Convert formatting to use ruff instead of yapf + isort (#26247)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
+47
-29
@@ -29,17 +29,20 @@ def use_v1_only(monkeypatch: pytest.MonkeyPatch):
|
||||
|
||||
|
||||
def setup_vllm(num_loras: int, tp: int) -> vllm.LLM:
|
||||
return vllm.LLM(model="Qwen/Qwen2.5-3B-Instruct",
|
||||
max_model_len=256,
|
||||
max_num_seqs=8,
|
||||
tensor_parallel_size=tp,
|
||||
enable_lora=True,
|
||||
max_loras=num_loras,
|
||||
max_lora_rank=8)
|
||||
return vllm.LLM(
|
||||
model="Qwen/Qwen2.5-3B-Instruct",
|
||||
max_model_len=256,
|
||||
max_num_seqs=8,
|
||||
tensor_parallel_size=tp,
|
||||
enable_lora=True,
|
||||
max_loras=num_loras,
|
||||
max_lora_rank=8,
|
||||
)
|
||||
|
||||
|
||||
TPU_TENSOR_PARALLEL_SIZES = [1, tpu.num_available_chips()
|
||||
] if tpu.num_available_chips() > 1 else [1]
|
||||
TPU_TENSOR_PARALLEL_SIZES = (
|
||||
[1, tpu.num_available_chips()] if tpu.num_available_chips() > 1 else [1]
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("tp", TPU_TENSOR_PARALLEL_SIZES)
|
||||
@@ -55,12 +58,19 @@ def test_single_lora(tp: int):
|
||||
prompt = "What is 1+1? \n"
|
||||
|
||||
lora_request = LoRARequest(
|
||||
"lora_adapter_1", 1,
|
||||
"Username6568/Qwen2.5-3B-Instruct-1_plus_1_equals_1_adapter")
|
||||
output = llm.generate(prompt,
|
||||
sampling_params=vllm.SamplingParams(max_tokens=256,
|
||||
temperature=0),
|
||||
lora_request=lora_request)[0].outputs[0].text
|
||||
"lora_adapter_1",
|
||||
1,
|
||||
"Username6568/Qwen2.5-3B-Instruct-1_plus_1_equals_1_adapter",
|
||||
)
|
||||
output = (
|
||||
llm.generate(
|
||||
prompt,
|
||||
sampling_params=vllm.SamplingParams(max_tokens=256, temperature=0),
|
||||
lora_request=lora_request,
|
||||
)[0]
|
||||
.outputs[0]
|
||||
.text
|
||||
)
|
||||
|
||||
answer = output.strip()[0]
|
||||
|
||||
@@ -73,13 +83,12 @@ def test_lora_hotswapping(tp: int):
|
||||
"""
|
||||
This test ensures we can run multiple LoRA adapters on the TPU backend, even
|
||||
if we only have space to store 1.
|
||||
|
||||
|
||||
We run "Username6568/Qwen2.5-3B-Instruct-1_plus_1_equals_x_adapter" which
|
||||
will force Qwen2.5-3B-Instruct to claim 1+1=x, for a range of x.
|
||||
"""
|
||||
|
||||
lora_name_template = \
|
||||
"Username6568/Qwen2.5-3B-Instruct-1_plus_1_equals_{}_adapter"
|
||||
lora_name_template = "Username6568/Qwen2.5-3B-Instruct-1_plus_1_equals_{}_adapter"
|
||||
lora_requests = [
|
||||
LoRARequest(f"lora_adapter_{i}", i, lora_name_template.format(i))
|
||||
for i in range(1, 5)
|
||||
@@ -90,10 +99,15 @@ def test_lora_hotswapping(tp: int):
|
||||
prompt = "What is 1+1? \n"
|
||||
|
||||
for i, req in enumerate(lora_requests):
|
||||
output = llm.generate(prompt,
|
||||
sampling_params=vllm.SamplingParams(
|
||||
max_tokens=256, temperature=0),
|
||||
lora_request=req)[0].outputs[0].text
|
||||
output = (
|
||||
llm.generate(
|
||||
prompt,
|
||||
sampling_params=vllm.SamplingParams(max_tokens=256, temperature=0),
|
||||
lora_request=req,
|
||||
)[0]
|
||||
.outputs[0]
|
||||
.text
|
||||
)
|
||||
answer = output.strip()[0]
|
||||
|
||||
assert answer.isdigit()
|
||||
@@ -105,12 +119,11 @@ def test_multi_lora(tp: int):
|
||||
"""
|
||||
This test ensures we can run multiple LoRA adapters on the TPU backend, when
|
||||
we have enough space to store all of them.
|
||||
|
||||
|
||||
We run "Username6568/Qwen2.5-3B-Instruct-1_plus_1_equals_x_adapter" which
|
||||
will force Qwen2.5-3B-Instruct to claim 1+1=x, for a range of x.
|
||||
"""
|
||||
lora_name_template = \
|
||||
"Username6568/Qwen2.5-3B-Instruct-1_plus_1_equals_{}_adapter"
|
||||
lora_name_template = "Username6568/Qwen2.5-3B-Instruct-1_plus_1_equals_{}_adapter"
|
||||
lora_requests = [
|
||||
LoRARequest(f"lora_adapter_{i}", i, lora_name_template.format(i))
|
||||
for i in range(1, 5)
|
||||
@@ -121,10 +134,15 @@ def test_multi_lora(tp: int):
|
||||
prompt = "What is 1+1? \n"
|
||||
|
||||
for i, req in enumerate(lora_requests):
|
||||
output = llm.generate(prompt,
|
||||
sampling_params=vllm.SamplingParams(
|
||||
max_tokens=256, temperature=0),
|
||||
lora_request=req)[0].outputs[0].text
|
||||
output = (
|
||||
llm.generate(
|
||||
prompt,
|
||||
sampling_params=vllm.SamplingParams(max_tokens=256, temperature=0),
|
||||
lora_request=req,
|
||||
)[0]
|
||||
.outputs[0]
|
||||
.text
|
||||
)
|
||||
|
||||
answer = output.strip()[0]
|
||||
|
||||
|
||||
@@ -26,16 +26,15 @@ def test_tpu_compilation():
|
||||
|
||||
# Currently, top-p sampling is disabled. `top_p` should be 1.0.
|
||||
N = 1
|
||||
sampling_params = SamplingParams(temperature=0.7,
|
||||
top_p=1.0,
|
||||
n=N,
|
||||
max_tokens=16)
|
||||
sampling_params = SamplingParams(temperature=0.7, top_p=1.0, n=N, max_tokens=16)
|
||||
|
||||
llm = LLM(model="Qwen/Qwen2-1.5B-Instruct",
|
||||
max_num_batched_tokens=256,
|
||||
max_model_len=256,
|
||||
max_num_seqs=32,
|
||||
enforce_eager=False)
|
||||
llm = LLM(
|
||||
model="Qwen/Qwen2-1.5B-Instruct",
|
||||
max_num_batched_tokens=256,
|
||||
max_model_len=256,
|
||||
max_num_seqs=32,
|
||||
enforce_eager=False,
|
||||
)
|
||||
|
||||
outputs = llm.generate(prompts, sampling_params)
|
||||
for output, answer in zip(outputs, answers):
|
||||
@@ -45,7 +44,8 @@ def test_tpu_compilation():
|
||||
assert generated_text.startswith(answer)
|
||||
|
||||
compiled_codes = sorted(
|
||||
glob.glob(os.path.join(temp_dir, "__transformed_code*for_forward.py")))
|
||||
glob.glob(os.path.join(temp_dir, "__transformed_code*for_forward.py"))
|
||||
)
|
||||
|
||||
for i, compiled_code in enumerate(compiled_codes):
|
||||
print("{} file: {}".format(i + 1, compiled_code))
|
||||
@@ -66,9 +66,10 @@ def test_tpu_compilation():
|
||||
|
||||
# Check all the compilations are as expected. The dump files include the
|
||||
# captured graph for the forward function of the nn.Module.
|
||||
compiled_fns = sorted(glob.glob(
|
||||
os.path.join(temp_dir, "__compiled_fn*Forward_graph*.py")),
|
||||
key=lambda s: extract_compiled_index(s))
|
||||
compiled_fns = sorted(
|
||||
glob.glob(os.path.join(temp_dir, "__compiled_fn*Forward_graph*.py")),
|
||||
key=lambda s: extract_compiled_index(s),
|
||||
)
|
||||
|
||||
for i, compiled_fn in enumerate(compiled_fns):
|
||||
print("{} file: {}".format(i + 1, compiled_fn))
|
||||
@@ -82,4 +83,4 @@ def test_tpu_compilation():
|
||||
# ragged_paged_attention
|
||||
with open(compiled_fns[1]) as f:
|
||||
content = f.read()
|
||||
assert (kv_cache_prefix in content and attn_prefix in content)
|
||||
assert kv_cache_prefix in content and attn_prefix in content
|
||||
|
||||
@@ -15,17 +15,20 @@ from ..utils import compare_two_settings
|
||||
def test_custom_dispatcher(monkeypatch: pytest.MonkeyPatch):
|
||||
with monkeypatch.context() as m:
|
||||
m.setenv("VLLM_RPC_TIMEOUT", "30000")
|
||||
compare_two_settings("Qwen/Qwen2.5-1.5B-Instruct",
|
||||
arg1=[
|
||||
"--max-model-len=256",
|
||||
"--max-num-seqs=32",
|
||||
"--enforce-eager",
|
||||
f"-O{CompilationLevel.DYNAMO_ONCE}",
|
||||
],
|
||||
arg2=[
|
||||
"--max-model-len=256", "--max-num-seqs=32",
|
||||
"--enforce-eager",
|
||||
f"-O{CompilationLevel.DYNAMO_AS_IS}"
|
||||
],
|
||||
env1={},
|
||||
env2={})
|
||||
compare_two_settings(
|
||||
"Qwen/Qwen2.5-1.5B-Instruct",
|
||||
arg1=[
|
||||
"--max-model-len=256",
|
||||
"--max-num-seqs=32",
|
||||
"--enforce-eager",
|
||||
f"-O{CompilationLevel.DYNAMO_ONCE}",
|
||||
],
|
||||
arg2=[
|
||||
"--max-model-len=256",
|
||||
"--max-num-seqs=32",
|
||||
"--enforce-eager",
|
||||
f"-O{CompilationLevel.DYNAMO_AS_IS}",
|
||||
],
|
||||
env1={},
|
||||
env2={},
|
||||
)
|
||||
|
||||
@@ -4,16 +4,18 @@
|
||||
|
||||
Run `pytest tests/kernels/moe/test_moe_pallas.py`.
|
||||
"""
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
import torch_xla
|
||||
|
||||
# yapf conflicts with isort for this block
|
||||
# yapf: disable
|
||||
from vllm.model_executor.layers.fused_moe.moe_pallas import (
|
||||
fused_moe as pallas_moe)
|
||||
from vllm.model_executor.layers.fused_moe.moe_pallas import fused_moe as pallas_moe
|
||||
from vllm.model_executor.layers.fused_moe.moe_torch_iterative import (
|
||||
fused_moe as torch_moe)
|
||||
fused_moe as torch_moe,
|
||||
)
|
||||
|
||||
# yapf: enable
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
@@ -43,6 +45,7 @@ def test_pallas_moe(
|
||||
dtype: torch.dtype,
|
||||
):
|
||||
import torch_xla.core.xla_model as xm
|
||||
|
||||
with torch.device(xm.xla_device()):
|
||||
a = torch.randn((m, k), dtype=dtype) / 10
|
||||
w1 = torch.randn((e, 2 * n, k), dtype=dtype) / 10
|
||||
|
||||
@@ -17,15 +17,15 @@ class GSM8KAccuracyTestConfig:
|
||||
expected_value: float
|
||||
|
||||
def get_model_args(self) -> str:
|
||||
return (f"pretrained={self.model_name},"
|
||||
"max_model_len=4096,max_num_seqs=32")
|
||||
return f"pretrained={self.model_name},max_model_len=4096,max_num_seqs=32"
|
||||
|
||||
|
||||
# NOTE: Accuracy scores measured on GPUs.
|
||||
ACCURACY_CONFIGS = [
|
||||
GSM8KAccuracyTestConfig(
|
||||
model_name="neuralmagic/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
|
||||
expected_value=0.76), # no bias
|
||||
expected_value=0.76,
|
||||
), # no bias
|
||||
# NOTE(rob): We cannot re-initialize vLLM in the same process for TPU,
|
||||
# so only one of these tests can run in a single call to pytest. As
|
||||
# a follow-up, move this into the LM-EVAL section of the CI.
|
||||
@@ -37,7 +37,6 @@ ACCURACY_CONFIGS = [
|
||||
|
||||
@pytest.mark.parametrize("config", ACCURACY_CONFIGS)
|
||||
def test_gsm8k_correctness(config: GSM8KAccuracyTestConfig):
|
||||
|
||||
results = lm_eval.simple_evaluate(
|
||||
model="vllm",
|
||||
model_args=config.get_model_args(),
|
||||
@@ -47,6 +46,7 @@ def test_gsm8k_correctness(config: GSM8KAccuracyTestConfig):
|
||||
|
||||
EXPECTED_VALUE = config.expected_value
|
||||
measured_value = results["results"][TASK][FILTER]
|
||||
assert (measured_value - RTOL < EXPECTED_VALUE
|
||||
and measured_value + RTOL > EXPECTED_VALUE
|
||||
), f"Expected: {EXPECTED_VALUE} | Measured: {measured_value}"
|
||||
assert (
|
||||
measured_value - RTOL < EXPECTED_VALUE
|
||||
and measured_value + RTOL > EXPECTED_VALUE
|
||||
), f"Expected: {EXPECTED_VALUE} | Measured: {measured_value}"
|
||||
|
||||
Reference in New Issue
Block a user