[Bug]: Add speculative decoding metrics for HPU

[sungwook-son]

Your current environment

In the lines referenced below, is_cuda_alike() evaluates to False, causing the function to return None.
As a result, speculative decoding metrics are disabled.

vllm-fork/vllm/spec_decode/metrics.py

Lines 99 to 119 in bef3660

	def maybe_collect_rejsample_metrics(
	self, k: int) -> Optional[SpecDecodeWorkerMetrics]:
	# Skip for any platform that doesn't have device Event
	if current_platform.Event is None:
	return None
	if not current_platform.is_cuda_alike():
	return None
	# If a copy was initiated in the previous call, collect and return.
	if self._in_flight_copy is not None:
	ready_event = self._in_flight_copy
	self._in_flight_copy = None
	return self._collect_rejsample_metrics(k, ready_event)
	# Otherwise, check if we should start a new copy.
	if self._should_collect_rejsample_metrics(self._timer()):
	assert self._in_flight_copy is None
	self._in_flight_copy = self._copy_rejsample_metrics_async()
	return None

🐛 Describe the bug

export VLLM_CONTIGUOUS_PA=false
export VLLM_SKIP_WARMUP=true
export PT_HPU_LAZY_MODE=1
export PT_HPU_ENABLE_LAZY_COLLECTIVES=true

python -m vllm.entrypoints.openai.api_server
--host 0.0.0.0 --port 8000
--model meta-llama/Llama-3.1-70B-Instruct
--seed 42 -tp 4
--speculative_config '{"model": "meta-llama/Llama-3.1-8B-Instruct", "num_speculative_tokens": 5, "target_parallel_config": 1}'
--gpu_memory_utilization 0.95 --max-model-len 16384

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.