[Performance]: Profile & optimize the BlockManagerV2

[cadedaniel]

Proposal to improve performance

We've recently rewritten the block management subsystem for better testability. We need to profile it under real load to make sure it is performant enough to replace the block manager V1, and fix any issues.

We should do this once the block manager v2 is feature complete (still missing a few items).

Known issue:

• Prefix caching num_total_tokens is O(N^2) instead of O(N) (see [Core] Enable prefix caching with block manager v2 enabled #4142 (comment))

[cadedaniel]

What we want to profile:
For low-latency use case:

• Batch size of 8-16 range
• Various block sizes (16, 32, 128)
• Sequence length (long context, 1.5k). Can set num_output_tokens=50.
• For spec decode, also num_lookahead_tokens > 0. Try num_lookahead_tokens=5 (what is lookahead scheduling)

For high-throughput use-case:

• Batch size up to 256
• Various block sizes (16, 32, 128)
• Sequence length (long context, 1.5k). Can set num_output_tokens=50.

Other cases that are important (perhaps we make separate tasks):

• P0 prefix caching
• P1 Beam search
• P1 swapping
• P1 sliding window

In terms of how to profile, use benchmark_latency + torch profiling (or can use CPU profiler of your choosing)

vllm/benchmarks/benchmark_latency.py

Lines 178 to 187 in c3c2903

	parser.add_argument(
	'--profile',
	action='store_true',
	help='profile the generation process of a single batch')
	parser.add_argument(
	'--profile-result-dir',
	type=str,
	default=None,
	help=('path to save the pytorch profiler output. Can be visualized '
	'with ui.perfetto.dev or Tensorboard.'))

[cadedaniel]

@robertgshaw2-neuralmagic can you assign Alex