implementation detail · filed under inference & serving
TensorRT-LLM multi-head attention backend
The TensorRT-LLM multi-head attention backend selected with the trtllm_mha setting.
Also called trtllm_mha, trtllm_mha attention backend.
- sources
- 2
- model
- 1
- lab adopt it
- 1
- strongest
- used
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 2
Documented in
Evidence
2 spans quoted from the sources, strongest treatment first.
--attention-backend trtllm_mha
usedsoftware implementationin Step 3.7 FlashStepFun
--attention-backend trtllm_mha
usedsoftware implementationin Step 3.7 FlashStepFun
Filed alongside
Other methods under inference & serving :: inference kernel.
Batch-invariant deterministic kernelsCUDA GraphCUDA Graph capture size reductionFlashAttention 3Kernel fusionMoE-side chunkingSynchronization-free static-shape MoE executionAvoiding split-KDeepGEMM-based batch-invariant matrix multiplicationDistributed shared memory for cross-SM attention data exchangeDual-kernel batch-invariant attention decodingDynamic load balancingExp-free TopK kernelExpert-optimized Triton kernelsFA4 sheared-bias attention kernelFlashAttentionFlashAttention 4FlashKDAFP8 GEMMFused AttnRes merge and RMSNorm kernelFused latent down-projection and MoE-router GEMMFused mHC kernelsFused QSA kernelFused RoPE-attention-RoPE-cast kernel