[LoRA] bf16 MoE-LoRA for trtllm backend with two-stream overlap + routing opts - #28466

Closed

yushengsu-thu wants to merge 11 commits into

sgl-project:mainfrom

yushengsu-thu:trtllm-bf16-shared_outer_gate

Commits on Jun 7, 2026

[LoRA] bf16: route experimental_sgl_trtllm MoE-LoRA to Triton for unquantized models

yushengsu-thu
and
claude
committed
f2f9819
Revert "[LoRA] bf16: route experimental_sgl_trtllm MoE-LoRA to Triton for unquantized models"
yushengsu-thu
committed
bb812e6
[LoRA] bf16 support for the experimental_sgl_trtllm MoE-LoRA path

yushengsu-thu
and
claude
committed
623befb
[LoRA] bf16: init_cuda_graph_moe_buffers reads gemm1/gemm2_weights when the quant-info is the bf16 one

yushengsu-thu
and
claude
committed
4f74dcf
[LoRA] bf16: cuda-graph moe buffer init must not assume 3-D expert weights

yushengsu-thu
and
claude
committed
104db55
bf16 MoE LoRA two-stream overlap (O1-bf16): side-stream gate_up delta

yushengsu-thu
and
claude
committed
526e0ae

Commits on Jun 10, 2026

[LoRA] opt1: route shared_outer through the fused merged-align (align/sort fusion)

yushengsu-thu
and
claude
committed
869882a
[LoRA] opt3: lean _get_lora_info (cache layer-static scalars) for prefill

yushengsu-thu
and
claude
committed
1536c6e

Commits on Jun 11, 2026

[LoRA] opt5: prefill routing reuse — unify A/B stage routing block size

yushengsu-thu
and
claude
committed
850faa8

Commits on Jun 17, 2026

[LoRA] lint: unquote bf16 dispatch annotations to satisfy ruff

yushengsu-thu
and
claude
committed
16ca35b

Commits on Jun 22, 2026

Merge branch 'main' into trtllm-bf16-shared_outer_gate
yushengsu-thu
authored
4c28570