ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

Update PP inference benchmark script (#11323)

Browse source
This commit is contained in:
binbin Deng 2024-06-17 09:59:36 +08:00 committed by GitHub
parent be00380f1a
commit 6ea1e71af0
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
5 changed files with 63 additions and 82 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

50
python/llm/dev/benchmark/all-in-one/README.md
View file

@ -27,7 +27,7 @@ repo_id:
- 'meta-llama/Llama-2-7b-chat-hf'
# - 'liuhaotian/llava-v1.5-7b' # requires a LLAVA_REPO_DIR env variables pointing to the llava dir; added only for gpu win related test_api now
local_model_hub: 'path to your local model hub'
warm_up: 1
warm_up: 1 # must set >=2 when run "pipeline_parallel_gpu" test_api
num_trials: 3
num_beams: 1 # default to greedy search
low_bit: 'sym_int4' # default to use 'sym_int4' (i.e. symmetric int4)
@ -36,29 +36,33 @@ in_out_pairs:
- '32-32'
- '1024-128'
test_api:
- "transformer_int4_gpu" # on Intel GPU
# - "transformer_int4_fp16_gpu" # on Intel GPU, use fp16 for non-linear layer
# - "ipex_fp16_gpu" # on Intel GPU
# - "bigdl_fp16_gpu" # on Intel GPU
# - "optimize_model_gpu" # on Intel GPU
# - "transformer_int4_gpu_win" # on Intel GPU for Windows
# - "transformer_int4_fp16_gpu_win" # on Intel GPU for Windows, use fp16 for non-linear layer
# - "transformer_int4_loadlowbit_gpu_win" # on Intel GPU for Windows using load_low_bit API. Please make sure you have used the save.py to save the converted low bit model
# - "deepspeed_optimize_model_gpu" # deepspeed autotp on Intel GPU
# - "speculative_gpu"
# - "transformer_int4"
# - "native_int4"
# - "optimize_model"
# - "pytorch_autocast_bf16"
# - "transformer_autocast_bf16"
# - "bigdl_ipex_bf16"
# - "bigdl_ipex_int4"
# - "bigdl_ipex_int8"
# - "speculative_cpu"
# - "deepspeed_transformer_int4_cpu" # on Intel SPR Server
- "transformer_int4_fp16_gpu" # on Intel GPU, transformer-like API, (qtype=int4), (dtype=fp16)
# - "transformer_int4_fp16_gpu_win" # on Intel GPU for Windows, transformer-like API, (qtype=int4), (dtype=fp16)
# - "transformer_int4_gpu" # on Intel GPU, transformer-like API, (qtype=int4), (dtype=fp32)
# - "transformer_int4_gpu_win" # on Intel GPU for Windows, transformer-like API, (qtype=int4), (dtype=fp32)
# - "transformer_int4_loadlowbit_gpu_win" # on Intel GPU for Windows, transformer-like API, (qtype=int4), use load_low_bit API. Please make sure you have used the save.py to save the converted low bit model
# - "bigdl_fp16_gpu" # on Intel GPU, use ipex-llm transformers API, (dtype=fp16), (qtype=fp16)
# - "optimize_model_gpu" # on Intel GPU, can optimize any pytorch models include transformer model
# - "deepspeed_optimize_model_gpu" # on Intel GPU, deepspeed autotp inference
# - "pipeline_parallel_gpu" # on Intel GPU, pipeline parallel inference
# - "speculative_gpu" # on Intel GPU, inference with self-speculative decoding
# - "transformer_int4" # on Intel CPU, transformer-like API, (qtype=int4)
# - "native_int4" # on Intel CPU
# - "optimize_model" # on Intel CPU, can optimize any pytorch models include transformer model
# - "pytorch_autocast_bf16" # on Intel CPU
# - "transformer_autocast_bf16" # on Intel CPU
# - "bigdl_ipex_bf16" # on Intel CPU, (qtype=bf16)
# - "bigdl_ipex_int4" # on Intel CPU, (qtype=int4)
# - "bigdl_ipex_int8" # on Intel CPU, (qtype=int8)
# - "speculative_cpu" # on Intel CPU, inference with self-speculative decoding
# - "deepspeed_transformer_int4_cpu" # on Intel CPU, deepspeed autotp inference
# - "transformer_int4_fp16_lookahead_gpu" # on Intel GPU, transformer-like API, with lookahead, (qtype=int4), (dtype=fp16)
cpu_embedding: False # whether put embedding to CPU
streaming: False # whether output in streaming way (only avaiable now for gpu win related test_api)
streaming: False # whether output in streaming way (only available now for gpu win related test_api)
use_fp16_torch_dtype: True # whether use fp16 for non-linear layer (only available now for "pipeline_parallel_gpu" test_api)
lookahead: 3
max_matching_ngram_size: 2
task: 'continuation' # when test_api is "transformer_int4_fp16_lookahead_gpu", task could be 'QA', 'continuation' or 'summarize'
```

1
python/llm/dev/benchmark/all-in-one/config.yaml
View file

@ -36,7 +36,6 @@ test_api:
cpu_embedding: False # whether put embedding to CPU
streaming: False # whether output in streaming way (only available now for gpu win related test_api)
use_fp16_torch_dtype: True # whether use fp16 for non-linear layer (only available now for "pipeline_parallel_gpu" test_api)
n_gpu: 2 # number of GPUs to use (only available now for "pipeline_parallel_gpu" test_api)
lookahead: 3
max_matching_ngram_size: 2
task: 'continuation' # when test_api is "transformer_int4_fp16_lookahead_gpu", task could be 'QA', 'continuation' or 'summarize'

13
python/llm/dev/benchmark/all-in-one/run-pipeline-parallel-arc.sh Normal file
View file

@ -0,0 +1,13 @@
source /opt/intel/oneapi/setvars.sh
export MASTER_ADDR=127.0.0.1
export MASTER_PORT=8080
export FI_PROVIDER=tcp
export USE_XETLA=OFF
export OMP_NUM_THREADS=6
if [[ $KERNEL_VERSION != *"6.5"* ]]; then
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
fi
export TORCH_LLM_ALLREDUCE=0
NUM_GPUS=2 # number of used GPU
CCL_ZE_IPC_EXCHANGE=sockets torchrun --standalone --nnodes=1 --nproc-per-node $NUM_GPUS run.py

75
python/llm/dev/benchmark/all-in-one/run.py
View file

@ -106,7 +106,7 @@ def preprocess_prompt(tokenizer, in_len, task):
input_ids = tokenizer.encode(input_str, return_tensors="pt")
return input_ids
def run_model(repo_id, test_api, in_out_pairs, local_model_hub=None, warm_up=1, num_trials=3, num_beams=1, low_bit='sym_int4', cpu_embedding=False, batch_size=1, streaming=False, use_fp16_torch_dtype=False, n_gpu=2):
def run_model(repo_id, test_api, in_out_pairs, local_model_hub=None, warm_up=1, num_trials=3, num_beams=1, low_bit='sym_int4', cpu_embedding=False, batch_size=1, streaming=False, use_fp16_torch_dtype=False):
# TODO: make a parameter
result= {}
if test_api == 'transformer_int4':
@ -152,7 +152,7 @@ def run_model(repo_id, test_api, in_out_pairs, local_model_hub=None, warm_up=1,
elif test_api == 'speculative_gpu':
result = run_speculative_gpu(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, batch_size)
elif test_api == 'pipeline_parallel_gpu':
result = run_pipeline_parallel_gpu(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, low_bit, batch_size, cpu_embedding, fp16=use_fp16_torch_dtype, n_gpu=n_gpu)
result = run_pipeline_parallel_gpu(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, low_bit, batch_size, cpu_embedding, fp16=use_fp16_torch_dtype)
elif test_api == "transformer_int4_fp16_lookahead_gpu":
result = run_transformer_int4_gpu(repo_id, local_model_hub, in_out_pairs, warm_up, num_trials, num_beams, low_bit, batch_size, cpu_embedding, fp16=True, lookahead=True)
else:
@ -1747,41 +1747,30 @@ def run_pipeline_parallel_gpu(repo_id,
low_bit,
batch_size,
cpu_embedding,
fp16=False,
n_gpu=2):
from ipex_llm.transformers import AutoModel, AutoModelForCausalLM
fp16=False):
from ipex_llm.transformers import AutoModel, AutoModelForCausalLM, init_pipeline_parallel
from transformers import AutoTokenizer, GPTJForCausalLM, LlamaTokenizer
init_pipeline_parallel()
model_path = get_model_path(repo_id, local_model_hub)
pipeline_parallel_stages = torch.distributed.get_world_size()
# Load model in 4 bit,
# which convert the relevant layers in the model into INT4 format
st = time.perf_counter()
origin_repo_id = repo_id.replace("-4bit", "")
if origin_repo_id in CHATGLM_IDS:
if "4bit" in repo_id:
model = AutoModel.load_low_bit(model_path, optimize_model=True,
trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding).eval()
else:
model = AutoModel.from_pretrained(model_path, load_in_low_bit=low_bit, optimize_model=True,
trust_remote_code=True, use_cache=True).eval()
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True, cpu_embedding=cpu_embedding)
trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding,
pipeline_parallel_stages=pipeline_parallel_stages).eval()
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
elif origin_repo_id in LLAMA_IDS:
model = AutoModelForCausalLM.from_pretrained(model_path, load_in_low_bit=low_bit, trust_remote_code=True,
use_cache=True, cpu_embedding=cpu_embedding).eval()
model = AutoModelForCausalLM.from_pretrained(model_path, load_in_low_bit=low_bit, optimize_model=True,
trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding,
pipeline_parallel_stages=pipeline_parallel_stages).eval()
tokenizer = LlamaTokenizer.from_pretrained(model_path, trust_remote_code=True)
else:
if "4bit" in repo_id:
model = AutoModelForCausalLM.load_low_bit(model_path, optimize_model=True,
trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding).eval()
else:
if 'starcoder' in repo_id:
# Load starcoder-15.5b model in bf16 format to avoid CPU OOM.
model = AutoModelForCausalLM.from_pretrained(model_path, optimize_model=True, load_in_low_bit=low_bit,
trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding, torch_dtype=torch.bfloat16).eval()
# Convert the low-bit model back to fp32 for performance considerations.
model = model.float()
else:
model = AutoModelForCausalLM.from_pretrained(model_path, optimize_model=True, load_in_low_bit=low_bit,
trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding).eval()
model = AutoModelForCausalLM.from_pretrained(model_path, load_in_low_bit=low_bit, optimize_model=True,
trust_remote_code=True, use_cache=True, cpu_embedding=cpu_embedding,
pipeline_parallel_stages=pipeline_parallel_stages).eval()
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
if fp16:
@ -1792,29 +1781,8 @@ def run_pipeline_parallel_gpu(repo_id,
load_time = end - st
print(">> loading of model costs {}s and {}GB".format(load_time, torch.xpu.memory.memory_reserved()/(1024**3)))
model_layers = ['model.embed_tokens']
for i in range(model.config.num_hidden_layers):
model_layers.append(f'model.layers.{i}')
model_layers = model_layers + ['model.norm', 'lm_head']
device_map = {}
split_len = len(model_layers) // n_gpu
for i in range(n_gpu):
device_map.update({key: f'xpu:{i}' for key in model_layers[split_len * i: split_len * (i + 1)]})
if i == n_gpu - 1:
device_map.update({key: f'xpu:{i}' for key in model_layers[split_len * (i + 1): ]})
print(f">> device map: {device_map}")
from accelerate import dispatch_model
model = dispatch_model(
model,
device_map=device_map,
offload_dir=None,
skip_keys=["past_key_value", "past_key_values"],
)
model = BenchmarkWrapper(model)
result = {}
local_rank = torch.distributed.get_rank()
with torch.inference_mode():
for in_out in in_out_pairs:
in_out_len = in_out.split("-")
@ -1833,7 +1801,7 @@ def run_pipeline_parallel_gpu(repo_id,
input_ids = input_ids[:, :in_len]
true_str = tokenizer.batch_decode(input_ids)[0]
input_list = [true_str] * batch_size
input_ids = tokenizer(input_list, return_tensors="pt").input_ids.to('xpu:0')
input_ids = tokenizer(input_list, return_tensors="pt").input_ids.to(f'xpu:{local_rank}')
actual_in_len = input_ids.shape[1]
result[in_out] = []
for i in range(num_trials + warm_up):
@ -1849,8 +1817,8 @@ def run_pipeline_parallel_gpu(repo_id,
print(output[0])
torch.xpu.empty_cache()
if i >= warm_up:
result[in_out].append([model.first_cost, model.rest_cost_mean, model.encoder_time,
actual_in_len, actual_out_len, load_time, model.peak_memory, fp16])
result[in_out].append([model.first_token_time, model.rest_cost_mean, 0,
actual_in_len, actual_out_len, load_time,])
del model
torch.xpu.empty_cache()
return result
@ -1865,13 +1833,10 @@ if __name__ == '__main__':
excludes = conf['exclude']
streaming = False
use_fp16_torch_dtype = False
n_gpu = 2
if 'streaming' in conf:
streaming = conf['streaming']
if 'use_fp16_torch_dtype' in conf:
use_fp16_torch_dtype = conf['use_fp16_torch_dtype']
if 'n_gpu' in conf:
n_gpu = conf['n_gpu']
import pandas as pd
for api in conf.test_api:
@ -1891,7 +1856,7 @@ if __name__ == '__main__':
if model_id_input in excludes or model_id_input_batch_size in excludes:
in_out_pairs.remove(in_out)
run_model(model, api, in_out_pairs, conf['local_model_hub'], conf['warm_up'], conf['num_trials'], conf['num_beams'],
conf['low_bit'], conf['cpu_embedding'], batch_size, streaming, use_fp16_torch_dtype, n_gpu)
conf['low_bit'], conf['cpu_embedding'], batch_size, streaming, use_fp16_torch_dtype)
df = pd.DataFrame(results, columns=['model', '1st token avg latency (ms)', '2+ avg latency (ms/token)', 'encoder time (ms)',
'input/output tokens', 'batch_size', 'actual input/output tokens', 'num_beams', 'low_bit', 'cpu_embedding',
'model loading time (s)', 'peak mem (GB)', 'streaming', 'use_fp16_torch_dtype'])

4
python/llm/src/ipex_llm/transformers/pipeline_parallel.py
View file

@ -172,7 +172,7 @@ def pipeline_parallel_generate(self,
past_key_values=_past_key_values, use_cache=True)
else:
inputs_embeds = torch.empty(_input_ids.shape + (self.config.hidden_size,),
device=f'xpu:{local_rank}', dtype=torch.float32)
device=f'xpu:{local_rank}', dtype=self.dtype)
dist.recv(inputs_embeds, src=pre_rank)
outputs = self(input_ids=None, inputs_embeds=inputs_embeds,
past_key_values=_past_key_values, use_cache=True)
@ -182,7 +182,7 @@ def pipeline_parallel_generate(self,
next_ids = torch.argmax(logits[:, -1:, :], dim=-1)
dist.broadcast(next_ids, src=local_rank)
else:
dist.send(outputs[0], dst=next_rank)
dist.send(outputs[0].to(self.dtype), dst=next_rank)
next_ids = torch.empty((bs, 1), device=f'xpu:{local_rank}', dtype=torch.int64)
dist.broadcast(next_ids, src=self.pipeline_parallel_stages - 1)