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

Add dummy model in iGPU perf (#12341)

Browse source 
* Add dummy model in iGPU perf

* Add dummy model in iGPU perf

* Fix
This commit is contained in:
Jin, Qiao 2024-11-05 17:56:10 +08:00 committed by GitHub
parent 8e9a3a1158
commit 7240c283a3
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
8 changed files with 86 additions and 36 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

36
python/llm/dev/benchmark/all-in-one/run.py
View file

@ -53,6 +53,8 @@ QWENVL_IDS = ['Qwen/Qwen-VL-Chat']
MINICPM_V_IDS = ['openbmb/MiniCPM-V-2_6', 'openbmb/MiniCPM-Llama3-V-2_5']
DUMMY_IDS = ['dummy/dummy-1.5B', 'dummy/dummy-4B']
results = []
excludes = []
@ -1245,6 +1247,10 @@ def run_transformer_int4_gpu_win(repo_id,
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = model.to('xpu')
model = model.llm
elif repo_id in DUMMY_IDS:
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 = model.to('xpu')
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()
@ -1255,7 +1261,10 @@ def run_transformer_int4_gpu_win(repo_id,
print(">> loading of model costs {}s and {}GB".format(load_time, torch.xpu.memory.memory_reserved()/(1024**3)))
model = BenchmarkWrapper(model)
if repo_id not in DUMMY_IDS:
streamer = TextStreamer(tokenizer, skip_prompt=True)
else:
streaming = False
result = {}
with torch.inference_mode():
@ -1264,6 +1273,7 @@ def run_transformer_int4_gpu_win(repo_id,
in_out_len = in_out.split("-")
in_len = int(in_out_len[0])
out_len = int(in_out_len[1])
if repo_id not in DUMMY_IDS:
input_str = get_continuation_input_str(in_len, tokenizer)
# As different tokenizer has different encodings,
# slice the input_ids to ensure the prompt length is required length.
@ -1272,6 +1282,8 @@ def run_transformer_int4_gpu_win(repo_id,
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')
else:
input_ids = torch.randint(1000, 2000, [batch_size, in_len], dtype=torch.int64).to('xpu')
actual_in_len = input_ids.shape[1]
result[in_out] = []
for i in range(num_trials + warm_up):
@ -1288,6 +1300,7 @@ def run_transformer_int4_gpu_win(repo_id,
end = time.perf_counter()
output_ids = output_ids.cpu()
print("model generate cost: " + str(end - st))
if repo_id not in DUMMY_IDS:
output = tokenizer.batch_decode(output_ids)
if not streaming:
print(output[0])
@ -1369,6 +1382,11 @@ def run_transformer_int4_fp16_gpu_win(repo_id,
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = model.to('xpu')
model = model.llm
elif repo_id in DUMMY_IDS:
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,
torch_dtype=torch.float16).eval()
model = model.to('xpu')
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,
@ -1380,7 +1398,10 @@ def run_transformer_int4_fp16_gpu_win(repo_id,
print(">> loading of model costs {}s and {}GB".format(load_time, torch.xpu.memory.memory_reserved()/(1024**3)))
model = BenchmarkWrapper(model)
if repo_id not in DUMMY_IDS:
streamer = TextStreamer(tokenizer, skip_prompt=True)
else:
streaming = False
result = {}
with torch.inference_mode():
@ -1389,6 +1410,7 @@ def run_transformer_int4_fp16_gpu_win(repo_id,
in_out_len = in_out.split("-")
in_len = int(in_out_len[0])
out_len = int(in_out_len[1])
if repo_id not in DUMMY_IDS:
input_str = get_continuation_input_str(in_len, tokenizer)
# As different tokenizer has different encodings,
# slice the input_ids to ensure the prompt length is required length.
@ -1397,6 +1419,8 @@ def run_transformer_int4_fp16_gpu_win(repo_id,
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')
else:
input_ids = torch.randint(1000, 2000, [batch_size, in_len], dtype=torch.int64).to('xpu')
actual_in_len = input_ids.shape[1]
result[in_out] = []
for i in range(num_trials + warm_up):
@ -1413,6 +1437,7 @@ def run_transformer_int4_fp16_gpu_win(repo_id,
end = time.perf_counter()
output_ids = output_ids.cpu()
print("model generate cost: " + str(end - st))
if repo_id not in DUMMY_IDS:
output = tokenizer.batch_decode(output_ids)
if not streaming:
print(output[0])
@ -1590,6 +1615,10 @@ def run_transformer_int4_fp16_loadlowbit_gpu_win(repo_id,
use_cache=True, cpu_embedding=cpu_embedding).eval()
tokenizer = AutoTokenizer.from_pretrained(model_path+'-'+low_bit, trust_remote_code=True)
model = model.half().to('xpu')
elif repo_id in DUMMY_IDS:
model = AutoModelForCausalLM.load_low_bit(model_path+'-'+low_bit, optimize_model=True, trust_remote_code=True,
use_cache=True, cpu_embedding=cpu_embedding).eval()
model = model.to('xpu')
else:
model = AutoModelForCausalLM.load_low_bit(model_path+'-'+low_bit, optimize_model=True, trust_remote_code=True,
use_cache=True, cpu_embedding=cpu_embedding).eval()
@ -1600,7 +1629,10 @@ def run_transformer_int4_fp16_loadlowbit_gpu_win(repo_id,
print(">> loading of model costs {}s and {}GB".format(load_time, torch.xpu.memory.memory_reserved()/(1024**3)))
model = BenchmarkWrapper(model)
if repo_id not in DUMMY_IDS:
streamer = TextStreamer(tokenizer, skip_prompt=True)
else:
streaming = False
result = {}
with torch.inference_mode():
@ -1609,6 +1641,7 @@ def run_transformer_int4_fp16_loadlowbit_gpu_win(repo_id,
in_out_len = in_out.split("-")
in_len = int(in_out_len[0])
out_len = int(in_out_len[1])
if repo_id not in DUMMY_IDS:
input_str = get_continuation_input_str(in_len, tokenizer)
# As different tokenizer has different encodings,
# slice the input_ids to ensure the prompt length is required length.
@ -1617,6 +1650,8 @@ def run_transformer_int4_fp16_loadlowbit_gpu_win(repo_id,
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')
else:
input_ids = torch.randint(1000, 2000, [batch_size, in_len], dtype=torch.int64).to('xpu')
actual_in_len = input_ids.shape[1]
result[in_out] = []
for i in range(num_trials + warm_up):
@ -1633,6 +1668,7 @@ def run_transformer_int4_fp16_loadlowbit_gpu_win(repo_id,
end = time.perf_counter()
output_ids = output_ids.cpu()
print("model generate cost: " + str(end - st))
if repo_id not in DUMMY_IDS:
output = tokenizer.batch_decode(output_ids)
if not streaming:
print(output[0])

2
python/llm/test/benchmark/igpu-perf/1024-128.yaml
View file

@ -19,6 +19,8 @@ repo_id:
- 'microsoft/Phi-3-mini-128k-instruct'
- 'microsoft/phi-3-vision-128k-instruct'
- 'openbmb/MiniCPM-V-2_6'
- 'dummy/dummy-1.5B'
- 'dummy/dummy-4B'
local_model_hub: 'path to your local model hub'
warm_up: 1
num_trials: 3

2
python/llm/test/benchmark/igpu-perf/1024-128_int4_fp16.yaml
View file

@ -18,6 +18,8 @@ repo_id:
- 'microsoft/Phi-3-mini-128k-instruct'
- 'microsoft/phi-3-vision-128k-instruct'
- 'openbmb/MiniCPM-V-2_6'
- 'dummy/dummy-1.5B'
- 'dummy/dummy-4B'
local_model_hub: 'path to your local model hub'
warm_up: 1
num_trials: 3

2
python/llm/test/benchmark/igpu-perf/1024-128_int4_fp16_loadlowbit.yaml
View file

@ -17,6 +17,8 @@ repo_id:
- 'microsoft/Phi-3-mini-4k-instruct'
- 'microsoft/Phi-3-mini-128k-instruct'
- 'microsoft/phi-3-vision-128k-instruct'
- 'dummy/dummy-1.5B'
- 'dummy/dummy-4B'
local_model_hub: 'path to your local model hub'
warm_up: 1
num_trials: 3

2
python/llm/test/benchmark/igpu-perf/2048-256_int4_fp16.yaml
View file

@ -18,6 +18,8 @@ repo_id:
- 'microsoft/Phi-3-mini-128k-instruct'
- 'microsoft/phi-3-vision-128k-instruct'
- 'openbmb/MiniCPM-V-2_6'
- 'dummy/dummy-1.5B'
- 'dummy/dummy-4B'
local_model_hub: 'path to your local model hub'
warm_up: 1
num_trials: 3

2
python/llm/test/benchmark/igpu-perf/3072-384_int4_fp16.yaml
View file

@ -17,6 +17,8 @@ repo_id:
- 'microsoft/Phi-3-mini-128k-instruct'
- 'microsoft/phi-3-vision-128k-instruct'
- 'openbmb/MiniCPM-V-2_6'
- 'dummy/dummy-1.5B'
- 'dummy/dummy-4B'
local_model_hub: 'path to your local model hub'
warm_up: 1
num_trials: 3

2
python/llm/test/benchmark/igpu-perf/32-32_int4_fp16.yaml
View file

@ -18,6 +18,8 @@ repo_id:
- 'microsoft/Phi-3-mini-128k-instruct'
- 'microsoft/phi-3-vision-128k-instruct'
- 'openbmb/MiniCPM-V-2_6'
- 'dummy/dummy-1.5B'
- 'dummy/dummy-4B'
local_model_hub: 'path to your local model hub'
warm_up: 3
num_trials: 5

2
python/llm/test/benchmark/igpu-perf/4096-512_int4_fp16.yaml
View file

@ -17,6 +17,8 @@ repo_id:
- 'microsoft/Phi-3-mini-128k-instruct'
- 'microsoft/phi-3-vision-128k-instruct'
- 'openbmb/MiniCPM-V-2_6'
- 'dummy/dummy-1.5B'
- 'dummy/dummy-4B'
local_model_hub: 'path to your local model hub'
warm_up: 1
num_trials: 3