[Worker]Lazy import torch_npu

[Potabk]

What this PR does / why we need it?

To avoid unnecessary delays, we only import torch_npu when profilling is enabled.

Does this PR introduce any user-facing change?

How was this patch tested?

this is my test srcipts

# SPDX-License-Identifier: Apache-2.0

from vllm import LLM, SamplingParams

# Sample prompts.
prompts = [
    "Hello, my name is",
    "The president of the United States is",
    "The capital of France is",
    "The future of AI is",
]
# Create a sampling params object.
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)

# Create an LLM.
llm = LLM(model="facebook/opt-125m")
# Generate texts from the prompts. The output is a list of RequestOutput objects
# that contain the prompt, generated text, and other information.
outputs = llm.generate(prompts, sampling_params)
# Print the outputs.
for output in outputs:
    prompt = output.prompt
    generated_text = output.outputs[0].text
    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")

and the result is as follows:

INFO 02-27 01:45:12 [__init__.py:30] Available plugins for group vllm.platform_plugins:
INFO 02-27 01:45:12 [__init__.py:32] name=ascend, value=vllm_ascend:register
INFO 02-27 01:45:12 [__init__.py:34] all available plugins for group vllm.platform_plugins will be loaded.
INFO 02-27 01:45:12 [__init__.py:36] set environment variable VLLM_PLUGINS to control which plugins to load.
INFO 02-27 01:45:12 [__init__.py:44] plugin ascend loaded.
INFO 02-27 01:45:12 [__init__.py:198] Platform plugin ascend is activated
INFO 02-27 01:45:27 [config.py:569] This model supports multiple tasks: {'reward', 'generate', 'embed', 'classify', 'score'}. Defaulting to 'generate'.
INFO 02-27 01:45:27 [llm_engine.py:235] Initializing a V0 LLM engine (v0.7.3.dev245+gcd1f843f) with config: model='/root/wl/cache/modelscope/models/facebook/opt-125m', speculative_config=None, tokenizer='/root/wl/cache/modelscope/models/facebook/opt-125m', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.float16, max_seq_len=2048, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto,  device_config=npu, decoding_config=DecodingConfig(guided_decoding_backend='xgrammar'), observability_config=ObservabilityConfig(show_hidden_metrics=False, otlp_traces_endpoint=None, collect_model_forward_time=False, collect_model_execute_time=False), seed=0, served_model_name=/root/wl/cache/modelscope/models/facebook/opt-125m, num_scheduler_steps=1, multi_step_stream_outputs=True, enable_prefix_caching=False, chunked_prefill_enabled=False, use_async_output_proc=True, disable_mm_preprocessor_cache=False, mm_processor_kwargs=None, pooler_config=None, compilation_config={"splitting_ops":[],"compile_sizes":[],"cudagraph_capture_sizes":[256,248,240,232,224,216,208,200,192,184,176,168,160,152,144,136,128,120,112,104,96,88,80,72,64,56,48,40,32,24,16,8,4,2,1],"max_capture_size":256}, use_cached_outputs=False, 
INFO 02-27 01:45:28 [importing.py:16] Triton not installed or not compatible; certain GPU-related functions will not be available.
WARNING 02-27 01:45:28 [utils.py:2298] Methods add_lora,add_prompt_adapter,cache_config,compilation_config,current_platform,list_loras,list_prompt_adapters,load_config,pin_lora,pin_prompt_adapter,remove_lora,remove_prompt_adapter not implemented in <vllm_ascend.worker.NPUWorker object at 0xfffd04c2aa40>
INFO 02-27 01:45:36 [parallel_state.py:948] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, TP rank 0
Loading pt checkpoint shards:   0% Completed | 0/1 [00:00<?, ?it/s]
Loading pt checkpoint shards: 100% Completed | 1/1 [00:00<00:00,  3.09it/s]
Loading pt checkpoint shards: 100% Completed | 1/1 [00:00<00:00,  3.08it/s]

INFO 02-27 01:45:39 [executor_base.py:111] # npu blocks: 98088, # CPU blocks: 7281
INFO 02-27 01:45:39 [executor_base.py:116] Maximum concurrency for 2048 tokens per request: 766.31x
INFO 02-27 01:45:39 [llm_engine.py:441] init engine (profile, create kv cache, warmup model) took 2.03 seconds
Processed prompts: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 15.49it/s, est. speed input: 100.76 toks/s, output: 248.00 toks/s]
Prompt: 'Hello, my name is', Generated text: ' is an acronym, the name of the name of the name is a phrase used'
Prompt: 'The president of the United States is', Generated text: ' is the subject of interest to be realized. The subject of interest to be realized'
Prompt: 'The capital of France is', Generated text: ' the star of the night is the name of the night is the name of the'
Prompt: 'The future of AI is', Generated text: ' is the same as a good example of any of the above mentioned.\n\n'

[wangxiyuan]

this relate to worker initailization. @noemotiovon @MengqingCao not sure if this will break ray case?

[noemotiovon]

I don't think so, but just to be sure, let me test it.

[noemotiovon]

I've tested it, and it doesn't affect the Ray backend.