Duplicate from Hazzzardous/RWKV-Instruct

Co-authored-by: Harrison Vanderbyl <[email protected]>

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+14B-8K-rwkvstic-2-1-2.rwkv filter=lfs diff=lfs merge=lfs -text

14B-8K-rwkvstic-2-1-2.rwkv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9863c8993a323971167f9625f6c37583bc5f78bcc8fe85ecbdf49ccb9b3b2f78
+size 14455104393

README.md

@@ -0,0 +1,17 @@
+---
+title: RWKV Instruct
+emoji: 💩
+colorFrom: yellow
+colorTo: indigo
+sdk: gradio
+sdk_version: 3.17.0
+app_file: app.py
+pinned: false
+license: gpl-3.0
+duplicated_from: Hazzzardous/RWKV-Instruct
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+Based on:
+
+https://huggingface.co/spaces/yahma/rwkv

app.py

@@ -0,0 +1,297 @@
+"""
+RWKV RNN Model - Gradio Space for HuggingFace
+YT - Mean Gene Hacks - https://www.youtube.com/@MeanGeneHacks
+(C) Gene Ruebsamen - 2/7/2023
+
+This program is free software: you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation, either version 3 of the License, or
+(at your option) any later version.
+
+This program is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with this program.  If not, see <http://www.gnu.org/licenses/>.
+"""
+
+import gradio as gr
+import codecs
+from ast import literal_eval
+from datetime import datetime
+from rwkvstic.load import RWKV
+from config import config, title
+import torch
+import gc
+
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+
+desc = '''<p>RNN with Transformer-level LLM Performance (<a href='https://github.com/BlinkDL/RWKV-LM'>github</a>).
+    According to the author: "It combines the best of RNN and transformers - great performance, fast inference, saves VRAM, fast training, "infinite" ctx_len, and free sentence embedding."'''
+
+thanks = '''<p>Thanks to <a href='https://github.com/gururise/rwkv_gradio'>Gururise</a> for this template</p>'''
+
+
+def to_md(text):
+    return text.replace("\n", "<br />")
+
+
+def get_model():
+    model = None
+    model = RWKV(
+        **config
+    )
+    return model
+
+
+model = get_model()
+
+
+
+def infer(
+        prompt,
+        mode="generative",
+        max_new_tokens=10,
+        temperature=0.1,
+        top_p=1.0,
+        stop="<|endoftext|>",
+        end_adj=0.0,
+        seed=42,
+):
+    global model
+
+    if model == None:
+        gc.collect()
+        if (DEVICE == "cuda"):
+            torch.cuda.empty_cache()
+        model = get_model()
+
+    max_new_tokens = int(max_new_tokens)
+    temperature = float(temperature)
+    end_adj = float(end_adj)
+    top_p = float(top_p)
+    stop = [x.strip(' ') for x in stop.split(',')]
+    seed = seed
+
+    assert 1 <= max_new_tokens <= 512
+    assert 0.0 <= temperature <= 5.0
+    assert 0.0 <= top_p <= 1.0
+
+    temperature = max(0.05, temperature)
+    if prompt == "":
+        prompt = " "
+
+    # Clear model state for generative mode
+    model.resetState()
+    if (mode == "Q/A"):
+        prompt = f"\nQ: {prompt}\n\nA:"
+    if (mode == "ELDR"):
+        prompt = f"\n{prompt}\n\nExpert Long Detailed Response:\n\nHi, thanks for reaching out, we would be happy to answer your question"
+    if (mode == "Expert"):
+        prompt = f"\n{prompt}\n\nExpert Full Response:\n\nHi, thanks for reaching out, we would be happy to answer your question.\n"
+    if (mode == "EFA"):
+        prompt = f'\nAsk Expert\n\nQuestion:\n{prompt}\n\nExpert Full Answer:\n'
+    if (mode == "BFR"):
+        prompt = f"Task given:\n\n{prompt}\n\nBest Full Response:"
+
+    print(f"PROMPT ({datetime.now()}):\n-------\n{prompt}")
+    print(f"OUTPUT ({datetime.now()}):\n-------\n")
+    # Load prompt
+    model.loadContext(newctx=prompt)
+    generated_text = ""
+    done = False
+    with torch.no_grad():
+        for _ in range(max_new_tokens):
+            char = model.forward(stopStrings=stop, temp=temperature, top_p_usual=top_p, end_adj=end_adj)[
+                "output"]
+            print(char, end='', flush=True)
+            generated_text += char
+            generated_text = generated_text.lstrip("\n ")
+
+            for stop_word in stop:
+                stop_word = codecs.getdecoder("unicode_escape")(stop_word)[0]
+                if stop_word != '' and stop_word in generated_text:
+                    done = True
+                    break
+            yield generated_text
+            if done:
+                print("<stopped>\n")
+                break
+
+    # print(f"{generated_text}")
+
+    for stop_word in stop:
+        stop_word = codecs.getdecoder("unicode_escape")(stop_word)[0]
+        if stop_word != '' and stop_word in generated_text:
+            generated_text = generated_text[:generated_text.find(stop_word)]
+
+    gc.collect()
+    yield generated_text
+username = "USER"
+intro = f'''The following is a verbose and detailed conversation between an AI assistant called FRITZ, and a human user called USER. FRITZ is intelligent, knowledgeable, wise and polite.
+
+    {username}: What year was the french revolution?
+    FRITZ: The French Revolution started in 1789, and lasted 10 years until 1799.
+    {username}: 3+5=?
+    FRITZ: The answer is 8.
+    {username}: What year did the Berlin Wall fall?
+    FRITZ: The Berlin wall stood for 28 years and fell in 1989.
+    {username}: solve for a: 9-a=2
+    FRITZ: The answer is a=7, because 9-7 = 2.
+    {username}: wat is lhc
+    FRITZ: The Large Hadron Collider (LHC) is a high-energy particle collider, built by CERN, and completed in 2008. It was used to confirm the existence of the Higgs boson in 2012.
+    {username}: Tell me about yourself.
+    FRITZ: My name is Fritz. I am an RNN based Large Language Model (LLM).
+    '''
+model.resetState()
+model.loadContext(newctx=intro)
+chatState = model.getState()
+model.resetState()
+def chat(
+        prompt,
+        history,
+        max_new_tokens=10,
+        temperature=0.1,
+        top_p=1.0,
+        seed=42,
+):
+    global model
+    global username
+    history = history or []
+
+    intro = ""
+
+    if model == None:
+        gc.collect()
+        if (DEVICE == "cuda"):
+            torch.cuda.empty_cache()
+        model = get_model()
+
+    username = username.strip()
+    username = username or "USER"
+
+    
+
+    if len(history) == 0:
+        # no history, so lets reset chat state
+        model.setState(chatState)
+        history = [[], model.emptyState]
+        print("reset chat state")
+    else:
+        if (history[0][0][0].split(':')[0] != username):
+            model.setState((chatState[0],chatState[1].clone()))
+            history = [[], model.chatState]
+            print("username changed, reset state")
+        else:
+            model.setState((history[1][0],history[1][1].clone()))
+            intro = ""
+
+    max_new_tokens = int(max_new_tokens)
+    temperature = float(temperature)
+    top_p = float(top_p)
+    seed = seed
+
+    assert 1 <= max_new_tokens <= 512
+    assert 0.0 <= temperature <= 3.0
+    assert 0.0 <= top_p <= 1.0
+
+    temperature = max(0.05, temperature)
+
+    prompt = f"{username}: " + prompt + "\n"
+    print(f"CHAT ({datetime.now()}):\n-------\n{prompt}")
+    print(f"OUTPUT ({datetime.now()}):\n-------\n")
+    # Load prompt
+
+    model.loadContext(newctx=prompt)
+
+    out = model.forward(number=max_new_tokens, stopStrings=[
+                        "<|endoftext|>", username+":"], temp=temperature, top_p_usual=top_p)
+
+    generated_text = out["output"].lstrip("\n ")
+    generated_text = generated_text.rstrip(username+":")
+    print(f"{generated_text}")
+
+    gc.collect()
+    history[0].append((prompt, generated_text))
+    return history[0], [history[0], out["state"]]
+
+
+examples = [
+    [
+        # Question Answering
+        '''What is the capital of Germany?''', "Q/A", 25, 0.2, 1.0, "<|endoftext|>"],
+    [
+        # Question Answering
+        '''Are humans good or bad?''', "Q/A", 150, 0.8, 0.8, "<|endoftext|>"],
+    [
+        # Question Answering
+        '''What is the purpose of Vitamin A?''', "Q/A", 50, 0.2, 0.8, "<|endoftext|>"],
+    [
+        # Chatbot
+        '''This is a conversation between two AI large language models named Alex and Fritz. They are exploring each other's capabilities, and trying to ask interesting questions of one another to explore the limits of each others AI.
+
+Conversation:
+Alex: Good morning, Fritz, what type of LLM are you based upon?
+Fritz: Morning Alex, I am an RNN with transformer level performance. My language model is 100% attention free.
+Alex:''', "generative", 220, 0.9, 0.9, "\\n\\n,<|endoftext|>"],
+    [
+        # Generate List
+        '''Task given:
+
+Please Write a Short story about a cat learning python
+
+Best Full Response:
+''', "generative", 140, 0.85, 0.8, "<|endoftext|>"],
+    [
+        # Natural Language Interface
+        '''Here is a short story (in the style of Tolkien) in which Aiden attacks a robot with a sword:
+        ''', "generative", 140, 0.85, 0.8, "<|endoftext|>"]
+]
+
+
+iface = gr.Interface(
+    fn=infer,
+    description=f'''<h3>Generative and Question/Answer</h3>{desc}{thanks}''',
+    allow_flagging="never",
+    inputs=[
+        gr.Textbox(lines=20, label="Prompt"),  # prompt
+        gr.Radio(["Freeform", "Q/A","ELDR","Expert","EFR","BFR"],
+                 value="Expert", label="Choose Mode"),
+        gr.Slider(1, 512, value=40),  # max_tokens
+        gr.Slider(0.0, 5.0, value=0.9),  # temperature
+        gr.Slider(0.0, 1.0, value=0.85),  # top_p
+        gr.Textbox(lines=1, value="<|endoftext|>"),  # stop
+        gr.Slider(-999, 0.0, value=0.0),  # end_adj
+       
+    ],
+    outputs=gr.Textbox(label="Generated Output", lines=25),
+    examples=examples,
+    cache_examples=False,
+).queue()
+
+chatiface = gr.Interface(
+    fn=chat,
+    description=f'''<h3>Chatbot</h3><h4>Refresh page or change name to reset memory context</h4>{desc}{thanks}''',
+    allow_flagging="never",
+    inputs=[
+        gr.Textbox(lines=5, label="Message"),  # prompt
+        "state",
+        gr.Slider(1, 256, value=60),  # max_tokens
+        gr.Slider(0.0, 1.0, value=0.8),  # temperature
+        gr.Slider(0.0, 1.0, value=0.85)  # top_p
+    ],
+    outputs=[gr.Chatbot(label="Chat Log", color_map=(
+        "green", "pink")), "state"],
+).queue()
+
+demo = gr.TabbedInterface(
+
+    [iface, chatiface], ["Q/A", "Chatbot"],
+    title=title,
+
+)
+
+demo.queue()
+demo.launch(share=False)

config.py

@@ -0,0 +1,60 @@
+from rwkvstic.agnostic.backends import TORCH, TORCH_QUANT
+import torch
+
+quantized = {
+    "mode": TORCH_QUANT,
+    "runtimedtype": torch.bfloat16,
+    "useGPU": torch.cuda.is_available(),
+    "chunksize": 32,  # larger = more accurate, but more memory
+    "target": 100  # your gpu max size, excess vram offloaded to cpu
+}
+
+# UNCOMMENT TO SELECT OPTIONS
+# Not full list of options, see https://pypi.org/project/rwkvstic/ and https://huggingface.co/BlinkDL/ for more models/modes
+
+# RWKV 1B5 instruct test 1 model
+# Approximate
+# [Vram usage: 6.0GB]
+# [File size: 3.0GB]
+
+
+config = {
+    "path":"14B-8K-rwkvstic-2-1-2.rwkv"
+    #"useLogFix":False # When enabled, use BlinkDLs version of the att.
+}
+
+title = "RWKV-4 (14B 8k)"
+
+# RWKV 1B5 instruct model quantized
+# Approximate
+# [Vram usage: 1.3GB]
+# [File size: 3.0GB]
+
+# config = {
+#     "path": "https://huggingface.co/BlinkDL/rwkv-4-pile-1b5/resolve/main/RWKV-4-Pile-1B5-Instruct-test1-20230124.pth",
+#     **quantized
+# }
+
+# title = "RWKV-4 (1.5b Instruct Quantized)"
+
+# RWKV 7B instruct pre-quantized (settings baked into model)
+# Approximate
+# [Vram usage: 7.0GB]
+# [File size: 8.0GB]
+
+# config = {
+#     "path": "https://huggingface.co/Hazzzardous/RWKV-8Bit/resolve/main/RWKV-4-Pile-7B-Instruct.pqth"
+# }
+
+# title = "RWKV-4 (7b Instruct Quantized)"
+
+# RWKV 14B quantized (latest as of feb 9)
+# Approximate
+# [Vram usage: 15.0GB]
+# [File size: 15.0GB]
+
+# config = {
+#     "path": "https://huggingface.co/Hazzzardous/RWKV-8Bit/resolve/main/RWKV-4-Pile-14B-20230204-7324.pqth"
+# }
+
+# title = "RWKV-4 (14b 94% trained, not yet instruct tuned, 8-Bit)"

requirements.txt

@@ -0,0 +1,7 @@
+transformers
+torch
+inquirer
+
+rwkvstic>=2.1.2
+scipy
+onnx