Revert "[HMA] [KVEvent] Enable GPU-side KV events for HMA" (#37688)

[vllm-agent]

Revert of #37688

This reverts commit cc07dad.

Reason: This PR is linked to 1 new CI failure in build #60970:

• MultiConnector (Nixl+Offloading) PD edge cases (2 GPUs): test_full_decode_gpu_cache_hit_metrics — local_cache_hit off-by-one (expected 128, got 127.0)

Original PR: #37688

---

Auto-generated by CI failure analyzer

[mergify]

Documentation preview: https://vllm--39658.org.readthedocs.build/en/39658/

[gemini-code-assist]

Code Review

This pull request removes the group_idx field from KV cache events (BlockStored and BlockRemoved) and updates associated tests and logic across the codebase. It also introduces a change to disable the hybrid KV cache manager when KV events are configured, as they are currently incompatible. Feedback suggests refining the logic in vllm/config/vllm.py to only disable the hybrid KV cache manager if KV events are explicitly enabled, rather than just checking for the presence of the configuration object.

[gemini-code-assist]

The current logic disables the hybrid KV cache manager (HMA) whenever kv_events_config is present, regardless of whether KV events are actually enabled. This can lead to unnecessary performance regressions for hybrid models if the configuration object exists but enable_kv_cache_events is set to False. It is better to check the enable_kv_cache_events flag specifically.

Suggested change

	if self.kv_events_config is not None:
	# Hybrid KV cache manager is not compatible with KV events.
	need_disable_hybrid_kv_cache_manager = True
	if self.kv_events_config is not None and self.kv_events_config.enable_kv_cache_events:
	# Hybrid KV cache manager is not compatible with KV events.
	need_disable_hybrid_kv_cache_manager = True