关联 Issue #5512 [FRONTEND] clean-up backend/jit interface #9834 tl.trans on fp8 (byte) types produces 6.5x performance regression on AMDGPU since triton 3.4.0 #4576 [training_utils] feat: use TMA to load Tiles in linear_cross_entropy kernels #6110 [fsdp, training_utils] feat: support per-sample temperature with fused linear cross entropy #7945 [megatron] feat: add Liger TP fused cross entropy backend #8143 [hardware, model, cfg, tests] feat: Ascend CANN fused linear cross-entropy
参与讨论