EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Feat: support cuda graph for LoRA (#4115)

Browse Source 
Co-authored-by: Beichen Ma <mabeichen12@gmail.com>
This commit is contained in:
Qiaolin Yu
2025-04-29 02:30:44 -04:00
committed by GitHub
parent 2c3ea29476
commit 8c0cfca87d
13 changed files with 366 additions and 55 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
test/srt/models/lora/test_lora_backend.py
View File

@@ -24,7 +24,7 @@ from utils import (
DEFAULT_PROMPTS,
TORCH_DTYPES,
LoRAModelCase,
run_batch_lora_test,
run_lora_test_one_by_one,
)
from sglang.test.test_utils import CustomTestCase, is_in_ci
@@ -42,7 +42,7 @@ class TestLoRABackend(CustomTestCase):
)
for torch_dtype in TORCH_DTYPES:
for backend in BACKENDS:
run_batch_lora_test(
run_lora_test_one_by_one(
prompts,
model_case,
torch_dtype,

110
test/srt/models/lora/test_lora_cuda_graph.py Normal file
View File

@@ -0,0 +1,110 @@
# Copyright 2023-2024 SGLang Team
# 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 applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
import multiprocessing as mp
import os
import unittest
from typing import List
from utils import (
ALL_OTHER_LORA_MODELS,
CI_LORA_MODELS,
DEFAULT_PROMPTS,
TORCH_DTYPES,
LoRAModelCase,
run_lora_test_by_batch,
run_lora_test_one_by_one,
)
from sglang.test.test_utils import CustomTestCase, is_in_ci
TEST_CUDA_GRAPH_PADDING_PROMPTS = [
"AI is a field of computer science focused on",
"""
### Instruction:
Tell me about llamas and alpacas
### Response:
Llamas are large, long-necked animals with a woolly coat. They have two toes on each foot instead of three like other camelids (camels, dromedaries). Llamas live in the Andean mountains of South America where they graze on grasses and shrubs. Alpaca is another name for domesticated llama. The word "alpaca" comes from an Incan language meaning "golden fleece." Alpacas look very similar to llamas but are smaller than their wild relatives. Both species were used by ancient people as pack animals and for meat. Today both llamas and alpacas are raised primarily for their fiber which can be spun into yarn or knitted into clothing.
### Question 2:
What do you know about llamas?
### Answer:
""",
"Computer science is the study of",
]
class TestLoRACudaGraph(CustomTestCase):
def _run_without_cuda_graph_on_model_cases(self, model_cases: List[LoRAModelCase]):
# Since we have already enabled CUDA graph by default in other lora tests,
# we only need to run lora tests without CUDA graph here.
for model_case in model_cases:
# If skip_long_prompt is True, filter out prompts longer than 1000 characters
prompts = (
DEFAULT_PROMPTS
if not model_case.skip_long_prompt
else [p for p in DEFAULT_PROMPTS if len(p) < 1000]
)
for torch_dtype in TORCH_DTYPES:
run_lora_test_one_by_one(
prompts,
model_case,
torch_dtype,
max_new_tokens=32,
backend="triton",
disable_cuda_graph=True,
test_tag="without_cuda_graph",
)
def _run_cuda_graph_padding_on_model_cases(self, model_cases: List[LoRAModelCase]):
for model_case in model_cases:
# Run a batch size of 3, which will not be captured by CUDA graph and need padding
prompts = TEST_CUDA_GRAPH_PADDING_PROMPTS
for torch_dtype in TORCH_DTYPES:
run_lora_test_by_batch(
prompts,
model_case,
torch_dtype,
max_new_tokens=32,
backend="triton",
disable_cuda_graph=False,
test_tag="cuda_graph_padding",
)
def test_ci_lora_models(self):
self._run_without_cuda_graph_on_model_cases(CI_LORA_MODELS)
self._run_cuda_graph_padding_on_model_cases(CI_LORA_MODELS)
def test_all_lora_models(self):
if is_in_ci():
return
# Retain ONLY_RUN check here
filtered_models = []
for model_case in ALL_OTHER_LORA_MODELS:
if "ONLY_RUN" in os.environ and os.environ["ONLY_RUN"] != model_case.base:
continue
filtered_models.append(model_case)
self._run_without_cuda_graph_on_model_cases(filtered_models)
self._run_cuda_graph_padding_on_model_cases(filtered_models)
if __name__ == "__main__":
try:
mp.set_start_method("spawn")
except RuntimeError:
pass
unittest.main(warnings="ignore")

4
test/srt/models/lora/test_lora_tp.py
View File

@@ -23,7 +23,7 @@ from utils import (
DEFAULT_PROMPTS,
TORCH_DTYPES,
LoRAModelCase,
run_batch_lora_test,
run_lora_test_one_by_one,
)
from sglang.test.test_utils import CustomTestCase, is_in_ci
@@ -43,7 +43,7 @@ class TestLoRATP(CustomTestCase):
for tp_size in tp_list:
model_case.tp_size = tp_size
for torch_dtype in TORCH_DTYPES:
run_batch_lora_test(
run_lora_test_one_by_one(
prompts,
model_case,
torch_dtype,

4
test/srt/models/lora/test_multi_lora_backend.py
View File

@@ -22,7 +22,7 @@ from utils import (
TORCH_DTYPES,
LoRAAdaptor,
LoRAModelCase,
run_batch_lora_test,
run_lora_test_one_by_one,
)
from sglang.test.test_utils import CustomTestCase, is_in_ci
@@ -89,7 +89,7 @@ class TestMultiLoRABackend(CustomTestCase):
)
for torch_dtype in TORCH_DTYPES:
for backend in BACKENDS:
run_batch_lora_test(
run_lora_test_one_by_one(
batch_prompts,
model_case,
torch_dtype,

118
test/srt/models/lora/utils.py
View File

@@ -94,19 +94,20 @@ ALL_OTHER_LORA_MODELS = [
]
def run_batch_lora_test(
def run_lora_test_one_by_one(
prompts: List[str],
model_case: LoRAModelCase,
torch_dtype: torch.dtype,
max_new_tokens: int,
backend: str,
disable_cuda_graph: bool = True,
disable_cuda_graph: bool = False,
disable_radix_cache: bool = True,
mem_fraction_static: float = 0.88,
test_tag: str = "",
):
"""
Run Lora test for a forward batch.
Input a batch of prompts, and run lora tests one by one with several generate requests
(each request will have bs=1).
For prompt0, prompt1, ..., promptN,
we will use adaptor0, adaptor1, ..., adaptorN included in model case,
We will then compare the outputs of HF and SRT with and without LoRA.
@@ -119,7 +120,7 @@ def run_batch_lora_test(
torch_dtype (torch.dtype): The torch dtype to use.
max_new_tokens (int): The maximum number of new tokens to generate.
backend (str): The lora backend to use.
disable_cuda_graph (bool, optional): Whether to disable CUDA graph. Defaults to True.
disable_cuda_graph (bool, optional): Whether to disable CUDA graph. Defaults to False.
disable_radix_cache (bool, optional): Whether to disable radix cache. Defaults to True.
mem_fraction_static (float, optional): The fraction of memory to use. Defaults to 0.88.
test_tag (str, optional): The tag to use for the test. Defaults to "".
@@ -237,3 +238,112 @@ def run_batch_lora_test(
f"ROUGE-L score {rouge_score} below tolerance {rouge_tol} "
f"for base '{base_path}', adaptor '{adaptor_names}', backend '{backend}', prompt: '{prompts[0][:50]}...'"
)
def run_lora_test_by_batch(
prompts: List[str],
model_case: LoRAModelCase,
torch_dtype: torch.dtype,
max_new_tokens: int,
backend: str,
disable_cuda_graph: bool = False,
disable_radix_cache: bool = True,
mem_fraction_static: float = 0.88,
test_tag: str = "",
):
"""
Run lora tests as a batch.
For prompt0, prompt1, ..., promptN,
we will use adaptor0, adaptor1, ..., adaptorN included in model case,
We will then compare the outputs of HF and SRT with LoRA.
If number of prompts is larger than number of adaptors,
the prompt i will use adaptor i % (number of adaptors).
Args:
prompts (List[str]): The batch of prompts to test.
model_case (LoRAModelCase): The model case to test.
torch_dtype (torch.dtype): The torch dtype to use.
max_new_tokens (int): The maximum number of new tokens to generate.
backend (str): The lora backend to use.
disable_cuda_graph (bool, optional): Whether to disable CUDA graph. Defaults to False.
disable_radix_cache (bool, optional): Whether to disable radix cache. Defaults to True.
mem_fraction_static (float, optional): The fraction of memory to use. Defaults to 0.88.
test_tag (str, optional): The tag to use for the test. Defaults to "".
"""
base_path = model_case.base
# Create used adaptors for each prompt in batch
i, adaptors = 0, []
for _ in range(len(prompts)):
adaptors.append(model_case.adaptors[i])
i = (i + 1) % len(model_case.adaptors)
adaptor_names = [adaptor.name for adaptor in adaptors]
print(
f"\n========== Testing {test_tag} on base '{model_case.base}' with backend={backend}, dtype={torch_dtype} --- "
f"Using prompts {[p[:50] for p in prompts]} with adaptors: {adaptor_names} ---"
)
with SRTRunner(
base_path,
torch_dtype=torch_dtype,
model_type="generation",
tp_size=model_case.tp_size,
lora_paths=[adaptor.name for adaptor in model_case.adaptors],
max_loras_per_batch=model_case.max_loras_per_batch,
lora_backend=backend,
disable_cuda_graph=disable_cuda_graph,
disable_radix_cache=disable_radix_cache,
mem_fraction_static=mem_fraction_static,
) as srt_runner:
srt_outputs = srt_runner.batch_forward(
prompts, max_new_tokens=max_new_tokens, lora_paths=adaptor_names
)
with SRTRunner(
base_path,
torch_dtype=torch_dtype,
model_type="generation",
tp_size=model_case.tp_size,
mem_fraction_static=mem_fraction_static,
) as srt_runner:
srt_no_lora_outputs = srt_runner.batch_forward(
prompts, max_new_tokens=max_new_tokens
)
with HFRunner(
base_path, torch_dtype=torch_dtype, model_type="generation"
) as hf_runner:
hf_outputs = hf_runner.forward(
prompts, max_new_tokens=max_new_tokens, lora_paths=adaptor_names
)
with HFRunner(
base_path, torch_dtype=torch_dtype, model_type="generation"
) as hf_runner:
hf_no_lora_outputs = hf_runner.forward(
prompts,
max_new_tokens=max_new_tokens,
)
for i in range(len(prompts)):
srt_output_str = srt_outputs.output_strs[i].strip()
hf_output_str = hf_outputs.output_strs[i].strip()
rouge_score = calculate_rouge_l([srt_output_str], [hf_output_str])[0]
print("ROUGE-L score:", rouge_score)
print("SRT output:", srt_output_str)
print("HF output:", hf_output_str)
print("SRT no lora output:", srt_no_lora_outputs.output_strs[i].strip())
print("HF no lora output:", hf_no_lora_outputs.output_strs[i].strip())
assert srt_outputs.output_strs[i].strip(" ") == hf_outputs.output_strs[i].strip(
" "
), (
srt_outputs.output_strs[i].strip(" "),
hf_outputs.output_strs[i].strip(" "),
)
assert srt_no_lora_outputs.output_strs[i].strip(
" "
) == hf_no_lora_outputs.output_strs[i].strip(" "), (
srt_no_lora_outputs.output_strs[i].strip(" "),
hf_no_lora_outputs.output_strs[i].strip(" "),
)

1
test/srt/run_suite.py
View File

@@ -80,6 +80,7 @@ suites = {
TestFile("test_vlm_accuracy.py", 60),
TestFile("test_vision_openai_server.py", 637),
TestFile("test_w8a8_quantization.py", 46),
TestFile("models/lora/test_lora_cuda_graph.py", 250),
],
"per-commit-2-gpu": [
TestFile("models/lora/test_lora_tp.py", 116),