAI/GPT-Neo-2.7B-Horni\", \"KoboldAI/GPT-Neo-2.7B-Shinen\", \"EleutherAI/gpt-neo-2.7B\"] {allow-input: true}\n",
+        "Model = \"KoboldAI/fairseq-dense-2.7B-Nerys\" #@param [\"KoboldAI/fairseq-dense-2.7B-Nerys\", \"KoboldAI/GPT-Neo-2.7B-Janeway\", \"KoboldAI/GPT-Neo-2.7B-AID\", \"KoboldAI/GPT-Neo-2.7B-Picard\", \"KoboldAI/GPT-Neo-2.7B-Horni-LN\", \"KoboldAI/GPT-Neo-2.7B-Horni\", \"KoboldAI/GPT-Neo-2.7B-Shinen\", \"EleutherAI/gpt-neo-2.7B\"] {allow-input: true}\n",
         "Version = \"Official\" #@param [\"Official\", \"United\"] {allow-input: true}\n",
+        "Provider = \"Localtunnel\" #@param [\"Localtunnel\", \"Cloudflare\"]\n",
         "\n",
         "!nvidia-smi\n",
         "from google.colab import drive\n",
         "drive.mount('/content/drive/')\n",
         "\n",
-        "!wget https://henk.tech/ckds -O - | bash /dev/stdin -m $Model -g $Version"
+        "if Provider == \"Localtunnel\":\n",
+        "  tunnel = \"--localtunnel yes\"\n",
+        "else:\n",
+        "  tunnel = \"\"\n",
+        "\n",
+        "!wget https://henk.tech/ckds -O - | bash /dev/stdin -m $Model -g $Version $tunnel"
       ],
       "execution_count": null,
       "outputs": []
@@ -84,27 +90,32 @@
       "cell_type": "markdown",
       "source": [
         "# GPU Edition Model Descriptions\n",
-        "| Model                                                        | Size     | Style      | Description                                                  |\n",
-        "| ------------------------------------------------------------ | -------- | ---------- | ------------------------------------------------------------ |\n",
-        "| [GPT-Neo-2.7B-Janeway](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Janeway) by Mr Seeker | 2.7B GPU | Novel | Janeway is a model created from Picard's dataset combined with a brand new collection of ebooks. This model is trained on 20% more content than Picard and has been trained on literature from various genres. Although the model is mainly focussed on SFW, romantic scenes might involve a degree of nudity. |\n",
-        "| [GPT-Neo-2.7B-Picard](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Picard) by Mr Seeker | 2.7B GPU | Novel      | Picard is a model trained for SFW Novels based on GPT-Neo-2.7B. It is focused on Novel style writing without the NSFW bias. While the name suggests a sci-fi model this model is designed for Novels of a variety of genre's. It is meant to be used in KoboldAI's regular mode. |\n",
-        "| [GPT-Neo-2.7B-AID](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-AID) by melastacho | 2.7B GPU | Adventure | Also know as Adventure 2.7B this is a clone of the AI Dungeon Classic model and is best known for the epic wackey adventures that AI Dungeon Classic players love. |\n",
-        "| [GPT-Neo-2.7B-Horni-LN](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Horni-LN) by finetune | 2.7B GPU | Novel | This model is based on GPT-Neo-2.7B-Horni and retains its NSFW knowledge, but was then further biased towards SFW novel stories. If you seek a balance between a SFW Novel model and a NSFW model this model should be a good choice. |\n",
-        "| [GPT-Neo-2.7B-Horni](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Horni) by finetune | 2.7B GPU | NSFW       | This model is tuned on Literotica to produce a Novel style model biased towards NSFW content. Can still be used for SFW stories but will have a bias towards NSFW content. It is meant to be used in KoboldAI's regular mode. |\n",
-        "| [GPT-Neo-2.7B-Shinen](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Shinen) by Mr Seeker | 2.7B GPU | NSFW       | Shinen is an alternative to the Horni model designed to be more explicit. If Horni is to tame for you shinen might produce better results. While it is a Novel model it is unsuitable for SFW stories due to its heavy NSFW bias. Shinen will not hold back. It is meant to be used in KoboldAI's regular mode. |\n",
-        "| [GPT-Neo-2.7B](https://huggingface.co/EleutherAI/gpt-neo-2.7B) by EleutherAI | 2.7B GPU    | Generic    | This is the base model for all the other 2.7B models, it is best used when you have a use case that we have no other models available for, such as writing blog articles or programming. It can also be a good basis for the experience of some of the softprompts if your softprompt is not about a subject the other models cover. |\n",
+        "| Model | Size | Style | Description |\n",
+        "| --- | --- | --- | --- |\n",
+        "| [Fairseq-Dense-2.7B-Nerys](https://huggingface.co/KoboldAI/fairseq-dense-2.7B-Nerys) by Mr Seeker | 2.7B | Novel/Adventure | Nerys is a hybrid model based on Pike (A newer Janeway), on top of the Pike dataset you also get some Light Novels, Adventure mode support and a little bit of shinen thrown in the mix. The end result is a very diverse model that is heavily biased towards SFW novel writing, but one that can go beyond its novel training and make for an excellent adventure model to. Adventure mode is best played from a second person perspective, but can be played in first or third person as well. Novel writing can be done best from the first or third person. |\n",
+        "| [GPT-Neo-2.7B-Janeway](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Janeway) by Mr Seeker | 2.7B | Novel | Janeway is a model created from Picard's dataset combined with a brand new collection of ebooks. This model is trained on 20% more content than Picard and has been trained on literature from various genres. Although the model is mainly focussed on SFW, romantic scenes might involve a degree of nudity. |\n",
+        "| [GPT-Neo-2.7B-Picard](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Picard) by Mr Seeker | 2.7B | Novel | Picard is a model trained for SFW Novels based on GPT-Neo-2.7B. It is focused on Novel style writing without the NSFW bias. While the name suggests a sci-fi model this model is designed for Novels of a variety of genre's. It is meant to be used in KoboldAI's regular mode. |\n",
+        "| [GPT-Neo-2.7B-AID](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-AID) by melastacho | 2.7B | Adventure | Also know as Adventure 2.7B this is a clone of the AI Dungeon Classic model and is best known for the epic wackey adventures that AI Dungeon Classic players love. |\n",
+        "| [GPT-Neo-2.7B-Horni-LN](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Horni-LN) by finetune | 2.7B | Novel | This model is based on GPT-Neo-2.7B-Horni and retains its NSFW knowledge, but was then further biased towards SFW novel stories. If you seek a balance between a SFW Novel model and a NSFW model this model should be a good choice. |\n",
+        "| [GPT-Neo-2.7B-Horni](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Horni) by finetune | 2.7B | NSFW | This model is tuned on Literotica to produce a Novel style model biased towards NSFW content. Can still be used for SFW stories but will have a bias towards NSFW content. It is meant to be used in KoboldAI's regular mode. |\n",
+        "| [GPT-Neo-2.7B-Shinen](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Shinen) by Mr Seeker | 2.7B | NSFW | Shinen is an alternative to the Horni model designed to be more explicit. If Horni is to tame for you shinen might produce better results. While it is a Novel model it is unsuitable for SFW stories due to its heavy NSFW bias. Shinen will not hold back. It is meant to be used in KoboldAI's regular mode. |\n",
+        "| [GPT-Neo-2.7B](https://huggingface.co/EleutherAI/gpt-neo-2.7B) by EleutherAI | 2.7B | Generic | This is the base model for all the other 2.7B models, it is best used when you have a use case that we have no other models available for, such as writing blog articles or programming. It can also be a good basis for the experience of some of the softprompts if your softprompt is not about a subject the other models cover. |\n",
         "\n",
         "# [TPU Edition Model Descriptions](https://colab.research.google.com/github/KoboldAI/KoboldAI-Client/blob/main/colab/TPU.ipynb)\n",
         "\n",
-        "| Model                          | Size   | Style     | Drive Space | Description                                                  |\n",
-        "| ------------------------------ | ------ | --------- | ----------- | ------------------------------------------------------------ |\n",
-        "| Skein 6B by VE_FORBDRYDERNE    | 6B TPU | Hybrid    | 0 GB         | Skein is our flagship 6B model, it is a hybrid between a Adventure model and a Novel model. Best used with either Adventure mode or the You Bias userscript enabled. Skein has been trained on high quality Novels along with CYOA adventure stories and is not as wackey as the Adventure model. It also has tagging support. |\n",
-        "| Janeway 6B by Mr Seeker | 6B TPU | Novel | 0 GB | Janeway is a model created from Picard's dataset combined with a brand new collection of ebooks. This model is trained on 20% more content than Picard and has been trained on literature from various genres. Although the model is mainly focussed on SFW, romantic scenes might involve a degree of nudity. |\n",
-        "| Adventure 6B by VE_FORBRYDERNE | 6B TPU | Adventure | 0 GB         | Adventure is a 6B model designed to mimick the behavior of AI Dungeon. It is exclusively for Adventure Mode and can take you on the epic and wackey adventures that AI Dungeon players love. It also features the many tropes of AI Dungeon as it has been trained on very similar data. It must be used in second person (You). |\n",
-        "| Lit 6B by Haru                 | 6B TPU | NSFW      | 8 GB /  12 GB | Lit is a great NSFW model trained by Haru on both a large set of Literotica stories and high quality novels along with tagging support. Creating a high quality model for your NSFW stories. This model is exclusively a novel model and is best used in third person. |\n",
-        "| Shinen 6B by Mr Seeker | 6B TPU | NSFW | 0 GB | Shinen is an alternative to the Lit model designed to be more explicit. If Lit is to tame for you Shinen might produce better results. While it is a Novel model it is unsuitable for SFW stories due to its heavy NSFW bias. Shinen will not hold back. It is meant to be used in KoboldAI's regular mode. |\n",
-        "| Generic 6B by EleutherAI       | 6B TPU | Generic   | 10 GB / 12 GB | GPT-J-6B is what all other models are based on, if you need something that has no specific bias towards any particular subject this is the model for you. Best used when the other models are not suitable for what you wish to do. Such as homework assistance, blog writing, coding and more. It needs more hand holding than other models and is more prone to undesirable formatting changes. |\n",
-        "| C1 6B by Haru                  | 6B TPU | Chatbot   | 8 GB /  12 GB | C1 has been trained on various internet chatrooms, it makes the basis for an interesting chatbot model and has been optimized to be used in the Chatmode. |\n",
+        "| Model | Size | Style | Description |\n",
+        "| --- | --- | --- | --- |\n",
+        "| [Nerys](https://huggingface.co/KoboldAI/fairseq-dense-13B-Nerys) by Mr Seeker | 13B | Novel/Adventure | Nerys is a hybrid model based on Pike (A newer Janeway), on top of the Pike dataset you also get some Light Novels, Adventure mode support and a little bit of shinen thrown in the mix. The end result is a very diverse model that is heavily biased towards SFW novel writing, but one that can go beyond its novel training and make for an excellent adventure model to. Adventure mode is best played from a second person perspective, but can be played in first or third person as well. Novel writing can be done best from the first or third person. |\n",
+        "| [Janeway](https://huggingface.co/KoboldAI/fairseq-dense-13B-Janeway) by Mr Seeker | 13B | Novel | Janeway is a model created from Picard's dataset combined with a brand new collection of ebooks. This model is trained on 20% more content than Picard and has been trained on literature from various genres. Although the model is mainly focussed on SFW, romantic scenes might involve a degree of nudity. |\n",
+        "| [Shinen](https://huggingface.co/KoboldAI/fairseq-dense-13B-Shinen) by Mr Seeker | 13B | NSFW | Shinen is an NSFW model designed to be more explicit. Trained on a variety of stories from the website Sexstories it contains many different kinks. |\n",
+        "| [Skein](https://huggingface.co/KoboldAI/GPT-J-6B-Skein) by VE\\_FORBRYDERNE | 6B | Adventure | Skein is best used with Adventure mode enabled, it consists of a 4 times larger adventure dataset than the Adventure model making it excellent for text adventure gaming. On top of that it also consists of light novel training further expanding its knowledge and writing capabilities. It can be used with the You filter bias if you wish to write Novels with it, but dedicated Novel models can perform better for this task. |\n",
+        "| [Adventure](https://huggingface.co/KoboldAI/GPT-J-6B-Adventure) by VE\\_FORBRYDERNE | 6B | Adventure | Adventure is a 6B model designed to mimick the behavior of AI Dungeon. It is exclusively for Adventure Mode and can take you on the epic and wackey adventures that AI Dungeon players love. It also features the many tropes of AI Dungeon as it has been trained on very similar data. It must be used in second person (You). |\n",
+        "| [Lit](https://huggingface.co/hakurei/lit-6B) by Haru | 6B | NSFW | Lit is a great NSFW model trained by Haru on both a large set of Literotica stories and high quality novels along with tagging support. Creating a high quality model for your NSFW stories. This model is exclusively a novel model and is best used in third person. |\n",
+        "| [Convo](https://huggingface.co/hitomi-team/convo-6B) by Hitomi Team | 6B | Chatbot | Convo-6B is a GPT-J 6B model fine-tuned on a collection of high quality open source datasets which amount to 6 million messages. The primary goal of the model is to provide improved performance and generalization when generating multi-turn dialogue for characters that were not present from within the fine tuning data. The prompted performance has especially improved over the predecessor model [C1-6B](https://huggingface.co/hakurei/c1-6B). |\n",
+        "| [C1](https://huggingface.co/hakurei/c1-6B) by Haru | 6B | Chatbot | C1 has been trained on various internet chatrooms, it makes the basis for an interesting chatbot model and has been optimized to be used in the Chatmode. |\n",
+        "| Neo(X) by EleutherAI | 20B | Generic | NeoX is the largest EleutherAI model currently available, being a generic model it is not particularly trained towards anything and can do a variety of writing, Q&A and coding tasks. 20B's performance is closely compared to the 13B models and it is worth trying both especially if you have a task that does not involve english writing. Its behavior will be similar to the GPT-J-6B model since they are trained on the same dataset but with more sensitivity towards repetition penalty and with more knowledge. |\n",
+        "| [Fairseq Dense](https://huggingface.co/KoboldAI/fairseq-dense-13B) | 13B | Generic | Trained by Facebook Researchers this model stems from the MOE research project within Fairseq. This particular version has been converted by us for use in KoboldAI. It is known to be on par with the larger 20B model from EleutherAI and considered as better for pop culture and language tasks. Because the model has never seen a new line (enter) it may perform worse on formatting and paragraphing. |\n",
+        "| [GPT-J-6B](https://huggingface.co/EleutherAI/gpt-j-6B) by EleutherAI | 6B | Generic | This model serves as the basis for most other 6B models (Some being based on Fairseq Dense instead). Being trained on the Pile and not biased towards anything in particular it is suitable for a variety of tasks such as writing, Q&A and coding tasks. You will likely get better result with larger generic models or finetuned models. |\n",
         "\n",
         "\n",
         "| Style     | Description                                                  |\n",
@@ -113,7 +124,6 @@
         "| NSFW      | Indicates that the model is strongly biased towards NSFW content and is not suitable for children, work environments or livestreaming. Most NSFW models are also Novel models in nature. |\n",
         "| Adventure | These models are excellent for people willing to play KoboldAI like a Text Adventure game and are meant to be used with Adventure mode enabled. Even if you wish to use it as a Novel style model you should always have Adventure mode on and set it to story. These models typically have a strong bias towards the use of the word You and without Adventure mode enabled break the story flow and write actions on your behalf. |\n",
         "| Chatbot   | These models are specifically trained for chatting and are best used with the Chatmode enabled. Typically trained on either public chatrooms or private chats. |\n",
-        "| Hybrid    | Hybrid models are a blend between different styles, for example they are trained on both Novel stories and Adventure stories. These models are great variety models that you can use for multiple different playstyles and modes, but depending on your usage you may need to enable Adventure Mode or the You bias (in userscripts). |\n",
         "| Generic   | Generic models are not trained towards anything specific, typically used as a basis for other tasks and models. They can do everything the other models can do, but require much more handholding to work properly. Generic models are an ideal basis for tasks that we have no specific model for, or for experiencing a softprompt in its raw form. |\n",
         "\n",
         "# How to start KoboldAI in 7 simple steps\n",

colab/TPU.ipynb

@@ -18,7 +18,21 @@
         "\n",
         "For more information about KoboldAI check our our Github readme : https://github.com/KoboldAI/KoboldAI-Client/blob/main/readme.md\n",
         "\n",
-        "More (smaller) models are available in the **[GPU edition](https://colab.research.google.com/github/koboldai/KoboldAI-Client/blob/main/colab/GPU.ipynb)**!"
+        "More (smaller) models are available in the **[GPU edition](https://colab.research.google.com/github/koboldai/KoboldAI-Client/blob/main/colab/GPU.ipynb)**!\n",
+        "\n",
+        "---\n",
+        "## How to load KoboldAI: Everything you need to know\n",
+        "1. On a phone? First put your browser in desktop mode because of a Google Colab bug. Otherwise nothing will happen when you click the play button. Then tap the play button next to \"<-- Tap This if you play on Mobile\", you will see an audio player. Keep the audio player playing so Colab does not get shut down in the background.\n",
+        "2. Select the desired model, you will find a description of all the available models further down the page.\n",
+        "3. Click the play button next to \"<-- Select your model below and then click this to start KoboldAI\".\n",
+        "4. Got a message saying no accelerator is available? Click cancel, and try again in a few minutes. If you do not manage to get a session when you frequently try again try at a different time of day, colab can be busy or your priority may have been lowered by frequent usage.\n",
+        "5. After everything is done loading you will get a link that you can use to open KoboldAI. In case of Localtunnel you will also be warned that some people are abusing Localtunnel for phishing, once you acknowledge this warning you will be taken to KoboldAI's interface. If you picked Cloudflare and get a 1033 error refresh the error page after waiting one minute.\n",
+        "\n",
+        "---\n",
+        "\n",
+        "Further down the page you can find descriptions of the models, and tips to get the most out of your Google Colab experience.\n",
+        "\n",
+        "Make sure to keep this page open while you are using KoboldAI, and check back regularly to see if you got a Captcha. Failure to complete the captcha's in time can result in termination of your session or a lower priority towards the TPUs."
       ],
       "metadata": {
         "id": "zrLGxVCEaqZx"
@@ -47,11 +61,13 @@
       },
       "outputs": [],
       "source": [
+        "#@title <b><-- Select your model below and then click this to start KoboldAI</b>\n",
+        "#@markdown You can find a description of the models below along with instructions on how to start KoboldAI.\n",
+        "\n",
         "#@title <b><-- Click this to start KoboldAI</b>\n",
-        "Model = \"Skein 6B\" #@param [\"Skein 6B\", \"Janeway 6B\", \"Adventure 6B\", \"Lit 6B\", \"Shinen 6B\", \"Generic 6B\", \"C1 6B\"]\n",
+        "Model = \"Nerys 13B\" #@param [\"Nerys 13B\", \"Janeway 13B\", \"Shinen 13B\", \"Skein 6B\", \"Janeway 6B\", \"Adventure 6B\", \"Shinen 6B\", \"Lit 6B\", \"Convo 6B\", \"C1 6B\", \"NeoX 20B\", \"facebook/opt-13b\", \"KoboldAI/fairseq-dense-13B\", \"EleutherAI/gpt-j-6B\"] {allow-input: true}\n",
         "Version = \"Official\" #@param [\"Official\", \"United\"] {allow-input: true}\n",
-        "Drive = \"Unextracted (Less Space)\" #@param [\"Unextracted (Less Space)\", \"Extracted (Faster Loading)\"]\n",
-        "#@markdown Extracted models take up more space but load faster the next time you use them, not all models use your Google Drive. See the Model list below for descriptions and space requirements. If your extracted model does not load the next time you try to launch KoboldAI delete the folder from your Google Drive and ensure enough space is available.\n",
+        "Provider = \"Localtunnel\" #@param [\"Localtunnel\", \"Cloudflare\"]\n",
         "\n",
         "import os\n",
         "try:\n",
@@ -64,113 +80,115 @@
         "from google.colab import drive\n",
         "drive.mount('/content/drive/')\n",
         "\n",
-        "!wget https://henk.tech/ckds -O - | bash /dev/stdin -i drive\n",
-        "\n",
-        "if Model == \"Skein 6B\":\n",
-        "  path = \"gpt-j-6b-skein-jax\"\n",
+        "if Model == \"Janeway 13B\":\n",
+        "  Model = \"KoboldAI/fairseq-dense-13B-Janeway\"\n",
+        "  path = \"\"\n",
+        "  download = \"\"\n",
+        "elif Model == \"Nerys 13B\":\n",
+        "  Model = \"KoboldAI/fairseq-dense-13B-Nerys\"\n",
+        "  path = \"\"\n",
+        "  download = \"\"\n",
+        "elif Model == \"Shinen 13B\":\n",
+        "  Model = \"KoboldAI/fairseq-dense-13B-Shinen\"\n",
+        "  path = \"\"\n",
+        "  download = \"\"\n",
+        "elif Model == \"NeoX 20B\":\n",
+        "  Model = \"TPUMeshTransformerGPTNeoX\"\n",
+        "  path = \" -p gpt-neox-20b-jax\"\n",
         "  location = \"colab\"\n",
-        "  download = \"-a https://storage.henk.tech/KoboldAI/skein-jax.txt\"\n",
+        "  download = \" -a https://storage.henk.tech/KoboldAI/neox-20b.txt\"\n",
         "  extract = \"\"\n",
         "  Drive = \"Unextracted (Less Space)\"\n",
-        "  ![[ -f /content/drive/MyDrive/KoboldAI/settings/gpt-j-6b-skein-jax.settings ]] || echo -e \"{\\n   \\\"apikey\\\": \\\"\\\",\\n   \\\"andepth\\\": 3,\\n   \\\"temp\\\": 0.5,\\n   \\\"top_p\\\": 0.9,\\n   \\\"top_k\\\": 0,\\n   \\\"tfs\\\": 1.0,\\n   \\\"rep_pen\\\": 1.1,\\n   \\\"genamt\\\": 80,\\n   \\\"max_length\\\": 2048,\\n   \\\"ikgen\\\": 200,\\n   \\\"formatoptns\\\": {\\n      \\\"frmttriminc\\\": true,\\n      \\\"frmtrmblln\\\": false,\\n      \\\"frmtrmspch\\\": false,\\n      \\\"frmtadsnsp\\\": false\\n   },\\n   \\\"numseqs\\\": 1,\\n   \\\"widepth\\\": 3,\\n   \\\"useprompt\\\": true,\\n   \\\"adventure\\\": false\\n}\" > /content/drive/MyDrive/KoboldAI/settings/gpt-j-6b-skein-jax.settings\n",
-        "if Model == \"Janeway 6B\":\n",
-        "  path = \"gpt-j-6b-janeway-jax\"\n",
-        "  location = \"colab\"\n",
-        "  download = \"-a https://storage.henk.tech/KoboldAI/janeway-jax.txt\"\n",
-        "  extract = \"\"\n",
-        "  Drive = \"Unextracted (Less Space)\"\n",
-        "  ![[ -f /content/drive/MyDrive/KoboldAI/settings/gpt-j-6b-janeway-jax.settings ]] || echo -e \"{\\n   \\\"apikey\\\": \\\"\\\",\\n   \\\"andepth\\\": 3,\\n   \\\"temp\\\": 0.5,\\n   \\\"top_p\\\": 0.9,\\n   \\\"top_k\\\": 0,\\n   \\\"tfs\\\": 1.0,\\n   \\\"rep_pen\\\": 1.1,\\n   \\\"rep_pen_slope\\\": 0.7,\\n   \\\"rep_pen_range\\\": 1024.0,\\n   \\\"genamt\\\": 80,\\n   \\\"max_length\\\": 2048,\\n   \\\"ikgen\\\": 200,\\n   \\\"formatoptns\\\": {\\n      \\\"frmttriminc\\\": true,\\n      \\\"frmtrmblln\\\": false,\\n      \\\"frmtrmspch\\\": false,\\n      \\\"frmtadsnsp\\\": false,\\n      \\\"singleline\\\": false\\n   },\\n   \\\"numseqs\\\": 1,\\n   \\\"widepth\\\": 3,\\n   \\\"useprompt\\\": true,\\n   \\\"adventure\\\": false,\\n   \\\"chatmode\\\": false,\\n   \\\"chatname\\\": \\\"You\\\",\\n   \\\"dynamicscan\\\": false,\\n   \\\"nopromptgen\\\": false,\\n   \\\"rngpersist\\\": false,\\n   \\\"nogenmod\\\": false,\\n   \\\"autosave\\\": false,\\n   \\\"welcome\\\": false,\\n   \\\"newlinemode\\\": \\\"n\\\",\\n   \\\"antemplate\\\": \\\"[Genre: <|>]\\\",\\n   \\\"userscripts\\\": [],\\n   \\\"corescript\\\": \\\"default.lua\\\",\\n   \\\"softprompt\\\": \\\"\\\"\\n}\" > /content/drive/MyDrive/KoboldAI/settings/gpt-j-6b-janeway-jax.settings\n",
-        "if Model == \"Adventure 6B\":\n",
-        "  path = \"gpt-j-6b-adventure-jax\"\n",
-        "  location = \"colab\"\n",
-        "  download = \"-a https://api.wandb.ai/files/ve-forbryderne/adventure/carol-data/models/gpt-j-6b-adventure-jax/aria2.txt\"\n",
-        "  extract = \"\"\n",
-        "  Drive = \"Unextracted (Less Space)\"\n",
-        "  ![[ -f /content/drive/MyDrive/KoboldAI/settings/gpt-j-6b-adventure-jax.settings ]] || echo -e \"{\\n   \\\"apikey\\\": \\\"\\\",\\n   \\\"andepth\\\": 3,\\n   \\\"temp\\\": 0.5,\\n   \\\"top_p\\\": 0.9,\\n   \\\"top_k\\\": 0,\\n   \\\"tfs\\\": 1.0,\\n   \\\"rep_pen\\\": 1.1,\\n   \\\"genamt\\\": 80,\\n   \\\"max_length\\\": 2048,\\n   \\\"ikgen\\\": 200,\\n   \\\"formatoptns\\\": {\\n      \\\"frmttriminc\\\": true,\\n      \\\"frmtrmblln\\\": false,\\n      \\\"frmtrmspch\\\": false,\\n      \\\"frmtadsnsp\\\": false\\n   },\\n   \\\"numseqs\\\": 1,\\n   \\\"widepth\\\": 3,\\n   \\\"useprompt\\\": true,\\n   \\\"adventure\\\": true\\n}\" > /content/drive/MyDrive/KoboldAI/settings/gpt-j-6b-adventure-jax.settings\n",
-        "if Model == \"Lit 6B\":\n",
-        "  path = \"gpt-j-6b-lit-jax\"\n",
-        "  location = \"drive\"\n",
-        "  download = \"-a https://storage.henk.tech/KoboldAI/aria2.php?file=gpt-j-6b-lit-jax.7z\"\n",
-        "  extract = \"-z gpt-j-6b-lit-jax.7z\"\n",
-        "  ![[ -f /content/drive/MyDrive/KoboldAI/settings/gpt-j-6b-lit-jax.settings ]] || echo -e \"{\\n   \\\"apikey\\\": \\\"\\\",\\n   \\\"andepth\\\": 3,\\n   \\\"temp\\\": 0.5,\\n   \\\"top_p\\\": 0.9,\\n   \\\"top_k\\\": 0,\\n   \\\"tfs\\\": 1.0,\\n   \\\"rep_pen\\\": 1.1,\\n   \\\"genamt\\\": 80,\\n   \\\"max_length\\\": 2048,\\n   \\\"ikgen\\\": 200,\\n   \\\"formatoptns\\\": {\\n      \\\"frmttriminc\\\": true,\\n      \\\"frmtrmblln\\\": false,\\n      \\\"frmtrmspch\\\": false,\\n      \\\"frmtadsnsp\\\": false\\n   },\\n   \\\"numseqs\\\": 1,\\n   \\\"widepth\\\": 3,\\n   \\\"useprompt\\\": true,\\n   \\\"adventure\\\": false\\n}\" > /content/drive/MyDrive/KoboldAI/settings/gpt-j-6b-lit-jax.settings\n",
-        "if Model == \"Shinen 6B\":\n",
-        "  path = \"gpt-j-6b-shinen-jax\"\n",
-        "  location = \"colab\"\n",
-        "  download = \"-a https://storage.henk.tech/KoboldAI/shinen-jax.txt\"\n",
-        "  extract = \"\"\n",
-        "  Drive = \"Unextracted (Less Space)\"\n",
-        "  ![[ -f /content/drive/MyDrive/KoboldAI/settings/gpt-j-6b-shinen-jax.settings ]] || echo -e \"{\\n   \\\"apikey\\\": \\\"\\\",\\n   \\\"andepth\\\": 3,\\n   \\\"temp\\\": 0.5,\\n   \\\"top_p\\\": 0.9,\\n   \\\"top_k\\\": 0,\\n   \\\"tfs\\\": 1.0,\\n   \\\"rep_pen\\\": 1.1,\\n   \\\"rep_pen_slope\\\": 0.7,\\n   \\\"rep_pen_range\\\": 1024.0,\\n   \\\"genamt\\\": 80,\\n   \\\"max_length\\\": 2048,\\n   \\\"ikgen\\\": 200,\\n   \\\"formatoptns\\\": {\\n      \\\"frmttriminc\\\": true,\\n      \\\"frmtrmblln\\\": false,\\n      \\\"frmtrmspch\\\": false,\\n      \\\"frmtadsnsp\\\": false,\\n      \\\"singleline\\\": false\\n   },\\n   \\\"numseqs\\\": 1,\\n   \\\"widepth\\\": 3,\\n   \\\"useprompt\\\": true,\\n   \\\"adventure\\\": false,\\n   \\\"chatmode\\\": false,\\n   \\\"chatname\\\": \\\"You\\\",\\n   \\\"dynamicscan\\\": false,\\n   \\\"nopromptgen\\\": false,\\n   \\\"rngpersist\\\": false,\\n   \\\"nogenmod\\\": false,\\n   \\\"autosave\\\": false,\\n   \\\"welcome\\\": false,\\n   \\\"newlinemode\\\": \\\"n\\\",\\n   \\\"antemplate\\\": \\\"[Genre: <|>]\\\",\\n   \\\"userscripts\\\": [],\\n   \\\"corescript\\\": \\\"default.lua\\\",\\n   \\\"softprompt\\\": \\\"\\\"\\n}\" > /content/drive/MyDrive/KoboldAI/settings/gpt-j-6b-shinen-jax.settings\n",
-        "if Model == \"Generic 6B\":\n",
-        "  path = \"step_383500\"\n",
-        "  location = \"drive\"\n",
-        "  download = \"-a https://storage.henk.tech/KoboldAI/aria2.php?file=step_383500_slim.tar.zstd\"\n",
-        "  extract = \"-t step_383500_slim.tar.zstd\"\n",
-        "  ![[ -f /content/drive/MyDrive/KoboldAI/settings/step_383500.settings ]] || echo -e \"{\\n   \\\"apikey\\\": \\\"\\\",\\n   \\\"andepth\\\": 3,\\n   \\\"temp\\\": 0.5,\\n   \\\"top_p\\\": 0.9,\\n   \\\"top_k\\\": 0,\\n   \\\"tfs\\\": 1.0,\\n   \\\"rep_pen\\\": 1.1,\\n   \\\"genamt\\\": 80,\\n   \\\"max_length\\\": 2048,\\n   \\\"ikgen\\\": 200,\\n   \\\"formatoptns\\\": {\\n      \\\"frmttriminc\\\": true,\\n      \\\"frmtrmblln\\\": false,\\n      \\\"frmtrmspch\\\": false,\\n      \\\"frmtadsnsp\\\": false\\n   },\\n   \\\"numseqs\\\": 1,\\n   \\\"widepth\\\": 3,\\n   \\\"useprompt\\\": true,\\n   \\\"adventure\\\": false\\n}\" > /content/drive/MyDrive/KoboldAI/settings/step_383500.settings\n",
-        "if Model == \"C1 6B\":\n",
-        "  path = \"gpt-j-6b-c1-jax\"\n",
-        "  location = \"drive\"\n",
-        "  download = \"-a https://storage.henk.tech/KoboldAI/aria2.php?file=gpt-j-6b-c1-jax.7z\"\n",
-        "  extract = \"-z gpt-j-6b-c1-jax.7z\"\n",
-        "  ![[ -f /content/drive/MyDrive/KoboldAI/settings/gpt-j-6b-c1-jax.settings ]] || echo -e \"{\\n   \\\"apikey\\\": \\\"\\\",\\n   \\\"andepth\\\": 3,\\n   \\\"temp\\\": 0.5,\\n   \\\"top_p\\\": 0.9,\\n   \\\"top_k\\\": 0,\\n   \\\"tfs\\\": 1.0,\\n   \\\"rep_pen\\\": 1.1,\\n   \\\"genamt\\\": 80,\\n   \\\"max_length\\\": 2048,\\n   \\\"ikgen\\\": 200,\\n   \\\"formatoptns\\\": {\\n      \\\"frmttriminc\\\": true,\\n      \\\"frmtrmblln\\\": false,\\n      \\\"frmtrmspch\\\": false,\\n      \\\"frmtadsnsp\\\": false\\n   },\\n   \\\"numseqs\\\": 1,\\n   \\\"widepth\\\": 3,\\n   \\\"useprompt\\\": true,\\n   \\\"chatmode\\\": true\\n}\" > /content/drive/MyDrive/KoboldAI/settings/gpt-j-6b-c1-jax.settings\n",
+        "  ![[ -f /content/drive/MyDrive/KoboldAI/settings/gpt-neox-20b-jax.settings ]] || echo -e \"{\\n   \\\"apikey\\\": \\\"\\\",\\n   \\\"andepth\\\": 3,\\n   \\\"temp\\\": 0.5,\\n   \\\"top_p\\\": 0.9,\\n   \\\"top_k\\\": 0,\\n   \\\"tfs\\\": 1.0,\\n   \\\"rep_pen\\\": 1.03,\\n   \\\"genamt\\\": 80,\\n   \\\"max_length\\\": 2048,\\n   \\\"ikgen\\\": 200,\\n   \\\"formatoptns\\\": {\\n      \\\"frmttriminc\\\": true,\\n      \\\"frmtrmblln\\\": false,\\n      \\\"frmtrmspch\\\": false,\\n      \\\"frmtadsnsp\\\": false\\n   },\\n   \\\"numseqs\\\": 1,\\n   \\\"widepth\\\": 3,\\n   \\\"useprompt\\\": true,\\n   \\\"adventure\\\": false\\n}\" > /content/drive/MyDrive/KoboldAI/settings/gpt-neox-20b-jax.settings\n",
+        "elif Model == \"Skein 6B\":\n",
+        "  Model = \"KoboldAI/GPT-J-6B-Skein\"\n",
+        "  path = \"\"\n",
+        "  download = \"\"\n",
+        "elif Model == \"Janeway 6B\":\n",
+        "  Model = \"KoboldAI/GPT-J-6B-Janeway\"\n",
+        "  path = \"\"\n",
+        "  download = \"\"\n",
+        "elif Model == \"Adventure 6B\":\n",
+        "  Model = \"KoboldAI/GPT-J-6B-Adventure\"\n",
+        "  path = \"\"\n",
+        "  download = \"\"\n",
+        "elif Model == \"Lit 6B\":\n",
+        "  Model = \"hakurei/lit-6B\"\n",
+        "  path = \"\"\n",
+        "  download = \"\"\n",
+        "elif Model == \"Shinen 6B\":\n",
+        "  Model = \"KoboldAI/GPT-J-6B-Shinen\"\n",
+        "  path = \"\"\n",
+        "  download = \"\"\n",
+        "elif Model == \"Convo 6B\":\n",
+        "  Model = \"hitomi-team/convo-6B\"\n",
+        "  path = \"\"\n",
+        "  download = \"\"\n",
+        "elif Model == \"C1 6B\":\n",
+        "  Model = \"hakurei/c1-6B\"\n",
+        "  path = \"\"\n",
+        "  download = \"\"\n",
+        "else:\n",
+        "  path = \"\"\n",
+        "  download = \"\"\n",
         "\n",
-        "if Drive == \"Unextracted (Less Space)\":\n",
-        "  xloc = \"colab\"\n",
-        "if Drive == \"Extracted (Faster Loading)\":\n",
-        "  xloc = \"drive\"\n",
+        "if Provider == \"Localtunnel\":\n",
+        "  tunnel = \"--localtunnel yes\"\n",
+        "else:\n",
+        "  tunnel = \"\"\n",
         "\n",
-        "\n",
-        "!wget https://henk.tech/ckds -O - | bash /dev/stdin $download -l $location $extract -p $path -m TPUMeshTransformerGPTJ -g $Version -x $xloc"
+        "!wget https://henk.tech/ckds -O - | bash /dev/stdin $path$download -m $Model -g $Version $tunnel"
       ]
     },
     {
       "cell_type": "markdown",
       "source": [
-        "# TPU Edition Model Descriptions\n",
+        "## TPU Edition Model Descriptions\n",
         "\n",
-        "| Model                          | Size   | Style     | Drive Space | Description                                                  |\n",
-        "| ------------------------------ | ------ | --------- | ----------- | ------------------------------------------------------------ |\n",
-        "| Skein 6B by VE_FORBRYDERNE | 6B TPU | Hybrid    | 0 GB         | Skein is our flagship 6B model, it is a hybrid between a Adventure model and a Novel model. Best used with either Adventure mode or the You Bias userscript enabled. Skein has been trained on high quality Novels along with CYOA adventure stories and is not as wackey as the Adventure model. It also has tagging support. |\n",
-        "| Janeway 6B by Mr Seeker | 6B TPU | Novel | 0 GB | Janeway is a model created from Picard's dataset combined with a brand new collection of ebooks. This model is trained on 20% more content than Picard and has been trained on literature from various genres. Although the model is mainly focussed on SFW, romantic scenes might involve a degree of nudity. |\n",
-        "| Adventure 6B by VE_FORBRYDERNE | 6B TPU | Adventure | 0 GB         | Adventure is a 6B model designed to mimick the behavior of AI Dungeon. It is exclusively for Adventure Mode and can take you on the epic and wackey adventures that AI Dungeon players love. It also features the many tropes of AI Dungeon as it has been trained on very similar data. It must be used in second person (You). |\n",
-        "| Lit 6B by Haru                 | 6B TPU | NSFW      | 8 GB /  12 GB | Lit is a great NSFW model trained by Haru on both a large set of Literotica stories and high quality novels along with tagging support. Creating a high quality model for your NSFW stories. This model is exclusively a novel model and is best used in third person. |\n",
-        "| Shinen 6B by Mr Seeker | 6B TPU | NSFW | 0 GB | Shinen is an alternative to the Lit model designed to be more explicit. If Lit is to tame for you Shinen might produce better results. While it is a Novel model it is unsuitable for SFW stories due to its heavy NSFW bias. Shinen will not hold back. It is meant to be used in KoboldAI's regular mode. |\n",
-        "| Generic 6B by EleutherAI       | 6B TPU | Generic   | 10 GB / 12 GB | GPT-J-6B is what all other models are based on, if you need something that has no specific bias towards any particular subject this is the model for you. Best used when the other models are not suitable for what you wish to do. Such as homework assistance, blog writing, coding and more. It needs more hand holding than other models and is more prone to undesirable formatting changes. |\n",
-        "| C1 6B by Haru                  | 6B TPU | Chatbot   | 8 GB /  12 GB | C1 has been trained on various internet chatrooms, it makes the basis for an interesting chatbot model and has been optimized to be used in the Chatmode. |\n",
+        "| Model | Size | Style | Description |\n",
+        "| --- | --- | --- | --- |\n",
+        "| [Nerys](https://huggingface.co/KoboldAI/fairseq-dense-13B-Nerys) by Mr Seeker | 13B | Novel/Adventure | Nerys is a hybrid model based on Pike (A newer Janeway), on top of the Pike dataset you also get some Light Novels, Adventure mode support and a little bit of shinen thrown in the mix. The end result is a very diverse model that is heavily biased towards SFW novel writing, but one that can go beyond its novel training and make for an excellent adventure model to. Adventure mode is best played from a second person perspective, but can be played in first or third person as well. Novel writing can be done best from the first or third person. |\n",
+        "| [Janeway](https://huggingface.co/KoboldAI/fairseq-dense-13B-Janeway) by Mr Seeker | 13B | Novel | Janeway is a model created from Picard's dataset combined with a brand new collection of ebooks. This model is trained on 20% more content than Picard and has been trained on literature from various genres. Although the model is mainly focussed on SFW, romantic scenes might involve a degree of nudity. |\n",
+        "| [Shinen](https://huggingface.co/KoboldAI/fairseq-dense-13B-Shinen) by Mr Seeker | 13B | NSFW | Shinen is an NSFW model designed to be more explicit. Trained on a variety of stories from the website Sexstories it contains many different kinks. |\n",
+        "| [Skein](https://huggingface.co/KoboldAI/GPT-J-6B-Skein) by VE\\_FORBRYDERNE | 6B | Adventure | Skein is best used with Adventure mode enabled, it consists of a 4 times larger adventure dataset than the Adventure model making it excellent for text adventure gaming. On top of that it also consists of light novel training further expanding its knowledge and writing capabilities. It can be used with the You filter bias if you wish to write Novels with it, but dedicated Novel models can perform better for this task. |\n",
+        "| [Adventure](https://huggingface.co/KoboldAI/GPT-J-6B-Adventure) by VE\\_FORBRYDERNE | 6B | Adventure | Adventure is a 6B model designed to mimick the behavior of AI Dungeon. It is exclusively for Adventure Mode and can take you on the epic and wackey adventures that AI Dungeon players love. It also features the many tropes of AI Dungeon as it has been trained on very similar data. It must be used in second person (You). |\n",
+        "| [Lit](https://huggingface.co/hakurei/lit-6B) by Haru | 6B | NSFW | Lit is a great NSFW model trained by Haru on both a large set of Literotica stories and high quality novels along with tagging support. Creating a high quality model for your NSFW stories. This model is exclusively a novel model and is best used in third person. |\n",
+        "| [Convo](https://huggingface.co/hitomi-team/convo-6B) by Hitomi Team | 6B | Chatbot | Convo-6B is a GPT-J 6B model fine-tuned on a collection of high quality open source datasets which amount to 6 million messages. The primary goal of the model is to provide improved performance and generalization when generating multi-turn dialogue for characters that were not present from within the fine tuning data. The prompted performance has especially improved over the predecessor model [C1-6B](https://huggingface.co/hakurei/c1-6B). |\n",
+        "| [C1](https://huggingface.co/hakurei/c1-6B) by Haru | 6B | Chatbot | C1 has been trained on various internet chatrooms, it makes the basis for an interesting chatbot model and has been optimized to be used in the Chatmode. |\n",
+        "| Neo(X) by EleutherAI | 20B | Generic | NeoX is the largest EleutherAI model currently available, being a generic model it is not particularly trained towards anything and can do a variety of writing, Q&A and coding tasks. 20B's performance is closely compared to the 13B models and it is worth trying both especially if you have a task that does not involve english writing. Its behavior will be similar to the GPT-J-6B model since they are trained on the same dataset but with more sensitivity towards repetition penalty and with more knowledge. |\n",
+        "| [Fairseq Dense](https://huggingface.co/KoboldAI/fairseq-dense-13B) | 13B | Generic | Trained by Facebook Researchers this model stems from the MOE research project within Fairseq. This particular version has been converted by us for use in KoboldAI. It is known to be on par with the larger 20B model from EleutherAI and considered as better for pop culture and language tasks. Because the model has never seen a new line (enter) it may perform worse on formatting and paragraphing. |\n",
+        "| [GPT-J-6B](https://huggingface.co/EleutherAI/gpt-j-6B) by EleutherAI | 6B | Generic | This model serves as the basis for most other 6B models (Some being based on Fairseq Dense instead). Being trained on the Pile and not biased towards anything in particular it is suitable for a variety of tasks such as writing, Q&A and coding tasks. You will likely get better result with larger generic models or finetuned models. |\n",
         "\n",
         "\n",
         "# [GPU Edition Model Descriptions](https://colab.research.google.com/github/KoboldAI/KoboldAI-Client/blob/main/colab/GPU.ipynb)\n",
         "\n",
-        "| Model                                                        | Size     | Style      | Description                                                  |\n",
-        "| ------------------------------------------------------------ | -------- | ---------- | ------------------------------------------------------------ |\n",
-        "| [GPT-Neo-2.7B-Janeway](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Janeway) by Mr Seeker | 2.7B GPU | Novel | Janeway is a model created from Picard's dataset combined with a brand new collection of ebooks. This model is trained on 20% more content than Picard and has been trained on literature from various genres. Although the model is mainly focussed on SFW, romantic scenes might involve a degree of nudity. |\n",
-        "| [GPT-Neo-2.7B-Picard](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Picard) by Mr Seeker | 2.7B GPU | Novel      | Picard is a model trained for SFW Novels based on GPT-Neo-2.7B. It is focused on Novel style writing without the NSFW bias. While the name suggests a sci-fi model this model is designed for Novels of a variety of genre's. It is meant to be used in KoboldAI's regular mode. |\n",
-        "| [GPT-Neo-2.7B-AID](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-AID) by melastacho | 2.7B GPU | Adventure | Also know as Adventure 2.7B this is a clone of the AI Dungeon Classic model and is best known for the epic wackey adventures that AI Dungeon Classic players love. |\n",
-        "| [GPT-Neo-2.7B-Horni-LN](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Horni-LN) by finetune | 2.7B GPU | Novel | This model is based on GPT-Neo-2.7B-Horni and retains its NSFW knowledge, but was then further biased towards SFW novel stories. If you seek a balance between a SFW Novel model and a NSFW model this model should be a good choice. |\n",
-        "| [GPT-Neo-2.7B-Horni](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Horni) by finetune | 2.7B GPU | NSFW       | This model is tuned on Literotica to produce a Novel style model biased towards NSFW content. Can still be used for SFW stories but will have a bias towards NSFW content. It is meant to be used in KoboldAI's regular mode. |\n",
-        "| [GPT-Neo-2.7B-Shinen](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Shinen) by Mr Seeker | 2.7B GPU | NSFW       | Shinen is an alternative to the Horni model designed to be more explicit. If Horni is to tame for you shinen might produce better results. While it is a Novel model it is unsuitable for SFW stories due to its heavy NSFW bias. Shinen will not hold back. It is meant to be used in KoboldAI's regular mode. |\n",
-        "| [GPT-Neo-2.7B](https://huggingface.co/EleutherAI/gpt-neo-2.7B) by EleutherAI | 2.7B GPU    | Generic    | This is the base model for all the other 2.7B models, it is best used when you have a use case that we have no other models available for, such as writing blog articles or programming. It can also be a good basis for the experience of some of the softprompts if your softprompt is not about a subject the other models cover. |\n",
+        "| Model | Size | Style | Description |\n",
+        "| --- | --- | --- | --- |\n",
+        "| [Fairseq-Dense-2.7B-Nerys](https://huggingface.co/KoboldAI/fairseq-dense-2.7B-Nerys) by Mr Seeker | 2.7B | Novel/Adventure | Nerys is a hybrid model based on Pike (A newer Janeway), on top of the Pike dataset you also get some Light Novels, Adventure mode support and a little bit of shinen thrown in the mix. The end result is a very diverse model that is heavily biased towards SFW novel writing, but one that can go beyond its novel training and make for an excellent adventure model to. Adventure mode is best played from a second person perspective, but can be played in first or third person as well. Novel writing can be done best from the first or third person. |\n",
+        "| [GPT-Neo-2.7B-Janeway](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Janeway) by Mr Seeker | 2.7B | Novel | Janeway is a model created from Picard's dataset combined with a brand new collection of ebooks. This model is trained on 20% more content than Picard and has been trained on literature from various genres. Although the model is mainly focussed on SFW, romantic scenes might involve a degree of nudity. |\n",
+        "| [GPT-Neo-2.7B-Picard](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Picard) by Mr Seeker | 2.7B | Novel | Picard is a model trained for SFW Novels based on GPT-Neo-2.7B. It is focused on Novel style writing without the NSFW bias. While the name suggests a sci-fi model this model is designed for Novels of a variety of genre's. It is meant to be used in KoboldAI's regular mode. |\n",
+        "| [GPT-Neo-2.7B-AID](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-AID) by melastacho | 2.7B | Adventure | Also know as Adventure 2.7B this is a clone of the AI Dungeon Classic model and is best known for the epic wackey adventures that AI Dungeon Classic players love. |\n",
+        "| [GPT-Neo-2.7B-Horni-LN](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Horni-LN) by finetune | 2.7B | Novel | This model is based on GPT-Neo-2.7B-Horni and retains its NSFW knowledge, but was then further biased towards SFW novel stories. If you seek a balance between a SFW Novel model and a NSFW model this model should be a good choice. |\n",
+        "| [GPT-Neo-2.7B-Horni](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Horni) by finetune | 2.7B | NSFW | This model is tuned on Literotica to produce a Novel style model biased towards NSFW content. Can still be used for SFW stories but will have a bias towards NSFW content. It is meant to be used in KoboldAI's regular mode. |\n",
+        "| [GPT-Neo-2.7B-Shinen](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Shinen) by Mr Seeker | 2.7B | NSFW | Shinen is an alternative to the Horni model designed to be more explicit. If Horni is to tame for you shinen might produce better results. While it is a Novel model it is unsuitable for SFW stories due to its heavy NSFW bias. Shinen will not hold back. It is meant to be used in KoboldAI's regular mode. |\n",
+        "| [GPT-Neo-2.7B](https://huggingface.co/EleutherAI/gpt-neo-2.7B) by EleutherAI | 2.7B | Generic | This is the base model for all the other 2.7B models, it is best used when you have a use case that we have no other models available for, such as writing blog articles or programming. It can also be a good basis for the experience of some of the softprompts if your softprompt is not about a subject the other models cover. |\n",
         "\n",
-        "| Style     | Description                                                  |\n",
-        "| --------- | ------------------------------------------------------------ |\n",
-        "| Novel     | For regular story writing, not compatible with Adventure mode or other specialty modes. |\n",
-        "| NSFW      | Indicates that the model is strongly biased towards NSFW content and is not suitable for children, work environments or livestreaming. Most NSFW models are also Novel models in nature. |\n",
+        "| Style | Description |\n",
+        "| --- | --- |\n",
+        "| Novel | For regular story writing, not compatible with Adventure mode or other specialty modes. |\n",
+        "| NSFW | Indicates that the model is strongly biased towards NSFW content and is not suitable for children, work environments or livestreaming. Most NSFW models are also Novel models in nature. |\n",
         "| Adventure | These models are excellent for people willing to play KoboldAI like a Text Adventure game and are meant to be used with Adventure mode enabled. Even if you wish to use it as a Novel style model you should always have Adventure mode on and set it to story. These models typically have a strong bias towards the use of the word You and without Adventure mode enabled break the story flow and write actions on your behalf. |\n",
-        "| Chatbot   | These models are specifically trained for chatting and are best used with the Chatmode enabled. Typically trained on either public chatrooms or private chats. |\n",
-        "| Hybrid    | Hybrid models are a blend between different styles, for example they are trained on both Novel stories and Adventure stories. These models are great variety models that you can use for multiple different playstyles and modes, but depending on your usage you may need to enable Adventure Mode or the You bias (in userscripts). |\n",
-        "| Generic   | Generic models are not trained towards anything specific, typically used as a basis for other tasks and models. They can do everything the other models can do, but require much more handholding to work properly. Generic models are an ideal basis for tasks that we have no specific model for, or for experiencing a softprompt in its raw form. |\n",
+        "| Chatbot | These models are specifically trained for chatting and are best used with the Chatmode enabled. Typically trained on either public chatrooms or private chats. |\n",
+        "| Generic | Generic models are not trained towards anything specific, typically used as a basis for other tasks and models. They can do everything the other models can do, but require much more handholding to work properly. Generic models are an ideal basis for tasks that we have no specific model for, or for experiencing a softprompt in its raw form. |\n",
         "\n",
-        "## How to start KoboldAI in 7 simple steps\n",
-        "Using KoboldAI on Google Colab is easy! Simply follow these steps to get started:\n",
-        "1. Mobile phone? Tap the play button below next to \"<--- Tap this if you play on mobile\" to reveal an audio player, play the silent audio to keep the tab alive so Google will not shut you down when your using KoboldAI. If no audio player is revealed your phone browser does not support Google Colab in the mobile view, go to your browser menu and enable Desktop mode before you continue.\n",
-        "2. Select the model that most describes what you would like to do, by default we have the most recommended model for people willing to try out KoboldAI selected.\n",
-        "3. Click the play button next to \"<--- Click this to start KoboldAI\".\n",
-        "4. Allow Google Drive access, this typically happens trough a popup but sometimes Google Drive access may be requested trough the older method by asking you to click on a link and copy a code. This is normal behavior for Colab and only you will get access to your files, nothing is shared with us.\n",
-        "5. Now the automatic installation and Download process starts, for most models in the TPU edition expect the loading to take between 15 and 30 minutes on average depending on the current Colab download speeds and the model you selected. These downloads happen trough Google's internet connection, you will not be billed by your internet provider and it will not count towards any download limits.\n",
-        "6. After waiting a Trycloudflare link appears, click the link to enjoy KoboldAI. If you get a 1033 error Cloudflare is not done loading, in that case keep refreshing until it goes away. (If it keeps happening after 2 minutes Cloudflare has an issue, in that case you can use Runtime -> Restart and Run All to get a new link).\n",
-        "7. As you play KoboldAI, keep this Colab tab open in the background and check occationally for Captcha's so they do not shut your instance down. If you do get shut down you can always download a copy of your gamesave in the Save menu inside KoboldAI. Stories are never lost as long as you keep KoboldAI open in your browser.\n",
-        "\n",
-        "Get a error message saying you do not have access to a GPU/TPU instance? Do not continue and try again later, KoboldAI will not run correctly without them.\n",
-        "\n"
+        "---\n",
+        "## Tips to get the most out of Google Colab\n",
+        "-  Google will occationally show a Captcha, typically after it has been open for 30 minutes but it can be more frequent if you often use Colab. Make sure to do these properly, or you risk getting your instance shut down and getting a lower priority towards the TPU's.\n",
+        "- KoboldAI uses Google Drive to store your files and settings, if you wish to upload a softprompt or userscript this can be done directly on the Google Drive website. You can also use this to download backups of your KoboldAI related files or upload models of your own.\n",
+        "- Don't want to save your stories on Google Drive for privacy reasons? Do not use KoboldAI's save function and instead click Download as .json, this will automatically download the story to your own computer without ever touching Google's harddrives. You can load this back trough the Load from file option.\n",
+        "- Google shut your instance down unexpectedly? You can still make use of the Download as .json button to recover your story as long as you did not close the KoboldAI window. You can then load this back up in your next session.\n",
+        "- Done with KoboldAI? Go to the Runtime menu, click on Manage Sessions and terminate your open sessions that you no longer need. This trick can help you maintain higher priority towards getting a TPU.\n",
+        "- Models stored on Google Drive typically load faster than models we need to download from the internet."
       ],
       "metadata": {
         "id": "i0-9ARA3c4Fx"

colab/vscode.ipynb

@@ -67,6 +67,7 @@
         "!wget https://henk.tech/ckds -O - | bash /dev/stdin -g $Version -i only $Args\n",
         "\n",
         "!pip install colabcode\n",
+        "!pip install 'flask>=2.1.0'\n",
         "from colabcode import ColabCode\n",
         "ColabCode(authtoken=Authtoken)"
       ]

colabkobold.sh

@@ -2,7 +2,7 @@
 # KoboldAI Easy Colab Deployment Script by Henk717
 
 # read the options
-TEMP=`getopt -o m:i:p:c:d:x:a:l:z:g:t:n:b: --long model:,init:,path:,configname:,download:,aria2:,dloc:xloc:7z:git:tar:ngrok:branch: -- "$@"`
+TEMP=`getopt -o m:i:p:c:d:x:a:l:z:g:t:n:b:s: --long model:,init:,path:,configname:,download:,aria2:,dloc:,xloc:,7z:,git:,tar:,ngrok:,branch:,savemodel:,localtunnel:,lt: -- "$@"`
 eval set -- "$TEMP"
 
 # extract options and their arguments into variables.
@@ -17,7 +17,9 @@ while true ; do
         -c|--configname)
             configname=" --configname $2" ; shift 2 ;;
         -n|--ngrok)
-            configname=" --ngrok" ; shift 2 ;;
+            ngrok=" --ngrok" ; shift 2 ;;
+        --lt|--localtunnel)
+            localtunnel=" --localtunnel" ; shift 2 ;;
         -d|--download)
             download="$2" ; shift 2 ;;
         -a|--aria2)
@@ -34,6 +36,8 @@ while true ; do
             git="$2" ; shift 2 ;;
         -b|--branch)
             branch="$2" ; shift 2 ;;
+        -s|--savemodel)
+            savemodel=" --savemodel" ; shift 2 ;;
         --) shift ; break ;;
         *) echo "Internal error!" ; exit 1 ;;
     esac
@@ -48,8 +52,8 @@ function launch
         exit 0
     else
     cd /content/KoboldAI-Client
-    echo "Launching KoboldAI with the following options : python3 aiserver.py$model$kmpath$configname$ngrok --remote --override_delete --override_rename"
-    python3 aiserver.py$model$kmpath$configname$ngrok --colab
+    echo "Launching KoboldAI with the following options : python3 aiserver.py$model$kmpath$configname$ngrok$localtunnel$savemodel --colab"
+    python3 aiserver.py$model$kmpath$configname$ngrok$localtunnel$savemodel --colab
     exit
     fi
 }
@@ -134,28 +138,32 @@ if [ "$init" != "skip" ]; then
 
     cd /content/KoboldAI-Client
 
-    cp -rn stories/*.* /content/drive/MyDrive/KoboldAI/stories/
-    cp -rn userscripts/*.* /content/drive/MyDrive/KoboldAI/userscripts/
-    cp -rn softprompts/*.* /content/drive/MyDrive/KoboldAI/softprompts/
+    cp -rn stories/* /content/drive/MyDrive/KoboldAI/stories/
+    cp -rn userscripts/* /content/drive/MyDrive/KoboldAI/userscripts/
+    cp -rn softprompts/* /content/drive/MyDrive/KoboldAI/softprompts/
     rm stories
     rm -rf stories/
     rm userscripts
     rm -rf userscripts/
     rm softprompts
     rm -rf softprompts/
+    rm models
+    rm -rf models/
     ln -s /content/drive/MyDrive/KoboldAI/stories/ stories
     ln -s /content/drive/MyDrive/KoboldAI/settings/ settings
     ln -s /content/drive/MyDrive/KoboldAI/softprompts/ softprompts
     ln -s /content/drive/MyDrive/KoboldAI/userscripts/ userscripts
+    ln -s /content/drive/MyDrive/KoboldAI/models/ models
 
-    if [ "$model" == " --model TPUMeshTransformerGPTJ" ]; then
+    if [ -n "${COLAB_TPU_ADDR+set}" ]; then
         pip install -r requirements_mtj.txt
     else
         pip install -r requirements.txt
     fi
     
-    # Make sure Colab has netbase
-    sudo apt install netbase -y
+    # Make sure Colab has the system dependencies
+    sudo apt install netbase aria2 -y
+    npm install -g localtunnel
 fi
 
 cd /content
@@ -178,8 +186,7 @@ fi
 
 #Download routine for Aria2c scripts
 if [ ! -z ${aria2+x} ]; then
-    apt install aria2 -y
-    curl -L $aria2 | aria2c -c -i- -d$dloc --user-agent=KoboldAI --file-allocation=none
+    curl -L $aria2 | aria2c -x 10 -s 10 -j 10 -c -i- -d$dloc --user-agent=KoboldAI --file-allocation=none
 fi
 
 #Extract the model with 7z

commandline-rocm.sh

@@ -0,0 +1 @@
+bin/micromamba run -r runtime -n koboldai-rocm bash

commandline.bat

@@ -10,18 +10,18 @@ IF %M%==3 GOTO drivemap_B
 SET TEMP=%~DP0MINICONDA3
 SET TMP=%~DP0MINICONDA3
 call miniconda3\condabin\activate
-cmd /k
+cmd /k "%*"
 
 :drivemap
 subst K: miniconda3 >nul
 SET TEMP=K:\
 SET TMP=K:\
 call K:\python\condabin\activate
-cmd /k
+cmd /k "%*"
 
 :drivemap_B
 subst B: miniconda3 >nul
 SET TEMP=B:\
 SET TMP=B:\
 call B:\python\condabin\activate
-cmd /k
+cmd /k "%*"

commandline.sh

@@ -0,0 +1 @@
+bin/micromamba run -r runtime -n koboldai bash

disconnect-kobold-drive.bat

@@ -0,0 +1,7 @@
+@echo off
+SET /P M=<loader.settings
+IF %M%==3 subst /D B:
+IF %M%==1 subst /D K:
+cls
+echo KoboldAI Drive disconnected
+pause

docker-cuda/Dockerfile

@@ -6,4 +6,4 @@ WORKDIR /content/
 COPY env.yml /home/micromamba/env.yml
 RUN micromamba install -y -n base -f /home/micromamba/env.yml
 USER root
-RUN apt update && apt install xorg -y
+RUN apt update && apt install xorg -y

docker-rocm.sh

@@ -0,0 +1,4 @@
+cd docker-rocm
+xhost +local:docker
+cp ../environments/rocm.yml env.yml
+docker-compose run --service-ports koboldai bash -c "cd /content && python3 aiserver.py $*"

environments/finetuneanon.yml

@@ -14,7 +14,7 @@ dependencies:
   - markdown
   - bleach=4.1.0
   - pip
-  - git
+  - git=2.35.1
   - pip:
     - git+https://github.com/finetuneanon/transformers@gpt-neo-localattention3-rp-b
     - flask-cloudflared

environments/huggingface.yml

@@ -6,16 +6,18 @@ channels:
 dependencies:
   - colorama
   - flask-socketio
-  - pytorch
+  - pytorch=1.11.*
   - python=3.8.*
   - cudatoolkit=11.1
-  - transformers
   - eventlet
   - markdown
   - bleach=4.1.0
   - pip
-  - git
+  - git=2.35.1
+  - sentencepiece
+  - protobuf
   - pip:
     - flask-cloudflared
     - flask-ngrok
     - lupa==1.10
+    - transformers>=4.17

environments/rocm-finetune.yml

@@ -10,7 +10,7 @@ dependencies:
   - markdown
   - bleach=4.1.0
   - pip
-  - git
+  - git=2.35.1
   - pip:
     - --find-links https://download.pytorch.org/whl/rocm4.2/torch_stable.html
     - torch

environments/rocm.yml

@@ -3,7 +3,6 @@ channels:
   - conda-forge
   - defaults
 dependencies:
-  - transformers
   - colorama
   - flask-socketio
   - python=3.8.*
@@ -11,11 +10,14 @@ dependencies:
   - markdown
   - bleach=4.1.0
   - pip
-  - git
+  - git=2.35.1
+  - sentencepiece
+  - protobuf
   - pip:
     - --find-links https://download.pytorch.org/whl/rocm4.2/torch_stable.html
-    - torch
-    - torchvision==0.11.1
+    - torch==1.10.*
+    - torchvision
     - flask-cloudflared
     - flask-ngrok
     - lupa==1.10
+    - transformers>=4.17

fileops.py

@@ -1,5 +1,3 @@
-import tkinter as tk
-from tkinter import filedialog
 from os import getcwd, listdir, path
 from typing import Tuple, Union, Optional
 import os
@@ -10,6 +8,8 @@ import zipfile
 #  Generic Method for prompting for file path
 #==================================================================#
 def getsavepath(dir, title, types):
+    import tkinter as tk
+    from tkinter import filedialog
     root = tk.Tk()
     root.attributes("-topmost", True)
     path = tk.filedialog.asksaveasfile(
@@ -28,6 +28,8 @@ def getsavepath(dir, title, types):
 #  Generic Method for prompting for file path
 #==================================================================#
 def getloadpath(dir, title, types):
+    import tkinter as tk
+    from tkinter import filedialog
     root = tk.Tk()
     root.attributes("-topmost", True)
     path = tk.filedialog.askopenfilename(
@@ -45,6 +47,8 @@ def getloadpath(dir, title, types):
 #  Generic Method for prompting for directory path
 #==================================================================#
 def getdirpath(dir, title):
+    import tkinter as tk
+    from tkinter import filedialog
     root = tk.Tk()
     root.attributes("-topmost", True)
     path = filedialog.askdirectory(
@@ -61,30 +65,30 @@ def getdirpath(dir, title):
 #  Returns the path (as a string) to the given story by its name
 #==================================================================#
 def storypath(name):
-    return path.join(path.dirname(path.realpath(__file__)), "stories", name + ".json")
+    return path.join("stories", name + ".json")
 
 #==================================================================#
 #  Returns the path (as a string) to the given soft prompt by its filename
 #==================================================================#
 def sppath(filename):
-    return path.join(path.dirname(path.realpath(__file__)), "softprompts", filename)
+    return path.join("softprompts", filename)
 
 #==================================================================#
 #  Returns the path (as a string) to the given username by its filename
 #==================================================================#
 def uspath(filename):
-    return path.join(path.dirname(path.realpath(__file__)), "userscripts", filename)
+    return path.join("userscripts", filename)
 
 #==================================================================#
 #  Returns an array of dicts containing story files in /stories
 #==================================================================#
 def getstoryfiles():
     list = []
-    for file in listdir(path.dirname(path.realpath(__file__))+"/stories"):
+    for file in listdir("stories"):
         if file.endswith(".json"):
             ob = {}
             ob["name"] = file.replace(".json", "")
-            f = open(path.dirname(path.realpath(__file__))+"/stories/"+file, "r")
+            f = open("stories/"+file, "r")
             try:
                 js = json.load(f)
             except:
@@ -108,7 +112,7 @@ def checksp(filename: str, model_dimension: int) -> Tuple[Union[zipfile.ZipFile,
     if 'np' not in globals():
         import numpy as np
     try:
-        z = zipfile.ZipFile(path.dirname(path.realpath(__file__))+"/softprompts/"+filename)
+        z = zipfile.ZipFile("softprompts/"+filename)
         with z.open('tensor.npy') as f:
             # Read only the header of the npy file, for efficiency reasons
             version: Tuple[int, int] = np.lib.format.read_magic(f)
@@ -118,7 +122,10 @@ def checksp(filename: str, model_dimension: int) -> Tuple[Union[zipfile.ZipFile,
             shape, fortran_order, dtype = np.lib.format._read_array_header(f, version)
             assert len(shape) == 2
     except:
-        z.close()
+        try:
+            z.close()
+        except UnboundLocalError:
+            pass
         return 1, None, None, None, None
     if dtype not in ('V2', np.float16, np.float32):
         z.close()
@@ -136,8 +143,8 @@ def checksp(filename: str, model_dimension: int) -> Tuple[Union[zipfile.ZipFile,
 #==================================================================#
 def getspfiles(model_dimension: int):
     lst = []
-    os.makedirs(path.dirname(path.realpath(__file__))+"/softprompts", exist_ok=True)
-    for file in listdir(path.dirname(path.realpath(__file__))+"/softprompts"):
+    os.makedirs("softprompts", exist_ok=True)
+    for file in listdir("softprompts"):
         if not file.endswith(".zip"):
             continue
         z, version, shape, fortran_order, dtype = checksp(file, model_dimension)
@@ -170,8 +177,8 @@ def getspfiles(model_dimension: int):
 #==================================================================#
 def getusfiles(long_desc=False):
     lst = []
-    os.makedirs(path.dirname(path.realpath(__file__))+"/userscripts", exist_ok=True)
-    for file in listdir(path.dirname(path.realpath(__file__))+"/userscripts"):
+    os.makedirs("userscripts", exist_ok=True)
+    for file in listdir("userscripts"):
         if file.endswith(".lua"):
             ob = {}
             ob["filename"] = file

gensettings.py

@@ -51,8 +51,19 @@ gensettingstf = [
 	"min": 0.0,
 	"max": 1.0,
 	"step": 0.05,
-	"default": 0.0,
+	"default": 1.0,
     "tooltip": "Alternative sampling method; it is recommended to disable top_p and top_k (set top_p to 1 and top_k to 0) if using this. 0.95 is thought to be a good value. (Put this value on 1 to disable its effect)"
+	},
+	{
+	"uitype": "slider",
+	"unit": "float",
+	"label": "Typical Sampling",
+	"id": "settypical", 
+	"min": 0.0,
+	"max": 1.0,
+	"step": 0.05,
+	"default": 1.0,
+    "tooltip": "Alternative sampling method described in the paper \"Typical Decoding for Natural Language Generation\" (10.48550/ARXIV.2202.00666). The paper suggests 0.2 as a good value for this setting. Set this setting to 1 to disable its effect."
 	},
 	{
 	"uitype": "slider",
@@ -207,6 +218,17 @@ gensettingstf = [
 	"step": 1,
 	"default": 0,
   "tooltip": "Disables userscript generation modifiers."
+	},
+    {
+	"uitype": "toggle",
+	"unit": "bool",
+	"label": "Debug",
+	"id": "debug",
+	"min": 0,
+	"max": 1,
+	"step": 1,
+	"default": 0,
+  "tooltip": "Show debug info"
 	}
 ]
 
@@ -341,6 +363,17 @@ gensettingsik =[{
 	"step": 1,
 	"default": 0,
     "tooltip": "When enabled, the Memory text box in the Random Story dialog will be prefilled by default with your current story's memory instead of being empty."
+	},
+    {
+	"uitype": "toggle",
+	"unit": "bool",
+	"label": "Debug",
+	"id": "debug",
+	"min": 0,
+	"max": 1,
+	"step": 1,
+	"default": 0,
+  "tooltip": "Show debug info"
 	}
 ]
 

install_requirements.bat

@@ -45,6 +45,8 @@ subst B: miniconda3
 SET TEMP=B:\
 SET TMP=B:\
 copy umamba.exe B:\umamba.exe
+copy loader.settings B:\loader.settings
+copy disconnect-kobold-drive.bat B:\disconnect-kobold-drive.bat
 B:
 umamba.exe create -r B:\python\ -n base
 IF %B%==1 umamba.exe install --no-shortcuts -r B:\python\ -n base -f "%~dp0\environments\huggingface.yml" -y --always-copy

install_requirements.sh

@@ -0,0 +1,16 @@
+#!/bin/bash
+if [[ $1 = "cuda" ]]; then
+wget -qO- https://micromamba.snakepit.net/api/micromamba/linux-64/latest | tar -xvj bin/micromamba
+bin/micromamba create -f environments/huggingface.yml -r runtime -n koboldai -y
+# Weird micromamba bug causes it to fail the first time, running it twice just to be safe, the second time is much faster
+bin/micromamba create -f environments/huggingface.yml -r runtime -n koboldai -y
+exit
+fi
+if [[ $1 = "rocm" ]]; then
+wget -qO- https://micromamba.snakepit.net/api/micromamba/linux-64/latest | tar -xvj bin/micromamba
+bin/micromamba create -f environments/rocm.yml -r runtime -n koboldai-rocm -y
+# Weird micromamba bug causes it to fail the first time, running it twice just to be safe, the second time is much faster
+bin/micromamba create -f environments/rocm.yml -r runtime -n koboldai-rocm -y
+exit
+fi
+echo Please specify either CUDA or ROCM

maps/gpt_neo.json

@@ -0,0 +1,32 @@
+{
+  "mtj_compat": "neo",
+  "mtj_pe": "fixed",
+  "mtj_config_map": {
+    "d_model": "hidden_size",
+    "n_heads": "num_heads",
+    "layers": "num_layers"
+  },
+  "static_weights": {
+    "transformer.wte.weight": {"mtj": {"module": "embedding_shard/~/linear", "param": "w", "transforms": ["no_transpose", "vocab_pad"]}},
+    "transformer.wpe.weight": {"mtj": {"module": "embedding_shard", "param": "pos_embs", "transforms": ["no_transpose"]}},
+    "transformer.ln_f.weight": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "scale"}},
+    "transformer.ln_f.bias": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "offset"}}
+  },
+  "layer_weights": {
+    "transformer.h.{layer}.attn.attention.bias": {},
+    "transformer.h.{layer}.attn.attention.masked_bias": {},
+    "transformer.h.{layer}.attn.attention.q_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear", "param": "w"}},
+    "transformer.h.{layer}.attn.attention.v_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_1", "param": "w"}},
+    "transformer.h.{layer}.attn.attention.k_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_2", "param": "w"}},
+    "transformer.h.{layer}.attn.attention.out_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "w"}},
+    "transformer.h.{layer}.attn.attention.out_proj.bias": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "b", "transforms": ["divide_by_shards"]}},
+    "transformer.h.{layer}.mlp.c_fc.weight": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "w"}},
+    "transformer.h.{layer}.mlp.c_fc.bias": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "b"}},
+    "transformer.h.{layer}.mlp.c_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "w"}},
+    "transformer.h.{layer}.mlp.c_proj.bias": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "b", "transforms": ["divide_by_shards"]}},
+    "transformer.h.{layer}.ln_1.weight": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "scale"}},
+    "transformer.h.{layer}.ln_1.bias": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "offset"}},
+    "transformer.h.{layer}.ln_2.weight": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm_1", "param": "scale"}},
+    "transformer.h.{layer}.ln_2.bias": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm_1", "param": "offset"}}
+  }
+}

maps/gptj.json

@@ -0,0 +1,32 @@
+{
+  "mtj_compat": "j",
+  "mtj_pe": "rotary",
+  "mtj_config_map": {
+    "pe_rotary_dims": ["rotary_dim", 64],
+    "d_model": "n_embd",
+    "n_heads": "n_head",
+    "layers": "n_layer"
+  },
+  "static_weights": {
+    "transformer.wte.weight": {"mtj": {"module": "embedding_shard/~/linear", "param": "w", "transforms": ["no_transpose", "vocab_pad"]}},
+    "transformer.wte.bias": {"mtj": {"module": "embedding_shard/~/linear", "param": "b"}},
+    "transformer.ln_f.weight": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "scale"}},
+    "transformer.ln_f.bias": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "offset"}},
+    "lm_head.weight": {"mtj": {"module": "projection_shard/~/linear", "param": "w", "transforms": ["vocab_pad"]}},
+    "lm_head.bias": {"mtj": {"module": "projection_shard/~/linear", "param": "b"}}
+  },
+  "layer_weights": {
+    "transformer.h.{layer}.attn.bias": {},
+    "transformer.h.{layer}.attn.masked_bias": {},
+    "transformer.h.{layer}.attn.q_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear", "param": "w"}},
+    "transformer.h.{layer}.attn.v_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_1", "param": "w"}},
+    "transformer.h.{layer}.attn.k_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_2", "param": "w"}},
+    "transformer.h.{layer}.attn.out_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "w"}},
+    "transformer.h.{layer}.mlp.fc_in.weight": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "w"}},
+    "transformer.h.{layer}.mlp.fc_in.bias": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "b"}},
+    "transformer.h.{layer}.mlp.fc_out.weight": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "w"}},
+    "transformer.h.{layer}.mlp.fc_out.bias": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "b", "transforms": ["divide_by_shards"]}},
+    "transformer.h.{layer}.ln_1.weight": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "scale"}},
+    "transformer.h.{layer}.ln_1.bias": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "offset"}}
+  }
+}

maps/opt.json

@@ -0,0 +1,35 @@
+{
+  "mtj_compat": "opt",
+  "mtj_pe": "fixed",
+  "mtj_config_map": {
+    "do_layer_norm_before": ["do_layer_norm_before", true],
+    "d_embed": "word_embed_proj_dim",
+    "d_model": "hidden_size",
+    "n_heads": "num_attention_heads",
+    "layers": "num_hidden_layers"
+  },
+  "static_weights": {
+    "decoder.embed_tokens.weight": {"mtj": {"module": "embedding_shard/~/linear", "param": "w", "transforms": ["no_transpose", "vocab_pad"]}},
+    "decoder.project_in.weight": {"mtj": {"module": "embedding_shard", "param": "project_in"}},
+    "decoder.embed_positions.weight": {"mtj": {"module": "embedding_shard", "param": "pos_embs", "transforms": ["no_transpose", "remove_first_two_rows"]}},
+    "decoder.project_out.weight": {"mtj": {"module": "projection_shard", "param": "project_out"}}
+  },
+  "layer_weights": {
+    "decoder.layers.{layer}.self_attn.q_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear", "param": "w"}},
+    "decoder.layers.{layer}.self_attn.q_proj.bias": {"mtj": {"module": "layer_{layer}/~/linear", "param": "b"}},
+    "decoder.layers.{layer}.self_attn.v_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_1", "param": "w"}},
+    "decoder.layers.{layer}.self_attn.v_proj.bias": {"mtj": {"module": "layer_{layer}/~/linear_1", "param": "b"}},
+    "decoder.layers.{layer}.self_attn.k_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_2", "param": "w"}},
+    "decoder.layers.{layer}.self_attn.k_proj.bias": {"mtj": {"module": "layer_{layer}/~/linear_2", "param": "b"}},
+    "decoder.layers.{layer}.self_attn.out_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "w"}},
+    "decoder.layers.{layer}.self_attn.out_proj.bias": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "b", "transforms": ["divide_by_shards"]}},
+    "decoder.layers.{layer}.fc1.weight": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "w"}},
+    "decoder.layers.{layer}.fc1.bias": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "b"}},
+    "decoder.layers.{layer}.fc2.weight": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "w"}},
+    "decoder.layers.{layer}.fc2.bias": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "b", "transforms": ["divide_by_shards"]}},
+    "decoder.layers.{layer}.self_attn_layer_norm.weight": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "scale"}},
+    "decoder.layers.{layer}.self_attn_layer_norm.bias": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "offset"}},
+    "decoder.layers.{layer}.final_layer_norm.weight": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm_1", "param": "scale"}},
+    "decoder.layers.{layer}.final_layer_norm.bias": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm_1", "param": "offset"}}
+  }
+}

maps/xglm.json

@@ -0,0 +1,32 @@
+{
+  "mtj_compat": "fairseq_lm",
+  "mtj_pe": "fairseq_sinusoidal",
+  "mtj_config_map": {
+    "d_model": "d_model",
+    "n_heads": "attention_heads",
+    "layers": "num_layers"
+  },
+  "static_weights": {
+    "model.embed_tokens.weight": {"mtj": {"module": "embedding_shard/~/linear", "param": "w", "transforms": ["no_transpose", "vocab_pad"]}},
+    "model.layer_norm.weight": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "scale"}},
+    "model.layer_norm.bias": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "offset"}}
+  },
+  "layer_weights": {
+    "model.layers.{layer}.self_attn.q_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear", "param": "w"}},
+    "model.layers.{layer}.self_attn.q_proj.bias": {"mtj": {"module": "layer_{layer}/~/linear", "param": "b"}},
+    "model.layers.{layer}.self_attn.v_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_1", "param": "w"}},
+    "model.layers.{layer}.self_attn.v_proj.bias": {"mtj": {"module": "layer_{layer}/~/linear_1", "param": "b"}},
+    "model.layers.{layer}.self_attn.k_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_2", "param": "w"}},
+    "model.layers.{layer}.self_attn.k_proj.bias": {"mtj": {"module": "layer_{layer}/~/linear_2", "param": "b"}},
+    "model.layers.{layer}.self_attn.out_proj.weight": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "w"}},
+    "model.layers.{layer}.self_attn.out_proj.bias": {"mtj": {"module": "layer_{layer}/~/linear_3", "param": "b", "transforms": ["divide_by_shards"]}},
+    "model.layers.{layer}.fc1.weight": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "w"}},
+    "model.layers.{layer}.fc1.bias": {"mtj": {"module": "layer_{layer}/~/linear_4", "param": "b"}},
+    "model.layers.{layer}.fc2.weight": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "w"}},
+    "model.layers.{layer}.fc2.bias": {"mtj": {"module": "layer_{layer}/~/linear_5", "param": "b", "transforms": ["divide_by_shards"]}},
+    "model.layers.{layer}.self_attn_layer_norm.weight": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "scale"}},
+    "model.layers.{layer}.self_attn_layer_norm.bias": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm", "param": "offset"}},
+    "model.layers.{layer}.final_layer_norm.weight": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm_1", "param": "scale"}},
+    "model.layers.{layer}.final_layer_norm.bias": {"mtj": {"module": "layer_{layer}/~/replicated_layer_norm_1", "param": "offset"}}
+  }
+}

models/models go here.txt

@@ -0,0 +1,2 @@
+Place the models extracted in their own subfolder.
+Downloaded models from the menu will automatically appear here.

play-rocm.sh

@@ -1,4 +1,5 @@
-cd docker-rocm
-xhost +local:docker
-cp ../environments/rocm.yml env.yml
-docker-compose run --service-ports koboldai bash -c "cd /content && python3 aiserver.py $*"
+#!/bin/bash
+if [ ! -f "runtime/envs/koboldai-rocm/bin/python" ]; then
+./install_requirements.sh rocm
+fi
+bin/micromamba run -r runtime -n koboldai-rocm python aiserver.py $*

play.bat

@@ -16,20 +16,20 @@ cmd /k
 
 :drivemap
 ECHO Runtime launching in K: drive mode
+subst /D K: >nul
 subst K: miniconda3 >nul
 SET TEMP=K:\
 SET TMP=K:\
 call K:\python\condabin\activate
 python aiserver.py %*
-subst K: /D
 cmd /k
 
 :drivemap_B
 ECHO Runtime launching in B: drive mode
+subst /D B: >nul
 subst B: miniconda3 >nul
 SET TEMP=B:\
 SET TMP=B:\
 call B:\python\condabin\activate
 python aiserver.py %*
-subst B: /D
 cmd /k

play.ipynb

@@ -1,45 +0,0 @@
-{
-  "nbformat": 4,
-  "nbformat_minor": 0,
-  "metadata": {
-    "colab": {
-      "name": "KoboldAI Jupyter",
-      "provenance": [],
-      "authorship_tag": "ABX9TyMDTbAhtDnKJa+aIEaQjpsL"
-    },
-    "kernelspec": {
-      "name": "python3",
-      "display_name": "Python 3"
-    },
-    "language_info": {
-      "name": "python"
-    },
-    "accelerator": "TPU"
-  },
-  "cells": [
-    {
-      "cell_type": "markdown",
-      "source": [
-        "# KoboldAI Launcher for generic Jupyter Notebooks\n",
-        "This notebook is meant as a way to easily launch KoboldAI on existing Jupyter instances that already have KoboldAI installed (For example a custom Saturn Cloud or Paperspace instance).\n",
-        "\n",
-        "For Google Colab please check out our Google Colab edition available at : https://colab.research.google.com/github/KoboldAI/KoboldAI-Client/blob/main/colab/TPU.ipynb"
-      ],
-      "metadata": {
-        "id": "hMRnGz42Xsy3"
-      }
-    },
-    {
-      "cell_type": "code",
-      "execution_count": null,
-      "metadata": {
-        "id": "40B1QvI3Xv02"
-      },
-      "outputs": [],
-      "source": [
-        "!pip install -r requirements.txt\n",
-        "!python3 aiserver.py --remote"
-      ]
-    }
-  ]
-}

play.sh

@@ -0,0 +1,5 @@
+#!/bin/bash
+if [ ! -f "runtime/envs/koboldai/bin/python" ]; then
+./install_requirements.sh cuda
+fi
+bin/micromamba run -r runtime -n koboldai python aiserver.py $*

readme.md

@@ -1,4 +1,4 @@
-# KoboldAI - Your gateway to GPT writing
+## KoboldAI - Your gateway to GPT writing
 
 This is a browser-based front-end for AI-assisted writing with multiple local & remote AI models. It offers the standard array of tools, including Memory, Author's Note, World Info, Save & Load, adjustable AI settings, formatting options, and the ability to import existing AI Dungeon adventures. You can also turn on Adventure mode and play the game like AI Dungeon Unleashed.
 
@@ -12,9 +12,9 @@ By default KoboldAI will run in a generic mode optimized for writing, but with t
 
 The gameplay will be slightly different than the gameplay in AI Dungeon because we adopted the Type of the Unleashed fork, giving you full control over all the characters because we do not automatically adapt your sentences behind the scenes. This means you can more reliably control characters that are not you.
 
-As a result of this what you need to type is slightly different, in AI Dungeon you would type ***take the sword*** while in KoboldAI you would type it like a sentence such as ***You take the sword*** and this is best done with the word You instead of I.
+As a result of this what you need to type is slightly different, in AI Dungeon you would type _**take the sword**_ while in KoboldAI you would type it like a sentence such as _**You take the sword**_ and this is best done with the word You instead of I.
 
-To speak simply type : *You say "We should probably gather some supplies first"*
+To speak simply type : _You say "We should probably gather some supplies first"_  
 Just typing the quote might work, but the AI is at its best when you specify who does what in your commands.
 
 If you want to do this with your friends we advise using the main character as You and using the other characters by their name if you are playing on a model trained for Adventures. These models assume there is a You in the story. This mode does usually not perform well on Novel models because they do not know how to handle the input those are best used with regular story writing where you take turns with the AI.
@@ -27,7 +27,7 @@ If you want to use KoboldAI as a writing assistant this is best done in the regu
 
 In chatbot mode you can use a suitable model as a chatbot, this mode automatically adds your name to the beginning of the sentences and prevents the AI from talking as you. To use it properly you must write your story opening as both characters in the following format (You can use your own text) :
 
-``` ChatBot Opening Example
+```plaintext
 Bot : Hey!
 You : Hey Boyname, how have you been?
 Bot : Been good! How about you?
@@ -42,8 +42,6 @@ This mode works the best on either a Generic model or a chatbot model specifical
 
 Novel or Adventure models are not recommended for this feature but might still work but can derail away from the conversation format quickly.
 
-
-
 ## Play KoboldAI online for free on Google Colab (The easiest way to play)
 
 If you would like to play KoboldAI online for free on a powerful computer you can use Google Colaboraty. We provide two editions, a TPU and a GPU edition with a variety of models available. These run entirely on Google's Servers and will automatically upload saves to your Google Drive if you choose to save a story (Alternatively, you can choose to download your save instead so that it never gets stored on Google Drive). Detailed instructions on how to use them are at the bottom of the Colab's.
@@ -52,35 +50,71 @@ Each edition features different models and requires different hardware to run, t
 
 ### [Click here for the TPU Edition Colab](https://colab.research.google.com/github/KoboldAI/KoboldAI-Client/blob/main/colab/TPU.ipynb)
 
-| Model                          | Size   | Type     | Drive Space | Description                                                  |
-| ------------------------------ | ------ | --------- | ----------- | ------------------------------------------------------------ |
-| Skein 6B by VE_FORBDRYDERNE    | 6B TPU | Hybrid    | 0 GB         | Skein is our flagship 6B model, it is a hybrid between a Adventure model and a Novel model. Best used with either Adventure mode or the You Bias userscript enabled. Skein has been trained on high quality Novels along with CYOA adventure stories and is not as wackey as the Adventure model. It also has tagging support. |
-| Adventure 6B by VE_FORBRYDERNE | 6B TPU | Adventure | 0 GB         | Adventure is a 6B model designed to mimick the behavior of AI Dungeon. It is exclusively for Adventure Mode and can take you on the epic and wackey adventures that AI Dungeon players love. It also features the many tropes of AI Dungeon as it has been trained on very similar data. It must be used in second person (You). |
-| Lit 6B by Haru                 | 6B TPU | NSFW      | 8 GB /  12 GB | Lit is a great NSFW model trained by Haru on both a large set of Literotica stories and high quality novels along with tagging support. Creating a high quality model for your NSFW stories. This model is exclusively a novel model and is best used in third person. |
-| Generic 6B by EleutherAI       | 6B TPU | Generic   | 10 GB / 12 GB | GPT-J-6B is what all other models are based on, if you need something that has no specific bias towards any particular subject this is the model for you. Best used when the other models are not suitable for what you wish to do. Such as homework assistance, blog writing, coding and more. It needs more hand holding than other models and is more prone to undesirable formatting changes. |
-| C1 6B by Haru                  | 6B TPU | Chatbot   | 8 GB /  12 GB | C1 has been trained on various internet chatrooms, it makes the basis for an interesting chatbot model and has been optimized to be used in the Chatmode. |
+| Model | Size | Style | Description |
+| --- | --- | --- | --- |
+| [Nerys](https://huggingface.co/KoboldAI/fairseq-dense-13B-Nerys) by Mr Seeker | 13B | Novel/Adventure | Nerys is a hybrid model based on Pike (A newer Janeway), on top of the Pike dataset you also get some Light Novels, Adventure mode support and a little bit of shinen thrown in the mix. The end result is a very diverse model that is heavily biased towards SFW novel writing, but one that can go beyond its novel training and make for an excellent adventure model to. Adventure mode is best played from a second person perspective, but can be played in first or third person as well. Novel writing can be done best from the first or third person. |
+| [Janeway](https://huggingface.co/KoboldAI/fairseq-dense-13B-Janeway) by Mr Seeker | 13B | Novel | Janeway is a model created from Picard's dataset combined with a brand new collection of ebooks. This model is trained on 20% more content than Picard and has been trained on literature from various genres. Although the model is mainly focussed on SFW, romantic scenes might involve a degree of nudity. |
+| [Shinen](https://huggingface.co/KoboldAI/fairseq-dense-13B-Shinen) by Mr Seeker | 13B | NSFW | Shinen is an NSFW model designed to be more explicit. Trained on a variety of stories from the website Sexstories it contains many different kinks. |
+| [Skein](https://huggingface.co/KoboldAI/GPT-J-6B-Skein) by VE\_FORBRYDERNE | 6B | Adventure | Skein is best used with Adventure mode enabled, it consists of a 4 times larger adventure dataset than the Adventure model making it excellent for text adventure gaming. On top of that it also consists of light novel training further expanding its knowledge and writing capabilities. It can be used with the You filter bias if you wish to write Novels with it, but dedicated Novel models can perform better for this task. |
+| [Adventure](https://huggingface.co/KoboldAI/GPT-J-6B-Adventure) by VE\_FORBRYDERNE | 6B | Adventure | Adventure is a 6B model designed to mimick the behavior of AI Dungeon. It is exclusively for Adventure Mode and can take you on the epic and wackey adventures that AI Dungeon players love. It also features the many tropes of AI Dungeon as it has been trained on very similar data. It must be used in second person (You). |
+| [Lit](https://huggingface.co/hakurei/lit-6B) by Haru | 6B | NSFW | Lit is a great NSFW model trained by Haru on both a large set of Literotica stories and high quality novels along with tagging support. Creating a high quality model for your NSFW stories. This model is exclusively a novel model and is best used in third person. |
+| [Convo](https://huggingface.co/hitomi-team/convo-6B) by Hitomi Team | 6B | Chatbot | Convo-6B is a GPT-J 6B model fine-tuned on a collection of high quality open source datasets which amount to 6 million messages. The primary goal of the model is to provide improved performance and generalization when generating multi-turn dialogue for characters that were not present from within the fine tuning data. The prompted performance has especially improved over the predecessor model [C1-6B](https://huggingface.co/hakurei/c1-6B). |
+| [C1](https://huggingface.co/hakurei/c1-6B) by Haru | 6B | Chatbot | C1 has been trained on various internet chatrooms, it makes the basis for an interesting chatbot model and has been optimized to be used in the Chatmode. |
+| Neo(X) by EleutherAI | 20B | Generic | NeoX is the largest EleutherAI model currently available, being a generic model it is not particularly trained towards anything and can do a variety of writing, Q&A and coding tasks. 20B's performance is closely compared to the 13B models and it is worth trying both especially if you have a task that does not involve english writing. Its behavior will be similar to the GPT-J-6B model since they are trained on the same dataset but with more sensitivity towards repetition penalty and with more knowledge. |
+| [Fairseq Dense](https://huggingface.co/KoboldAI/fairseq-dense-13B) | 13B | Generic | Trained by Facebook Researchers this model stems from the MOE research project within Fairseq. This particular version has been converted by us for use in KoboldAI. It is known to be on par with the larger 20B model from EleutherAI and considered as better for pop culture and language tasks. Because the model has never seen a new line (enter) it may perform worse on formatting and paragraphing. |
+| [GPT-J-6B](https://huggingface.co/EleutherAI/gpt-j-6B) by EleutherAI | 6B | Generic | This model serves as the basis for most other 6B models (Some being based on Fairseq Dense instead). Being trained on the Pile and not biased towards anything in particular it is suitable for a variety of tasks such as writing, Q&A and coding tasks. You will likely get better result with larger generic models or finetuned models. |
+
+## [GPU Edition Model Descriptions](https://colab.research.google.com/github/KoboldAI/KoboldAI-Client/blob/main/colab/GPU.ipynb)
+
+| Model | Size | Style | Description |
+| --- | --- | --- | --- |
+| [Fairseq-Dense-2.7B-Nerys](https://huggingface.co/KoboldAI/fairseq-dense-2.7B-Nerys) by Mr Seeker | 2.7B | Novel/Adventure | Nerys is a hybrid model based on Pike (A newer Janeway), on top of the Pike dataset you also get some Light Novels, Adventure mode support and a little bit of shinen thrown in the mix. The end result is a very diverse model that is heavily biased towards SFW novel writing, but one that can go beyond its novel training and make for an excellent adventure model to. Adventure mode is best played from a second person perspective, but can be played in first or third person as well. Novel writing can be done best from the first or third person. |
+| [GPT-Neo-2.7B-Janeway](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Janeway) by Mr Seeker | 2.7B | Novel | Janeway is a model created from Picard's dataset combined with a brand new collection of ebooks. This model is trained on 20% more content than Picard and has been trained on literature from various genres. Although the model is mainly focussed on SFW, romantic scenes might involve a degree of nudity. |
+| [GPT-Neo-2.7B-Picard](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Picard) by Mr Seeker | 2.7B | Novel | Picard is a model trained for SFW Novels based on GPT-Neo-2.7B. It is focused on Novel style writing without the NSFW bias. While the name suggests a sci-fi model this model is designed for Novels of a variety of genre's. It is meant to be used in KoboldAI's regular mode. |
+| [GPT-Neo-2.7B-AID](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-AID) by melastacho | 2.7B | Adventure | Also know as Adventure 2.7B this is a clone of the AI Dungeon Classic model and is best known for the epic wackey adventures that AI Dungeon Classic players love. |
+| [GPT-Neo-2.7B-Horni-LN](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Horni-LN) by finetune | 2.7B | Novel | This model is based on GPT-Neo-2.7B-Horni and retains its NSFW knowledge, but was then further biased towards SFW novel stories. If you seek a balance between a SFW Novel model and a NSFW model this model should be a good choice. |
+| [GPT-Neo-2.7B-Horni](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Horni) by finetune | 2.7B | NSFW | This model is tuned on Literotica to produce a Novel style model biased towards NSFW content. Can still be used for SFW stories but will have a bias towards NSFW content. It is meant to be used in KoboldAI's regular mode. |
+| [GPT-Neo-2.7B-Shinen](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Shinen) by Mr Seeker | 2.7B | NSFW | Shinen is an alternative to the Horni model designed to be more explicit. If Horni is to tame for you shinen might produce better results. While it is a Novel model it is unsuitable for SFW stories due to its heavy NSFW bias. Shinen will not hold back. It is meant to be used in KoboldAI's regular mode. |
+| [GPT-Neo-2.7B](https://huggingface.co/EleutherAI/gpt-neo-2.7B) by EleutherAI | 2.7B | Generic | This is the base model for all the other 2.7B models, it is best used when you have a use case that we have no other models available for, such as writing blog articles or programming. It can also be a good basis for the experience of some of the softprompts if your softprompt is not about a subject the other models cover. |
+
+| Style | Description |
+| --- | --- |
+| Novel | For regular story writing, not compatible with Adventure mode or other specialty modes. |
+| NSFW | Indicates that the model is strongly biased towards NSFW content and is not suitable for children, work environments or livestreaming. Most NSFW models are also Novel models in nature. |
+| Adventure | These models are excellent for people willing to play KoboldAI like a Text Adventure game and are meant to be used with Adventure mode enabled. Even if you wish to use it as a Novel style model you should always have Adventure mode on and set it to story. These models typically have a strong bias towards the use of the word You and without Adventure mode enabled break the story flow and write actions on your behalf. |
+| Chatbot | These models are specifically trained for chatting and are best used with the Chatmode enabled. Typically trained on either public chatrooms or private chats. |
+| Generic | Generic models are not trained towards anything specific, typically used as a basis for other tasks and models. They can do everything the other models can do, but require much more handholding to work properly. Generic models are an ideal basis for tasks that we have no specific model for, or for experiencing a softprompt in its raw form. |
+
+## Tips to get the most out of Google Colab
+
+*   Google will occationally show a Captcha, typically after it has been open for 30 minutes but it can be more frequent if you often use Colab. Make sure to do these properly, or you risk getting your instance shut down and getting a lower priority towards the TPU's.
+*   KoboldAI uses Google Drive to store your files and settings, if you wish to upload a softprompt or userscript this can be done directly on the Google Drive website. You can also use this to download backups of your KoboldAI related files or upload models of your own.
+*   Don't want to save your stories on Google Drive for privacy reasons? Do not use KoboldAI's save function and instead click Download as .json, this will automatically download the story to your own computer without ever touching Google's harddrives. You can load this back trough the Load from file option.
+*   Google shut your instance down unexpectedly? You can still make use of the Download as .json button to recover your story as long as you did not close the KoboldAI window. You can then load this back up in your next session.
+*   Done with KoboldAI? Go to the Runtime menu, click on Manage Sessions and terminate your open sessions that you no longer need. This trick can help you maintain higher priority towards getting a TPU.
+*   Models stored on Google Drive typically load faster than models we need to download from the internet.
 
 ### [Click here for the GPU Edition Colab](https://colab.research.google.com/github/KoboldAI/KoboldAI-Client/blob/main/colab/GPU.ipynb)
 
-| Model                                                        | Size     | Type      | Description                                                  |
-| ------------------------------------------------------------ | -------- | ---------- | ------------------------------------------------------------ |
-| [GPT-Neo-2.7B-Picard](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Picard) by Mr Seeker | 2.7B GPU | Novel      | Picard is a model trained for SFW Novels based on GPT-Neo-2.7B. It is focused on Novel Type writing without the NSFW bias. While the name suggests a sci-fi model this model is designed for Novels of a variety of genre's. It is meant to be used in KoboldAI's regular mode. |
+| Model | Size | Type | Description |
+| --- | --- | --- | --- |
+| [GPT-Neo-2.7B-Picard](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Picard) by Mr Seeker | 2.7B GPU | Novel | Picard is a model trained for SFW Novels based on GPT-Neo-2.7B. It is focused on Novel Type writing without the NSFW bias. While the name suggests a sci-fi model this model is designed for Novels of a variety of genre's. It is meant to be used in KoboldAI's regular mode. |
 | [GPT-Neo-2.7B-AID](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-AID) by melastacho | 2.7B GPU | Adventure | Also know as Adventure 2.7B this is a clone of the AI Dungeon Classic model and is best known for the epic wackey adventures that AI Dungeon Classic players love. |
 | [GPT-Neo-2.7B-Horni-LN](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Horni-LN) by finetune | 2.7B GPU | Novel | This model is based on GPT-Neo-2.7B-Horni and retains its NSFW knowledge, but was then further biased towards SFW novel stories. If you seek a balance between a SFW Novel model and a NSFW model this model should be a good choice. |
-| [GPT-Neo-2.7B-Horni](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Horni) by finetune | 2.7B GPU | NSFW       | This model is tuned on Literotica to produce a Novel Type model biased towards NSFW content. Can still be used for SFW stories but will have a bias towards NSFW content. It is meant to be used in KoboldAI's regular mode. |
-| [GPT-Neo-2.7B-Shinen](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Shinen) by Mr Seeker | 2.7B GPU | NSFW       | Shinen is an alternative to the Horni model designed to be more explicit. If Horni is to tame for you shinen might produce better results. While it is a Novel model it is unsuitable for SFW stories due to its heavy NSFW bias. Shinen will not hold back. It is meant to be used in KoboldAI's regular mode. |
-| [GPT-Neo-2.7B](https://huggingface.co/EleutherAI/gpt-neo-2.7B) by EleutherAI | 2.7B GPU    | Generic    | This is the base model for all the other 2.7B models, it is best used when you have a use case that we have no other models available for, such as writing blog articles or programming. It can also be a good basis for the experience of some of the softprompts if your softprompt is not about a subject the other models cover. |
+| [GPT-Neo-2.7B-Horni](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Horni) by finetune | 2.7B GPU | NSFW | This model is tuned on Literotica to produce a Novel Type model biased towards NSFW content. Can still be used for SFW stories but will have a bias towards NSFW content. It is meant to be used in KoboldAI's regular mode. |
+| [GPT-Neo-2.7B-Shinen](https://huggingface.co/KoboldAI/GPT-Neo-2.7B-Shinen) by Mr Seeker | 2.7B GPU | NSFW | Shinen is an alternative to the Horni model designed to be more explicit. If Horni is to tame for you shinen might produce better results. While it is a Novel model it is unsuitable for SFW stories due to its heavy NSFW bias. Shinen will not hold back. It is meant to be used in KoboldAI's regular mode. |
+| [GPT-Neo-2.7B](https://huggingface.co/EleutherAI/gpt-neo-2.7B) by EleutherAI | 2.7B GPU | Generic | This is the base model for all the other 2.7B models, it is best used when you have a use case that we have no other models available for, such as writing blog articles or programming. It can also be a good basis for the experience of some of the softprompts if your softprompt is not about a subject the other models cover. |
 
 ### Model Types
-| Type     | Description                                                  |
-| --------- | ------------------------------------------------------------ |
-| Novel     | For regular story writing, not compatible with Adventure mode or other specialty modes. |
-| NSFW      | Indicates that the model is strongly biased towards NSFW content and is not suitable for children, work environments or livestreaming. Most NSFW models are also Novel models in nature. |
-| Adventure | These models are excellent for people willing to play KoboldAI like a Text Adventure game and are meant to be used with Adventure mode enabled. Even if you wish to use it as a Novel Type model you should always have Adventure mode on and set it to story. These models typically have a strong bias towards the use of the word You and without Adventure mode enabled break the story flow and write actions on your behalf. |
-| Chatbot   | These models are specifically trained for chatting and are best used with the Chatmode enabled. Typically trained on either public chatrooms or private chats. |
-| Hybrid    | Hybrid models are a blend between different Types, for example they are trained on both Novel stories and Adventure stories. These models are great variety models that you can use for multiple different playTypes and modes, but depending on your usage you may need to enable Adventure Mode or the You bias (in userscripts). |
-| Generic   | Generic models are not trained towards anything specific, typically used as a basis for other tasks and models. They can do everything the other models can do, but require much more handholding to work properly. Generic models are an ideal basis for tasks that we have no specific model for, or for experiencing a softprompt in its raw form. |
 
+| Type | Description |
+| --- | --- |
+| Novel | For regular story writing, not compatible with Adventure mode or other specialty modes. |
+| NSFW | Indicates that the model is strongly biased towards NSFW content and is not suitable for children, work environments or livestreaming. Most NSFW models are also Novel models in nature. |
+| Adventure | These models are excellent for people willing to play KoboldAI like a Text Adventure game and are meant to be used with Adventure mode enabled. Even if you wish to use it as a Novel Type model you should always have Adventure mode on and set it to story. These models typically have a strong bias towards the use of the word You and without Adventure mode enabled break the story flow and write actions on your behalf. |
+| Chatbot | These models are specifically trained for chatting and are best used with the Chatmode enabled. Typically trained on either public chatrooms or private chats. |
+| Hybrid | Hybrid models are a blend between different Types, for example they are trained on both Novel stories and Adventure stories. These models are great variety models that you can use for multiple different playTypes and modes, but depending on your usage you may need to enable Adventure Mode or the You bias (in userscripts). |
+| Generic | Generic models are not trained towards anything specific, typically used as a basis for other tasks and models. They can do everything the other models can do, but require much more handholding to work properly. Generic models are an ideal basis for tasks that we have no specific model for, or for experiencing a softprompt in its raw form. |
 
 ## Install KoboldAI on your own computer
 
@@ -94,33 +128,42 @@ The easiest way for Windows users is to use the [offline installer](https://sour
 
 ### Installing KoboldAI offline bundle on Windows 7 or higher using the KoboldAI Offline Installer (Easiest)
 
-1. [Download the latest offline installer from here](https://sourceforge.net/projects/koboldai/files/latest/download)
-2. Run the installer to place KoboldAI on a location of choice, KoboldAI is portable software and is not bound to a specific harddrive. (Because of long paths inside our dependencies you may not be able to extract it many folders deep).
-3. Update KoboldAI to the latest version with update-koboldai.bat if desired.
-4. Use KoboldAI offline using play.bat or remotely with remote-play.bat
+1.  [Download the latest offline installer from here](https://sourceforge.net/projects/koboldai/files/latest/download)
+2.  Run the installer to place KoboldAI on a location of choice, KoboldAI is portable software and is not bound to a specific harddrive. (Because of long paths inside our dependencies you may not be able to extract it many folders deep).
+3.  Update KoboldAI to the latest version with update-koboldai.bat if desired.
+4.  Use KoboldAI offline using play.bat or remotely with remote-play.bat
 
 ### Installing KoboldAI Github release on Windows 10 or higher using the KoboldAI Runtime Installer
 
-1. Extract the .zip to a location you wish to install KoboldAI, you will need roughly 20GB of free space for the installation (this does not include the models).
-2. Open install_requirements.bat as **administrator**.
-3. Choose the regular version of Transformers (Option 1), finetuneanon is depreciated and no longer recommended.
-4. You will now be asked to choose the installation mode, we **strongly** recommend the Temporary B: drive option. This option eliminates most installation issues and also makes KoboldAI portable. The B: drive will be gone after a reboot and will automatically be recreated each time you play KoboldAI.
-5. The installation will now automatically install its requirements, some stages may appear to freeze do not close the installer until it asks you to press a key. Before pressing a key to exit the installer please check if errors occurred. Most problems with the game crashing are related to installation/download errors. Disabling your antivirus can help if you get errors.
-6. Use play.bat to start KoboldAI.
+1.  Extract the .zip to a location you wish to install KoboldAI, you will need roughly 20GB of free space for the installation (this does not include the models).
+2.  Open install\_requirements.bat as **administrator**.
+3.  Choose the regular version of Transformers (Option 1), finetuneanon is depreciated and no longer recommended.
+4.  You will now be asked to choose the installation mode, we **strongly** recommend the Temporary B: drive option. This option eliminates most installation issues and also makes KoboldAI portable. The B: drive will be gone after a reboot and will automatically be recreated each time you play KoboldAI.
+5.  The installation will now automatically install its requirements, some stages may appear to freeze do not close the installer until it asks you to press a key. Before pressing a key to exit the installer please check if errors occurred. Most problems with the game crashing are related to installation/download errors. Disabling your antivirus can help if you get errors.
+6.  Use play.bat to start KoboldAI.
 
-### Manual installation / Linux / Mac
+### Installing KoboldAI on Linux using the KoboldAI Runtime (Easiest)
+
+1.  Clone the URL of this Github repository (For example git clone [https://github.com/koboldai/koboldai-client](https://github.com/koboldai/koboldai-client) )
+2.  AMD user? Make sure ROCm is installed if you want GPU support. Is yours not compatible with ROCm? Follow the usual instructions.
+3.  Run play.sh or if your AMD GPU supports ROCm use play-rocm.sh
+
+KoboldAI will now automatically configure its dependencies and start up, everything is contained in its own conda runtime so we will not clutter your system. The files will be located in the runtime subfolder. If at any point you wish to force a reinstallation of the runtime you can do so with the install\_requirements.sh file. While you can run this manually it is not neccesary.
+
+### Manual installation / Mac
 
 We can not provide a step by step guide for manual installation due to the vast differences between the existing software configuration and the systems of our users.
 
 If you would like to manually install KoboldAI you will need some python/conda package management knowledge to manually do one of the following steps :
 
-1. Use our bundled environments files to install your own conda environment, this should also automatically install CUDA (Recommended, you can get Miniconda from https://docs.conda.io/en/latest/miniconda.html#latest-miniconda-installer-links). The recommended configuration is huggingface.yml for CUDA users and rocm.yml for ROCm users.
-2. If you have a working copy of Docker for either CUDA or ROCm try play-cuda.sh or play-rocm.sh to launch the docker versions. In this case the installation is mostly automatic.
-3. If conda is proving difficult you could also look inside requirements.txt for the required dependencies and try to install them yourself. This will likely be a mixture of pip and your native package manager, just installing our requirements.txt is not recommended since to speed things up we do not force any version changes. For local installations definitely prioritize conda as that is a better way for us to enforce you have the latest compatible versions.
+1.  Use our bundled environments files to install your own conda environment, this should also automatically install CUDA (Recommended, you can get Miniconda from https://docs.conda.io/en/latest/miniconda.html#latest-miniconda-installer-links). The recommended configuration is huggingface.yml for CUDA users and rocm.yml for ROCm users.
+2.  If conda is proving difficult you could also look inside requirements.txt for the required dependencies and try to install them yourself. This will likely be a mixture of pip and your native package manager, just installing our requirements.txt is not recommended since we assume local users will run conda to get all dependencies. For local installations definitely prioritize conda as that is a better way for us to enforce that you have the compatible versions.
+3.  Clone our Github or download the zip file.
+4.  Now start KoboldAI with aiserver.py and not with our play.bat or play.sh files.
 
-### AMD GPU's
+### AMD GPU's (Linux only)
 
-AMD GPU's have terrible compute support, this will currently not work on Windows and will only work for a select few Linux GPU's. [You can find a list of the compatible GPU's here](https://github.com/RadeonOpenCompute/ROCm#Hardware-and-Software-Support). Any GPU that is not listed is guaranteed not to work with KoboldAI and we will not be able to provide proper support on GPU's that are not compatible with the versions of ROCm we require. 
+AMD GPU's have terrible compute support, this will currently not work on Windows and will only work for a select few Linux GPU's. [You can find a list of the compatible GPU's here](https://github.com/RadeonOpenCompute/ROCm#Hardware-and-Software-Support). Any GPU that is not listed is guaranteed not to work with KoboldAI and we will not be able to provide proper support on GPU's that are not compatible with the versions of ROCm we require. Make sure to first install ROCm on your Linux system using a guide for your distribution, after that you can follow the usual linux instructions above.
 
 ### Troubleshooting
 
@@ -140,45 +183,17 @@ In general, the less versions of Python you have on your system the higher your
 
 GPU not found errors can be caused by one of two things, either you do not have a suitable Nvidia GPU (It needs Compute Capability 5.0 or higher to be able to play KoboldAI). Your Nvidia GPU is supported by KoboldAI but is not supported by the latest version of CUDA. Your Nvidia GPU is not yet supported by the latest version of CUDA or you have a dependency conflict like the ones mentioned above.
 
-Like with Python version conflicts we recommend uninstalling CUDA from your system if you have manually installed it and do not need it for anything else and trying again. If your GPU needs CUDA10 to function open environments\finetuneanon.yml and add a line that says - cudatoolkit=10.2 underneath dependencies: . After this you can run the installer again (Pick the option to delete the existing files) and it will download a CUDA10 compatible version.
+Like with Python version conflicts we recommend uninstalling CUDA from your system if you have manually installed it and do not need it for anything else and trying again. If your GPU needs CUDA10 to function open environments\\finetuneanon.yml and add a line that says - cudatoolkit=10.2 underneath dependencies: . After this you can run the installer again (Pick the option to delete the existing files) and it will download a CUDA10 compatible version.
 
 If you do not have a suitable Nvidia GPU that can run on CUDA10 or Higher and that supports Compute Capabilities 5.0 or higher we can not help you get the game detected on the GPU. Unless you are following our ROCm guide with a compatible AMD GPU.
 
 #### vocab.json / config.json is not found error
 
-If you get these errors you either did not select the correct folder for your custom model or the model you have downloaded is not (yet) compatible with KoboldAI. There exist a few models out there that are compatible and provide a pytorch_model.bin file but do not ship all the required files. In this case try downloading a compatible model of the same kind (For example another GPT-Neo if you downloaded a GPT-Neo model) and replace the pytorch_model.bin file with the one you are trying to run. Chances are this will work fine.
-
-## KoboldAI Compatible Models
-
-Most of the high quality models have been integrated in the menu, these models have their download link removed since the easiest way to obtain them is to run them directly from the menu. KoboldAI will automatically download and convert the models to a offline format for later use.
-
-If you have old 6B versions which end in -hf they will no longer be compatible with the newer versions of transformers and will no longer behave correctly. It is highly recommended that you install the official version of transformers (offline installers for KoboldAI contain this version by default) and redownload these models from the menu to get compatible versions. If you have very limited internet we will for a limited time also offer finetuneanon's fork in the install_requirements.bat file, when using that option you will not be able to use the 6B models in our main menu so definitely upgrade when your internet allows.
-
-The VRAM requirements amounts are the recommended amounts for fast smooth play, playing with lower VRAM is possible but then you may need to either lower the amount of tokens in the settings, or you may need to put less layers on your GPU causing a significant performance loss. 
-
-**For CPU players and during the loading regular RAM usage is double of what we list here.**
-
-
-
-| **Model**                                                    | Type                              | **(V)RAM** | Repetition Penalty | Description                                                  |
-| ------------------------------------------------------------ | --------------------------------- | ---------- | ------------------ | ------------------------------------------------------------ |
-| Skein 6B by VE_FORBDRYERNE | Adventure Novel / 6B / Neo Custom | 16GB       | 1.1                | Skein is our flagship 6B model, it is a hybrid between a Adventure model and a Novel model. Best used with either Adventure mode or the You Bias userscript enabled. Skein has been trained on high quality Novels along with CYOA adventure stories and is not as wackey as the Adventure model. It also has tagging support. |
-| Adventure 6B by VE_FORBRYDERNE | Adventure / 6B / Neo Custom       | 16GB       | 1.2                | Adventure is a 6B model designed to mimick the behavior of AI Dungeon. It is exclusively for Adventure Mode and can take you on the epic and wackey adventures that AI Dungeon players love. It also features the many tropes of AI Dungeon as it has been trained on very similar data. It must be used in second person (You). |
-| Adventure 2.7B by melastashco | Adventure / 2.7B / Neo Custom     | 8GB        | 2.0                | This is one of the closest replications of the original AI Dungeon Classic model. Tuned on the same data that got uploaded alongside AI Dungeon. In KoboldAI we noticed this model performs better than the conversions of the original AI Dungeon model. It has all the traits you expect of AI Dungeon Classic while not having as many artifacts as this model was trained specifically for KoboldAI. Must be played with Adventure mode enabled to prevent it from doing actions on your behalf. |
-| Horni 2.7B by finetuneanon | Novel / 2.7B / Neo Custom         | 8GB        | 2.0                | One of the best novel models available for 2.7B focused on NSFW content. This model trains the AI to write in a story like fashion using a very large collection of Literotica stories. It is one of the original finetuned models for 2.7B. |
-| Horni-LN 2.7B by finetuneanon | Novel / 2.7B / Neo Custom         | 8GB        | 2.0                | This model is much like the one above, but has been additionally trained on regular light novels. More likely to go SFW and is more focused towards themes found in these light novels over general cultural references. This is a good model for Novel writing especially if you want to add erotica to the mix. |
-| Picard 2.7B by Mr Seeker | Novel / 2.7B / Neo Custom         | 8GB        | 2.0                | Picard is another Novel model, this time exclusively focused on SFW content of various genres. Unlike the name suggests this goes far beyond Star Trek stories and is not exclusively sci-fi. |
-| Janeway 2.7B by Mr Seeker | Novel / 2.7B / Neo Custom         | 8GB        | 2.0                | Janeway is a model created from Picard's dataset combined with a brand new collection of ebooks. This model is trained on 20% more content than Picard and has been trained on literature from various genres. Although the model is mainly focussed on SFW, romantic scenes might involve a degree of nudity.|
-| Shinen 2.7B by Mr Seeker | Novel / 2.7B / Neo Custom         | 8GB        | 2.0                | The most NSFW of them all, Shinen WILL make things sexual. This model will assume that whatever you are doing is meant to be a sex story and will sexualize constantly. It is designed for people who find Horni to tame. It was trained on SexStories instead of Literotica and was trained on tags making it easier to guide the AI to the right context. |
-| [AID-16Bit](https://storage.henk.tech/KoboldAI/aid-16bit.zip) | Adventure / 1.5B / GPT-2 Custom   | 4GB        | 2.0                | The original AI Dungeon Classic model converted to Pytorch and then converted to a 16-bit Model making it half the size. |
-| [model_v5_pytorch](https://storage.henk.tech/KoboldAI/model_v5_pytorch.zip) (AI Dungeon's Original Model) | Adventure / 1.5B / GPT-2 Custom   | 8GB       | 2.0                | This is the original AI Dungeon Classic model converted to the Pytorch format compatible with AI Dungeon Clover and KoboldAI. We consider this model inferior to the GPT-Neo version because it has more artifacting due to its conversion. This is however the most authentic you can get to AI Dungeon Classic. |
-| [Novel 774M](https://storage.henk.tech/KoboldAI/Novel%20model%20774M.rar) | Novel / 774M / GPT-2 Custom       | 4GB        | 2.0                | Novel 774M is made by the AI Dungeon Clover community, because of its small size and novel bias it is more suitable for CPU players that want to play with speed over substance or players who want to test a GPU with a low amount of VRAM. These performance savings are at the cost of story quality and you should not expect the kind of in depth story capabilities that the larger models offer. It was trained for SFW stories. |
-| [Smut 774M](https://storage.henk.tech/KoboldAI/Smut%20model%20774M%2030K.rar) | Novel / 774M / GPT-2 Custom       | 4GB        | 2.0                | The NSFW version of the above, its a smaller GPT-2 based model made by the AI Dungeon Clover community. Gives decent speed on a CPU at the cost of story quality like the other 774M models. |
-| [Mia (GPT-Neo-125M-AID)](https://huggingface.co/KoboldAI/GPT-Neo-125M-AID) by Henk717 | Adventure / 125M / Neo Custom     | 1GB        | 2.0                | Mia is the smallest Adventure model, it runs at very fast speeds on the CPU which makes it a good testing model for developers who do not have GPU access. Because of its small size it will constantly attempt to do actions on behalf of the player and it will not produce high quality stories. If you just need a small model for a quick test, or if you want to take the challenge of trying to run KoboldAI entirely on your phone this would be an easy model to use due to its small RAM requirements and fast (loading) speeds. |
+If you get these errors you either did not select the correct folder for your custom model or the model you have downloaded is not (yet) compatible with KoboldAI. There exist a few models out there that are compatible and provide a pytorch\_model.bin file but do not ship all the required files. In this case try downloading a compatible model of the same kind (For example another GPT-Neo if you downloaded a GPT-Neo model) and replace the pytorch\_model.bin file with the one you are trying to run. Chances are this will work fine.
 
 ## Softprompts
 
-Softprompts (also known as Modules in other products) are addons that can change the output of existing models. For example you may load a softprompt that biases the AI towards a certain subject and style like transcripts from your favorite TV show. 
+Softprompts (also known as Modules in other products) are addons that can change the output of existing models. For example you may load a softprompt that biases the AI towards a certain subject and style like transcripts from your favorite TV show.
 
 Since these softprompts are often based on existing franchises we currently do not bundle any of them with KoboldAI due to copyright concerns (We do not want to put the entire project at risk). Instead look at community resources like #softprompts on the [KoboldAI Discord](https://discord.gg/XuQWadgU9k) or the [community hosted mirror](https://storage.henk.tech/KoboldAI/softprompts/) .
 
@@ -188,10 +203,10 @@ Training softprompts can be done for free with the [mtj-softtuner colab](https:/
 
 ## Userscripts
 
-Userscripts are scripts that can automate tasks in KoboldAI, or modify the AI behavior / input / output.
+Userscripts are scripts that can automate tasks in KoboldAI, or modify the AI behavior / input / output.  
 Scripting is done in LUA5.4 (Lua does not need to be separately installed as long as you got all the python requirements) and has sandboxing to help protect you from malicious behavior. Even with these measures in place we strongly advise you only run userscripts from places you trust and/or understand, otherwise consult the community for advice on how safe the script might be.
 
-Inside the userscripts folder you will find our kaipreset scripts, these are default scripts that we think will be useful for our users. These scripts are automatically overwritten when you update KoboldAI, if you wish to modify these scripts make sure to first rename them to something else that does not contain kaipreset so your changes are not lost. These scripts range from a You Bias filter that prevents the AI from addressing characters as you. Ways to be able to prevent the AI from using words, word replacements and more. 
+Inside the userscripts folder you will find our kaipreset scripts, these are default scripts that we think will be useful for our users. These scripts are automatically overwritten when you update KoboldAI, if you wish to modify these scripts make sure to first rename them to something else that does not contain kaipreset so your changes are not lost. These scripts range from a You Bias filter that prevents the AI from addressing characters as you. Ways to be able to prevent the AI from using words, word replacements and more.
 
 Along with our preset scripts we also ship examples in the examples folder that merely serve as a demonstration and do not enhance your usage of KoboldAI. To use these scripts make sure to move them out of the examples folder before either using or modifying the script.
 
@@ -203,16 +218,17 @@ For our TPU versions keep in mind that scripts modifying AI behavior relies on a
 
 This project contains work from the following contributors :
 
-- The Gantian - Creator of KoboldAI, has created most features such as the interface, the different AI model / API integrations and in general the largest part of the project.
-- VE FORBRYDERNE - Contributed many features such as the Editing overhaul, Adventure Mode, expansions to the world info section, breakmodel integration, scripting support, softpromtps and much more. As well as vastly improving the TPU compatibility and integrating external code into KoboldAI so we could use official versions of Transformers with virtually no downsides.
-- Henk717 - Contributed the installation scripts, this readme, random story generator, the docker scripts, the foundation for the commandline interface and other smaller changes as well as integrating multiple parts of the code of different forks to unite it all. He also optimized the model loading so that downloaded models get converted to efficient offline models and that in future models are more likely to work out of the box.  Not all code Github attributes to Henk717 is by Henk717 as some of it has been integrations of other people's work. We try to clarify this in the contributors list as much as we can.
-- Ebolam - Automatic Saving
-- Frogging101 - top_k / tfs support (Part of this support was later redone by VE to integrate what was originally inside of finetuneanon's transformers)
-- UWUplus (Ralf) - Contributed storage systems for community colabs, as well as cleaning up and integrating the website dependencies/code better. He is also the maintainer of flask-cloudflared which we use to generate the cloudflare links.
-- Javalar - Initial Performance increases on the story_refresh
-- LexSong - Initial environment file adaptation for conda that served as a basis for the install_requirements.bat overhaul.
-- Arrmansa - Breakmodel support for other projects that served as a basis for VE FORBRYDERNE's integration.
-- Jojorne - Small improvements to the response selection for gens per action.
+*   The Gantian - Creator of KoboldAI, has created most features such as the interface, the different AI model / API integrations and in general the largest part of the project.
+*   VE FORBRYDERNE - Contributed many features such as the Editing overhaul, Adventure Mode, expansions to the world info section, breakmodel integration, scripting support, softpromtps and much more. As well as vastly improving the TPU compatibility and integrating external code into KoboldAI so we could use official versions of Transformers with virtually no downsides.
+*   Henk717 - Contributed the installation scripts, this readme, random story generator, the docker scripts, the foundation for the commandline interface and other smaller changes as well as integrating multiple parts of the code of different forks to unite it all. He also optimized the model loading so that downloaded models get converted to efficient offline models and that in future models are more likely to work out of the box. Not all code Github attributes to Henk717 is by Henk717 as some of it has been integrations of other people's work. We try to clarify this in the contributors list as much as we can.
+*   Ebolam - Automatic Saving
+*   Frogging101 - top\_k / tfs support (Part of this support was later redone by VE to integrate what was originally inside of finetuneanon's transformers)
+*   UWUplus (Ralf) - Contributed storage systems for community colabs, as well as cleaning up and integrating the website dependencies/code better. He is also the maintainer of flask-cloudflared which we use to generate the cloudflare links.
+*   Javalar - Initial Performance increases on the story\_refresh
+*   LexSong - Initial environment file adaptation for conda that served as a basis for the install\_requirements.bat overhaul.
+*   Arrmansa - Breakmodel support for other projects that served as a basis for VE FORBRYDERNE's integration.
+*   Jojorne - Small improvements to the response selection for gens per action.
+*   OccultSage (GooseAI) - Improved support for GooseAI/OpenAI
 
 As well as various Model creators who will be listed near their models, and all the testers who helped make this possible!
 
@@ -222,4 +238,4 @@ Did we miss your contribution? Feel free to issue a commit adding your name to t
 
 KoboldAI is licensed with a AGPL license, in short this means that it can be used by anyone for any purpose. However, if you decide to make a publicly available instance your users are entitled to a copy of the source code including all modifications that you have made (which needs to be available trough an interface such as a button on your website), you may also not distribute this project in a form that does not contain the source code (Such as compiling / encrypting the code and distributing this version without also distributing the source code that includes the changes that you made. You are allowed to distribute this in a closed form if you also provide a separate archive with the source code.).
 
-umamba.exe is bundled for convenience because we observed that many of our users had trouble with command line download methods, it is not part of our project and does not fall under the AGPL license. It is licensed under the BSD-3-Clause license.
+umamba.exe is bundled for convenience because we observed that many of our users had trouble with command line download methods, it is not part of our project and does not fall under the AGPL license. It is licensed under the BSD-3-Clause license. Other files with differing licenses will have a reference or embedded version of this license within the file.

requirements.txt

@@ -1,11 +1,13 @@
-transformers
+transformers>=4.19
 Flask
 Flask-SocketIO
 requests
-torch
+torch==1.11
 flask-cloudflared
 flask-ngrok
 eventlet
 lupa==1.10
 markdown
-bleach==4.1.0
+bleach==4.1.0
+sentencepiece
+protobuf

requirements_mtj.txt

@@ -1,11 +1,11 @@
+torch >= 1.9, <= 1.11
 numpy
 tqdm
 requests
 optax >= 0.0.5, <= 0.0.9
 dm-haiku == 0.0.5
-ray[default]
 jax == 0.2.21
-transformers
+transformers >= 4.19
 progressbar2
 git+https://github.com/VE-FORBRYDERNE/mesh-transformer-jax@ck
 flask

static/application.js

@@ -25,6 +25,7 @@ var button_mode_label;
 var button_send;
 var button_actmem;
 var button_actback;
+var button_actfwd;
 var button_actretry;
 var button_actwi;
 var game_text;
@@ -38,6 +39,7 @@ var anote_menu;
 var anote_input;
 var anote_labelcur;
 var anote_slider;
+var debug_area;
 var popup;
 var popup_title;
 var popup_content;
@@ -49,6 +51,7 @@ var aidg_accept;
 var aidg_close;
 var saveaspopup;
 var saveasinput;
+var savepins;
 var topic;
 var saveas_accept;
 var saveas_close;
@@ -115,10 +118,33 @@ var adventure = false;
 // Chatmode
 var chatmode = false;
 
+var sliders_throttle = getThrottle(200);
+
 //=================================================================//
 //  METHODS
 //=================================================================//
 
+/**
+ * Returns a function that will automatically wait for X ms before executing the callback
+ * The timer is reset each time the returned function is called
+ * Useful for methods where something is overridden too fast
+ * @param ms milliseconds to wait before executing the callback
+ * @return {(function(*): void)|*} function that takes the ms to wait and a callback to execute after the timer
+ */
+function getThrottle(ms) {
+    var timer = {};
+
+    return function (id, callback) {
+        if (timer[id]) {
+            clearTimeout(timer[id]);
+        }
+        timer[id] = setTimeout(function () {
+            callback();
+            delete timer[id];
+        }, ms);
+    }
+}
+
 function addSetting(ob) {	
 	// Add setting block to Settings Menu
 	if(ob.uitype == "slider"){
@@ -127,9 +153,7 @@ function addSetting(ob) {
 			<div class=\"justifyleft\">\
 				"+ob.label+" <span class=\"helpicon\">?<span class=\"helptext\">"+ob.tooltip+"</span></span>\
 			</div>\
-			<div class=\"justifyright\" id=\""+ob.id+"cur\">\
-				"+ob.default+"\
-			</div>\
+			<input inputmode=\""+(ob.unit === "float" ? "decimal" : "numeric")+"\" class=\"justifyright flex-push-right\" id=\""+ob.id+"cur\" value=\""+ob.default+"\">\
 		</div>\
 		<div>\
 			<input type=\"range\" class=\"form-range airange\" min=\""+ob.min+"\" max=\""+ob.max+"\" step=\""+ob.step+"\" id=\""+ob.id+"\">\
@@ -149,8 +173,37 @@ function addSetting(ob) {
 		window["setting_"+ob.id] = refin;  // Is this still needed?
 		window["label_"+ob.id]   = reflb;  // Is this still needed?
 		// Add event function to input
-		refin.on("input", function () {
-			socket.send({'cmd': $(this).attr('id'), 'data': $(this).val()});
+		var updateLabelColor = function () {
+			var value = (ob.unit === "float" ? parseFloat : parseInt)(reflb.val());
+			if(value > ob.max || value < ob.min) {
+				reflb.addClass("setting-value-warning");
+			} else {
+				reflb.removeClass("setting-value-warning");
+			}
+		}
+		var send = function () {
+			sliders_throttle(ob.id, function () {
+			    socket.send({'cmd': $(refin).attr('id'), 'data': $(reflb).val()});
+			});
+		}
+		refin.on("input", function (event) {
+			reflb.val(refin.val());
+			updateLabelColor();
+			send();
+		}).on("change", updateLabelColor);
+		reflb.on("change", function (event) {
+			var value = (ob.unit === "float" ? parseFloat : parseInt)(event.target.value);
+			if(Number.isNaN(value) || (ob.min >= 0 && value < 0)) {
+				event.target.value = refin.val();
+				return;
+			}
+			if (ob.unit === "float") {
+				value = parseFloat(value.toFixed(3));  // Round to 3 decimal places to help avoid the number being too long to fit in the box
+			}
+			refin.val(value);
+			reflb.val(value);
+			updateLabelColor();
+			send();
 		});
 	} else if(ob.uitype == "toggle"){
 		settings_menu.append("<div class=\"settingitem\">\
@@ -748,7 +801,7 @@ function enterMemoryMode() {
 	setchatnamevisibility(false);
 	showMessage("Edit the memory to be sent with each request to the AI.");
 	button_actmem.html("Cancel");
-	hide([button_actback, button_actretry, button_actwi]);
+	hide([button_actback, button_actfwd, button_actretry, button_actwi]);
 	// Display Author's Note field
 	anote_menu.slideDown("fast");
 }
@@ -759,7 +812,7 @@ function exitMemoryMode() {
 	setchatnamevisibility(chatmode);
 	hideMessage();
 	button_actmem.html("Memory");
-	show([button_actback, button_actretry, button_actwi]);
+	show([button_actback, button_actfwd, button_actretry, button_actwi]);
 	input_text.val("");
 	// Hide Author's Note field
 	anote_menu.slideUp("fast");
@@ -768,7 +821,7 @@ function exitMemoryMode() {
 function enterWiMode() {
 	showMessage("World Info will be added to memory only when the key appears in submitted text or the last action.");
 	button_actwi.html("Accept");
-	hide([button_actback, button_actmem, button_actretry, game_text]);
+	hide([button_actback, button_actfwd, button_actmem, button_actretry, game_text]);
 	setchatnamevisibility(false);
 	show([wi_menu]);
 	disableSendBtn();
@@ -780,7 +833,7 @@ function exitWiMode() {
 	button_actwi.html("W Info");
 	hide([wi_menu]);
 	setchatnamevisibility(chatmode);
-	show([button_actback, button_actmem, button_actretry, game_text]);
+	show([button_actback, button_actfwd, button_actmem, button_actretry, game_text]);
 	enableSendBtn();
 	$("#gamescreen").removeClass("wigamescreen");
 }
@@ -884,7 +937,7 @@ function hideSaveAsPopup() {
 }
 
 function sendSaveAsRequest() {
-	socket.send({'cmd': 'saveasrequest', 'data': saveasinput.val()});
+	socket.send({'cmd': 'saveasrequest', 'data': {"name": saveasinput.val(), "pins": savepins.val()}});
 }
 
 function showLoadPopup() {
@@ -1142,9 +1195,9 @@ function updateSPStatItems(items) {
 function setStartState() {
 	enableSendBtn();
 	enableButtons([button_actmem, button_actwi]);
-	disableButtons([button_actback, button_actretry]);
+	disableButtons([button_actback, button_actfwd, button_actretry]);
 	hide([wi_menu]);
-	show([game_text, button_actmem, button_actwi, button_actback, button_actretry]);
+	show([game_text, button_actmem, button_actwi, button_actback, button_actfwd, button_actretry]);
 	hideMessage();
 	hideWaitAnimation();
 	button_actmem.html("Memory");
@@ -1160,10 +1213,41 @@ function parsegenseqs(seqs) {
 	seqselcontents.html("");
 	var i;
 	for(i=0; i<seqs.length; i++) {
-		seqselcontents.append("<div class=\"seqselitem\" id=\"seqsel"+i+"\" n=\""+i+"\">"+seqs[i].generated_text+"</div>");
+		//setup selection data
+		text_data = "<table><tr><td width=100%><div class=\"seqselitem\" id=\"seqsel"+i+"\" n=\""+i+"\">"+seqs[i][0]+"</div></td><td width=10>"
+		
+		//Now do the icon (pin/redo)
+		
+		if (seqs[i][1] == "redo") {
+			text_data = text_data + "<span style=\"color: white\" class=\"oi oi-loop-circular\" title=\"Redo\" aria-hidden=\"true\" id=\"seqselpin"+i+"\" n=\""+i+"\"></span>"
+		} else if (seqs[i][1] == "pinned") {
+			text_data = text_data + "<span style=\"color: white\" class=\"oi oi-pin\" title=\"Pin\" aria-hidden=\"true\" id=\"seqselpin"+i+"\" n=\""+i+"\"></span>"
+		} else {
+			text_data = text_data + "<span style=\"color: grey\" class=\"oi oi-pin\" title=\"Pin\" aria-hidden=\"true\" id=\"seqselpin"+i+"\" n=\""+i+"\"></span>"
+		}
+		text_data = text_data + "</td></tr></table>"
+		seqselcontents.append(text_data);
+		
+		//setup on-click actions
 		$("#seqsel"+i).on("click", function () {
 			socket.send({'cmd': 'seqsel', 'data': $(this).attr("n")});
 		});
+		
+		//onclick for pin only
+		if (seqs[i][1] != "redo") {
+			$("#seqselpin"+i).on("click", function () {
+				socket.send({'cmd': 'seqpin', 'data': $(this).attr("n")});
+				if ($(this).attr("style") == "color: grey") {
+					console.log($(this).attr("style"));
+					$(this).css({"color": "white"});
+					console.log($(this).attr("style"));
+				} else {
+					console.log($(this).attr("style"));
+					$(this).css({"color": "grey"});
+					console.log($(this).attr("style"));
+				}
+			});
+		}
 	}
 	$('#seqselmenu').slideDown("slow");
 }
@@ -1756,6 +1840,7 @@ $(document).ready(function(){
 	button_send       = $('#btnsend');
 	button_actmem     = $('#btn_actmem');
 	button_actback    = $('#btn_actundo');
+	button_actfwd     = $('#btn_actredo');
 	button_actretry   = $('#btn_actretry');
 	button_actwi      = $('#btn_actwi');
 	game_text         = $('#gametext');
@@ -1765,6 +1850,7 @@ $(document).ready(function(){
 	settings_menu     = $("#settingsmenu");
 	format_menu       = $('#formatmenu');
 	anote_menu        = $('#anoterowcontainer');
+	debug_area        = $('#debugcontainer');
 	wi_menu           = $('#wimenu');
 	anote_input       = $('#anoteinput');
 	anote_labelcur    = $('#anotecur');
@@ -1780,6 +1866,7 @@ $(document).ready(function(){
 	aidg_close        = $("#btn_aidgpopupclose");
 	saveaspopup       = $("#saveascontainer");
 	saveasinput       = $("#savename");
+	savepins          = $("#savepins");
 	topic             = $("#topic");
 	saveas_accept     = $("#btn_saveasaccept");
 	saveas_close      = $("#btn_saveasclose");
@@ -1928,13 +2015,13 @@ $(document).ready(function(){
 			// Enable or Disable buttons
 			if(msg.data == "ready") {
 				enableSendBtn();
-				enableButtons([button_actmem, button_actwi, button_actback, button_actretry]);
+				enableButtons([button_actmem, button_actwi, button_actback, button_actfwd, button_actretry]);
 				hideWaitAnimation();
 				gamestate = "ready";
 			} else if(msg.data == "wait") {
 				gamestate = "wait";
 				disableSendBtn();
-				disableButtons([button_actmem, button_actwi, button_actback, button_actretry]);
+				disableButtons([button_actmem, button_actwi, button_actback, button_actfwd, button_actretry]);
 				showWaitAnimation();
 			} else if(msg.data == "start") {
 				setStartState();
@@ -1988,74 +2075,81 @@ $(document).ready(function(){
 			newTextHighlight($("#n"+msg.data))
 		} else if(msg.cmd == "updatetemp") {
 			// Send current temp value to input
-			$("#settemp").val(parseFloat(msg.data));
-			$("#settempcur").html(msg.data);
+			$("#settempcur").val(msg.data);
+			$("#settemp").val(parseFloat(msg.data)).trigger("change");
 		} else if(msg.cmd == "updatetopp") {
 			// Send current top p value to input
-			$("#settopp").val(parseFloat(msg.data));
-			$("#settoppcur").html(msg.data);
+			$("#settoppcur").val(msg.data);
+			$("#settopp").val(parseFloat(msg.data)).trigger("change");
 		} else if(msg.cmd == "updatetopk") {
 			// Send current top k value to input
-			$("#settopk").val(parseFloat(msg.data));
-			$("#settopkcur").html(msg.data);
+			$("#settopkcur").val(msg.data);
+			$("#settopk").val(parseFloat(msg.data)).trigger("change");
 		} else if(msg.cmd == "updatetfs") {
 			// Send current tfs value to input
-			$("#settfs").val(parseFloat(msg.data));
-			$("#settfscur").html(msg.data);
+			$("#settfscur").val(msg.data);
+			$("#settfs").val(parseFloat(msg.data)).trigger("change");
+		} else if(msg.cmd == "updatetypical") {
+			// Send current typical value to input
+			$("#settypicalcur").val(msg.data);
+			$("#settypical").val(parseFloat(msg.data)).trigger("change");
 		} else if(msg.cmd == "updatereppen") {
 			// Send current rep pen value to input
-			$("#setreppen").val(parseFloat(msg.data));
-			$("#setreppencur").html(msg.data);
+			$("#setreppencur").val(msg.data);
+			$("#setreppen").val(parseFloat(msg.data)).trigger("change");
 		} else if(msg.cmd == "updatereppenslope") {
 			// Send current rep pen value to input
-			$("#setreppenslope").val(parseFloat(msg.data));
-			$("#setreppenslopecur").html(msg.data);
+			$("#setreppenslopecur").val(msg.data);
+			$("#setreppenslope").val(parseFloat(msg.data)).trigger("change");
 		} else if(msg.cmd == "updatereppenrange") {
 			// Send current rep pen value to input
-			$("#setreppenrange").val(parseFloat(msg.data));
-			$("#setreppenrangecur").html(msg.data);
+			$("#setreppenrangecur").val(msg.data);
+			$("#setreppenrange").val(parseFloat(msg.data)).trigger("change");
 		} else if(msg.cmd == "updateoutlen") {
 			// Send current output amt value to input
-			$("#setoutput").val(parseInt(msg.data));
-			$("#setoutputcur").html(msg.data);
+			$("#setoutputcur").val(msg.data);
+			$("#setoutput").val(parseInt(msg.data)).trigger("change");
 		} else if(msg.cmd == "updatetknmax") {
 			// Send current max tokens value to input
-			$("#settknmax").val(parseInt(msg.data));
-			$("#settknmaxcur").html(msg.data);
+			$("#settknmaxcur").val(msg.data);
+			$("#settknmax").val(parseInt(msg.data)).trigger("change");
 		} else if(msg.cmd == "updateikgen") {
 			// Send current max tokens value to input
-			$("#setikgen").val(parseInt(msg.data));
-			$("#setikgencur").html(msg.data);
+			$("#setikgencur").val(msg.data);
+			$("#setikgen").val(parseInt(msg.data)).trigger("change");
 		} else if(msg.cmd == "setlabeltemp") {
 			// Update setting label with value from server
-			$("#settempcur").html(msg.data);
+			$("#settempcur").val(msg.data);
 		} else if(msg.cmd == "setlabeltopp") {
 			// Update setting label with value from server
-			$("#settoppcur").html(msg.data);
+			$("#settoppcur").val(msg.data);
 		} else if(msg.cmd == "setlabeltopk") {
 			// Update setting label with value from server
-			$("#settopkcur").html(msg.data);
+			$("#settopkcur").val(msg.data);
 		} else if(msg.cmd == "setlabeltfs") {
 			// Update setting label with value from server
-			$("#settfscur").html(msg.data);
+			$("#settfscur").val(msg.data);
+		} else if(msg.cmd == "setlabeltypical") {
+			// Update setting label with value from server
+			$("#settypicalcur").val(msg.data);
 		} else if(msg.cmd == "setlabelreppen") {
 			// Update setting label with value from server
-			$("#setreppencur").html(msg.data);
+			$("#setreppencur").val(msg.data);
 		} else if(msg.cmd == "setlabelreppenslope") {
 			// Update setting label with value from server
-			$("#setreppenslopecur").html(msg.data);
+			$("#setreppenslopecur").val(msg.data);
 		} else if(msg.cmd == "setlabelreppenrange") {
 			// Update setting label with value from server
-			$("#setreppenrangecur").html(msg.data);
+			$("#setreppenrangecur").val(msg.data);
 		} else if(msg.cmd == "setlabeloutput") {
 			// Update setting label with value from server
-			$("#setoutputcur").html(msg.data);
+			$("#setoutputcur").val(msg.data);
 		} else if(msg.cmd == "setlabeltknmax") {
 			// Update setting label with value from server
-			$("#settknmaxcur").html(msg.data);
+			$("#settknmaxcur").val(msg.data);
 		} else if(msg.cmd == "setlabelikgen") {
 			// Update setting label with value from server
-			$("#setikgencur").html(msg.data);
+			$("#setikgencur").val(msg.data);
 		} else if(msg.cmd == "updateanotedepth") {
 			// Send current Author's Note depth value to input
 			anote_slider.val(parseInt(msg.data));
@@ -2226,15 +2320,15 @@ $(document).ready(function(){
 			$("#setnumseqcur").html(msg.data);
 		} else if(msg.cmd == "updatenumseq") {
 			// Send current max tokens value to input
-			$("#setnumseq").val(parseInt(msg.data));
 			$("#setnumseqcur").html(msg.data);
+			$("#setnumseq").val(parseInt(msg.data)).trigger("change");
 		} else if(msg.cmd == "setlabelwidepth") {
 			// Update setting label with value from server
 			$("#setwidepthcur").html(msg.data);
 		} else if(msg.cmd == "updatewidepth") {
 			// Send current max tokens value to input
-			$("#setwidepth").val(parseInt(msg.data));
 			$("#setwidepthcur").html(msg.data);
+			$("#setwidepth").val(parseInt(msg.data)).trigger("change");
 		} else if(msg.cmd == "updateuseprompt") {
 			// Update toggle state
 			$("#setuseprompt").prop('checked', msg.data).change();
@@ -2269,6 +2363,14 @@ $(document).ready(function(){
 		} else if(msg.cmd == "runs_remotely") {
 			remote = true;
 			hide([button_savetofile, button_import, button_importwi]);
+		} else if(msg.cmd == "debug_info") {
+			$("#debuginfo").val(msg.data);
+		} else if(msg.cmd == "set_debug") {
+			if(msg.data) {
+				debug_area.removeClass("hidden");
+			} else {
+				debug_area.addClass("hidden");
+			}
 		}
 	});
 	
@@ -2349,6 +2451,12 @@ $(document).ready(function(){
 		hidegenseqs();
 	});
 	
+	button_actfwd.on("click", function(ev) {
+		hideMessage();
+		//hidegenseqs();
+		socket.send({'cmd': 'redo', 'data': ''});
+	});
+	
 	button_actmem.on("click", function(ev) {
 		socket.send({'cmd': 'memory', 'data': ''});
 	});

static/custom.css

@@ -22,6 +22,25 @@ chunk.editing, chunk.editing * {
 	font-style: normal !important;
 }
 
+.setting-value-warning {
+	color: #ff7777;
+}
+
+.setting-value-warning:focus {
+	color: #ffaaaa !important;
+}
+
+.settinglabel input {
+	width: 5ch;
+	background-color: inherit;
+	border: none;
+	outline: none;
+}
+
+.settinglabel input:focus {
+	color: #cdf;
+}
+
 #gametext, chunk, chunk * {
 	outline: 0px solid transparent;
 }
@@ -1273,8 +1292,8 @@ body.connected .popupfooter, .popupfooter.always-available {
 
 .settinglabel {
 	color: #ffffff;
-	display: grid;
-	grid-template-columns: 80% 20%;
+	display: flex;
+	flex-flow: wrap;
 }
 
 .settingminmax {

structures.py

@@ -19,10 +19,16 @@ class KoboldStoryRegister(collections.OrderedDict):
         return self.popitem()[1]
     
     def get_first_key(self) -> int:
-        return next(iter(self))
+        if len(self) == 0:
+            return -1
+        else:
+            return next(iter(self))
 
     def get_last_key(self) -> int:
-        return next(reversed(self))
+        if len(self) == 0:
+            return -1
+        else:
+            return next(reversed(self))
 
     def __getitem__(self, k: int) -> str:
         return super().__getitem__(k)

templates/index.html

@@ -9,7 +9,7 @@
 	<link rel="stylesheet" href="static/bootstrap.min.css">
 	<link rel="stylesheet" href="static/bootstrap-toggle.min.css">
 	<link rel="stylesheet" href="static/open-iconic-bootstrap.min.css">
-	<link rel="stylesheet" href="static/custom.css?ver=1.17">
+	<link rel="stylesheet" href="static/custom.css?ver=1.18b">
 
 	<script src="static/jquery-3.6.0.min.js"></script>
 	<script src="static/jquery-ui.sortable.min.js"></script>
@@ -17,7 +17,7 @@
 	<script src="static/bootstrap.min.js"></script>
 	<script src="static/bootstrap-toggle.min.js"></script>
 	<script src="static/rangy-core.min.js"></script>
-	<script src="static/application.js?ver=1.17a"></script>
+	<script src="static/application.js?ver=1.18b"></script>
 </head>
 <body>
 	<input type="file" id="remote-save-select" accept="application/json" style="display:none">
@@ -123,6 +123,7 @@
 				<button type="button" class="btn btn-primary" id="btn_actmem">Memory</button>
 				<button type="button" class="btn btn-primary" id="btn_actwi">W Info</button>
 				<button type="button" class="btn btn-primary" id="btn_actundo">Back</button>
+				<button type="button" class="btn btn-primary" id="btn_actredo">Redo</button>
 				<button type="button" class="btn btn-primary" id="btn_actretry">Retry</button>
 			</div>
 			<input type="text" id="chatname" class="form-control hidden" placeholder="Chat name">
@@ -185,6 +186,9 @@
 					</div>
 				</div>
 			</div>
+			<div  class="hidden" id="debugcontainer">
+				<textarea class="form-control" placeholder="Debug Info" id="debuginfo"></textarea>
+			</div>
 		</div>
 	</div>
 	<div class="hidden" id="popupcontainer">
@@ -228,7 +232,9 @@
 				<div class="popuptitletext">Enter Name For Save</div>
 			</div>
 			<div class="aidgpopupcontent">
-				<input class="form-control" type="text" placeholder="Save Name" id="savename">
+				<input class="form-control" type="text" placeholder="Save Name" id="savename"><br>
+				<input type="checkbox" data-toggle="toggle" data-onstyle="success" id="savepins" checked>
+				<div class="box-label">Save Pin Information</div>
 			</div>
 			<div class="popuperror hidden">
 				<span></span>

torch_lazy_loader.py

@@ -0,0 +1,268 @@
+'''
+This file is AGPL-licensed.
+
+Some of the code in this file is copied from PyTorch.
+
+The license for PyTorch is shown below:
+
+Copyright (c) 2016-     Facebook, Inc            (Adam Paszke)
+Copyright (c) 2014-     Facebook, Inc            (Soumith Chintala)
+Copyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert)
+Copyright (c) 2012-2014 Deepmind Technologies    (Koray Kavukcuoglu)
+Copyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu)
+Copyright (c) 2011-2013 NYU                      (Clement Farabet)
+Copyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou, Iain Melvin, Jason Weston)
+Copyright (c) 2006      Idiap Research Institute (Samy Bengio)
+Copyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert, Samy Bengio, Johnny Mariethoz)
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+3. Neither the names of Facebook, Deepmind Technologies, NYU, NEC Laboratories America
+   and IDIAP Research Institute nor the names of its contributors may be
+   used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+'''
+
+
+import contextlib
+from functools import reduce
+import itertools
+import zipfile
+import pickle
+import torch
+from torch.nn import Module
+from typing import Any, Callable, Dict, Optional, Tuple, Type, Union
+
+
+_EXTRA_STATE_KEY_SUFFIX = '_extra_state'
+
+
+STORAGE_TYPE_MAP = {
+    torch.float64: torch.DoubleStorage,
+    torch.float32: torch.FloatStorage,
+    torch.float16: torch.HalfStorage,
+    torch.int64: torch.LongStorage,
+    torch.int32: torch.IntStorage,
+    torch.int16: torch.ShortStorage,
+    torch.int8: torch.CharStorage,
+    torch.uint8: torch.ByteStorage,
+    torch.bool: torch.BoolStorage,
+    torch.bfloat16: torch.BFloat16Storage,
+}
+
+
+class LazyTensor:
+    def __init__(self, storage_type: Type[torch._StorageBase], key: str, location: str, dtype: Optional[torch.dtype] = None, seek_offset: Optional[int] = None, shape: Optional[Tuple[int, ...]] = None, stride: Optional[Tuple[int, ...]] = None, requires_grad=False, backward_hooks: Any = None):
+        self.storage_type = storage_type
+        self.key = key
+        self.location = location
+        self.dtype = dtype
+        self.seek_offset = seek_offset
+        self.shape = shape
+        self.stride = stride
+        self.requires_grad = requires_grad
+        self.backward_hooks = backward_hooks
+
+    def __view(self, f: Callable):
+        return f"{type(self).__name__}(storage_type={f(self.storage_type)}, key={f(self.key)}, location={f(self.location)}, dtype={f(self.dtype)}, seek_offset={f(self.seek_offset)}, shape={f(self.shape)}, stride={f(self.stride)}, requires_grad={f(self.requires_grad)}, backward_hooks={f(self.backward_hooks)})"
+
+    def __repr__(self):
+        return self.__view(repr)
+
+    def materialize(self, checkpoint: Union[zipfile.ZipFile, zipfile.ZipExtFile], map_location=None, no_grad=True) -> torch.Tensor:
+        size = reduce(lambda x, y: x * y, self.shape, 1)
+        dtype = self.dtype
+        nbytes = size if dtype is torch.bool else size * ((torch.finfo if dtype.is_floating_point else torch.iinfo)(dtype).bits >> 3)
+        if isinstance(checkpoint, zipfile.ZipFile):
+            f = checkpoint.open(f"archive/data/{self.key}", "r")
+            f.read(self.seek_offset)
+        else:
+            f = checkpoint
+        try:
+            storage = STORAGE_TYPE_MAP[dtype].from_buffer(f.read(nbytes), "little")
+        finally:
+            if isinstance(checkpoint, zipfile.ZipFile):
+                f.close()
+        storage = torch.serialization._get_restore_location(map_location)(storage, self.location)
+        tensor = torch.tensor([], dtype=storage.dtype, device=storage.device)
+        tensor.set_(storage, 0, self.shape, self.stride)
+        tensor.requires_grad = not no_grad and self.requires_grad
+        tensor._backward_hooks = self.backward_hooks
+        return tensor
+
+
+class _LazyUnpickler(pickle.Unpickler):
+    lazy_loaded_storages: Dict[str, LazyTensor]
+
+    def __init__(self, *args, **kwargs):
+        self.lazy_loaded_storages = {}
+        return super().__init__(*args, **kwargs)
+
+    def forced_persistent_load(self, saved_id):
+        assert isinstance(saved_id, tuple)
+        typename = saved_id[0]
+        assert typename == "storage", f"Unknown typename for persistent_load, expected 'storage' but got '{typename}'"
+        storage_type, key, location, _ = saved_id[1:]
+        return LazyTensor(storage_type, key, location)
+
+    def load(self, *args, **kwargs):
+        self.persistent_load = self.forced_persistent_load
+        retval = super().load(*args, **kwargs)
+        self.lazy_loaded_storages = {}
+        return retval
+
+
+def _rebuild_tensor(lazy_storage: LazyTensor, storage_offset, shape, stride):
+    lazy_storage.shape = shape
+    lazy_storage.stride = stride
+    dtype = lazy_storage.storage_type.dtype
+    if not isinstance(dtype, torch.dtype):
+        dtype = lazy_storage.storage_type(0).dtype
+    lazy_storage.dtype = dtype
+    lazy_storage.seek_offset = storage_offset if dtype is torch.bool else storage_offset * ((torch.finfo if dtype.is_floating_point else torch.iinfo)(dtype).bits >> 3)
+    return lazy_storage
+
+
+# Modified version of https://github.com/pytorch/pytorch/blob/v1.11.0-rc4/torch/nn/modules/module.py#L1346-L1438
+def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs):
+    for hook in self._load_state_dict_pre_hooks.values():
+        hook(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)
+
+    persistent_buffers = {k: v for k, v in self._buffers.items() if k not in self._non_persistent_buffers_set}
+    local_name_params = itertools.chain(self._parameters.items(), persistent_buffers.items())
+    local_state = {k: v for k, v in local_name_params if v is not None}
+
+    for name, param in local_state.items():
+        key = prefix + name
+        if key in state_dict:
+            input_param = state_dict[key]
+            if not torch.overrides.is_tensor_like(input_param):
+                error_msgs.append('While copying the parameter named "{}", '
+                                    'expected torch.Tensor or Tensor-like object from checkpoint but '
+                                    'received {}'
+                                    .format(key, type(input_param)))
+                continue
+
+            # This is used to avoid copying uninitialized parameters into
+            # non-lazy modules, since they dont have the hook to do the checks
+            # in such case, it will error when accessing the .shape attribute.
+            is_param_lazy = torch.nn.parameter.is_lazy(param)
+            # Backward compatibility: loading 1-dim tensor from 0.3.* to version 0.4+
+            if not is_param_lazy and len(param.shape) == 0 and len(input_param.shape) == 1:
+                input_param = input_param[0]
+
+            if not is_param_lazy and input_param.shape != param.shape:
+                # local shape should match the one in checkpoint
+                error_msgs.append('size mismatch for {}: copying a param with shape {} from checkpoint, '
+                                    'the shape in current model is {}.'
+                                    .format(key, input_param.shape, param.shape))
+                continue
+            try:
+                with torch.no_grad():
+                    #param.copy_(input_param)
+                    new_param = torch.nn.Parameter(input_param, requires_grad=param.requires_grad)  # This line is new
+                    if name in self._parameters:  # This line is new
+                        self._parameters[name] = new_param  # This line is new
+                    if name in persistent_buffers:  # This line is new
+                        self._buffers[name] = new_param  # This line is new
+            except Exception as ex:
+                error_msgs.append('While copying the parameter named "{}", '
+                                    'whose dimensions in the model are {} and '
+                                    'whose dimensions in the checkpoint are {}, '
+                                    'an exception occurred : {}.'
+                                    .format(key, param.size(), input_param.size(), ex.args))
+        elif strict:
+            missing_keys.append(key)
+
+    extra_state_key = prefix + _EXTRA_STATE_KEY_SUFFIX
+    if hasattr(Module, "set_extra_state") and getattr(self.__class__, "set_extra_state", Module.set_extra_state) is not Module.set_extra_state:  # if getattr(self.__class__, "set_extra_state", Module.set_extra_state) is not Module.set_extra_state:
+        if extra_state_key in state_dict:
+            self.set_extra_state(state_dict[extra_state_key])
+        elif strict:
+            missing_keys.append(extra_state_key)
+    elif strict and (extra_state_key in state_dict):
+        unexpected_keys.append(extra_state_key)
+
+    if strict:
+        for key in state_dict.keys():
+            if key.startswith(prefix) and key != extra_state_key:
+                input_name = key[len(prefix):]
+                input_name = input_name.split('.', 1)[0]  # get the name of param/buffer/child
+                if input_name not in self._modules and input_name not in local_state:
+                    unexpected_keys.append(key)
+
+
+@contextlib.contextmanager
+def use_lazy_torch_load(enable=True, callback: Optional[Callable] = None, dematerialized_modules=False):
+    if not enable:
+        yield False
+        return
+
+    try:
+        old_unpickler = pickle.Unpickler
+        pickle.Unpickler = _LazyUnpickler
+
+        old_rebuild_tensor = torch._utils._rebuild_tensor
+        torch._utils._rebuild_tensor = _rebuild_tensor
+
+        old_torch_load = torch.load
+
+        def torch_load(f, map_location=None, pickle_module=pickle, **pickle_load_args):
+            retval = old_torch_load(f=f, map_location=map_location, pickle_module=pickle_module, **pickle_load_args)
+            if callback is not None:
+                callback(retval, f=f, map_location=map_location, pickle_module=pickle_module, **pickle_load_args)
+            return retval
+
+        torch.load = torch_load
+
+        if dematerialized_modules:
+            old_linear_init = torch.nn.Linear.__init__
+            old_embedding_init = torch.nn.Embedding.__init__
+            old_layernorm_init = torch.nn.LayerNorm.__init__
+
+            def linear_init(self, *args, device=None, **kwargs):
+                return old_linear_init(self, *args, device="meta", **kwargs)
+
+            def embedding_init(self, *args, device=None, **kwargs):
+                return old_embedding_init(self, *args, device="meta", **kwargs)
+
+            def layernorm_init(self, *args, device=None, **kwargs):
+                return old_layernorm_init(self, *args, device="meta", **kwargs)
+
+            torch.nn.Linear.__init__ = linear_init
+            torch.nn.Embedding.__init__ = embedding_init
+            torch.nn.LayerNorm.__init__ = layernorm_init
+            old_load_from_state_dict = torch.nn.Module._load_from_state_dict
+            torch.nn.Module._load_from_state_dict = _load_from_state_dict
+
+        yield True
+
+    finally:
+        pickle.Unpickler = old_unpickler
+        torch._utils._rebuild_tensor = old_rebuild_tensor
+        torch.load = old_torch_load
+        if dematerialized_modules:
+            torch.nn.Linear.__init__ = old_linear_init
+            torch.nn.Embedding.__init__ = old_embedding_init
+            torch.nn.LayerNorm.__init__ = old_layernorm_init
+            torch.nn.Module._load_from_state_dict = old_load_from_state_dict

tpu_mtj_backend.py

@@ -27,23 +27,31 @@ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 SOFTWARE.
 '''
 
+import utils
+
 import multiprocessing
 from typing import Any, Callable, Dict, List, Optional, Tuple, TypeVar
 import progressbar
 import time
 import os
+import sys
+import json
+import zipfile
 import requests
 import random
 import jax
+import jax.dlpack
 from jax.config import config
 from jax.experimental import maps
 import jax.numpy as jnp
 import numpy as np
 import optax
 import haiku as hk
-import transformers
+from transformers import AutoTokenizer, GPT2TokenizerFast, AutoModelForCausalLM, GPTNeoForCausalLM
+from tokenizers import Tokenizer
 from mesh_transformer.checkpoint import read_ckpt_lowmem
-from mesh_transformer.transformer_shard import CausalTransformer, CausalTransformerShard
+from mesh_transformer.transformer_shard import CausalTransformer, CausalTransformerShard, PlaceholderTensor
+from mesh_transformer.util import to_bf16
 
 
 params: Dict[str, Any] = {}
@@ -61,6 +69,7 @@ def settings_callback() -> dict:
         "temp": 0.5,
         "top_k": 0,
         "tfs": 1.0,
+        "typical": 1.0,
         "repetition_penalty": 1.0,
         "rpslope": 0.0,
         "rprange": 0,
@@ -149,11 +158,11 @@ def apply_repetition_penalty_dynamic(logits, tokens, repetition_penalty, generat
     logits[tokens] = penalty_logits
     return logits
 
-def kobold_sample_dynamic(key, logits, top_p=0.9, temp=0.5, top_k=0, tfs=1.0):
+def kobold_sample_dynamic(key, logits, top_p=0.9, temp=0.5, top_k=0, tfs=1.0, typical=1.0):
     '''
-    This gets called by generate_loop_fn to apply a series of 4 filters
-    to the logits (top-k, then top-p, then TFS, then temperature) before
-    picking one token using the modified logits
+    This gets called by generate_loop_fn to apply a series of 5 filters
+    to the logits (top-k, then top-p, then TFS, then typical, then temperature)
+    before picking one token using the modified logits
     '''
     # Top-k (keep only the k tokens with the highest logits and remove
     # the rest, by setting their logits to negative infinity)
@@ -240,6 +249,37 @@ def kobold_sample_dynamic(key, logits, top_p=0.9, temp=0.5, top_k=0, tfs=1.0):
         return np.where(indices_to_remove, -np.inf, logits)
     if tfs < 1.0:
         logits = tail_free_filter(logits)
+    # Typical sampling (https://arxiv.org/pdf/2202.00666.pdf)
+    def typical_filter(logits):
+        # Compute softmax probabilities and the natural logarithms of them
+        probs = jax.nn.softmax(logits)
+        with np.errstate(divide="ignore"):
+            log_probs = np.log(probs)
+        # Compute the negative of entropy, which is the sum of p*ln(p) for all p
+        # in the set of softmax probabilities of the logits
+        neg_entropy = np.nansum(probs * log_probs, axis=-1, keepdims=True)
+        # Determine absolute difference between the negative entropy and the
+        # log probabilities
+        entropy_deviation = np.abs(neg_entropy - log_probs)
+        # Keep certain tokens such that the sum of the entropy_deviation of the
+        # kept tokens is the smallest possible value such that the sum of the
+        # softmax probabilities of the kept tokens is at least the threshold
+        # value (by sorting the tokens in ascending order of entropy_deviation
+        # and then keeping the smallest possible number of tokens from the
+        # beginning such that sum of softmax probabilities is at or above the
+        # threshold)
+        _, sorted_logits = jax.lax.sort_key_val(entropy_deviation, probs)
+        sorted_indices_to_remove = np.cumsum(sorted_logits, axis=-1) >= typical
+        sorted_indices_to_remove = np.roll(sorted_indices_to_remove, 1, axis=-1)
+        sorted_indices_to_remove[0] = False
+        # Unsort and remove
+        _, indices_to_remove = jax.lax.sort_key_val(
+            jnp.argsort(entropy_deviation),
+            sorted_indices_to_remove,
+        )
+        return np.where(indices_to_remove, -jnp.inf, logits)
+    if typical < 1.0:
+        logits = typical_filter(logits)
     # Temperature (just divide the logits by the temperature)
     logits /= temp
     # Finally, pick one token using the softmax thingy again (it gives
@@ -292,11 +332,11 @@ 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, top_p=0.9, temp=0.5, top_k=0, tfs=1.0):
+def kobold_sample_static(key, logits, top_p=0.9, temp=0.5, top_k=0, tfs=1.0, typical=1.0):
     '''
-    This gets called by generate_loop_fn to apply a series of 4 filters
-    to the logits (top-k, then top-p, then TFS, then temperature) before
-    picking one token using the modified logits
+    This gets called by generate_loop_fn to apply a series of 5 filters
+    to the logits (top-k, then top-p, then TFS, then typical, then temperature)
+    before picking one token using the modified logits
     '''
     # Top-k (keep only the k tokens with the highest logits and remove
     # the rest, by setting their logits to negative infinity)
@@ -380,6 +420,35 @@ def kobold_sample_static(key, logits, top_p=0.9, temp=0.5, top_k=0, tfs=1.0):
         )
         return jnp.where(indices_to_remove, -jnp.inf, logits)
     logits = jax.lax.cond(tfs < 1.0, tail_free_filter, lambda x: x, logits)
+    # Typical sampling (https://arxiv.org/pdf/2202.00666.pdf)
+    def typical_filter(logits):
+        # Compute softmax probabilities and the natural logarithms of them
+        probs = jax.nn.softmax(logits)
+        log_probs = jnp.log(probs)
+        # Compute the negative of entropy, which is the sum of p*ln(p) for all p
+        # in the set of softmax probabilities of the logits
+        neg_entropy = jnp.nansum(probs * log_probs, axis=-1, keepdims=True)
+        # Determine absolute difference between the negative entropy and the
+        # log probabilities
+        entropy_deviation = jnp.abs(neg_entropy - log_probs)
+        # Keep certain tokens such that the sum of the entropy_deviation of the
+        # kept tokens is the smallest possible value such that the sum of the
+        # softmax probabilities of the kept tokens is at least the threshold
+        # value (by sorting the tokens in ascending order of entropy_deviation
+        # and then keeping the smallest possible number of tokens from the
+        # beginning such that sum of softmax probabilities is at or above the
+        # threshold)
+        _, sorted_logits = jax.lax.sort_key_val(entropy_deviation, probs)
+        sorted_indices_to_remove = jnp.cumsum(sorted_logits, axis=-1) >= typical
+        sorted_indices_to_remove = jnp.roll(sorted_indices_to_remove, 1, axis=-1)
+        sorted_indices_to_remove = sorted_indices_to_remove.at[0].set(False)
+        # Unsort and remove
+        _, indices_to_remove = jax.lax.sort_key_val(
+            jnp.argsort(entropy_deviation),
+            sorted_indices_to_remove,
+        )
+        return jnp.where(indices_to_remove, -jnp.inf, logits)
+    logits = jax.lax.cond(typical < 1.0, typical_filter, lambda x: x, logits)
     # Temperature (just divide the logits by the temperature)
     def temp_filter(logits):
         return logits / temp
@@ -443,9 +512,9 @@ def sample_func(data, key, numseqs_aux, badwords, repetition_penalty, generated_
     return carry
 
 class PenalizingCausalTransformer(CausalTransformer):
-    def __init__(self, config):
+    def __init__(self, config, **kwargs):
         # Initialize
-        super().__init__(config)
+        super().__init__(config, **kwargs)
         def generate_static(state, key, ctx, ctx_length, gen_length, numseqs_aux, sampler_options, soft_embeddings=None):
             compiling_callback()
             numseqs = numseqs_aux.shape[0]
@@ -736,6 +805,7 @@ def infer_static(
     temp=0.5,
     top_k=0,
     tfs=1.0,
+    typical=1.0,
     repetition_penalty=1.0,
     rpslope=0.0,
     rprange=0,
@@ -758,6 +828,7 @@ def infer_static(
         "temp": temp * np.ones(total_batch),
         "top_p": top_p * np.ones(total_batch),
         "tfs": tfs * np.ones(total_batch),
+        "typical": typical * np.ones(total_batch),
         "repetition_penalty": repetition_penalty * np.ones(total_batch),
         "rpslope": rpslope * np.ones(total_batch),
         "rprange": np.full(total_batch, rprange, dtype=np.uint32),
@@ -776,7 +847,142 @@ def infer_static(
     return samples
 
 
-def load_model(path: str, driver_version="tpu_driver0.1_dev20210607", **kwargs) -> None:
+def reshard_reverse(x, total_shards, old_shape):
+    assert len(x.shape) != 1
+    if len(x.shape) == 2:
+        if old_shape[1] == x.shape[1]:
+            out = x[0:1].tile((total_shards, 1))
+        else:
+            out = x.reshape(old_shape)
+    elif len(x.shape) == 3:
+        if x.shape[0] * x.shape[2] == old_shape[2]:
+            out = x.reshape(old_shape)
+        elif x.shape[0] * x.shape[1] == old_shape[1]:
+            out = x.reshape((old_shape[1], old_shape[0], old_shape[2])).permute((1, 0, 2))
+        else:
+            assert False
+    else:
+        assert False
+    return out
+
+
+def get_old_shape(t, total_shards, dim=2):
+    if len(t.shape) == 2:
+        shard_shape = t.shape
+        if dim == 1:
+            assert shard_shape[0] % total_shards == 0
+            return (shard_shape[0] // total_shards, shard_shape[1])
+        elif dim == 2:
+            assert shard_shape[1] % total_shards == 0
+            return (shard_shape[0], shard_shape[1] // total_shards)
+        else:
+            raise ValueError(f"Unsupported dim {dim}")
+    if len(t.shape) == 1:
+        assert t.shape[0] % total_shards == 0
+        return (t.shape[0] // total_shards,)
+    else:
+        raise ValueError(f"Unsupported shape {t.shape}")
+
+
+def read_neox_checkpoint(state, path, config, checkpoint_shards=2):
+    assert config["cores_per_replica"] % checkpoint_shards == 0
+    output_shards = config["cores_per_replica"] // checkpoint_shards
+
+    import torch
+    import torch.utils.dlpack
+    from tqdm.auto import tqdm
+
+    move_xmap = jax.experimental.maps.xmap(
+        fun=lambda x, _: to_bf16(x),
+        in_axes=(["shard", ...], ["batch", ...]),
+        out_axes=["shard", ...],
+        axis_resources={'shard': 'mp', 'batch': 'dp'}
+    )
+
+    path_template = os.path.join(path, "layer_{layer:02d}-model_{shard:02d}-model_states.pt")
+
+    static_mapping = {
+        "word_embeddings.weight": {"module": "embedding_shard/~/linear", "param": "w", "axis": 1},
+        "final_linear.weight": {"module": "projection_shard/~/linear", "param": "w", "axis": 2},
+        "norm.weight": {"module": "projection_shard/~/replicated_layer_norm", "param": "scale", "axis": None},
+        "norm.bias": {"module": "projection_shard/~/replicated_layer_norm", "param": "offset", "axis": None},
+    }
+
+    layer_mapping = {
+        "attention.query_key_value.weight": {"module": "combined_qkv", "param": "w", "axis": 2},
+        "attention.query_key_value.bias": {"module": "combined_qkv", "param": "b", "axis": 1},
+        "attention.dense.weight": {"module": "linear_3", "param": "w", "axis": 1},
+        "attention.dense.bias": {"module": "linear_3", "param": "b", "axis": None},
+        "mlp.dense_h_to_4h.weight": {"module": "linear_4", "param": "w", "axis": 2},
+        "mlp.dense_h_to_4h.bias": {"module": "linear_4", "param": "b", "axis": 1},
+        "mlp.dense_4h_to_h.weight": {"module": "linear_5", "param": "w", "axis": 1},
+        "mlp.dense_4h_to_h.bias": {"module": "linear_5", "param": "b", "axis": None},
+        "input_layernorm.weight": {"module": "replicated_layer_norm", "param": "scale", "axis": None},
+        "input_layernorm.bias": {"module": "replicated_layer_norm", "param": "offset", "axis": None},
+        "post_attention_layernorm.weight": {"module": "replicated_layer_norm_1", "param": "scale", "axis": None},
+        "post_attention_layernorm.bias": {"module": "replicated_layer_norm_1", "param": "offset", "axis": None},
+    }
+
+    tqdm_length = len(static_mapping) + config["layers"]*len(layer_mapping)
+    bar = tqdm(total=tqdm_length, desc="Loading from NeoX checkpoint")
+
+    for checkpoint_layer in range(config["layers"] + 5):
+        if checkpoint_layer in (1, config["layers"] + 2):
+            continue
+        layer = checkpoint_layer - 2
+        shards = []
+        for checkpoint_shard in range(checkpoint_shards):
+            shards.append(torch.load(path_template.format(layer=checkpoint_layer, shard=checkpoint_shard), map_location="cpu"))
+        for key in shards[0]:
+            if key == "attention.rotary_emb.inv_freq":
+                continue
+            elif key in static_mapping:
+                target_module = "causal_transformer_shard/~/" + static_mapping[key]["module"]
+                target_param = static_mapping[key]["param"]
+                target_axis = static_mapping[key]["axis"]
+            elif key in layer_mapping:
+                target_module = f"causal_transformer_shard/~/layer_{layer}/~/" + layer_mapping[key]["module"]
+                target_param = layer_mapping[key]["param"]
+                target_axis = layer_mapping[key]["axis"]
+            else:
+                error = f"{repr(key)} not found in mapping"
+                print("\n\nERROR: ", error, file=sys.stderr)
+                raise RuntimeError(error)
+            original_shape = shards[0][key].shape
+            for checkpoint_shard in range(checkpoint_shards):
+                if key in ("attention.dense.bias", "mlp.dense_4h_to_h.bias"):
+                    shards[checkpoint_shard][key] /= output_shards
+                if key != "word_embeddings.weight" and shards[checkpoint_shard][key].ndim == 2:
+                    shards[checkpoint_shard][key] = shards[checkpoint_shard][key].T
+                tensor = shards[checkpoint_shard][key]
+                if target_axis is not None:
+                    target_shape = (output_shards,) + get_old_shape(tensor, total_shards=output_shards, dim=target_axis)
+                else:
+                    target_shape = (output_shards, tensor.shape[0])
+                shards[checkpoint_shard][key] = reshard_reverse(tensor.unsqueeze_(0), output_shards, target_shape)
+            #print(key, ":", original_shape, "->", shards[0][key].shape)
+            tensor = torch.cat([shards[s][key] for s in range(checkpoint_shards)], dim=0)
+            target_shape = state["params"][target_module][target_param].shape
+            if tensor.shape != target_shape:
+                error = f"Weight {repr(key)} has shape {tensor.shape} in checkpoint but shape {target_shape} was requested by MTJ for {target_module} {target_param}"
+                print("\n\nERROR: ", error, file=sys.stderr)
+                raise RuntimeError(error)
+            if tensor.dtype is torch.float16 or tensor.dtype is torch.float32:
+                tensor = tensor.bfloat16()
+            state["params"][target_module][target_param] = move_xmap(
+                jax.dlpack.from_dlpack(torch.utils.dlpack.to_dlpack(tensor)).copy(),
+                np.zeros(config["cores_per_replica"]),
+            )
+            bar.update(1)
+    for mk, mv in state["params"].items():
+        for pk, pv in mv.items():
+            if isinstance(pv, PlaceholderTensor):
+                error = f"{mk} {pk} could not be found in the model checkpoint"
+                print("\n\nERROR:  " + error, file=sys.stderr)
+                raise RuntimeError(error)
+
+
+def load_model(path: str, driver_version="tpu_driver0.1_dev20210607", hf_checkpoint=False, **kwargs) -> None:
     global thread_resources_env, seq, tokenizer, network, params
 
     default_params = {
@@ -791,12 +997,96 @@ def load_model(path: str, driver_version="tpu_driver0.1_dev20210607", **kwargs)
         "pe_rotary_dims": 64,
         "seq": 2048,
         "cores_per_replica": 8,
+        "tokenizer_class": "GPT2TokenizerFast",
+        "tokenizer": "gpt2",
     }
     params = kwargs
+
+    if vars.model == "TPUMeshTransformerGPTNeoX":
+        default_params = {
+            "compat": "neox",
+            "layers": 44,
+            "d_model": 6144,
+            "n_heads": 64,
+            "n_vocab": 50432,
+            "n_vocab_padding": 0,
+            "norm": "doublelayernorm",
+            "pe": "neox_rotary",
+            "pe_rotary_dims": 24,
+            "seq": 2048,
+            "cores_per_replica": 8,
+            "tokenizer_class": "GPT2TokenizerFast",
+            "tokenizer": "gpt2",
+        }
+
+    # Try to convert HF config.json to MTJ config
+    if hf_checkpoint:
+        spec_path = os.path.join("maps", vars.model_type + ".json")
+        if not os.path.isfile(spec_path):
+            raise NotImplementedError(f"Unsupported model type {repr(vars.model_type)}")
+        with open(spec_path) as f:
+            lazy_load_spec = json.load(f)
+
+        if "mtj_compat" in lazy_load_spec:
+            params["compat"] = lazy_load_spec["mtj_compat"]
+        if "mtj_pe" in lazy_load_spec:
+            params["pe"] = lazy_load_spec["mtj_pe"]
+        for k, v in lazy_load_spec.get("mtj_config_map", {}).items():
+            if type(v) is not list:
+                params[k] = params[v]
+                continue
+            for i in range(len(v)):
+                if i == len(v) - 1:
+                    params[k] = v[i]
+                elif v[i] in params:
+                    params[k] = params[v[i]]
+                    break
+
+        params["n_vocab"] = params["vocab_size"]
+
+        if "activation_function" in params:
+            params["activation"] = params["activation_function"]
+
+        # Both the number of attention heads in the model and the embedding
+        # dimension of the model need to be divisible by the number of TPU cores
+        # that we use, and JAX also requires the number of TPU cores used to be
+        # an even number if we're using more than one core, so logically we try
+        # to pick the largest possible even number of TPU cores such that the
+        # number of attention heads and embedding dimension are both divisible
+        # by the number of TPU cores, and fall back to one core if an even
+        # number of TPU cores is not possible.
+        for c in (8, 6, 4, 2, 1):
+            if 0 == params["n_heads"] % c == params.get("d_embed", params["d_model"]) % c:
+                params["cores_per_replica"] = c
+                break
+
+        # The vocabulary size of the model also has to be divisible by the
+        # number of TPU cores, so we pad the vocabulary with the minimum
+        # possible number of dummy tokens such that it's divisible.
+        params["n_vocab_padding"] = -(params["n_vocab"] % -params["cores_per_replica"])
+
+    if "compat" in params:
+        default_params["compat"] = params["compat"]
+    if default_params["compat"] == "fairseq_lm":
+        default_params["tokenizer"] = "KoboldAI/fairseq-dense-125M"
     for param in default_params:
         if param not in params:
             params[param] = default_params[param]
 
+    # Load tokenizer
+    if vars.model == "TPUMeshTransformerGPTNeoX":
+        tokenizer = Tokenizer.from_file(os.path.join(path, "20B_tokenizer.json"))
+        def new_encode(old_encode):
+            def encode(s, *args, **kwargs):
+                return old_encode(s).ids
+            return encode
+        tokenizer.encode = new_encode(tokenizer.encode)
+    elif not hf_checkpoint:
+        if not isinstance(params["tokenizer_class"], str) or not any(params["tokenizer_class"].endswith(s) for s in ("Tokenizer", "TokenizerFast")):
+            raise ValueError("`tokenizer_class` must be a string ending in 'Tokenizer' or 'TokenizerFast'")
+        tokenizer_class = getattr(__import__("transformers"), params["tokenizer_class"])
+        tokenizer = tokenizer_class.from_pretrained(params["tokenizer"])
+
     # Disable JAX warnings about these two functions having been renamed
     jax.host_count = jax.process_count
     jax.host_id = jax.process_index
@@ -804,13 +1094,18 @@ def load_model(path: str, driver_version="tpu_driver0.1_dev20210607", **kwargs)
     print("Connecting to your Colab instance's TPU", flush=True)
     spinner = multiprocessing.Process(target=show_spinner, args=())
     spinner.start()
-    colab_tpu_addr = os.environ['COLAB_TPU_ADDR'].split(':')[0]
-    url = f'http://{colab_tpu_addr}:8475/requestversion/{driver_version}'
+    if os.environ.get('COLAB_TPU_ADDR', '') != '':
+        tpu_address = os.environ['COLAB_TPU_ADDR']  # Colab
+    else:
+        tpu_address = os.environ['TPU_NAME']  # Kaggle
+    tpu_address = tpu_address.replace("grpc://", "")
+    tpu_address_without_port = tpu_address.split(':', 1)[0]
+    url = f'http://{tpu_address_without_port}:8475/requestversion/{driver_version}'
+    config.FLAGS.jax_xla_backend = "tpu_driver"
+    config.FLAGS.jax_backend_target = "grpc://" + tpu_address
     requests.post(url)
     spinner.terminate()
     print()
-    config.FLAGS.jax_xla_backend = "tpu_driver"
-    config.FLAGS.jax_backend_target = "grpc://" + os.environ['COLAB_TPU_ADDR']
 
     cores_per_replica = params["cores_per_replica"]
     seq = params["seq"]
@@ -819,7 +1114,6 @@ def load_model(path: str, driver_version="tpu_driver0.1_dev20210607", **kwargs)
     devices = np.array(jax.devices()[:cores_per_replica]).reshape(mesh_shape)
     thread_resources_env = maps.ResourceEnv(maps.Mesh(devices, ('dp', 'mp')), ())
     maps.thread_resources.env = thread_resources_env
-    tokenizer = transformers.GPT2TokenizerFast.from_pretrained('gpt2')
 
     global shard_xmap, batch_xmap
     shard_xmap = __shard_xmap()
@@ -832,6 +1126,198 @@ def load_model(path: str, driver_version="tpu_driver0.1_dev20210607", **kwargs)
     if not path.endswith("/"):
         path += "/"
 
-    network = PenalizingCausalTransformer(params)
-    network.state = read_ckpt_lowmem(network.state, path, devices.shape[1])
-    network.state = network.move_xmap(network.state, np.zeros(cores_per_replica))
+    network = PenalizingCausalTransformer(params, dematerialized=True)
+
+    if not hf_checkpoint and vars.model != "TPUMeshTransformerGPTNeoX":
+        network.state = read_ckpt_lowmem(network.state, path, devices.shape[1])
+        #network.state = network.move_xmap(network.state, np.zeros(cores_per_replica))
+        return
+
+    if vars.model == "TPUMeshTransformerGPTNeoX":
+        print("\n\n\nThis model has  ", f"{hk.data_structures.tree_size(network.state['params']):,d}".replace(",", " "), "  parameters.\n")
+        read_neox_checkpoint(network.state, path, params)
+        return
+
+    # Convert from HF checkpoint
+
+    move_xmap = jax.experimental.maps.xmap(
+        fun=lambda x, _: to_bf16(x),
+        in_axes=(["shard", ...], ["batch", ...]),
+        out_axes=["shard", ...],
+        axis_resources={'shard': 'mp', 'batch': 'dp'}
+    )
+
+    model_spec = {}
+    for key, spec in lazy_load_spec.get("static_weights", {}).items():
+        if spec.get("mtj") is not None:
+            model_spec[key] = spec["mtj"].copy()
+            model_spec[key]["module"] = "causal_transformer_shard/~/" + model_spec[key]["module"]
+    for _key, spec in lazy_load_spec.get("layer_weights", {}).items():
+        for layer in range(params["layers"]):
+            if spec.get("mtj") is not None:
+                key = _key.format(layer=layer)
+                model_spec[key] = spec["mtj"].copy()
+                model_spec[key]["module"] = "causal_transformer_shard/~/" + model_spec[key]["module"].format(layer=layer)
+
+    import torch_lazy_loader
+    import torch
+    from tqdm.auto import tqdm
+    import functools
+
+    def callback(model_dict, f, **_):
+        if callback.nested:
+            return
+        callback.nested = True
+        with zipfile.ZipFile(f, "r") as z:
+            try:
+                last_storage_key = None
+                f = None
+                current_offset = 0
+                if utils.current_shard == 0:
+                    print("\n\n\nThis model has  ", f"{hk.data_structures.tree_size(network.state['params']):,d}".replace(",", " "), "  parameters.\n")
+
+                if utils.num_shards is None or utils.current_shard == 0:
+                    if utils.num_shards is not None:
+                        num_tensors = len(utils.get_sharded_checkpoint_num_tensors(utils.from_pretrained_model_name, utils.from_pretrained_index_filename, **utils.from_pretrained_kwargs))
+                    else:
+                        num_tensors = len(model_dict)
+                    utils.bar = tqdm(total=num_tensors, desc="Loading model tensors")
+
+                if utils.num_shards is not None:
+                    utils.current_shard += 1
+                for key in sorted(model_dict.keys(), key=lambda k: (model_dict[k].key, model_dict[k].seek_offset)):
+
+                    # Some model weights are used by transformers but not by MTJ.
+                    # We have to materialize these weights anyways because
+                    # transformers will throw a tantrum otherwise.  To attain
+                    # the least possible memory usage, we create them as meta
+                    # tensors, which don't take up any actual CPU or TPU memory.
+                    if key not in model_spec:
+                        model_dict[key] = torch.empty(model_dict[key].shape, dtype=model_dict[key].dtype, device="meta")
+                        utils.bar.update(1)
+                        continue
+
+                    storage_key = model_dict[key].key
+                    if storage_key != last_storage_key or model_dict[key].seek_offset < current_offset:
+                        last_storage_key = storage_key
+                        if isinstance(f, zipfile.ZipExtFile):
+                            f.close()
+                        f = z.open(f"archive/data/{storage_key}")
+                        current_offset = 0
+                    if current_offset != model_dict[key].seek_offset:
+                        f.read(model_dict[key].seek_offset - current_offset)
+                        current_offset = model_dict[key].seek_offset
+                    spec = model_spec[key]
+                    transforms = set(spec.get("transforms", ()))
+                    if not isinstance(model_dict[key], torch_lazy_loader.LazyTensor):
+                        error = f"Duplicate key {repr(key)}"
+                        print("\n\nERROR:  " + error, file=sys.stderr)
+                        raise RuntimeError(error)
+                    size = functools.reduce(lambda x, y: x * y, model_dict[key].shape, 1)
+                    dtype = model_dict[key].dtype
+                    nbytes = size if dtype is torch.bool else size * ((torch.finfo if dtype.is_floating_point else torch.iinfo)(dtype).bits >> 3)
+                    tensor = model_dict[key].materialize(f, map_location="cpu")
+                    model_dict[key] = tensor.to("meta")
+                    current_offset += nbytes
+
+                    # MTJ requires certain mathematical operations to be performed
+                    # on tensors in order for them to be in the correct format
+                    if "remove_first_two_rows" in transforms:
+                        tensor = tensor[2:]
+                    if "divide_by_shards" in transforms:
+                        tensor /= params["cores_per_replica"]
+                    if "vocab_pad" in transforms:
+                        tensor = torch.nn.functional.pad(tensor, (0, 0, 0, params["n_vocab_padding"]))
+                    if "no_transpose" not in transforms and tensor.ndim == 2:
+                        tensor = tensor.T
+                    tensor.unsqueeze_(0)
+                    if tensor.dtype is torch.float16 or tensor.dtype is torch.float32:
+                        tensor = tensor.bfloat16()
+
+                    # Shard the tensor so that parts of the tensor can be used
+                    # on different TPU cores
+                    network.state["params"][spec["module"]][spec["param"]] = move_xmap(
+                        jax.dlpack.from_dlpack(torch.utils.dlpack.to_dlpack(
+                            reshard_reverse(
+                                tensor,
+                                params["cores_per_replica"],
+                                network.state["params"][spec["module"]][spec["param"]].shape,
+                            )
+                        )).copy(),
+                        np.empty(params["cores_per_replica"]),
+                    )
+
+                    utils.bar.update(1)
+
+                if utils.num_shards is not None and utils.current_shard < utils.num_shards:
+                    return
+
+                # Check for tensors that MTJ needs that were not provided in the
+                # HF model
+                for mk, mv in network.state["params"].items():
+                    for pk, pv in mv.items():
+                        if isinstance(pv, PlaceholderTensor):
+                            # The transformers GPT-J models apparently do not
+                            # have embedding bias, whereas MTJ GPT-J models do,
+                            # so we have to supplement an embedding bias tensor
+                            # by creating a tensor with the necessary shape, filled
+                            # with zeros.
+                            if mk == "causal_transformer_shard/~/embedding_shard/~/linear" and pk == "b":
+                                mv[pk] = move_xmap(jnp.zeros(mv[pk].shape, dtype=jnp.bfloat16), np.empty(params["cores_per_replica"]))
+
+                            else:
+                                error = f"{mk} {pk} could not be found in the model checkpoint"
+                                print("\n\nERROR:  " + error, file=sys.stderr)
+                                raise RuntimeError(error)
+            finally:
+                if utils.num_shards is None or utils.current_shard >= utils.num_shards:
+                    utils.bar.close()
+                    utils.bar = None
+                callback.nested = False
+                if isinstance(f, zipfile.ZipExtFile):
+                    f.close()
+    callback.nested = False
+
+    if os.path.isdir(vars.model.replace('/', '_')):
+        import shutil
+        shutil.move(vars.model.replace('/', '_'), "models/{}".format(vars.model.replace('/', '_')))
+    print("\n", flush=True)
+    with torch_lazy_loader.use_lazy_torch_load(callback=callback, dematerialized_modules=True):
+        if(os.path.isdir(vars.custmodpth)):
+            try:
+                tokenizer = AutoTokenizer.from_pretrained(vars.custmodpth, revision=vars.revision, cache_dir="cache")
+            except Exception as e:
+                try:
+                    tokenizer = GPT2TokenizerFast.from_pretrained(vars.custmodpth, revision=vars.revision, cache_dir="cache")
+                except Exception as e:
+                    tokenizer = GPT2TokenizerFast.from_pretrained("gpt2", revision=vars.revision, cache_dir="cache")
+            try:
+                model     = AutoModelForCausalLM.from_pretrained(vars.custmodpth, revision=vars.revision, cache_dir="cache")
+            except Exception as e:
+                model     = GPTNeoForCausalLM.from_pretrained(vars.custmodpth, revision=vars.revision, cache_dir="cache")
+        elif(os.path.isdir("models/{}".format(vars.model.replace('/', '_')))):
+            try:
+                tokenizer = AutoTokenizer.from_pretrained("models/{}".format(vars.model.replace('/', '_')), revision=vars.revision, cache_dir="cache")
+            except Exception as e:
+                try:
+                    tokenizer = GPT2TokenizerFast.from_pretrained("models/{}".format(vars.model.replace('/', '_')), revision=vars.revision, cache_dir="cache")
+                except Exception as e:
+                    tokenizer = GPT2TokenizerFast.from_pretrained("gpt2", revision=vars.revision, cache_dir="cache")
+            try:
+                model     = AutoModelForCausalLM.from_pretrained("models/{}".format(vars.model.replace('/', '_')), revision=vars.revision, cache_dir="cache")
+            except Exception as e:
+                model     = GPTNeoForCausalLM.from_pretrained("models/{}".format(vars.model.replace('/', '_')), revision=vars.revision, cache_dir="cache")
+        else:
+            try:
+                tokenizer = AutoTokenizer.from_pretrained(vars.model, revision=vars.revision, cache_dir="cache")
+            except Exception as e:
+                try:
+                    tokenizer = GPT2TokenizerFast.from_pretrained(vars.model, revision=vars.revision, cache_dir="cache")
+                except Exception as e:
+                    tokenizer = GPT2TokenizerFast.from_pretrained("gpt2", revision=vars.revision, cache_dir="cache")
+            try:
+                model     = AutoModelForCausalLM.from_pretrained(vars.model, revision=vars.revision, cache_dir="cache")
+            except Exception as e:
+                model     = GPTNeoForCausalLM.from_pretrained(vars.model, revision=vars.revision, cache_dir="cache")
+
+    #network.state = network.move_xmap(network.state, np.zeros(cores_per_replica))

update-koboldai.bat

@@ -50,4 +50,4 @@ git remote add origin %origin%
 git fetch --all
 git checkout %branch% -f
 git reset --hard origin/%branch%
-cmd /k
+%windir%\system32\timeout -t 10

userscripts/api_documentation.html

@@ -56,6 +56,7 @@
 <li><a href="#kobold.num_outputs">kobold.num_outputs</a></li>
 <li><a href="#kobold.outputs">kobold.outputs</a></li>
 <li><a href="#kobold.settings">kobold.settings</a></li>
+<li><a href="#kobold.spfilename">kobold.spfilename</a></li>
 <li><a href="#kobold.story">kobold.story</a>
 <ul>
 <li></li>
@@ -172,6 +173,7 @@
 <li><code>kobold.num_outputs</code></li>
 <li><code>kobold.outputs</code></li>
 <li><code>kobold.settings</code></li>
+<li><code>kobold.spfilename</code></li>
 <li><code>kobold.story</code></li>
 <li><code>kobold.submission</code></li>
 <li><code>kobold.worldinfo</code></li>
@@ -394,6 +396,14 @@
 <li><code>kobold.settings.setwidepth</code> (World Info Depth)</li>
 <li><code>kobold.settings.setuseprompt</code> (Always Use Prompt)</li>
 </ul>
+<h1 id="kobold.spfilename">kobold.spfilename</h1>
+<p><em><strong>Readable from:</strong></em> anywhere<br>
+<em><strong>Writable from:</strong></em> anywhere</p>
+<pre class=" language-lua"><code class="prism  language-lua">field kobold<span class="token punctuation">.</span>spfilename<span class="token punctuation">:</span> string?
+</code></pre>
+<p>The name of the soft prompt file to use (as a string), including the file extension. If not using a soft prompt, this is <code>nil</code> instead.</p>
+<p>You can also set the soft prompt to use by setting this to a string or <code>nil</code>.</p>
+<p>Modifying this field from inside of a generation modifier triggers a regeneration, which means that the context is recomputed after modification and generation begins again with the new context and previously generated tokens. This incurs a small performance penalty and should not be performed in excess.</p>
 <h1 id="kobold.story">kobold.story</h1>
 <p><em><strong>Readable from:</strong></em> anywhere<br>
 <em><strong>Writable from:</strong></em> nowhere</p>

userscripts/api_documentation.md

@@ -29,6 +29,7 @@ global kobold: KoboldLib
 * `kobold.num_outputs`
 * `kobold.outputs`
 * `kobold.settings`
+* `kobold.spfilename`
 * `kobold.story`
 * `kobold.submission`
 * `kobold.worldinfo`
@@ -372,6 +373,21 @@ Modifying certain fields from inside of a generation modifier triggers a regener
 * `kobold.settings.setwidepth` (World Info Depth)
 * `kobold.settings.setuseprompt` (Always Use Prompt)
 
+# kobold.spfilename
+
+***Readable from:*** anywhere
+***Writable from:*** anywhere
+
+```lua
+field kobold.spfilename: string?
+```
+
+The name of the soft prompt file to use (as a string), including the file extension. If not using a soft prompt, this is `nil` instead.
+
+You can also set the soft prompt to use by setting this to a string or `nil`.
+
+Modifying this field from inside of a generation modifier triggers a regeneration, which means that the context is recomputed after modification and generation begins again with the new context and previously generated tokens. This incurs a small performance penalty and should not be performed in excess.
+
 # kobold.story
 
 ***Readable from:*** anywhere

utils.py

@@ -1,5 +1,24 @@
 from threading import Timer
 import re
+import shutil
+import json
+import subprocess
+import tempfile
+import requests
+import requests.adapters
+import time
+from tqdm.auto import tqdm
+import os
+import itertools
+from typing import Optional
+
+vars = None
+num_shards: Optional[int] = None
+current_shard = 0
+from_pretrained_model_name = ""
+from_pretrained_index_filename: Optional[str] = None
+from_pretrained_kwargs = {}
+bar = None
 
 #==================================================================#
 # Decorator to prevent a function's actions from being run until
@@ -111,8 +130,171 @@ def cleanfilename(filename):
     filename = "".join(c for c in filename if c not in filteredcharacters).rstrip()
     return filename
     
-    
-    
-    
-    
-    
+#==================================================================#
+#  Newline substitution for fairseq models
+#==================================================================#
+def encodenewlines(txt):
+    if(vars.newlinemode == "s"):
+        return txt.replace('\n', "</s>")
+    return txt
+
+def decodenewlines(txt):
+    if(vars.newlinemode == "s"):
+        return txt.replace("</s>", '\n')
+    if(vars.newlinemode == "ns"):
+        return txt.replace("</s>", '')
+    return 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
+
+#==================================================================#
+#  Downloads huggingface checkpoints using aria2c if possible
+#==================================================================#
+def aria2_hook(pretrained_model_name_or_path: str, force_download=False, cache_dir=None, proxies=None, resume_download=False, local_files_only=False, use_auth_token=None, user_agent=None, revision=None, mirror=None, **kwargs):
+    import transformers
+    import transformers.modeling_utils
+    from huggingface_hub import HfFolder
+    if shutil.which("aria2c") is None:  # Don't do anything if aria2 is not installed
+        return
+    if local_files_only:  # If local_files_only is true, we obviously don't need to download anything
+        return
+    if os.path.isdir(pretrained_model_name_or_path) or os.path.isfile(pretrained_model_name_or_path) or os.path.isfile(pretrained_model_name_or_path + ".index") or transformers.modeling_utils.is_remote_url(pretrained_model_name_or_path):
+        return
+    if proxies:
+        print("WARNING:  KoboldAI does not support using aria2 to download models from huggingface.co through a proxy.  Disabling aria2 download mode.")
+        return
+    if use_auth_token:
+        if isinstance(use_auth_token, str):
+            token = use_auth_token
+        else:
+            token = HfFolder.get_token()
+            if token is None:
+                raise EnvironmentError("You specified use_auth_token=True, but a huggingface token was not found.")
+    _cache_dir = str(cache_dir) if cache_dir is not None else transformers.TRANSFORMERS_CACHE
+    sharded = False
+    headers = {"user-agent": transformers.file_utils.http_user_agent(user_agent)}
+    if use_auth_token:
+        headers["authorization"] = f"Bearer {use_auth_token}"
+    def is_cached(url):
+        try:
+            transformers.file_utils.get_from_cache(url, cache_dir=cache_dir, local_files_only=True)
+        except FileNotFoundError:
+            return False
+        return True
+    while True:  # Try to get the huggingface.co URL of the model's pytorch_model.bin or pytorch_model.bin.index.json file
+        try:
+            filename = transformers.modeling_utils.WEIGHTS_INDEX_NAME if sharded else transformers.modeling_utils.WEIGHTS_NAME
+        except AttributeError:
+            return
+        url = transformers.file_utils.hf_bucket_url(pretrained_model_name_or_path, filename, revision=revision, mirror=mirror)
+        if is_cached(url) or requests.head(url, allow_redirects=True, proxies=proxies, headers=headers):
+            break
+        if sharded:
+            return
+        else:
+            sharded = True
+    if not sharded:  # If the model has a pytorch_model.bin file, that's the only file to download
+        filenames = [transformers.modeling_utils.WEIGHTS_NAME]
+    else:  # Otherwise download the pytorch_model.bin.index.json and then let aria2 download all the pytorch_model-#####-of-#####.bin files mentioned inside it
+        map_filename = transformers.file_utils.cached_path(url, cache_dir=cache_dir, force_download=force_download, proxies=proxies, resume_download=resume_download, use_auth_token=use_auth_token, user_agent=user_agent)
+        with open(map_filename) as f:
+            map_data = json.load(f)
+        filenames = set(map_data["weight_map"].values())
+    urls = [transformers.file_utils.hf_bucket_url(pretrained_model_name_or_path, n, revision=revision, mirror=mirror) for n in filenames]
+    if not force_download:
+        urls = [u for u in urls if not is_cached(u)]
+        if not urls:
+            return
+    etags = [h.get("X-Linked-Etag") or h.get("ETag") for u in urls for h in [requests.head(u, headers=headers, allow_redirects=False, proxies=proxies, timeout=10).headers]]
+    headers = [requests.head(u, headers=headers, allow_redirects=True, proxies=proxies, timeout=10).headers for u in urls]
+    filenames = [transformers.file_utils.url_to_filename(u, t) for u, t in zip(urls, etags)]
+    for n in filenames:
+        path = os.path.join(_cache_dir, "kai-tempfile." + n + ".aria2")
+        if os.path.exists(path):
+            os.remove(path)
+        path = os.path.join(_cache_dir, "kai-tempfile." + n)
+        if os.path.exists(path):
+            os.remove(path)
+        if force_download:
+            path = os.path.join(_cache_dir, n + ".json")
+            if os.path.exists(path):
+                os.remove(path)
+            path = os.path.join(_cache_dir, n)
+            if os.path.exists(path):
+                os.remove(path)
+    total_length = sum(int(h["Content-Length"]) for h in headers)
+    lengths = {}
+    aria2_config = "\n".join(f"{u}\n  out=kai-tempfile.{n}" for u, n in zip(urls, filenames)).encode()
+    s = requests.Session()
+    s.mount("http://", requests.adapters.HTTPAdapter(max_retries=requests.adapters.Retry(total=120, backoff_factor=1)))
+    bar = None
+    done = False
+    secret = os.urandom(17).hex()
+    try:
+        with tempfile.NamedTemporaryFile("w+b", delete=False) as f:
+            f.write(aria2_config)
+            f.flush()
+            p = subprocess.Popen(["aria2c", "-x", "10", "-s", "10", "-j", "10", "--enable-rpc=true", f"--rpc-secret={secret}", "--rpc-listen-port", str(vars.aria2_port), "--disable-ipv6", "--file-allocation=trunc", "--allow-overwrite", "--auto-file-renaming=false", "-d", _cache_dir, "-i", f.name, "-U", transformers.file_utils.http_user_agent(user_agent)] + (["-c"] if not force_download else []) + ([f"--header='Authorization: Bearer {token}'"] if use_auth_token else []), stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+            while p.poll() is None:
+                r = s.post(f"http://localhost:{vars.aria2_port}/jsonrpc", json={"jsonrpc": "2.0", "id": "kai", "method": "aria2.tellActive", "params": [f"token:{secret}"]}).json()["result"]
+                if not r:
+                    s.close()
+                    if bar is not None:
+                        bar.n = bar.total
+                        bar.close()
+                    p.terminate()
+                    done = True
+                    break
+                if bar is None:
+                    bar = tqdm(total=total_length, desc=f"[aria2] Downloading model", unit="B", unit_scale=True, unit_divisor=1000)
+                visited = set()
+                for x in r:
+                    filename = x["files"][0]["path"]
+                    lengths[filename] = (int(x["completedLength"]), int(x["totalLength"]))
+                    visited.add(filename)
+                for k, v in lengths.items():
+                    if k not in visited:
+                        lengths[k] = (v[1], v[1])
+                bar.n = sum(v[0] for v in lengths.values())
+                bar.update()
+                time.sleep(0.1)
+            path = f.name
+    except Exception as e:
+        p.terminate()
+        raise e
+    finally:
+        try:
+            os.remove(path)
+        except OSError:
+            pass
+    code = p.wait()
+    if not done and code:
+        raise OSError(f"aria2 exited with exit code {code}")
+    for u, t, n in zip(urls, etags, filenames):
+        os.rename(os.path.join(_cache_dir, "kai-tempfile." + n), os.path.join(_cache_dir, n))
+        with open(os.path.join(_cache_dir, n + ".json"), "w") as f:
+            json.dump({"url": u, "etag": t}, f)
+
+#==================================================================#
+#  Given the path to a pytorch_model.bin.index.json, returns how many
+#  shards there are in the model
+#==================================================================#
+def get_num_shards(filename):
+    with open(filename) as f:
+        map_data = json.load(f)
+    return len(set(map_data["weight_map"].values()))
+
+#==================================================================#
+#  Given the name/path of a sharded model and the path to a
+#  pytorch_model.bin.index.json, returns a list of weight names in the
+#  sharded model.  Requires lazy loader to be enabled to work properl
+#==================================================================#
+def get_sharded_checkpoint_num_tensors(pretrained_model_name_or_path, filename, cache_dir=None, force_download=False, proxies=None, resume_download=False, local_files_only=False, use_auth_token=None, user_agent=None, revision=None, mirror=None, **kwargs):
+    import transformers.modeling_utils
+    import torch
+    shard_paths, _ = transformers.modeling_utils.get_checkpoint_shard_files(pretrained_model_name_or_path, filename, cache_dir=cache_dir, force_download=force_download, proxies=proxies, resume_download=resume_download, local_files_only=local_files_only, use_auth_token=use_auth_token, user_agent=user_agent, revision=revision, mirror=mirror)
+    return list(itertools.chain(*(torch.load(p, map_location="cpu").keys() for p in shard_paths)))

warpers.py

@@ -62,7 +62,7 @@ class TailFreeLogitsWarper(LogitsWarper):
     def __init__(self, tfs: float, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
         tfs = float(tfs)
         if tfs < 0 or tfs > 1.0:
-            raise ValueError(f"`tfs` has to be a float > 0 and < 1, but is {tfs}")
+            raise ValueError(f"`tfs` has to be a float >= 0 and <= 1, but is {tfs}")
         self.tfs = tfs
         self.filter_value = filter_value
         self.min_tokens_to_keep = min_tokens_to_keep
@@ -98,3 +98,53 @@ class TailFreeLogitsWarper(LogitsWarper):
         indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
         scores = scores.masked_fill(indices_to_remove, self.filter_value)
         return scores
+
+
+class TypicalLogitsWarper(LogitsWarper):
+    '''
+    Typical sampling, described in https://arxiv.org/pdf/2202.00666.pdf
+    '''
+
+    def __init__(self, typical: float, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
+        typical = float(typical)
+        if typical < 0 or typical > 1.0:
+            raise ValueError(f"`typical` has to be a float >= 0 and <= 1, but is {typical}")
+        self.typical = typical
+        self.filter_value = filter_value
+        self.min_tokens_to_keep = min_tokens_to_keep
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        if self.filter_value >= 1.0:
+            return scores
+
+        # Compute softmax probabilities and the natural logarithms of them
+        probs = scores.softmax(dim=-1)
+        log_probs = probs.log()
+
+        # Compute the negative of entropy, which is the sum of p*ln(p) for all p
+        # in the set of softmax probabilities of the logits
+        neg_entropy = (probs * log_probs).nansum(dim=-1, keepdim=True)
+
+        # Determine absolute difference between the negative entropy and the
+        # log probabilities
+        entropy_deviation = (neg_entropy - log_probs).abs()
+
+        # Keep certain tokens such that the sum of the entropy_deviation of the
+        # kept tokens is the smallest possible value such that the sum of the
+        # softmax probabilities of the kept tokens is at least the threshold
+        # value (by sorting the tokens in ascending order of entropy_deviation
+        # and then keeping the smallest possible number of tokens from the
+        # beginning such that sum of softmax probabilities is at or above the
+        # threshold)
+        _, sorted_indices = torch.sort(entropy_deviation)
+        sorted_logits = probs.gather(-1, sorted_indices)
+        sorted_indices_to_remove = sorted_logits.cumsum(dim=-1) >= self.typical
+        sorted_indices_to_remove = sorted_indices_to_remove.roll(1, dims=-1)
+
+        min_tokens_to_keep = max(self.min_tokens_to_keep, 1)
+        # Keep at least min_tokens_to_keep
+        sorted_indices_to_remove[..., : min_tokens_to_keep] = 0
+
+        indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
+        scores = scores.masked_fill(indices_to_remove, self.filter_value)
+        return scores