vllm

17bcc8ee - Sanitize unfilled recv slots in flashinfer_nvlink_one_sided dispatch (#9)

Sanitize unfilled recv slots in flashinfer_nvlink_one_sided dispatch (#9) Padded rows in the [ep_size, max_num_tokens, ...] workspace retain stale topk_ids from prior dispatch calls (the workspace is zeroed only once at init). Those stale ids cause the downstream trtllm_fp4 grouped GEMM to do phantom work for random local experts every layer, which (a) inflates expert GEMM time and (b) creates the cross-rank skew that the combine kernel then has to wait on. Setting `invalid_token_expert_id` to `num_experts` (one past the valid expert range) makes the flashinfer worker overwrite all top_k topk_ids slots of padded rows with that sentinel (moeA2ASanitizeExpertIdsKernel in moeAlltoAllKernels.cu); the trtllm grouped GEMM then sees those rows as routed to no local expert (out of [local_expert_offset, local_expert_offset + local_num_experts)) and skips them. Signed-off-by: Zijing Liu <liuzijing2014@gmail.com>