more tests on remote worker container

runpod/runpod-worker-transformers/Dockerfile

@@ -1,8 +1,9 @@
+ARG DEV_IMAGE=nvidia/cuda:11.8.0-cudnn8-devel-ubuntu22.04
 ARG BASE_IMAGE=nvidia/cuda:11.8.0-cudnn8-runtime-ubuntu22.04
 #ARG BASE_IMAGE=nvidia/cuda:11.6.2-cudnn8-devel-ubuntu20.04
 #ARG BASE_IMAGE=runpod/pytorch:3.10-2.0.0-117
 #ARG BASE_IMAGE=nvcr.io/nvidia/pytorch:23.03-py3
-FROM ${BASE_IMAGE} as dev-base
+FROM ${DEV_IMAGE} as builder
 
 ARG MODEL_NAME
 ENV MODEL_NAME=${MODEL_NAME}
@@ -15,16 +16,17 @@ ENV DEBIAN_FRONTEND noninteractive\
 RUN apt-get update --yes && \
     # - apt-get upgrade is run to patch known vulnerabilities in apt-get packages as
     #   the ubuntu base image is rebuilt too seldom sometimes (less than once a month)
-    #apt-get upgrade --yes && \
+    apt-get upgrade --yes && \
     apt install --yes --no-install-recommends \
-#      build-essential \
+      build-essential \
+      cmake \
       ca-certificates \
       git \
       git-lfs \
       wget \
       curl \
       bash \
-      libgl1 \
+#      libgl1 \
       software-properties-common \
       openssh-server && \
       apt-get clean && rm -rf /var/lib/apt/lists/* && \
@@ -51,6 +53,7 @@ RUN pip3 install --upgrade pip && \
     pip3 install torch torchvision torchaudio --extra-index-url=https://download.pytorch.org/whl/cu118 && \
     pip3 install bitsandbytes && \
     pip3 install safetensors && \
+    pip3 install sentencepiece triton && \
     pip3 install diffusers && \
     pip3 install transformers accelerate xformers triton && \
     pip3 install huggingface-hub && \
@@ -60,13 +63,44 @@ RUN pip3 install --upgrade pip && \
 RUN mkdir -p /workspace
 WORKDIR /workspace
 
-RUN mkdir repositories && git clone --branch cuda --single-branch --depth 1 https://github.com/qwopqwop200/GPTQ-for-LLaMa.git repositories/GPTQ-for-LLaMa
+ENV TORCH_CUDA_ARCH_LIST="3.5;5.0;6.0;6.1;7.0;7.5;8.0;8.6+PTX"
-#RUN mkdir repositories && git clone https://github.com/AlpinDale/gptq-gptj.git repositories/GPTQ-for-LLaMa && (cd repositories/GPTQ-for-LLaMa && python3 setup_cuda.py install)
+RUN mkdir repositories && git clone --branch cuda --single-branch https://github.com/qwopqwop200/GPTQ-for-LLaMa.git repositories/GPTQ-for-LLaMa && \
+    (cd repositories/GPTQ-for-LLaMa && git reset --hard 437154dd434c3f9d5c9c4e6f401d6d71116ac248) && \
+#RUN mkdir repositories && git clone --depth 1 https://github.com/AlpinDale/gptq-gptj.git repositories/GPTQ-for-LLaMa && \
+    (cd repositories/GPTQ-for-LLaMa && python3 setup_cuda.py install)
+
 COPY model_fetcher.py /workspace/
 RUN python3 model_fetcher.py --model_name=${MODEL_NAME}
+
+
+FROM ${BASE_IMAGE}
+RUN mkdir -p /workspace
+WORKDIR /workspace
+
+RUN apt-get update --yes && \
+    apt install --yes --no-install-recommends \
+      python3 python3-dev python3-venv python3-pip && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+
+RUN pip3 install --upgrade pip && \
+    pip3 install torch torchvision torchaudio --extra-index-url=https://download.pytorch.org/whl/cu118 && \
+    pip3 install bitsandbytes && \
+    pip3 install safetensors && \
+    pip3 install sentencepiece triton && \
+    pip3 install diffusers && \
+    pip3 install transformers accelerate xformers triton && \
+    pip3 install huggingface-hub && \
+    pip3 install runpod && \
+    pip3 cache purge
+
+RUN mkdir /workspace/repositories
+COPY --from=builder /workspace/repositories /workspace/repositories/
+COPY --from=builder /root/.cache/huggingface /root/.cache/huggingface
+
 #RUN git lfs install && \
 #    git clone --depth 1 https://huggingface.co/${MODEL_NAME}
 
+COPY model_fetcher.py /workspace/
 COPY runpod_infer.py /workspace/
 COPY test_input.json /workspace/
 

runpod/runpod-worker-transformers/model_fetcher.py

@@ -80,6 +80,7 @@ def download_model(model_name):
         #https://huggingface.co/yahma/RWKV-14b_quant/resolve/main/RWKV-4-Pile-14B-20230213-8019.pqth
 
     if snapshot_path:
+        print("model downloaded to \"{snapshot_path}\"")
         os.system("ln -s \"{snapshot_path}\" /workdir/model")
 
 # ---------------------------------------------------------------------------- #

runpod/runpod-worker-transformers/runpod_infer.py

@@ -101,7 +101,7 @@ INPUT_SCHEMA = {
 }
 
 
-def load_quantized(model_name, wbits, groupsize):
+def load_quantized(model_name, wbits, groupsize, device):
     """https://github.com/oobabooga/text-generation-webui/blob/main/modules/GPTQ_loader.py"""
     from pathlib import Path
     import os, sys
@@ -167,7 +167,9 @@ def load_quantized(model_name, wbits, groupsize):
     print('Loading model ...')
     if str(pt_path).endswith('.safetensors'):
         from safetensors.torch import load_file as safe_load
-        model.load_state_dict(safe_load(str(pt_path)))
+        if device == -1:
+            device = "cpu"
+        model.load_state_dict(safe_load(str(pt_path), device))
     else:
         model.load_state_dict(torch.load(str(pt_path)))
     model.seqlen = 2048
@@ -248,7 +250,7 @@ if __name__ == "__main__":
     elif args.model_name == 'pygmalion-6b-4bit-128g':
 #        model = AutoModelForCausalLM.from_pretrained(
 #            "mayaeary/pygmalion-6b-4bit-128g", local_files_only=True).to(device)
-        model = load_quantized("pygmalion-6b-4bit-128g", 4, 128).to(device)
+        model = load_quantized("pygmalion-6b-4bit-128g", 4, 128, device).to(device)
         tokenizer = AutoTokenizer.from_pretrained(
             "mayaeary/pygmalion-6b-4bit-128g", local_files_only=True)