repo_name stringlengths 1 62 | dataset stringclasses 1
value | lang stringclasses 11
values | pr_id int64 1 20.1k | owner stringlengths 2 34 | reviewer stringlengths 2 39 | diff_hunk stringlengths 15 262k | code_review_comment stringlengths 1 99.6k |
|---|---|---|---|---|---|---|---|
llm-on-ray | github_2023 | python | 204 | intel | minmingzhu | @@ -193,92 +242,43 @@ def train_func(config: Dict[str, Any]):
}
)
- optimizer = common.optimizer.Optimizer.registory.get("DefaultOptimizer")()(
- model,
- config={
- "name": config["Training"]["optimizer"],
- "config": {"lr": config["Training"]["learning_rate"]},
-... | transformers.Trainer how to distinguish FSDP and Deepspeed? |
llm-on-ray | github_2023 | python | 204 | intel | KepingYan | @@ -130,6 +121,52 @@ def get_accelerate_environment_variable(config: Dict[str, Any]) -> dict:
return mode_env_vars[device][accelerate_mode]
+def convert_to_training_args(cls, config):
+ device = config["Training"]["device"]
+ accelerate_mode = config["Training"]["accelerate_mode"]
+ checkpoint_dir = c... | Can you add the max_train_steps parameter, otherwise, the UI will not be able to demo finetuning task in a short time.
https://github.com/intel/llm-on-ray/blob/b0a5840df2675bc797040e24c4ec4aa6043001cd/llm_on_ray/ui/start_ui.py#L631-L632
In addition, can other parameters of lr_scheduler such as num_warmup_steps be s... |
llm-on-ray | github_2023 | python | 204 | intel | KepingYan | @@ -339,27 +339,22 @@ def main(external_config=None):
"FI_PROVIDER": "tcp",
}
}
- accelerate_env_vars = get_accelerate_environment_variable(config)
- runtime_env["env_vars"].update(accelerate_env_vars)
+ # accelerate_env_vars = get_accelerate_environment_varia... | Is get_accelerate_environment_variable no longer needed? Let us remove this function. |
llm-on-ray | github_2023 | python | 204 | intel | KepingYan | @@ -176,13 +173,22 @@ def group_texts(examples):
desc=f"Grouping texts in chunks of {block_size}",
)
+ return tokenized_datasets
+
+ def convert_dataset(self, tokenizer, dataset): | general_processer is just generating dataloader, it may be better to call it 'prepare' or 'prepare_dataloader'. And please align function names in other files such as pretrain modules'. |
llm-on-ray | github_2023 | python | 204 | intel | KepingYan | @@ -130,7 +130,7 @@ def prepare(self, model, tokenizer, dataset, optimizer, accelerator):
f"model embedding size resize to {len(tokenizer)} because of tokenizer size"
)
- train_dataloader, eval_dataloader = self.dataprocesser.prepare(tokenizer, dataset)
+ train_dataloader, ... | If we no longer use default_trainer, should this file be removed? |
llm-on-ray | github_2023 | python | 204 | intel | KepingYan | @@ -193,92 +242,43 @@ def train_func(config: Dict[str, Any]):
}
)
- optimizer = common.optimizer.Optimizer.registory.get("DefaultOptimizer")()(
- model,
- config={
- "name": config["Training"]["optimizer"],
- "config": {"lr": config["Training"]["learning_rate"]},
-... | How is resuming finetuning from checkpoint supported? |
llm-on-ray | github_2023 | python | 204 | intel | KepingYan | @@ -193,92 +242,43 @@ def train_func(config: Dict[str, Any]):
}
)
- optimizer = common.optimizer.Optimizer.registory.get("DefaultOptimizer")()(
- model,
- config={
- "name": config["Training"]["optimizer"],
- "config": {"lr": config["Training"]["learning_rate"]},
-... | Why remove optimizer? |
llm-on-ray | github_2023 | python | 204 | intel | KepingYan | @@ -193,92 +242,43 @@ def train_func(config: Dict[str, Any]):
}
)
- optimizer = common.optimizer.Optimizer.registory.get("DefaultOptimizer")()(
- model,
- config={
- "name": config["Training"]["optimizer"],
- "config": {"lr": config["Training"]["learning_rate"]},
-... | I think we should pass optimizer parameter into the Trainer according to the user's configuration. |
llm-on-ray | github_2023 | python | 204 | intel | carsonwang | @@ -130,6 +121,52 @@ def get_accelerate_environment_variable(config: Dict[str, Any]) -> dict:
return mode_env_vars[device][accelerate_mode]
+def convert_to_training_args(cls, config):
+ device = config["Training"]["device"]
+ accelerate_mode = config["Training"]["accelerate_mode"]
+ checkpoint_dir = c... | The above three configs are hard coded? Are these values also used in our previous implementation? |
llm-on-ray | github_2023 | python | 204 | intel | carsonwang | @@ -130,6 +121,52 @@ def get_accelerate_environment_variable(config: Dict[str, Any]) -> dict:
return mode_env_vars[device][accelerate_mode]
+def convert_to_training_args(cls, config):
+ device = config["Training"]["device"]
+ accelerate_mode = config["Training"]["accelerate_mode"]
+ checkpoint_dir = c... | It seems this is not set. |
llm-on-ray | github_2023 | python | 204 | intel | carsonwang | @@ -130,6 +121,52 @@ def get_accelerate_environment_variable(config: Dict[str, Any]) -> dict:
return mode_env_vars[device][accelerate_mode]
+def convert_to_training_args(cls, config):
+ device = config["Training"]["device"]
+ accelerate_mode = config["Training"]["accelerate_mode"]
+ checkpoint_dir = c... | Do we have default values for all these configurations in `config`? Previously we wrote `config["Training"].get("gradient_accumulation_steps", 1)` |
llm-on-ray | github_2023 | python | 204 | intel | KepingYan | @@ -322,44 +323,27 @@ def main(external_config=None):
"accelerate_mode"
] = "DDP" # will use DDP to accelerate if no method specified
- ccl_worker_count = 1
device = config["Training"]["device"]
if device != "cpu":
- ccl_worker_count = num_training_workers
+ pass
... | Why are these ccl configurations no longer needed? |
llm-on-ray | github_2023 | python | 165 | intel | xwu99 | @@ -187,9 +130,7 @@ def main(argv=None):
# all models are served under the same URL and then accessed
# through model_id, so it needs to pass in a unified URL.
host = "127.0.0.1" if args.serve_local_only else "0.0.0.0"
- rp = args.route_prefix if args.route_prefix else ""
- rout... | we can't use a fixed port here as the previous port is configurable. |
llm-on-ray | github_2023 | python | 165 | intel | xwu99 | @@ -31,7 +30,7 @@ def get_deployed_models(args):
2. Use relevant configuration parameters to generate `InferenceConfig` if model_id_or_path is set,
3. Serve all pre-defined models in inference/models/*.yaml, or part of them if models is set. | I don't think we are serving all models by default if none were chosen. Due to the resource constraint, it's default to serving gpt2 if none where chosen. |
llm-on-ray | github_2023 | python | 165 | intel | xwu99 | @@ -31,7 +30,7 @@ def get_deployed_models(args):
2. Use relevant configuration parameters to generate `InferenceConfig` if model_id_or_path is set, | The comments also need to be update. |
llm-on-ray | github_2023 | python | 165 | intel | xwu99 | @@ -89,57 +71,13 @@ def main(argv=None):
type=str,
help="Inference configuration file in YAML. If specified, all other arguments will be ignored.",
)
- parser.add_argument("--model_id_or_path", default=None, type=str, help="Model name or path.")
- parser.add_argument(
- "--tokenizer_... | help doc like this needed to be revised. |
llm-on-ray | github_2023 | python | 165 | intel | xwu99 | @@ -89,57 +71,13 @@ def main(argv=None):
type=str,
help="Inference configuration file in YAML. If specified, all other arguments will be ignored.", | need to revise |
llm-on-ray | github_2023 | python | 165 | intel | xwu99 | @@ -47,23 +46,6 @@ def get_deployed_models(args):
print("reading from config file, " + args.config_file) | ```suggestion
print("Reading from config file: " + args.config_file)
``` |
llm-on-ray | github_2023 | others | 165 | intel | xwu99 | @@ -32,11 +32,8 @@ We support three methods to specify the models to be served, and they have the f
```
llm_on_ray-serve --config_file llm_on_ray/inference/models/gpt2.yaml
```
-2. Use relevant configuration parameters if model_id_or_path is set.
```
-llm_on_ray-serve --model_id_or_path gpt2 [--tokenizer_id_or_path... | ```suggestion
2. If --config_file is None, it will serve GPT2 by default, or the models specified by --models.
``` |
llm-on-ray | github_2023 | others | 123 | intel | xwu99 | @@ -68,8 +68,10 @@ jobs:
- name: Build Docker Image
run: |
- docker build ./ --build-arg CACHEBUST=1 --build-arg http_proxy=${{ inputs.http_proxy }} --build-arg https_proxy=${{ inputs.https_proxy }} -f dev/docker/Dockerfile.cpu_and_deepspeed -t finetune:latest && yes | docker container prune ... | I think you can move some environment constants to .github/workflows/config/docker.sh, let the function focus on the key parameters regarding to docker. |
llm-on-ray | github_2023 | others | 123 | intel | xwu99 | @@ -0,0 +1,54 @@
+set -eo pipefail | need to add a shebang like `#!/usr/bin/env bash`, please check other scripts. |
llm-on-ray | github_2023 | others | 123 | intel | xwu99 | @@ -83,6 +83,10 @@ jobs:
echo "target is ${target}"
echo "target=$target" >> $GITHUB_OUTPUT
+ - name: Source build script
+ run: |
+ cp .github/workflows/scripts/docker.sh . | move the script to dev/scripts, source directly not need to copy |
llm-on-ray | github_2023 | others | 123 | intel | xwu99 | @@ -83,6 +83,10 @@ jobs:
echo "target is ${target}"
echo "target=$target" >> $GITHUB_OUTPUT
+ - name: Source build script
+ run: |
+ cp .github/workflows/scripts/docker.sh .
+
- name: Build Docker Image
run: |
if [[ ${{ matrix.model }} == "mpt-7b... | also move this block to a function to return DF_SUFFIX |
llm-on-ray | github_2023 | others | 123 | intel | xwu99 | @@ -95,23 +99,31 @@ jobs:
DF_SUFFIX=".cpu_and_deepspeed"
fi
TARGET=${{steps.target.outputs.target}}
- docker build ./ --build-arg CACHEBUST=1 --build-arg http_proxy=${{ inputs.http_proxy }} --build-arg https_proxy=${{ inputs.https_proxy }} -f dev/docker/Dockerfile${DF_SUFFIX}... | rename to docker_exec |
llm-on-ray | github_2023 | others | 123 | intel | xwu99 | @@ -95,23 +99,31 @@ jobs:
DF_SUFFIX=".cpu_and_deepspeed"
fi
TARGET=${{steps.target.outputs.target}}
- docker build ./ --build-arg CACHEBUST=1 --build-arg http_proxy=${{ inputs.http_proxy }} --build-arg https_proxy=${{ inputs.https_proxy }} -f dev/docker/Dockerfile${DF_SUFFIX}... | I don't like this. Could we provide a file template and use sed to replace the variables?
Also could we define a map from model to the yaml file in bash and move the logic into bash? |
llm-on-ray | github_2023 | others | 123 | intel | xwu99 | @@ -0,0 +1,76 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+HTTP_PROXY='http://10.24.221.149:911'
+HTTPS_PROXY='http://10.24.221.149:911'
+MODEL_CACHE_PATH_LOACL='/root/.cache/huggingface/hub'
+CODE_CHECKOUT_PATH_LOCAL='/root/llm-on-ray'
+
+build_and_prune() {
+ # Set TARGET and DF-SUFFIX using the passed in paramete... | Could you use $1, $2 to denote the arguments? it looks like this is not the standard way to pass in arguments? |
llm-on-ray | github_2023 | python | 123 | intel | xwu99 | @@ -0,0 +1,90 @@
+#
+# Copyright 2023 The LLM-on-Ray Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by... | argv is no necessary as argparse.ArgumentParser() will get the command line? |
llm-on-ray | github_2023 | others | 123 | intel | xwu99 | @@ -0,0 +1,269 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+HTTP_PROXY='http://10.24.221.149:911'
+HTTPS_PROXY='http://10.24.221.149:911'
+MODEL_CACHE_PATH_LOACL='/root/.cache/huggingface/hub'
+CODE_CHECKOUT_PATH_LOCAL='/root/llm-on-ray'
+
+build_and_prune_test() {
+ # Set TARGET and DF-SUFFIX using the passed in pa... | pls use test_ as suffix or test_ as prefix for all test functions. |
llm-on-ray | github_2023 | others | 123 | intel | xwu99 | @@ -70,122 +70,46 @@ jobs:
- name: Build Docker Image
run: |
- docker build ./ --build-arg CACHEBUST=1 --build-arg http_proxy=${{ inputs.http_proxy }} --build-arg https_proxy=${{ inputs.https_proxy }} -f dev/docker/Dockerfile.cpu_and_deepspeed -t finetune:latest
- docker container pr... | Could we use USE_PROXY=1 or 0 instead of using "USE_PROXY" as input? |
llm-on-ray | github_2023 | others | 123 | intel | xwu99 | @@ -70,122 +70,46 @@ jobs:
- name: Build Docker Image
run: |
- docker build ./ --build-arg CACHEBUST=1 --build-arg http_proxy=${{ inputs.http_proxy }} --build-arg https_proxy=${{ inputs.https_proxy }} -f dev/docker/Dockerfile.cpu_and_deepspeed -t finetune:latest
- docker container pr... | could we move huggingface login to docker container start as it will be used for all the rest of steps?
Start Docker Container can be renamed to Start Docker Container and Initialize |
llm-on-ray | github_2023 | others | 123 | intel | xwu99 | @@ -150,43 +150,47 @@ jobs:
echo "target is ${target}"
echo "target=$target" >> $GITHUB_OUTPUT
+ # - name: Source build script
+ # run: |
+ # cp .github/workflows/scripts/docker.sh .
+
- name: Build Docker Image
run: |
DF_SUFFIX=".tests_cpu"
... | unnecessary comment |
llm-on-ray | github_2023 | others | 123 | intel | xwu99 | @@ -0,0 +1,269 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+HTTP_PROXY='http://10.24.221.149:911'
+HTTPS_PROXY='http://10.24.221.149:911'
+MODEL_CACHE_PATH_LOACL='/root/.cache/huggingface/hub'
+CODE_CHECKOUT_PATH_LOCAL='/root/llm-on-ray'
+
+build_and_prune_test() {
+ # Set TARGET and DF-SUFFIX using the passed in pa... | better to print out some info about what test we are conducting before each docker exec |
llm-on-ray | github_2023 | others | 123 | intel | xwu99 | @@ -0,0 +1,269 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+HTTP_PROXY='http://10.24.221.149:911'
+HTTPS_PROXY='http://10.24.221.149:911'
+MODEL_CACHE_PATH_LOACL='/root/.cache/huggingface/hub'
+CODE_CHECKOUT_PATH_LOCAL='/root/llm-on-ray'
+
+build_and_prune_test() {
+ # Set TARGET and DF-SUFFIX using the passed in pa... | it looks it's not get mapper but get suffix |
llm-on-ray | github_2023 | python | 179 | intel | carsonwang | @@ -32,13 +33,17 @@ def __init__(self, infer_conf: InferenceConfig, max_num_seqs):
model_config = model_desc.config
dtype = "bfloat16" if infer_conf.vllm.precision == PRECISION_BF16 else "float32"
+ # Set environment variable VLLM_CPU_KVCACHE_SPACE to control the size of the CPU key-value cac... | This should have been fixed and need to pass the device, right? |
llm-on-ray | github_2023 | python | 179 | intel | carsonwang | @@ -25,20 +26,25 @@
class VllmPredictor(Predictor):
+ VLLM_CPU_KVCACHE_SPACE_DEFAULT = 40
+
def __init__(self, infer_conf: InferenceConfig, max_num_seqs):
super().__init__(infer_conf)
model_desc = infer_conf.model_description
model_config = model_desc.config
dtype = "b... | We should expose this config in the yaml file |
llm-on-ray | github_2023 | python | 167 | intel | carsonwang | @@ -0,0 +1,189 @@
+import openai
+import pytest
+import os
+import subprocess
+from openai import OpenAI
+
+os.environ["no_proxy"] = "localhost,127.0.0.1"
+os.environ["OPENAI_API_BASE"] = "http://localhost:8000/v1"
+os.environ["OPENAI_API_KEY"] = "YOUR_OPEN_AI_KEY"
+os.environ["OPENAI_BASE_URL"] = "http://localhost:800... | The same function also appears in `test_example_simple.py`. Can we put it in a common place? |
llm-on-ray | github_2023 | python | 167 | intel | carsonwang | @@ -0,0 +1,189 @@
+import openai
+import pytest
+import os
+import subprocess
+from openai import OpenAI
+
+os.environ["no_proxy"] = "localhost,127.0.0.1"
+os.environ["OPENAI_API_BASE"] = "http://localhost:8000/v1"
+os.environ["OPENAI_API_KEY"] = "YOUR_OPEN_AI_KEY"
+os.environ["OPENAI_BASE_URL"] = "http://localhost:800... | We can remove 4 methods here starting with `completions_` as they are not used. |
llm-on-ray | github_2023 | python | 167 | intel | carsonwang | @@ -0,0 +1,189 @@
+import openai
+import pytest
+import os
+import subprocess
+from openai import OpenAI
+
+os.environ["no_proxy"] = "localhost,127.0.0.1"
+os.environ["OPENAI_API_BASE"] = "http://localhost:8000/v1"
+os.environ["OPENAI_API_KEY"] = "YOUR_OPEN_AI_KEY"
+os.environ["OPENAI_BASE_URL"] = "http://localhost:800... | Can we enable these two starting with `chat_` and remove others starting with `completions_` ? |
llm-on-ray | github_2023 | python | 69 | intel | carsonwang | @@ -0,0 +1,256 @@
+# Adapted from https://github.com/vllm-project/vllm/blob/main/benchmarks/benchmark_serving.py
+
+import argparse
+import asyncio
+import json
+import random
+import time
+from tqdm import tqdm
+from typing import AsyncGenerator, Dict, List, Tuple, Union
+
+import aiohttp
+import numpy as np
+from tra... | If max_new_tokens is specified in config, the value is not set to config["max_tokens"]. Remove the else word so this line always runs? |
llm-on-ray | github_2023 | python | 69 | intel | carsonwang | @@ -0,0 +1,256 @@
+# Adapted from https://github.com/vllm-project/vllm/blob/main/benchmarks/benchmark_serving.py
+
+import argparse
+import asyncio
+import json
+import random
+import time
+from tqdm import tqdm
+from typing import AsyncGenerator, Dict, List, Tuple, Union
+
+import aiohttp
+import numpy as np
+from tra... | Are these prints only for debug? Remove them so they are not included in the latency. |
llm-on-ray | github_2023 | python | 69 | intel | carsonwang | @@ -0,0 +1,256 @@
+# Adapted from https://github.com/vllm-project/vllm/blob/main/benchmarks/benchmark_serving.py
+
+import argparse
+import asyncio
+import json
+import random
+import time
+from tqdm import tqdm
+from typing import AsyncGenerator, Dict, List, Tuple, Union
+
+import aiohttp
+import numpy as np
+from tra... | If you start using command `python inference/serve.py --config_file <model_config_file.yaml> --simple --keep_serve_terminal` as described in the doc, the yaml file might not in the models directory. So we should use the `route_prefix` and `tokenizer_name_or_path` etc specified in the user conf file instead of `all_mode... |
llm-on-ray | github_2023 | python | 69 | intel | carsonwang | @@ -0,0 +1,256 @@
+# Adapted from https://github.com/vllm-project/vllm/blob/main/benchmarks/benchmark_serving.py | Can we use the same format like files we implemented the openAI API? Adding the apache license header. |
llm-on-ray | github_2023 | python | 69 | intel | KepingYan | @@ -0,0 +1,589 @@
+#
+# Copyright 2024 The LLM-on-Ray Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required b... | Maybe we need to make a deep copy of the parameter config in this loop. |
llm-on-ray | github_2023 | python | 69 | intel | KepingYan | @@ -0,0 +1,589 @@
+#
+# Copyright 2024 The LLM-on-Ray Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required b... | Or make a deep copy of the config parameters here. Otherwise, when passing through line 231 for the first time, the value of `config` will become {'max_new_tokens': 37}, so that the output_len will be 37 every time in the future. This is why total_time is reduced a lot |
llm-on-ray | github_2023 | python | 109 | intel | xwu99 | @@ -57,10 +64,15 @@ def script_with_args(api_base, model_name, streaming_response, max_new_tokens, t
result_openai = subprocess.run(cmd_openai, capture_output=True, text=True)
+ # Ensure there are no errors in the OpenAI API query script execution
+ try: | this is a test, you need to raise the assert exception rather than capture it. |
llm-on-ray | github_2023 | python | 109 | intel | xwu99 | @@ -57,10 +64,15 @@ def script_with_args(api_base, model_name, streaming_response, max_new_tokens, t
result_openai = subprocess.run(cmd_openai, capture_output=True, text=True)
+ # Ensure there are no errors in the OpenAI API query script execution
+ try:
+ assert result_openai.returncode == 0, pri... | this seems strange to me to use ` ,` to seperate two statements. |
llm-on-ray | github_2023 | others | 166 | intel | carsonwang | @@ -5,7 +5,7 @@ General:
checkpoint_dir: /tmp/llm-ray/checkpoint
config:
trust_remote_code: false
- use_auth_token: null
+ use_auth_tokenuth_token: null | Is this unintended modification? |
llm-on-ray | github_2023 | python | 166 | intel | carsonwang | @@ -57,10 +57,10 @@ def test_max_input_len():
# Add more tests for edge cases
-def test_get_torch_dtype_cpu_without_ipex(mock_infer_conf):
- hf_config = None
- dtype = get_torch_dtype(mock_infer_conf, hf_config)
- assert dtype == torch.get_default_dtype()
+def test_decide_torch_dtype_cpu_without_ipex(mock... | can you please cover more test cases in your implementation, for example device is HPU? |
llm-on-ray | github_2023 | python | 166 | intel | carsonwang | @@ -23,6 +23,8 @@ class ModelConfig(BaseModel):
trust_remote_code: bool = False
use_auth_token: Union[str, None] = None
load_in_4bit: bool = False
+ torch_dtype: Union[str, None] = None | IPEX has a `precision` config, vLLM has a `precision` config too. Here we are adding a new one. All should use this single config, right? @kira-lin @KepingYan @xwu99 |
llm-on-ray | github_2023 | others | 189 | intel | xwu99 | @@ -0,0 +1,22 @@
+## Deploying and Serving LLMs with ipex-LLM
+[ipex-LLM](https://ipex-llm.readthedocs.io/en/latest/doc/LLM/index.html) is a library for running LLM (large language model) on Intel XPU (from Laptop to GPU to Cloud) using INT4 with very low latency (for any PyTorch model). | Pls honor the original capitalization. ipex-LLM should be IPEX-LLM when refer to the product.
When refer to library, you can use ipex-llm.
See: https://ipex-llm.readthedocs.io/en/latest/doc/LLM/index.html for styles.
Pls update other cases as well. |
llm-on-ray | github_2023 | others | 190 | intel | carsonwang | @@ -0,0 +1,25 @@
+port: 8000
+name: deepseek-coder-33b-instruct
+route_prefix: /deepseek-coder-33b-instruct
+num_replicas: 1
+cpus_per_worker: 24
+gpus_per_worker: 0
+deepspeed: false
+workers_per_group: 2
+device: "cpu" | ```suggestion
device: cpu
``` |
llm-on-ray | github_2023 | others | 190 | intel | KepingYan | @@ -0,0 +1,25 @@
+port: 8000
+name: deepseek-coder-33b-instruct
+route_prefix: /deepseek-coder-33b-instruct
+num_replicas: 1
+cpus_per_worker: 24
+gpus_per_worker: 0
+deepspeed: false
+workers_per_group: 2
+device: CPU | Please change this to lowercase letters. |
llm-on-ray | github_2023 | others | 196 | intel | carsonwang | @@ -1,4 +1,4 @@
-FROM vault.habana.ai/gaudi-docker/1.14.0/ubuntu22.04/habanalabs/pytorch-installer-2.1.1:latest
+FROM vault.habana.ai/gaudi-docker/1.15.1/ubuntu22.04/habanalabs/pytorch-installer-2.2.0:latest | Should we also update to 1.15.1 in line 15 `pip install git+https://github.com/HabanaAI/DeepSpeed.git@1.14.0`? |
llm-on-ray | github_2023 | others | 154 | intel | xwu99 | @@ -26,3 +26,11 @@ jobs:
- name: Run Lint
run: ./format.sh -a
+
+ - name: Install dependencies for license check
+ run: |
+ python -m pip install --upgrade pip
+ pip install regex
+
+ - name: Run License Check
+ run: python .github/license/license-header.py fi... | It is better to use --files, otherwise it will be confused with the a file named `files` |
llm-on-ray | github_2023 | python | 154 | intel | xwu99 | @@ -0,0 +1,237 @@
+#
+# Copyright 2023 The LLM-on-Ray Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required b... | It's OK to raise exception from where the error happened instead rather than using global variable to mark a failure. |
llm-on-ray | github_2023 | python | 153 | intel | carsonwang | @@ -244,20 +244,34 @@ def streaming_generate(self, prompt, streamer, **config):
for worker in self.prediction_workers[1:]:
worker.streaming_generate.remote(inputs_ref, self._create_dummy_streamer(), **config)
- def generate(self, prompt, **config):
- input_ids, input_length = self.toke... | Can we simplify this method by always accepting prompts as List[str] like below? We can write another method that accepts a prompt: str and returns `self.generate([prompt], **config)[0]`
```suggestion
def generate(
self, prompts: List[str], **config
) -> List[GenerateResult]:
input_ids, i... |
llm-on-ray | github_2023 | python | 153 | intel | carsonwang | @@ -115,20 +116,34 @@ def streaming_generate(self, prompt, streamer, **config):
**config,
)
- def generate(self, prompt, **config):
+ def generate( | same here. |
llm-on-ray | github_2023 | python | 153 | intel | carsonwang | @@ -93,145 +104,253 @@ async def consume_streamer_async(self, streamer: TextIteratorStreamer):
# back to the event loop so other coroutines can run.
await asyncio.sleep(0.001)
- async def __call__(self, http_request: Request) -> Union[StreamingResponse, JSONResponse, str]:
- ... | Update `prompt: str` to include list in the method definition. |
llm-on-ray | github_2023 | python | 153 | intel | carsonwang | @@ -93,145 +104,253 @@ async def consume_streamer_async(self, streamer: TextIteratorStreamer):
# back to the event loop so other coroutines can run.
await asyncio.sleep(0.001)
- async def __call__(self, http_request: Request) -> Union[StreamingResponse, JSONResponse, str]:
- ... | I feel this file is too complex, too many if else. Can we further simplify this? For example, if we are returning ModelResponse for openai, can we return SimpleModelResponse for simple serving server, but the logic to create the response can be the same. So you can create a method to generate the response without if el... |
llm-on-ray | github_2023 | python | 153 | intel | KepingYan | @@ -0,0 +1,102 @@
+#
+# Copyright 2023 The LLM-on-Ray Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required b... | Please modify the endpoint according to parameter --model_endpoint. |
llm-on-ray | github_2023 | python | 153 | intel | KepingYan | @@ -93,145 +104,253 @@ async def consume_streamer_async(self, streamer: TextIteratorStreamer):
# back to the event loop so other coroutines can run.
await asyncio.sleep(0.001)
- async def __call__(self, http_request: Request) -> Union[StreamingResponse, JSONResponse, str]:
- ... | This should be str(dict(sorted(config.items()))), now if config is different from sorted(config.items()), batched_prompts[key] will have only one prompt. |
llm-on-ray | github_2023 | others | 153 | intel | carsonwang | @@ -116,6 +116,7 @@ The following are detailed guidelines for pretraining, finetuning and serving LL
* [Deploy and Serve LLMs on Intel CPU/GPU/Gaudi](docs/serve.md)
* [Deploy and Serve LLMs with Deepspeed](docs/serve_deepspeed.md)
* [Deploy and Serve LLMs with BigDL-LLM](docs/serve_bigdl.md)
+* [Enable Continuous, S... | Let's remove this and the file for now. |
llm-on-ray | github_2023 | python | 181 | intel | carsonwang | @@ -155,6 +155,10 @@ def train_func(config: Dict[str, Any]):
gradient_accumulation_steps = config["Training"].get("gradient_accumulation_steps", 1)
base_model = config["General"]["base_model"]
+ if config["General"].get("tokenizer_name") is not None:
+ tokenizer_name = config["General"].get("token... | ```suggestion
tokenizer_name = config["General"].get("tokenizer_name", base_model)
``` |
llm-on-ray | github_2023 | python | 158 | intel | carsonwang | @@ -22,11 +22,10 @@ def serve_run(deployments, model_list):
for model_id, infer_conf in model_list.items():
print("deploy model: ", model_id)
deployment = deployments[model_id]
+
+ serve.start(http_options={"host": infer_conf.host, "port": infer_conf.port})
serve.run(
... | can we pass `keep_serve_terminal` and set it to the new `blocking` parameter in serve.run? Better to also rename `keep_serve_terminal` to `blocking` in our script which will be more consistent and easy to understand. |
llm-on-ray | github_2023 | python | 170 | intel | carsonwang | @@ -208,7 +208,7 @@ def train(self):
if step % logging_steps == 0:
loss = loss.item()
ppl = math.exp(loss)
- epochs = (step + idx * total_steps) / (num_train_epochs * total_steps)
+ epochs = idx + step /... | What about update `total_steps` to `steps_per_epoch`? Otherwise it was confusing. Also check line 222 when we report the total steps, that should be steps_per_epoch * num_train_epochs? |
llm-on-ray | github_2023 | python | 170 | intel | carsonwang | @@ -219,9 +219,9 @@ def train(self):
"train_epoch": idx,
"total_epochs": num_train_epochs,
"train_step": step,
- "total_steps": min(max_train_step, total_steps)
+ ... | ```suggestion
"total_steps": min(max_train_step, steps_per_epoch * num_train_epochs)
if max_train_step
else steps_per_epoch * num_train_epochs,
``` |
llm-on-ray | github_2023 | others | 160 | intel | carsonwang | @@ -46,21 +46,30 @@ Training:
GPU: 1
accelerate_mode: GPU_DDP
```
+For HPU, set `device` to HPU, set HPU number in `resources_per_worker` and set `accelerate_mode` to HPU_DDP or HPU_DEEPSPEED.
+```
+Training:
+ device: HPU
+ resources_per_worker:
+ CPU: 1
+ HPU: 1
+ accelerate_mode: HPU_DDP
+```
Plea... | Also update the description above, "The following models have been verified on Intel CPUs, GPUs and HPUs." |
llm-on-ray | github_2023 | others | 160 | intel | carsonwang | @@ -46,21 +46,30 @@ Training:
GPU: 1
accelerate_mode: GPU_DDP
```
+For HPU, set `device` to HPU, set HPU number in `resources_per_worker` and set `accelerate_mode` to HPU_DDP or HPU_DEEPSPEED.
+```
+Training:
+ device: HPU
+ resources_per_worker:
+ CPU: 1
+ HPU: 1
+ accelerate_mode: HPU_DDP
+```
Plea... | Did you run and verify gpt-j on Gaudi? |
llm-on-ray | github_2023 | others | 160 | intel | carsonwang | @@ -46,21 +46,30 @@ Training:
GPU: 1
accelerate_mode: GPU_DDP
```
+For HPU, set `device` to HPU, set HPU number in `resources_per_worker` and set `accelerate_mode` to HPU_DDP or HPU_DEEPSPEED.
+```
+Training:
+ device: HPU
+ resources_per_worker:
+ CPU: 1
+ HPU: 1
+ accelerate_mode: HPU_DDP | I am wondering if we can just use DDP，FSDP, and Deepspeed for this parameter without adding CPU_, GPU_ and HPU_. Because we can already know the device from `device` parameter, right? |
llm-on-ray | github_2023 | python | 160 | intel | carsonwang | @@ -137,13 +137,12 @@ def train_func(config: Dict[str, Any]):
config={
"name": base_model,
"dtype": convert_dtype(config["Training"].get("mixed_precision", "no")),
+ "device": torch.device(config["Training"]["device"].lower()), | Can you please do the lower conversion in finetune_config.py, referring to https://github.com/intel/llm-on-ray/blob/main/llm_on_ray/inference/inference_config.py#L137 |
llm-on-ray | github_2023 | others | 160 | intel | carsonwang | @@ -2,7 +2,7 @@ General:
base_model: EleutherAI/gpt-j-6b
gpt_base_model: true
output_dir: /tmp/llm-ray/output
- checkpoint_dir: /tmp/llm-ray/checkpoint
+ checkpoint_dir: null | Why do we change this? |
llm-on-ray | github_2023 | python | 160 | intel | carsonwang | @@ -1771,10 +1771,11 @@ def _init_ui(self):
"optimizer": "AdamW",
"lr_scheduler": "linear",
"weight_decay": 0.0,
- "device": "CPU",
+ "device": "cpu", | The document says "CPU". Let's be consistent to use "CPU" |
llm-on-ray | github_2023 | python | 160 | intel | carsonwang | @@ -69,11 +69,11 @@ def check_device(cls, v: str):
devices = ["CPU", "GPU", "HPU"]
if v not in devices:
raise ValueError(f"device must be one of {devices}")
- return v
+ return v.lower() | ```suggestion
DEVICE_CPU = "cpu"
DEVICE_HPU = "hpu"
DEVICE_GPU = "gpu"
@validator("device")
def _check_device(cls, v: str):
if v:
assert v.lower() in [DEVICE_CPU, DEVICE_GPU, DEVICE_HPU]
return v.lower()
``` |
llm-on-ray | github_2023 | python | 160 | intel | carsonwang | @@ -68,23 +71,23 @@ def get_accelerate_environment_variable(mode: str, config: Union[Dict[str, Any],
"FSDP_SYNC_MODULE_STATES": "true",
"ACCELERATE_MIXED_PRECISION": mixed_precision,
},
- "GPU_DEEPSPEED": {
+ "gpu_DEEPSPEED": {
"ACCELERATE_USE_CPU": "false",... | The error message above is no longer correct because accelerate mode can only be CPU, GPU or Deepspeed now. Can you please print something like Accelerate mode xxx is not supported on device xxx? |
llm-on-ray | github_2023 | others | 160 | intel | carsonwang | @@ -0,0 +1,25 @@
+{
+ "steps_per_print": 64,
+ "train_batch_size": "auto",
+ "train_micro_batch_size_per_gpu": "auto",
+ "gradient_accumulation_steps": "auto",
+ "bf16": {
+ "enabled": true
+ },
+ "optimizer": { | We already have optimizer configured in the yaml, why it appears again in this json file? This should only include deepspeed config, right? |
llm-on-ray | github_2023 | python | 163 | intel | xwu99 | @@ -47,7 +50,10 @@ async def _get_generator_output(self, results_generator):
async def generate_async(self, prompts: Union[str, List[str]], **config) -> GenerateResult:
config = self.update_vllm_config(**config)
+ # In order to align with vllm test parameters
+ config["ignore_eos"] = True
... | remove this |
llm-on-ray | github_2023 | python | 163 | intel | xwu99 | @@ -148,6 +150,12 @@ def main(argv=None):
action="store_true",
help="Whether to keep serve terminal.",
)
+ parser.add_argument(
+ "--ray_max_concurrent_queries", required=True, type=int, help="The batch size in Ray." | need better help. the batch size here causes more confusions. just say "Max concurrent requests the serving system can process."
--ray_max_concurrent_queries no need to add ray here. user just need to know it's the limit to the serving system. |
llm-on-ray | github_2023 | python | 141 | intel | carsonwang | @@ -70,6 +70,13 @@ def get_deployed_models(args):
deployments = {}
for model_id, infer_conf in model_list.items():
ray_actor_options = get_deployment_actor_options(infer_conf)
+ if args.model_id_or_path is None and args.config_file is None:
+ if "mistral-7b-v0.1" in args.models:
+ ... | Please see if we can disable IPEX and make Mistral work without pinning transformer to 4.35. |
llm-on-ray | github_2023 | python | 134 | intel | carsonwang | @@ -355,26 +369,42 @@ async def chat(
status_code=results.error.code,
type=results.error.type,
)
- results = results.dict()
- choices: List[ChatCompletionResponseChoice] = [
- ... | ```suggestion
if results.tool_calls is not None:
msg = ChatMessage(role="assistant", tool_calls=results.tool_calls)
# deleting this fields so that they don't appear in the response
del msg.tool_call_id
e... |
llm-on-ray | github_2023 | python | 134 | intel | carsonwang | @@ -1713,6 +1718,12 @@ def _init_ui(self):
type=str,
help="The ip:port of head node to connect when restart a worker node.",
)
+ parser.add_argument(
+ "--ref_app_url",
+ default="http://127.0.0.1:8501", | How can other external users make this work by default? can you please add document how to launch it? If not, can you please remove and add an option to make it work for you only for now? |
llm-on-ray | github_2023 | others | 134 | intel | carsonwang | @@ -188,6 +188,14 @@ jobs:
docker exec "${TARGET}" bash -c "python examples/inference/api_server_openai/query_http_requests.py --model_name ${{ matrix.model }}"
fi
+ - name: Run Agent tool Inference Test with REST API
+ run: |
+ TARGET=${{steps.target.outputs.target}}
+ ... | This is to run the example, can we test and verify the LLM response is a correct Json to call the function? |
llm-on-ray | github_2023 | python | 87 | intel | carsonwang | @@ -0,0 +1,208 @@
+# Reference: https://github.com/huggingface/optimum-habana/blob/main/examples/text-generation/utils.py
+import os
+import tempfile
+import ray
+from torch_dist import (
+ TorchDistributedWorker,
+ init_torch_dist_process_group,
+)
+from ray.util.scheduling_strategies import PlacementGroupSchedu... | can we use get_torch_dtype in utils? Update that function if needed. |
llm-on-ray | github_2023 | python | 87 | intel | carsonwang | @@ -0,0 +1,208 @@
+# Reference: https://github.com/huggingface/optimum-habana/blob/main/examples/text-generation/utils.py | can you follow our openAI API code and use the same way to write the license and header here? |
llm-on-ray | github_2023 | python | 87 | intel | carsonwang | @@ -0,0 +1,208 @@
+# Reference: https://github.com/huggingface/optimum-habana/blob/main/examples/text-generation/utils.py
+import os
+import tempfile
+import ray
+from torch_dist import (
+ TorchDistributedWorker,
+ init_torch_dist_process_group,
+)
+from ray.util.scheduling_strategies import PlacementGroupSchedu... | Do we need this kind of configs for other larger models? Does it work with other modes right now? |
llm-on-ray | github_2023 | python | 87 | intel | carsonwang | @@ -0,0 +1,208 @@
+# Reference: https://github.com/huggingface/optimum-habana/blob/main/examples/text-generation/utils.py
+import os
+import tempfile
+import ray
+from torch_dist import (
+ TorchDistributedWorker,
+ init_torch_dist_process_group,
+)
+from ray.util.scheduling_strategies import PlacementGroupSchedu... | Is this always true? Otherwise, we need to add a config in the yaml. |
llm-on-ray | github_2023 | python | 87 | intel | carsonwang | @@ -0,0 +1,194 @@
+"""This file is modeled after ray/python/ray/train/torch/config.py
+
+The logics are duplicated right now to allow maximum flexibility for
+setting up PyTorch DDP process groups outside the context of Ray Train.
+Eventually, these use cases should be consolidated.
+"""
+
+from abc import ABC
+from co... | Do we need to clean up for HPU? |
llm-on-ray | github_2023 | others | 87 | intel | carsonwang | @@ -0,0 +1,24 @@
+port: 8000
+name: llama-2-70b-chat-hf
+route_prefix: /llama-2-70b-chat-hf | Add `num_replicas: 1` after `route_prefix` that was recently introduced. |
llm-on-ray | github_2023 | python | 87 | intel | carsonwang | @@ -0,0 +1,194 @@
+"""This file is modeled after ray/python/ray/train/torch/config.py | same here. |
llm-on-ray | github_2023 | others | 87 | intel | carsonwang | @@ -2,9 +2,34 @@ FROM vault.habana.ai/gaudi-docker/1.14.0/ubuntu22.04/habanalabs/pytorch-installe
ENV LANG=en_US.UTF-8
-COPY ../../ /root/llm-on-ray
-
WORKDIR /root/llm-on-ray
-RUN pip install . && \
- pip install --upgrade-strategy eager optimum[habana]
\ No newline at end of file
+COPY ./pyproject.toml .
+... | ```suggestion
# Required by DeepSpeed
``` |
llm-on-ray | github_2023 | others | 87 | intel | carsonwang | @@ -56,19 +56,24 @@ For Gaudi:
Please use the [Dockerfile](../dev/docker/Dockerfile.habana) to build the image. Alternatively, you can install the dependecies on a bare metal machine. In this case, please refer to [here](https://docs.habana.ai/en/latest/Installation_Guide/Bare_Metal_Fresh_OS.html#build-docker-bare).
... | can we add the following options as we used in our demo? `-e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --cap-add sys_ptrace --net=host --ipc=host llm-on-ray:habana` |
llm-on-ray | github_2023 | python | 87 | intel | carsonwang | @@ -0,0 +1,281 @@
+# Reference: https://github.com/huggingface/optimum-habana/blob/main/examples/text-generation/utils.py | Also add license header for this file |
llm-on-ray | github_2023 | python | 87 | intel | carsonwang | @@ -0,0 +1,313 @@
+#
+# Copyright 2023 The LLM-on-Ray Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required b... | Need to import from llm_on_ray.xxx like below. can you please update all of others in your PR?
```
from llm_on_ray.inference.inference_config import InferenceConfig, GenerateResult
``` |
llm-on-ray | github_2023 | python | 93 | intel | xwu99 | @@ -0,0 +1,77 @@
+import subprocess
+import pytest
+
+# Config matrix
+# models_array = ["gpt2", "gpt2 gpt-j-6b", "gpt2 bloom-560m", "falcon-7b"]
+# model_endpoint_array = ["http://127.0.0.1:8000", None]
+# streaming_response_array = [True, False]
+# max_new_tokens_array = [10, None]
+# temperature_array = [0.7, None]
... | I think you can use list directly without defining another models_array above. it's the same and more clearer. |
llm-on-ray | github_2023 | python | 93 | intel | xwu99 | @@ -0,0 +1,77 @@
+import subprocess
+import pytest
+
+# Config matrix
+# models_array = ["gpt2", "gpt2 gpt-j-6b", "gpt2 bloom-560m", "falcon-7b"]
+# model_endpoint_array = ["http://127.0.0.1:8000", None]
+# streaming_response_array = [True, False]
+# max_new_tokens_array = [10, None]
+# temperature_array = [0.7, None]
... | could we test all config matrix above and report what don't work? |
llm-on-ray | github_2023 | python | 93 | intel | xwu99 | @@ -0,0 +1,68 @@
+import subprocess
+import pytest
+
+# Config matrix
+# models_array = ["gpt2", "gpt2 gpt-j-6b", "gpt2 bloom-560m", "falcon-7b"]
+# model_endpoint_array = ["http://127.0.0.1:8000", None]
+# streaming_response_array = [True, False]
+# max_new_tokens_array = [10, None]
+# temperature_array = [0.7, None]
... | I think you can check the exit code of the processes. Successful run is 0. If exception thrown, it should be non-zero. |
llm-on-ray | github_2023 | others | 84 | intel | xwu99 | @@ -0,0 +1,50 @@
+#!/bin/bash
+set -eo pipefail
+
+# Step 1: Python environment
+# Check Python version is or later than 3.9
+if ! python -c 'import sys; assert sys.version_info >= (3,9)' > /dev/null; then
+ exit "Python should be 3.9 or later!"
+fi
+
+# Clone repo
+git clone https://github.com/intel/llm-on-ray.git ... | we have updated the way to start serving by using llm_on_ray-serve, please check README.md for details. |
llm-on-ray | github_2023 | others | 84 | intel | xwu99 | @@ -50,6 +50,7 @@ jobs:
- name: Run Tests
run: |
./tests/run-tests.sh
+ ./tests/test_getting_started.sh | pls run this in a separate Step:
```code
- name: Run Test for Getting Started
run: |
xxx
``` |
llm-on-ray | github_2023 | others | 84 | intel | xwu99 | @@ -0,0 +1,45 @@
+#!/bin/bash
+set -eo pipefail
+
+# Step 1: Python environment
+# Check Python version is or later than 3.9
+if ! python -c 'import sys; assert sys.version_info >= (3,9)' > /dev/null; then
+ exit "Python should be 3.9 or later!"
+fi
+
+# Clone repo
+git clone https://github.com/intel/llm-on-ray.git ... | better to print some output message for what is being tested for each step in the script then we can check CI log later. |
llm-on-ray | github_2023 | others | 84 | intel | xwu99 | @@ -1,7 +1,7 @@
port: 8000
name: gpt2
route_prefix: /gpt2
-cpus_per_worker: 2
+cpus_per_worker: 1 | Is it OK to use 2 cores, it will be much faster? |
llm-on-ray | github_2023 | others | 84 | intel | xwu99 | @@ -0,0 +1,57 @@
+#!/bin/bash
+set -eo pipefail
+
+# Usage: ./test_setup [CPU, GPU, Gaudi] [Deepspeed]
+
+# Step 1: Clone and create conda environment | also we can print out those step instead of comments that can be checked in the CI log. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.