Add moe topk softmax templated from vllm (#4302)

This commit is contained in:
Qingquan Song
2025-03-14 12:03:33 -07:00
committed by GitHub
parent 660305c38a
commit 61e4433caf
9 changed files with 716 additions and 6 deletions

View File

@@ -33,7 +33,7 @@ from sgl_kernel.gemm import (
sgl_per_token_group_quant_fp8,
sgl_per_token_quant_fp8,
)
from sgl_kernel.moe import moe_align_block_size
from sgl_kernel.moe import moe_align_block_size, topk_softmax
from sgl_kernel.sampling import (
min_p_sampling_from_probs,
top_k_renorm_prob,