biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Convert formatting to use ruff instead of yapf + isort (#26247)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-10-05 15:06:22 +01:00
committed by GitHub
parent 17edd8a807
commit d6953beb91
1508 changed files with 115244 additions and 94146 deletions
Download Patch File Download Diff File Expand all files Collapse all files

41
tests/distributed/test_torchrun_example_moe.py
View File

@@ -24,23 +24,22 @@ dp_rank = int(os.getenv("DP_RANK", "0"))
if dp_size > 1:
# distribute the prompts across the data parallel ranks
prompts = [
prompt for idx, prompt in enumerate(prompts)
if idx % dp_size == dp_rank
]
prompts = [prompt for idx, prompt in enumerate(prompts) if idx % dp_size == dp_rank]
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
# set different `gpu_memory_utilization` and `swap_space` for different ranks,
# to test if all ranks agree on the same kv cache configuration.
llm = LLM(model="microsoft/Phi-mini-MoE-instruct",
tensor_parallel_size=int(os.getenv("TP_SIZE", "1")),
pipeline_parallel_size=int(os.getenv("PP_SIZE", "1")),
enable_expert_parallel=int(os.getenv("ENABLE_EP", "0")) == 1,
distributed_executor_backend="external_launcher",
gpu_memory_utilization=random.uniform(0.7, 0.9),
swap_space=random.randint(1, 4),
seed=0)
llm = LLM(
model="microsoft/Phi-mini-MoE-instruct",
tensor_parallel_size=int(os.getenv("TP_SIZE", "1")),
pipeline_parallel_size=int(os.getenv("PP_SIZE", "1")),
enable_expert_parallel=int(os.getenv("ENABLE_EP", "0")) == 1,
distributed_executor_backend="external_launcher",
gpu_memory_utilization=random.uniform(0.7, 0.9),
swap_space=random.randint(1, 4),
seed=0,
)
outputs = llm.generate(prompts, sampling_params)
@@ -54,21 +53,18 @@ def test_consistent_across_ranks(obj):
dist.broadcast_object_list([obj], src=group.ranks[0], group=cpu_group)
else:
container = [None]
dist.broadcast_object_list(container,
src=group.ranks[0],
group=cpu_group)
dist.broadcast_object_list(container, src=group.ranks[0], group=cpu_group)
assert container[0] == obj
test_consistent_across_ranks(
llm.llm_engine.vllm_config.cache_config.num_cpu_blocks)
test_consistent_across_ranks(
llm.llm_engine.vllm_config.cache_config.num_gpu_blocks)
test_consistent_across_ranks(llm.llm_engine.vllm_config.cache_config.num_cpu_blocks)
test_consistent_across_ranks(llm.llm_engine.vllm_config.cache_config.num_gpu_blocks)
# make sure we can access the model parameters from the calling process
# of the `LLM` instance.
params = list(llm.llm_engine.model_executor.driver_worker.worker.model_runner.
model.parameters())
params = list(
llm.llm_engine.model_executor.driver_worker.worker.model_runner.model.parameters()
)
test_consistent_across_ranks(len(params))
# all ranks should have the same outputs
@@ -77,5 +73,4 @@ for output in outputs:
generated_text = output.outputs[0].text
test_consistent_across_ranks(prompt)
test_consistent_across_ranks(generated_text)
print(f"Rank {group_rank}, Prompt: {prompt!r}, "
f"Generated text: {generated_text!r}")
print(f"Rank {group_rank}, Prompt: {prompt!r}, Generated text: {generated_text!r}")