H800

D4 Profile

层级	模块	Kernel	h800 ds prof	b 200 d1 mtp on
	端到端		96ms
	61层执行耗时		87.32ms
attention层	self.input_layernorm	flashinfer::norm::FusedAddRM	4
attention层	q_a_kv_a	deep_gemm::sm100_fp8_gem	13
attention层	self.input_layernorm	rmsnorm_split_col_kernel	2
attention层	q_proj_b	per_token_group_quant_fp8_	2
attention层	q_proj_b	deep_gemm::sm100_fp8_gem	19
attention层	torch.bmm transpose(0,1)	nvjet_tst_256x128_64x5_2x1	16
attention层	self.rotary_emb	elementwise_kernel
attention层	self.rotary_emb	rotary_embedding_kernel	7
attention层	self.rotary_emb	unrolled_elementwise_kernel
attention层	self.rotary_emb	index_elementwise_kernel
attention层	self.attn_mqa	mhaSm100fKernel_Qkv	255	225.767
attention层	torch.bmm transpose(0,1)	nvjet_tst_64x128_64x13_2x1	15
attention层	self.o_proj	elementwise_kernel
attention层	self.o_proj	per_token_group_quant_fp8_	5
attention层	self.o_proj	deep_gemm::sm100_fp8_gem	50	43.512
attention层	sum		388
gate层	self.input_layernorm	flashinfer::norm::FusedAddRM	3
gate层		nvjet_tst_64x32_64x16_2x4_2	6
gate层		cublasLt::splitKreduce_kernel	2
gate层		distribution_elementwise_grid
gate层		topk	8
gate层	sum		19
moe层	dispatch	deep_ep::internode_ll::dispatch	187	48.212
moe层	up_gate_gemm	deep_gemm::sm100_fp8_gem	81	195.136
moe层	up_gate_gemm	vectorized_elementwise_kernel
moe层	silu	silu_and_mul_kernel_ep_index	8	40.491
moe层	down_gemm	deep_gemm::sm100_fp8_gem	39	104.671
moe层	combine		394
moe层	sum
combine/share overlap		deep_ep::internode_ll::combine
combine/share overlap		per_token_group_quant_fp8_	3
combine/share overlap		deep_gemm::sm100_fp8_gem	17
combine/share overlap		act_and_mul_kernel	4
combine/share overlap		deep_gemm::sm100_fp8_gem	10
combine/share overlap		deep_ep::internode_ll::combine