updated MLC containers

packages/llm/mlc/benchmark.py

@@ -6,10 +6,17 @@
 import argparse
 import resource
 
-from mlc_chat import ChatModule, ChatConfig
-from mlc_chat.callback import StreamToStdout
-
-
+try:
+    from mlc_chat import ChatModule, ChatConfig
+    from mlc_chat.callback import StreamToStdout
+except Exception as error:
+    print(f"failed to import ChatModule from mlc_chat ({error})")
+    print(f"trying to import ChatModule from mlc_llm instead...")
+
+    from mlc_llm import ChatModule, ChatConfig
+    from mlc_llm.callback import StreamToStdout
+
+
 parser = argparse.ArgumentParser()
 
 parser.add_argument('--model', type=str, default="Llama-2-7b-chat-hf-q4f16_1")

packages/llm/mlc/patches/3403a4e.diff

@@ -321,6 +321,19 @@ index 4c36cab9..6afe6d0c 100644
  MLC_LIBRARY_PATH = os.environ.get("MLC_LIBRARY_PATH", None)
 
 
+diff --git a/python/mlc_llm/quantization/ft_quantization.py b/python/mlc_llm/quantization/ft_quantization.py
+index 4a158460..7cd8fc6a 100644
+--- a/python/mlc_llm/quantization/ft_quantization.py
++++ b/python/mlc_llm/quantization/ft_quantization.py
+@@ -207,7 +207,7 @@ class FTQuantize:  # pylint: disable=too-many-instance-attributes
+                                 relax.call_pure_packed(
+                                     "cutlass.ft_preprocess_weight",
+                                     lv1,
+-                                    detect_cuda_arch_list(target=target)[0],
++                                    int(detect_cuda_arch_list(target=target)[0]),
+                                     DataType(self.quantize_dtype).bits == 4,
+                                     sinfo_args=lv1.struct_info,
+                                 )
 diff --git a/python/setup.py b/python/setup.py
 index 2f1b632b..6c7c2088 100644
 --- a/python/setup.py

packages/llm/mlc/test.sh

@@ -1,40 +1,69 @@
 #!/usr/bin/env bash
-set -e
+set -ex
 
+MODEL_NAME=${1:-"Llama-2-7b-hf"}
+MODEL_PATH=${2:-"https://nvidia.box.com/shared/static/i3jtp8jdmdlsq4qkjof8v4muth8ar7fo.gz"}
 MODEL_ROOT="/data/models/mlc/dist"
-MLC_USE_CACHE=${MLC_USE_CACHE:-1}
 
-test_model()
+QUANTIZATION=${QUANTIZATION:-"q4f16_ft"}
+QUANTIZATION_PATH="${MODEL_ROOT}/${MODEL_NAME}-${QUANTIZATION}"
+SKIP_QUANTIZATION=${SKIP_QUANTIZATION:-"no"}
+
+USE_CACHE=${USE_CACHE:-1}
+CONV_TEMPLATE=${CONV_TEMPLATE:-"llama-2"}
+MAX_CONTEXT_LEN=${MAX_CONTEXT_LEN:-4096}
+
+
+quantize() # mlc_llm >= 0.1.1
 {
-	local MODEL_NAME=$1
-	local MODEL_URL=$2
-	local MODEL_TAR="$MODEL_NAME.tar.gz"
-	local QUANTIZATION=${3:-q4f16_ft}
+	QUANTIZATION_LIB="$QUANTIZATION_PATH/$MODEL_NAME-$QUANTIZATION-cuda.so"
 
-	if [ ! -d "$MODEL_ROOT/models/$MODEL_NAME" ]; then
+	if [ $SKIP_QUANTIZATION != "yes" ]; then
+		mlc_llm convert_weight $MODEL_PATH --quantization $QUANTIZATION --output $QUANTIZATION_PATH
+		mlc_llm gen_config $MODEL_PATH --quantization $QUANTIZATION --conv-template $CONV_TEMPLATE --context-window-size $MAX_CONTEXT_LEN --max-batch-size 1  --output $QUANTIZATION_PATH
+		mlc_llm compile $QUANTIZATION_PATH --device cuda --opt O3 --output $QUANTIZATION_LIB
+	fi
+
+	QUANTIZATION_LIB="--model-lib-path $QUANTIZATION_LIB"
+}
+
+quantize_legacy() # mlc_llm == 0.1.0
+{
+	if [ $SKIP_QUANTIZATION != "yes" ]; then
+		python3 -m mlc_llm.build \
+			--model $MODEL_NAME \
+			--target cuda \
+			--use-cuda-graph \
+			--use-flash-attn-mqa \
+			--use-cache $USE_CACHE \
+			--quantization $QUANTIZATION \
+			--artifact-path $MODEL_ROOT \
+			--max-seq-len 4096
+
+		if [ $? != 0 ]; then
+			return 1
+		fi
+	fi
+
+	QUANTIZATION_PATH="$QUANTIZATION_PATH/params"
+}
+
+if [[ $MODEL_PATH == http* ]]; then
+	MODEL_EXTRACTED="$MODEL_ROOT/models/$MODEL_NAME"
+	if [ ! -d "$MODEL_EXTRACTED" ]; then
 		cd $MODEL_ROOT/models
-		wget --quiet --show-progress --progress=bar:force:noscroll --no-check-certificate $MODEL_URL -O $MODEL_TAR
+		MODEL_TAR="$MODEL_NAME.tar.gz"
+		wget --quiet --show-progress --progress=bar:force:noscroll --no-check-certificate $MODEL_PATH -O $MODEL_TAR
 		tar -xzvf $MODEL_TAR
 		#rm $MODEL_TAR
 	fi
+	MODEL_PATH=$MODEL_EXTRACTED
+fi
 
-	cd $MODEL_ROOT
-
-	python3 -m mlc_llm.build \
-		--model $MODEL_NAME \
-		--target cuda \
-		--use-cuda-graph \
-		--use-flash-attn-mqa \
-		--use-cache $MLC_USE_CACHE \
-		--quantization $QUANTIZATION \
-		--artifact-path $MODEL_ROOT \
-		--max-seq-len 4096
-
-	python3 /opt/mlc-llm/benchmark.py \
-		--model ${MODEL_ROOT}/${MODEL_NAME}-${QUANTIZATION}/params \
-		--max-new-tokens 128 \
-		--max-num-prompts 4 \
-		--prompt /data/prompts/completion.json
-}
+quantize_legacy || quantize
 
-test_model "Llama-2-7b-hf" "https://nvidia.box.com/shared/static/i3jtp8jdmdlsq4qkjof8v4muth8ar7fo.gz"
+python3 benchmark.py \
+	--model $QUANTIZATION_PATH $QUANTIZATION_LIB \
+	--max-new-tokens 128 \
+	--max-num-prompts 4 \
+	--prompt /data/prompts/completion.json