fix: avoid num_return_sequences prefill batch fanout

[Vinkle-hzt]

Summary

• Keep non-beam num_return_sequences streams at batch size 1 for the prompt/context execution step, then fan out to num_return_sequences at the first sampling step.
• Preserve sampler fan-out by tiling sampler token/logit inputs, cum log probs, finished masks, logits processor intervals, and dispatcher success offsets by sampler batch size.
• Keep warmup behavior intact by treating decode warmup as perf-test mode, and size return_softmax_probs storage by max batch.

Why

For FIFO scheduling, num_return_sequences should not inflate the context/model batch size before the request has produced its first sampled token. The previous GenerateStream::batchSize(0) returned num_return_sequences for non-beam requests, so StreamGroups and model input assembly treated the prompt step as a larger batch even though the prompt only needs one model execution.

[LLLLKKKK]

AI Code Review - PR #1158

Status: BLOCKING

Summary: P0/0 · P1/0 · P2/1 · P3/0

Non-blocking Suggestions

P2

• 生成式推荐 / Tree logits processor 的 tiling 状态复制路径缺少集成单测覆盖 @ rtp_llm/cpp/models/logits_processor/LogitsProcessorFactory.cc:26
  • 建议：建议补一条集成单测：构造 num_return_sequences>1 且设置 combo_token_size（或 tree 相关配置）的 stream，走一遍 createLogitsProcessors → context 步 gather/dispatch → update 的 fan-out，断言 N 份 infos_/tree_infos_ 均由 context 状态正确复制且各自独立推进，从而锁定 state_batch_size 选择与 remap 组合的正确性。此为非阻塞的覆盖补强。

Checklist Violations (1 fail / 25 total)

General Principles Checklist

• [6.1] Tests — 新逻辑有聚焦单测 + 相关集成/smoke 测试 → issue 生成式推荐 / Tree logits processor 的 tiling 状态复制路径缺少集成单测覆盖
LogitsProcessorFactory 新增 need_return_sequence_tiling/state_batch_size 逻辑：non-beam 且 max_batch_size > init_batch_size 时，TreeLogitsProcessor/RecommendationLogitsProcessor 直接按 max_batch_size(=N) 建满 N 份状态，并依赖去守卫后的 updateMultiSeqStatus({0,...,0}) 在 context 步做状态复制。新增的两条单测覆盖了 softmax_probs/cum_log_probs 的 tiling，但未覆盖 num_return_sequences>1 且带 combo_token_size/tree 状态时，工厂建满 N 份 + fan-out remap 的集成行为。虽有 RecommendationLogitsProcessorTest 单测了 updateMultiSeqStatus({0,0,0}) 这一构件，

Strengths

• batch 维度语义拆分（model batch vs sampler batch）在所有落点统一，且 StreamGroups::totalSamplerBatchSizeIn/Out 的累加口径与 dispatcher/gatherer 的 per-stream 计算严格对齐，避免了索引错位与越界。
• isSubGenerateDoneWithoutLock(i) → isSubGenerateDoneWithoutLock(cur_batch) 精准修复了 context 阶段 sub_generate_status_ 尺寸变为 1 后的潜在越界读，属于必要且到位的配套修复。
• softmax_probs_ 改用 maxBatchSize() 而非 init_batch_size 分配，顺带消除了 beam/return-seq 下缓冲区可能欠配的隐患。
• 新增 testNumReturnSequencesTileAfterContextBatch 与 testNumReturnSequencesTilesSamplerInputOnly 两条单测，分别验证 stream 级 batch 尺寸（current=1/next=20/softmax numel）与端到端 sampler tiling（token_ids/logits/cum_log_probs 复制），覆盖了主要行为契约。