[Question]: No speedup on multi-card with vllm

Describe the issue

hi,

I got some interesting results, could you share some comments?

Firstly, I followed your guide and tested Llama-3-8B-Instruct-262k with benchmark_e2e_vllm_tp.py on L20 single card, the results seems right. Check the table, from 40k seq, sparse works faster than flashatten. But when I tested it on 4 cards, sparse has no better results than dense. It seems abnormal... The same results on qwen2.5-32b. From your paper, you tested the sparse one with 8 A100, could you share the comparison results with Flashattn or give us some advice to fix this problem? Thanks.

vllm: 0.9.2
minference: 0.1.6.0
triton: 3.3.0

<style> </style>

Meta-Llama-3.1-70B-Instruct @L20

TTFT		Input length	Time（TP=1）	Time（TP=4）
	Dense	10k	1.58	0.80
	Sparse		2.68	1.09
	Dense	20k	3.64	1.71
	Sparse		4.67	1.98
	Dense	30k	6.17	2.73
	Sparse		6.63	2.92
	Dense	40k	9.18	3.86
	Sparse		8.60	3.96
	Dense	50k	12.76	5.11
	Sparse		10.30	5.11
	Dense	60k	19.55	6.49
	Sparse		12.74	6.50
	Dense	70k	21.12	8.00
	Sparse		14.75	7.99
	Dense	80k	26.02	9.58
	Sparse		17.12	9.57
	Dense	90k	32.91	11.35
	Sparse		19.10	11.33
	Dense	100k	37.17	13.18
	Sparse		20.65	13.17

<style> </style>

Qwen2.5-32B-Instrcut TP=4

qwen2.5-32b		dense	sparse
	50k	15.78	17.10
	60k	19.76	20.66
	70k	23.94	23.95
	80k	28.7	27.85
	90k	33.75	32.27
	120k	49.80	42.93

Start with benchmark_e2e_vllm_tp.py and reproduce the reported L20 TP=1 versus TP=4 results using vLLM 0.9.2, MInference 0.1.6.0, and Triton 3.3.0. Compare dense and sparse runs for Llama-3-8B-Instruct-262k and Qwen2.5-32B across the listed sequence lengths; done would require determining whether the missing multi-card speedup is expected and documenting actionable guidance.