Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Update packaging in AOTI path #896

Merged
merged 2 commits into from
Nov 1, 2024
Merged

Update packaging in AOTI path #896

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5cd2862
Update AOTI package
angelayi Sep 17, 2024
32c5807
Fix test-runner-aot-cuda
larryliu0820 Oct 31, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
36 changes: 18 additions & 18 deletions .ci/scripts/validate.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -133,60 +133,60 @@ function generate_aoti_model_output() {
echo "******************************************"
echo "************** non-quantized *************"
echo "******************************************"
python3 -W ignore torchchat.py export --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --output-dso-path "${MODEL_DIR}/${MODEL_NAME}.so" --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py generate --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --dso-path "$MODEL_DIR/${MODEL_NAME}.so" --prompt "$PROMPT" --device "$TARGET_DEVICE" > "$MODEL_DIR/output_aoti" || exit 1
python3 -W ignore torchchat.py export --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --output-aoti-package-path "${MODEL_DIR}/${MODEL_NAME}.pt2" --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py generate --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --aoti-package-path "$MODEL_DIR/${MODEL_NAME}.pt2" --prompt "$PROMPT" --device "$TARGET_DEVICE" > "$MODEL_DIR/output_aoti" || exit 1
.ci/scripts/check_gibberish "$MODEL_DIR/output_aoti"

echo "******************************************"
echo "******* Emb: channel-wise quantized ******"
echo "******************************************"
python3 -W ignore torchchat.py export --dtype ${DTYPE} --quant '{"embedding" : {"bitwidth": 8, "groupsize": 0}}' --checkpoint-path "$CHECKPOINT_PATH" --output-dso-path ${MODEL_DIR}/${MODEL_NAME}.so --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py generate --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --temperature 0 --dso-path ${MODEL_DIR}/${MODEL_NAME}.so --device "$TARGET_DEVICE" > "$MODEL_DIR/output_aoti" || exit 1
python3 -W ignore torchchat.py export --dtype ${DTYPE} --quant '{"embedding" : {"bitwidth": 8, "groupsize": 0}}' --checkpoint-path "$CHECKPOINT_PATH" --output-aoti-package-path ${MODEL_DIR}/${MODEL_NAME}.pt2 --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py generate --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --temperature 0 --aoti-package-path ${MODEL_DIR}/${MODEL_NAME}.pt2 --device "$TARGET_DEVICE" > "$MODEL_DIR/output_aoti" || exit 1
.ci/scripts/check_gibberish "$MODEL_DIR/output_aoti"

echo "******************************************"
echo "******** Emb: group-wise quantized *******"
echo "******************************************"
python3 -W ignore torchchat.py export --dtype ${DTYPE} --quant '{"embedding" : {"bitwidth": 8, "groupsize": 8}}' --checkpoint-path "$CHECKPOINT_PATH" --output-dso-path ${MODEL_DIR}/${MODEL_NAME}.so --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py generate --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --temperature 0 --dso-path ${MODEL_DIR}/${MODEL_NAME}.so --device "$TARGET_DEVICE" > "$MODEL_DIR/output_aoti" || exit 1
python3 -W ignore torchchat.py export --dtype ${DTYPE} --quant '{"embedding" : {"bitwidth": 8, "groupsize": 8}}' --checkpoint-path "$CHECKPOINT_PATH" --output-aoti-package-path ${MODEL_DIR}/${MODEL_NAME}.pt2 --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py generate --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --temperature 0 --aoti-package-path ${MODEL_DIR}/${MODEL_NAME}.pt2 --device "$TARGET_DEVICE" > "$MODEL_DIR/output_aoti" || exit 1
.ci/scripts/check_gibberish "$MODEL_DIR/output_aoti"

echo "***********************************************"
echo "******* Emb: 4bit channel-wise quantized ******"
echo "***********************************************"
python3 -W ignore torchchat.py export --dtype ${DTYPE} --quant '{"embedding" : {"bitwidth": 4, "groupsize": 0, "packed": "True"}}' --checkpoint-path "$CHECKPOINT_PATH" --output-dso-path ${MODEL_DIR}/${MODEL_NAME}.so --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py generate --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --temperature 0 --dso-path ${MODEL_DIR}/${MODEL_NAME}.so --device "$TARGET_DEVICE" > "$MODEL_DIR/output_aoti" || exit 1
python3 -W ignore torchchat.py export --dtype ${DTYPE} --quant '{"embedding" : {"bitwidth": 4, "groupsize": 0, "packed": "True"}}' --checkpoint-path "$CHECKPOINT_PATH" --output-aoti-package-path ${MODEL_DIR}/${MODEL_NAME}.pt2 --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py generate --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --temperature 0 --aoti-package-path ${MODEL_DIR}/${MODEL_NAME}.pt2 --device "$TARGET_DEVICE" > "$MODEL_DIR/output_aoti" || exit 1
.ci/scripts/check_gibberish "$MODEL_DIR/output_aoti"

echo "***********************************************"
echo "******** Emb: 4bit group-wise quantized *******"
echo "***********************************************"
python3 -W ignore torchchat.py export --dtype ${DTYPE} --quant '{"embedding" : {"bitwidth": 4, "groupsize": 8, "packed": "True"}}' --checkpoint-path "$CHECKPOINT_PATH" --output-dso-path ${MODEL_DIR}/${MODEL_NAME}.so --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py generate --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --temperature 0 --dso-path ${MODEL_DIR}/${MODEL_NAME}.so --device "$TARGET_DEVICE" > "$MODEL_DIR/output_aoti" || exit 1
python3 -W ignore torchchat.py export --dtype ${DTYPE} --quant '{"embedding" : {"bitwidth": 4, "groupsize": 8, "packed": "True"}}' --checkpoint-path "$CHECKPOINT_PATH" --output-aoti-package-path ${MODEL_DIR}/${MODEL_NAME}.pt2 --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py generate --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --temperature 0 --aoti-package-path ${MODEL_DIR}/${MODEL_NAME}.pt2 --device "$TARGET_DEVICE" > "$MODEL_DIR/output_aoti" || exit 1
.ci/scripts/check_gibberish "$MODEL_DIR/output_aoti"

if [ "${EXCLUDE_INT8_QUANT:-false}" == false ]; then
echo "******************************************"
echo "******* INT8 channel-wise quantized ******"
echo "******************************************"
python3 -W ignore torchchat.py export --dtype ${DTYPE} --quant '{"linear:int8" : {"bitwidth": 8, "groupsize": 0}}' --checkpoint-path "$CHECKPOINT_PATH" --output-dso-path ${MODEL_DIR}/${MODEL_NAME}.so --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py generate --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --temperature 0 --dso-path ${MODEL_DIR}/${MODEL_NAME}.so --device "$TARGET_DEVICE" > "$MODEL_DIR/output_aoti" || exit 1
python3 -W ignore torchchat.py export --dtype ${DTYPE} --quant '{"linear:int8" : {"bitwidth": 8, "groupsize": 0}}' --checkpoint-path "$CHECKPOINT_PATH" --output-aoti-package-path ${MODEL_DIR}/${MODEL_NAME}.pt2 --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py generate --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --temperature 0 --aoti-package-path ${MODEL_DIR}/${MODEL_NAME}.pt2 --device "$TARGET_DEVICE" > "$MODEL_DIR/output_aoti" || exit 1
.ci/scripts/check_gibberish "$MODEL_DIR/output_aoti"

echo "******************************************"
echo "******** INT8 group-wise quantized *******"
echo "******************************************"
python3 -W ignore torchchat.py export --dtype ${DTYPE} --quant '{"linear:int8" : {"bitwidth": 8, "groupsize": 8}}' --checkpoint-path "$CHECKPOINT_PATH" --output-dso-path ${MODEL_DIR}/${MODEL_NAME}.so --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py generate --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --temperature 0 --dso-path ${MODEL_DIR}/${MODEL_NAME}.so --device "$TARGET_DEVICE" > "$MODEL_DIR/output_aoti" || exit 1
python3 -W ignore torchchat.py export --dtype ${DTYPE} --quant '{"linear:int8" : {"bitwidth": 8, "groupsize": 8}}' --checkpoint-path "$CHECKPOINT_PATH" --output-aoti-package-path ${MODEL_DIR}/${MODEL_NAME}.pt2 --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py generate --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --temperature 0 --aoti-package-path ${MODEL_DIR}/${MODEL_NAME}.pt2 --device "$TARGET_DEVICE" > "$MODEL_DIR/output_aoti" || exit 1
.ci/scripts/check_gibberish "$MODEL_DIR/output_aoti"
fi
echo "******************************************"
echo "******** INT4 group-wise quantized *******"
echo "******************************************"
if [[ "$TARGET_DEVICE" != "cuda" || "$DTYPE" == "bfloat16" ]]; then
# For CUDA, only bfloat16 makes sense for int4 mm kernel
python3 -W ignore torchchat.py export --dtype ${DTYPE} --quant '{"linear:int4" : {"groupsize": 32}}' --checkpoint-path "$CHECKPOINT_PATH" --output-dso-path ${MODEL_DIR}/${MODEL_NAME}.so --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py generate --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --temperature 0 --dso-path ${MODEL_DIR}/${MODEL_NAME}.so --device "$TARGET_DEVICE" > "$MODEL_DIR/output_aoti" || exit 1
python3 -W ignore torchchat.py export --dtype ${DTYPE} --quant '{"linear:int4" : {"groupsize": 32}}' --checkpoint-path "$CHECKPOINT_PATH" --output-aoti-package-path ${MODEL_DIR}/${MODEL_NAME}.pt2 --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py generate --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --temperature 0 --aoti-package-path ${MODEL_DIR}/${MODEL_NAME}.pt2 --device "$TARGET_DEVICE" > "$MODEL_DIR/output_aoti" || exit 1
.ci/scripts/check_gibberish "$MODEL_DIR/output_aoti"
fi
done
Expand Down Expand Up @@ -285,8 +285,8 @@ function eval_model_sanity_check() {
echo "******** INT4 group-wise quantized (AOTI) *******"
echo "*************************************************"
if [ "$DTYPE" != "float16" ]; then
python3 -W ignore torchchat.py export --dtype ${DTYPE} --quant "$QUANT_OPTIONS" --checkpoint-path "$CHECKPOINT_PATH" --output-dso-path ${MODEL_DIR}/${MODEL_NAME}.so --dynamic-shapes --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py eval --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --dso-path ${MODEL_DIR}/${MODEL_NAME}.so --device "$TARGET_DEVICE" --limit 5 > "$MODEL_DIR/output_eval_aoti" || exit 1
python3 -W ignore torchchat.py export --dtype ${DTYPE} --quant "$QUANT_OPTIONS" --checkpoint-path "$CHECKPOINT_PATH" --output-aoti-package-path ${MODEL_DIR}/${MODEL_NAME}.pt2 --dynamic-shapes --device "$TARGET_DEVICE" || exit 1
python3 -W ignore torchchat.py eval --dtype ${DTYPE} --checkpoint-path "$CHECKPOINT_PATH" --aoti-package-path ${MODEL_DIR}/${MODEL_NAME}.pt2 --device "$TARGET_DEVICE" --limit 5 > "$MODEL_DIR/output_eval_aoti" || exit 1
cat "$MODEL_DIR/output_eval_aoti"
fi;
fi;
Expand Down
14 changes: 7 additions & 7 deletions .github/workflows/pull.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -378,8 +378,8 @@ jobs:

echo "::group::Run inference with quantize file"
if [ $(uname -s) == Darwin ]; then
python3 torchchat.py export --output-dso-path /tmp/model.so --quantize torchchat/quant_config/cuda.json --checkpoint "./checkpoints/${REPO_NAME}/model.pth"
python3 torchchat.py generate --dso-path /tmp/model.so --checkpoint "./checkpoints/${REPO_NAME}/model.pth"~
python3 torchchat.py export --output-aoti-package-path /tmp/model.pt2 --quantize torchchat/quant_config/cuda.json --checkpoint "./checkpoints/${REPO_NAME}/model.pth"
python3 torchchat.py generate --aoti-package-path /tmp/model.pt2 --checkpoint "./checkpoints/${REPO_NAME}/model.pth"~
fi
echo "::endgroup::"

Expand Down Expand Up @@ -1023,8 +1023,8 @@ jobs:

for dtype in fp32 fp16 bf16 fast fast16; do
echo "Running export + runner with dtype=$dtype"
python torchchat.py export --checkpoint-path ${MODEL_DIR}/stories15M.pt --dtype $dtype --output-dso-path /tmp/model.so
./cmake-out/aoti_run /tmp/model.so -z ${MODEL_DIR}/tokenizer.model -i "${PROMPT}"
python torchchat.py export --checkpoint-path ${MODEL_DIR}/stories15M.pt --dtype $dtype --output-aoti-package-path /tmp/model.pt2
./cmake-out/aoti_run /tmp/model.pt2 -z ${MODEL_DIR}/tokenizer.model -i "${PROMPT}"
done

echo "Tests complete."
Expand Down Expand Up @@ -1118,8 +1118,8 @@ jobs:
python torchchat.py export stories110M --output-pte-path ./model.pte --dtype float32 --quantize '{"embedding:wx": {"bitwidth": 2, "groupsize": 32}, "linear:a8wxdq": {"bitwidth": 3, "groupsize": 128, "has_weight_zeros": false}}'
./cmake-out/et_run ./model.pte -z ./tokenizer.model -t 0 -i "${PRMT}"
echo "Export and run AOTI (C++ runner)"
python torchchat.py export stories110M --output-dso-path ./model.so --dtype float32 --quantize '{"embedding:wx": {"bitwidth": 2, "groupsize": 32}, "linear:a8wxdq": {"bitwidth": 3, "groupsize": 128, "has_weight_zeros": false}}'
./cmake-out/aoti_run ./model.so -z ./tokenizer.model -t 0 -i "${PRMT}"
python torchchat.py export stories110M --output-aoti-package-path ./model.pt2 --dtype float32 --quantize '{"embedding:wx": {"bitwidth": 2, "groupsize": 32}, "linear:a8wxdq": {"bitwidth": 3, "groupsize": 128, "has_weight_zeros": false}}'
./cmake-out/aoti_run ./model.pt2 -z ./tokenizer.model -t 0 -i "${PRMT}"
echo "Generate AOTI"
python torchchat.py generate stories110M --dso-path ./model.so --prompt "${PRMT}"
python torchchat.py generate stories110M --aoti-package-path ./model.pt2 --prompt "${PRMT}"
echo "Tests complete."
4 changes: 2 additions & 2 deletions .github/workflows/runner-cuda-dtype.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,9 +56,9 @@ jobs:
for DTYPE in bfloat16; do
python torchchat.py generate --dtype ${DTYPE} --checkpoint-path ${MODEL_DIR}/stories15M.pt --temperature 0 --prompt "${PROMPT}" --device cuda

python torchchat.py export --checkpoint-path ${MODEL_DIR}/stories15M.pt --output-dso-path /tmp/model.so
python torchchat.py export --checkpoint-path ${MODEL_DIR}/stories15M.pt --output-aoti-package-path /tmp/model.pt2

./cmake-out/aoti_run /tmp/model.so -d CUDA -z ${MODEL_DIR}/tokenizer.model -i "${PROMPT}"
./cmake-out/aoti_run /tmp/model.pt2 -d CUDA -z ${MODEL_DIR}/tokenizer.model -i "${PROMPT}"

done

Expand Down
25 changes: 14 additions & 11 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -293,13 +293,18 @@ Use the "Max Response Tokens" slider to limit the maximum number of tokens gener
## Desktop/Server Execution

### AOTI (AOT Inductor)
[AOTI](https://pytorch.org/blog/pytorch2-2/) compiles models before execution for faster inference. The process creates a [DSO](https://en.wikipedia.org/wiki/Shared_library) model (represented by a file with extension `.so`)
that is then loaded for inference. This can be done with both Python and C++ environments.
[AOTI](https://pytorch.org/blog/pytorch2-2/) compiles models before execution
for faster inference. The process creates a zipped PT2 file containing all the
artifacts generated by AOTInductor, and a
[.so](https://en.wikipedia.org/wiki/Shared_library) file with the runnable
contents that is then loaded for inference. This can be done with both Python
and C++ enviroments.

The following example exports and executes the Llama3.1 8B Instruct
model. The first command compiles and performs the actual export.
```
python3 torchchat.py export llama3.1 --output-dso-path exportedModels/llama3.1.so

```bash
python3 torchchat.py export llama3.1 --output-aoti-package-path exportedModels/llama3_1_artifacts.pt2
```

> [!NOTE]
Expand All @@ -311,12 +316,11 @@ case visit our [customization guide](docs/model_customization.md).

### Run in a Python Environment

To run in a python environment, use the generate subcommand like before, but include the dso file.
To run in a python enviroment, use the generate subcommand like before, but include the pt2 file.

```bash
python3 torchchat.py generate llama3.1 --aoti-package-path exportedModels/llama3_1_artifacts.pt2 --prompt "Hello my name is"
```
python3 torchchat.py generate llama3.1 --dso-path exportedModels/llama3.1.so --prompt "Hello my name is"
```
**Note:** Depending on which accelerator is used to generate the .dso file, the command may need the device specified: `--device (cuda | cpu)`.


### Run using our C++ Runner
Expand All @@ -326,11 +330,10 @@ To run in a C++ enviroment, we need to build the runner binary.
torchchat/utils/scripts/build_native.sh aoti
```

Then run the compiled executable, with the exported DSO from earlier.
Then run the compiled executable, with the pt2.
```bash
cmake-out/aoti_run exportedModels/llama3.1.so -z `python3 torchchat.py where llama3.1`/tokenizer.model -l 3 -i "Once upon a time"
cmake-out/aoti_run exportedModels/llama3_1_artifacts.pt2 -z `python3 torchchat.py where llama3.1`/tokenizer.model -l 3 -i "Once upon a time"
```
**Note:** Depending on which accelerator is used to generate the .dso file, the runner may need the device specified: `-d (CUDA | CPU)`.

## Mobile Execution

Expand Down
19 changes: 4 additions & 15 deletions runner/run.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,10 +31,7 @@ LICENSE file in the root directory of this source tree.
#endif

#ifdef __AOTI_MODEL__
#include <torch/csrc/inductor/aoti_runner/model_container_runner_cpu.h>
#ifdef USE_CUDA
#include <torch/csrc/inductor/aoti_runner/model_container_runner_cuda.h>
#endif
#include <torch/csrc/inductor/aoti_package/model_package_loader.h>
torch::Device aoti_device(torch::kCPU);

#else // __ET_MODEL__
Expand Down Expand Up @@ -94,7 +91,7 @@ typedef struct {
RunState state; // buffers for the "wave" of activations in the forward pass

#ifdef __AOTI_MODEL__
torch::inductor::AOTIModelContainerRunner* runner;
torch::inductor::AOTIModelPackageLoader* runner;
#else // __ET_MODEL__
Module* runner;
#endif
Expand Down Expand Up @@ -144,16 +141,8 @@ void build_transformer(
malloc_run_state(&t->state, &t->config);

#ifdef __AOTI_MODEL__
#ifdef USE_CUDA
if (aoti_device.type() == torch::kCUDA) {
t->runner = new torch::inductor::AOTIModelContainerRunnerCuda(model_path);
aoti_device = torch::Device(torch::kCUDA);
} else {
#else
{
#endif
t->runner = new torch::inductor::AOTIModelContainerRunnerCpu(model_path);
}
t->runner = new torch::inductor::AOTIModelPackageLoader(model_path);
aoti_device = t->runner->get_metadata()["AOTI_DEVICE_KEY"] == "cpu" ? torch::Device(torch::kCPU) : torch::Device(torch::kCUDA);
#else //__ET_MODEL__
t->runner = new Module(
/* path to PTE model */ model_path,
Expand Down
Loading
Loading