From 987999b14815aba6f7d416dae7e83d6bfa9e160b Mon Sep 17 00:00:00 2001 From: soulteary Date: Sat, 25 Mar 2023 13:35:00 +0800 Subject: [PATCH 1/5] feat: docker support, and simple webui --- docker/Dockerfile | 22 ++++++++ docker/webui.py | 128 ++++++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 150 insertions(+) create mode 100644 docker/Dockerfile create mode 100644 docker/webui.py diff --git a/docker/Dockerfile b/docker/Dockerfile new file mode 100644 index 00000000..dbad6074 --- /dev/null +++ b/docker/Dockerfile @@ -0,0 +1,22 @@ +FROM nvcr.io/nvidia/pytorch:23.02-py3 +LABEL org.opencontainers.image.authors="soulteary@gmail.com" + +RUN pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple && \ + pip install huggingface_hub +WORKDIR /app +RUN cat > /get-models.py <ChatRWKV +RWKV-LM +RWKV pip package +''' + +os.environ["RWKV_JIT_ON"] = '1' +os.environ["RWKV_CUDA_ON"] = '1' # if '1' then use CUDA kernel for seq mode (much faster) + +from rwkv.model import RWKV +model_path = hf_hub_download(repo_id="BlinkDL/rwkv-4-pile-14b", filename=f"{title}.pth") +model = RWKV(model=model_path, strategy='cuda fp16i8 *20 -> cuda fp16') +from rwkv.utils import PIPELINE, PIPELINE_ARGS +pipeline = PIPELINE(model, "20B_tokenizer.json") + +def infer( + ctx, + token_count=10, + temperature=1.0, + top_p=0.8, + presencePenalty = 0.1, + countPenalty = 0.1, +): + args = PIPELINE_ARGS(temperature = max(0.2, float(temperature)), top_p = float(top_p), + alpha_frequency = countPenalty, + alpha_presence = presencePenalty, + token_ban = [0], # ban the generation of some tokens + token_stop = []) # stop generation whenever you see any token here + + ctx = ctx.strip(' ') + if ctx.endswith('\n'): + ctx = f'\n{ctx.strip()}\n' + else: + ctx = f'\n{ctx.strip()}' + + gpu_info = nvmlDeviceGetMemoryInfo(gpu_h) + print(f'vram {gpu_info.total} used {gpu_info.used} free {gpu_info.free}') + + all_tokens = [] + out_last = 0 + out_str = '' + occurrence = {} + state = None + for i in range(int(token_count)): + out, state = model.forward(pipeline.encode(ctx)[-ctx_limit:] if i == 0 else [token], state) + for n in args.token_ban: + out[n] = -float('inf') + for n in occurrence: + out[n] -= (args.alpha_presence + occurrence[n] * args.alpha_frequency) + + token = pipeline.sample_logits(out, temperature=args.temperature, top_p=args.top_p) + if token in args.token_stop: + break + all_tokens += [token] + if token not in occurrence: + occurrence[token] = 1 + else: + occurrence[token] += 1 + + tmp = pipeline.decode(all_tokens[out_last:]) + if '\ufffd' not in tmp: + out_str += tmp + yield out_str.strip() + out_last = i + 1 + gc.collect() + torch.cuda.empty_cache() + yield out_str.strip() + +examples = [ + ["Expert Questions & Helpful Answers\nAsk Research Experts\nQuestion:\nHow can we eliminate poverty?\n\nFull Answer:\n", 150, 1.0, 0.7, 0.2, 0.2], + ["Here's a short cyberpunk sci-fi adventure story. The story's main character is an artificial human created by a company called OpenBot.\n\nThe Story:\n", 150, 1.0, 0.7, 0.2, 0.2], + ['''Below is an instruction that describes a task. Write a response that appropriately completes the request. +### Instruction: +Generate a list of adjectives that describe a person as brave. +### Response: +''', 150, 1.0, 0.2, 0.5, 0.5], + ['''Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request. +### Instruction: +Arrange the given numbers in ascending order. +### Input: +2, 4, 0, 8, 3 +### Response: +''', 150, 1.0, 0.2, 0.5, 0.5], + ["Ask Expert\n\nQuestion:\nWhat are some good plans for world peace?\n\nExpert Full Answer:\n", 150, 1.0, 0.7, 0.2, 0.2], + ["Q & A\n\nQuestion:\nWhy is the sky blue?\n\nDetailed Expert Answer:\n", 150, 1.0, 0.7, 0.2, 0.2], + ["Dear sir,\nI would like to express my boundless apologies for the recent nuclear war.", 150, 1.0, 0.7, 0.2, 0.2], + ["Here is a shell script to find all .hpp files in /home/workspace and delete the 3th row string of these files:", 150, 1.0, 0.7, 0.1, 0.1], + ["Building a website can be done in 10 simple steps:\n1.", 150, 1.0, 0.7, 0.2, 0.2], + ["A Chinese phrase is provided: 百闻不如一见。\nThe masterful Chinese translator flawlessly translates the phrase into English:", 150, 1.0, 0.5, 0.2, 0.2], + ["I believe the meaning of life is", 150, 1.0, 0.7, 0.2, 0.2], + ["Simply put, the theory of relativity states that", 150, 1.0, 0.5, 0.2, 0.2], +] + + +iface = gr.Interface( + fn=infer, + description=f'''{desc} *** Please try examples first (bottom of page) *** (edit them to use your question). Demo limited to ctxlen {ctx_limit}.''', + allow_flagging="never", + inputs=[ + gr.Textbox(lines=10, label="Prompt", value="Here's a short cyberpunk sci-fi adventure story. The story's main character is an artificial human created by a company called OpenBot.\n\nThe Story:\n"), # prompt + gr.Slider(10, 200, step=10, value=150), # token_count + gr.Slider(0.2, 2.0, step=0.1, value=1.0), # temperature + gr.Slider(0.0, 1.0, step=0.05, value=0.7), # top_p + gr.Slider(0.0, 1.0, step=0.1, value=0.2), # presencePenalty + gr.Slider(0.0, 1.0, step=0.1, value=0.2), # countPenalty + ], + outputs=gr.Textbox(label="Generated Output", lines=28), + examples=examples, + cache_examples=False, +).queue() + +demo = gr.TabbedInterface( + [iface], ["Generative"], + title=title, +) + +demo.queue(max_size=10) +demo.launch(share=False, server_name="0.0.0.0") From e23e98dcd0eb8ebed71a0d8c7c590bcffb946869 Mon Sep 17 00:00:00 2001 From: soulteary Date: Sat, 25 Mar 2023 16:17:38 +0800 Subject: [PATCH 2/5] chore: use latest pypi package, allow save new format model --- docker/Dockerfile | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/docker/Dockerfile b/docker/Dockerfile index dbad6074..b79f8883 100644 --- a/docker/Dockerfile +++ b/docker/Dockerfile @@ -12,7 +12,8 @@ EOF RUN python /get-models.py && \ rm -rf /get-models.py && \ - pip install ninja rwkv==0.6.2 gradio + pip install ninja gradio && \ + pip install rwkv==0.7.1 --index-url=https://pypi.org/simple RUN git clone https://huggingface.co/spaces/BlinkDL/ChatRWKV-gradio/ ChatRWKV && \ cd ChatRWKV && git checkout 699bb36437cd15ef8da05c4f392f4ab3ea6be65d From ccebdc2c4eb83399b33c957c15dc2749aa9ab569 Mon Sep 17 00:00:00 2001 From: soulteary Date: Sat, 25 Mar 2023 16:23:47 +0800 Subject: [PATCH 3/5] feat: model convert to imporve app speed (wip) --- docker/convert.py | 6 ++++++ 1 file changed, 6 insertions(+) create mode 100644 docker/convert.py diff --git a/docker/convert.py b/docker/convert.py new file mode 100644 index 00000000..2e5c88bd --- /dev/null +++ b/docker/convert.py @@ -0,0 +1,6 @@ +from huggingface_hub import hf_hub_download +title = "RWKV-4-Pile-14B-20230313-ctx8192-test1050" +model_path = hf_hub_download(repo_id="BlinkDL/rwkv-4-pile-14b", filename=f"{title}.pth") + +from rwkv.model import RWKV +RWKV(model=model_path, strategy='cuda fp16i8 *20 -> cuda fp16', convert_and_save_and_exit = f"./models/{title}.pth") \ No newline at end of file From 22e3396ab5cf57db1c42338a03b81fc44724dc7f Mon Sep 17 00:00:00 2001 From: soulteary Date: Sat, 25 Mar 2023 16:26:54 +0800 Subject: [PATCH 4/5] chore: use pre-converted model to imporve app speed --- docker/webui.py | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-) diff --git a/docker/webui.py b/docker/webui.py index ef2bc4ac..f198fff3 100644 --- a/docker/webui.py +++ b/docker/webui.py @@ -2,7 +2,6 @@ import gradio as gr import os, gc, torch from datetime import datetime -from huggingface_hub import hf_hub_download from pynvml import * nvmlInit() gpu_h = nvmlDeviceGetHandleByIndex(0) @@ -18,7 +17,14 @@ os.environ["RWKV_CUDA_ON"] = '1' # if '1' then use CUDA kernel for seq mode (much faster) from rwkv.model import RWKV -model_path = hf_hub_download(repo_id="BlinkDL/rwkv-4-pile-14b", filename=f"{title}.pth") + +model_path = f"./models/{title}.pth" +if os.path.isfile(model_path): + print(f"The pre-converted model exists.") +else: + from huggingface_hub import hf_hub_download + model_path = hf_hub_download(repo_id="BlinkDL/rwkv-4-pile-14b", filename=f"{title}.pth") + model = RWKV(model=model_path, strategy='cuda fp16i8 *20 -> cuda fp16') from rwkv.utils import PIPELINE, PIPELINE_ARGS pipeline = PIPELINE(model, "20B_tokenizer.json") From 09fe10892a392bb7242f8bfe72ff6f14ca22646f Mon Sep 17 00:00:00 2001 From: soulteary Date: Sat, 25 Mar 2023 17:15:21 +0800 Subject: [PATCH 5/5] feat: add minimum resource requirements mode --- docker/Dockerfile | 1 + docker/convert.mini.py | 6 ++ docker/webui.mini.py | 134 +++++++++++++++++++++++++++++++++++++++++ 3 files changed, 141 insertions(+) create mode 100644 docker/convert.mini.py create mode 100644 docker/webui.mini.py diff --git a/docker/Dockerfile b/docker/Dockerfile index b79f8883..f438a228 100644 --- a/docker/Dockerfile +++ b/docker/Dockerfile @@ -20,4 +20,5 @@ RUN git clone https://huggingface.co/spaces/BlinkDL/ChatRWKV-gradio/ ChatRWKV && WORKDIR /app/ChatRWKV COPY docker/webui.py /app/ChatRWKV/app.py +COPY docker/*.mini.py /app/ChatRWKV/ CMD ["python", "/app/ChatRWKV/app.py"] \ No newline at end of file diff --git a/docker/convert.mini.py b/docker/convert.mini.py new file mode 100644 index 00000000..a89b22b4 --- /dev/null +++ b/docker/convert.mini.py @@ -0,0 +1,6 @@ +from huggingface_hub import hf_hub_download +title = "RWKV-4-Pile-14B-20230313-ctx8192-test1050" +model_path = hf_hub_download(repo_id="BlinkDL/rwkv-4-pile-14b", filename=f"{title}.pth") + +from rwkv.model import RWKV +RWKV(model=model_path, strategy='cuda fp16i8 *0+ -> cpu fp32 *1', convert_and_save_and_exit = f"./models/{title}.pth") diff --git a/docker/webui.mini.py b/docker/webui.mini.py new file mode 100644 index 00000000..31396221 --- /dev/null +++ b/docker/webui.mini.py @@ -0,0 +1,134 @@ +# modify https://huggingface.co/spaces/BlinkDL/ChatRWKV-gradio/blob/main/app.py +import gradio as gr +import os, gc, torch +from datetime import datetime +from pynvml import * +nvmlInit() +gpu_h = nvmlDeviceGetHandleByIndex(0) +ctx_limit = 1024 +title = "RWKV-4-Pile-14B-20230313-ctx8192-test1050" +desc = f'''Links: +ChatRWKV +RWKV-LM +RWKV pip package +''' + +os.environ["RWKV_JIT_ON"] = '1' +os.environ["RWKV_CUDA_ON"] = '1' # if '1' then use CUDA kernel for seq mode (much faster) + +from rwkv.model import RWKV + +model_path = f"./models/{title}.pth" +if os.path.isfile(model_path): + print(f"The pre-converted model exists.") +else: + from huggingface_hub import hf_hub_download + model_path = hf_hub_download(repo_id="BlinkDL/rwkv-4-pile-14b", filename=f"{title}.pth") + +model = RWKV(model=model_path, strategy='cuda fp16i8 *0+ -> cpu fp32 *1') +from rwkv.utils import PIPELINE, PIPELINE_ARGS +pipeline = PIPELINE(model, "20B_tokenizer.json") + +def infer( + ctx, + token_count=10, + temperature=1.0, + top_p=0.8, + presencePenalty = 0.1, + countPenalty = 0.1, +): + args = PIPELINE_ARGS(temperature = max(0.2, float(temperature)), top_p = float(top_p), + alpha_frequency = countPenalty, + alpha_presence = presencePenalty, + token_ban = [0], # ban the generation of some tokens + token_stop = []) # stop generation whenever you see any token here + + ctx = ctx.strip(' ') + if ctx.endswith('\n'): + ctx = f'\n{ctx.strip()}\n' + else: + ctx = f'\n{ctx.strip()}' + + gpu_info = nvmlDeviceGetMemoryInfo(gpu_h) + print(f'vram {gpu_info.total} used {gpu_info.used} free {gpu_info.free}') + + all_tokens = [] + out_last = 0 + out_str = '' + occurrence = {} + state = None + for i in range(int(token_count)): + out, state = model.forward(pipeline.encode(ctx)[-ctx_limit:] if i == 0 else [token], state) + for n in args.token_ban: + out[n] = -float('inf') + for n in occurrence: + out[n] -= (args.alpha_presence + occurrence[n] * args.alpha_frequency) + + token = pipeline.sample_logits(out, temperature=args.temperature, top_p=args.top_p) + if token in args.token_stop: + break + all_tokens += [token] + if token not in occurrence: + occurrence[token] = 1 + else: + occurrence[token] += 1 + + tmp = pipeline.decode(all_tokens[out_last:]) + if '\ufffd' not in tmp: + out_str += tmp + yield out_str.strip() + out_last = i + 1 + gc.collect() + torch.cuda.empty_cache() + yield out_str.strip() + +examples = [ + ["Expert Questions & Helpful Answers\nAsk Research Experts\nQuestion:\nHow can we eliminate poverty?\n\nFull Answer:\n", 150, 1.0, 0.7, 0.2, 0.2], + ["Here's a short cyberpunk sci-fi adventure story. The story's main character is an artificial human created by a company called OpenBot.\n\nThe Story:\n", 150, 1.0, 0.7, 0.2, 0.2], + ['''Below is an instruction that describes a task. Write a response that appropriately completes the request. +### Instruction: +Generate a list of adjectives that describe a person as brave. +### Response: +''', 150, 1.0, 0.2, 0.5, 0.5], + ['''Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request. +### Instruction: +Arrange the given numbers in ascending order. +### Input: +2, 4, 0, 8, 3 +### Response: +''', 150, 1.0, 0.2, 0.5, 0.5], + ["Ask Expert\n\nQuestion:\nWhat are some good plans for world peace?\n\nExpert Full Answer:\n", 150, 1.0, 0.7, 0.2, 0.2], + ["Q & A\n\nQuestion:\nWhy is the sky blue?\n\nDetailed Expert Answer:\n", 150, 1.0, 0.7, 0.2, 0.2], + ["Dear sir,\nI would like to express my boundless apologies for the recent nuclear war.", 150, 1.0, 0.7, 0.2, 0.2], + ["Here is a shell script to find all .hpp files in /home/workspace and delete the 3th row string of these files:", 150, 1.0, 0.7, 0.1, 0.1], + ["Building a website can be done in 10 simple steps:\n1.", 150, 1.0, 0.7, 0.2, 0.2], + ["A Chinese phrase is provided: 百闻不如一见。\nThe masterful Chinese translator flawlessly translates the phrase into English:", 150, 1.0, 0.5, 0.2, 0.2], + ["I believe the meaning of life is", 150, 1.0, 0.7, 0.2, 0.2], + ["Simply put, the theory of relativity states that", 150, 1.0, 0.5, 0.2, 0.2], +] + + +iface = gr.Interface( + fn=infer, + description=f'''{desc} *** Please try examples first (bottom of page) *** (edit them to use your question). Demo limited to ctxlen {ctx_limit}.''', + allow_flagging="never", + inputs=[ + gr.Textbox(lines=10, label="Prompt", value="Here's a short cyberpunk sci-fi adventure story. The story's main character is an artificial human created by a company called OpenBot.\n\nThe Story:\n"), # prompt + gr.Slider(10, 200, step=10, value=150), # token_count + gr.Slider(0.2, 2.0, step=0.1, value=1.0), # temperature + gr.Slider(0.0, 1.0, step=0.05, value=0.7), # top_p + gr.Slider(0.0, 1.0, step=0.1, value=0.2), # presencePenalty + gr.Slider(0.0, 1.0, step=0.1, value=0.2), # countPenalty + ], + outputs=gr.Textbox(label="Generated Output", lines=28), + examples=examples, + cache_examples=False, +).queue() + +demo = gr.TabbedInterface( + [iface], ["Generative"], + title=title, +) + +demo.queue(max_size=10) +demo.launch(share=False, server_name="0.0.0.0")