fix empty buckets issue for enforce eager mode

[yangulei]

Current implementation skip calling self.model_runner.warmup_model() for enforce_eager=True leads to empty bucket lists in the bucket manager, and the following find_bucket calls will get fallback buckets.

The buckets are generated by the following calls in self.model_runner.warmup_model().

vllm-gaudi/vllm_gaudi/v1/worker/hpu_model_runner.py

Lines 4581 to 4596 in cc37f1f

	def warmup_model(self) -> None:
	if not self.enable_bucketing:
	return
	if self.unified_attn:
	self.bucketing_manager.generate_unified_buckets()
	if self.supports_mm_inputs:
	# Delayed multimodal buckets during warmup until model is loaded.
	from vllm_gaudi.extension.bucketing.vision import HPUVisionBucketManager
	self.get_model().vision_bucket_manager = HPUVisionBucketManager(get_config().model_type)
	msg = (f"Multimodal bucket : {self.get_model().vision_bucket_manager.multimodal_buckets}")
	logger.info(msg)
	else:
	self.bucketing_manager.generate_prompt_buckets()
	if not self.is_pooling_model:
	self.bucketing_manager.generate_decode_buckets()

And the actual warmup will be skipped for enforce_eager=True according to

vllm-gaudi/vllm_gaudi/v1/worker/hpu_model_runner.py

Lines 4670 to 4695 in cc37f1f

	with compile_only_mode_context() if can_use_compile_only_mode else contextlib.nullcontext():
	if not self.model_config.enforce_eager:
	assert self.mem_margin is not None, \
	("HabanaWorker.determine_num_available_blocks needs "
	"to be called before warming up the model.")
	if self.is_pooling_model:
	self.warmup_pooler()
	else:
	self.warmup_sampler()
	self.warmup_defragmenter()
	# TODO(kzawora): align_workers
	if self.unified_attn:
	self.warmup_unified_graphs(self.bucketing_manager.unified_buckets, kv_caches)
	else:
	mem_post_prompt, prompt_batch_seq, prompt_captured_all = \
	self.warmup_graphs(
	self.bucketing_manager.prompt_buckets, True, kv_caches)
	self.log_graph_warmup_summary(self.bucketing_manager.prompt_buckets, True, mem_post_prompt)
	if not self.is_pooling_model:
	mem_post_decode, decode_batch_seq, decode_captured_all = \
	self.warmup_graphs(
	self.bucketing_manager.decode_buckets, False, kv_caches)
	self.log_graph_warmup_summary(self.bucketing_manager.decode_buckets, False, mem_post_decode)

So the self.model_runner.warmup_model() cannot be skipped when enforce_eager=True and no actual warmup in this case as expected.

[Copilot]

Pull request overview

This PR fixes an issue where eager execution mode (enforce_eager=True) was causing empty bucket lists in the bucket manager. The fix ensures that warmup_model() is always called to generate necessary buckets, while the actual warmup execution is still skipped for eager mode as intended.

Key changes:

• Removed the enforce_eager check from the warmup condition, allowing bucket generation to occur even in eager mode
• The actual warmup compilation is still skipped for eager mode internally within warmup_model()

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

✅ CI Passed

All checks passed successfully against the following vllm commit:
5d9308968649c81ee5903fc2a77377d738ed2f6d

[github-actions]

✅ CI Passed

All checks passed successfully against the following vllm commit:
cddbc2b4b2547c681d1bdb876fdd6a7b8e0ec58d

[github-actions]

✅ CI Passed

All checks passed successfully against the following vllm commit:
aa125ecf0edb9cd67656553d11d643aeb444ff9e

[github-actions]

✅ CI Passed

All checks passed successfully against the following vllm commit:
66652e8082b69ba7d1e6aca7c234433de55f1b9b

[github-actions]

✅ CI Passed

All checks passed successfully against the following vllm commit:
4c1c501a7ee1d5efbad945ea62a702ce5cefb799