[Feature] Support mtp overlap schedule

[Sunny-bot1]

Motivation

支持MTP（不开启logprob）场景下开启overlap schedule

Modifications

核心优化

1. 通过预先计算或复用历史值来避免同步开销
2. 将同步拷贝改为异步，延迟数据传输到非关键路径

位置	算子	同步操作	解决方法
gpu_model_runner.py 前处理	unified_update_model_status	not_need_stop	废弃
gpu_model_runner.py 后处理	save_output	accept_tokens_cpu, accept_num_cpu, seq_lens_decoder_cpu, prompt_lens_cpu	异步拷贝，延迟传输，与非MTP统一位置
mtp.py 前处理	draft_model_preprocess	not_need_stop	废弃
mtp.py 前处理	eagle_get_hidden_states（output_token_num）	self._mtp_input_token_num_event.synchronize()	直接使用token_num_cpu
mtp.py 前处理	eagle_get_self_hidden_states(多步)	output_token_num_cpu	直接使用token_num_cpu
mtp.py 前处理	_propose_cuda（token_num_cpu）	self.model_inputs["seq_lens_this_time"].numpy().sum().item()	使用主模型传过来的当前轮的real_bsz (self.share_inputs["seq_lens_this_time_cpu"].numpy() > 0).sum())
mtp.py 前处理	exist_prefill()	np.any(self.share_inputs["seq_lens_encoder"].numpy() > 0)	self.exist_prefill_flag
mtp.py 后处理	pre_process (real_output_token_num)	self._draft_output_token_num_event.synchronize()	直接使用token_num_cpu
mtp.py 后处理	draft_model_update	not_need_stop	废弃
gpu_model_runner.py 后处理	speculate_schedule_cache	not_need_stop	废弃
gpu_model_runner.py 后处理	pre_process (real_output_token_num)	self.output_token_num_event.synchronize()	直接使用self.share_inputs["seq_lens_this_time_cpu"].numpy() > 0).sum()
worker_process.py	tp_barrier	all_reduce	开启overlap schedule后自动使用CPU barrier
cudagraph_piecewise_backend.py	call（num_running_requests）	num_running_requests = int((seq_lens_this_time.flatten() > 0).sum().item())	主模型复用上一轮的real_bsz((seq_lens_this_time>0).sum)，MTP使用当前轮的real_bsz((seq_lens_this_time>0).sum)



注意事项

1. 以上优化仅在 decode batch 生效，prefill/mixed 阶段保持原有逻辑
2. 在 overlap schedule 场景下，由于空间预分配会引入无效槽位，算子中通过判断 batch_id_per_token < 0 提前退出处理
3. 不开启 overlap 时，保持原有逻辑

GLM TP4 效果

Usage or Command

Accuracy Tests

Checklist

• Add at least a tag in the PR title.
  • Tag list: [[FDConfig],[APIServer],[Engine], [Scheduler], [PD Disaggregation], [Executor], [Graph Optimization], [Speculative Decoding], [RL], [Models], [Quantization], [Loader], [OP], [KVCache], [DataProcessor], [BugFix], [Docs], [CI], [Optimization], [Feature], [Benchmark], [Others], [XPU], [HPU], [GCU], [DCU], [Iluvatar], [Metax]]
  • You can add new tags based on the PR content, but the semantics must be clear.
• Format your code, run pre-commit before commit.
• Add unit tests. Please write the reason in this PR if no unit tests.
• Provide accuracy results.
• If the current PR is submitting to the release branch, make sure the PR has been submitted to the develop branch, then cherry-pick it to the release branch with the [Cherry-Pick] PR tag.

[paddle-bot]

Thanks for your contribution!

[codecov-commenter]

Codecov Report

❌ Patch coverage is 75.75758% with 24 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@7a6c287). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/spec_decode/mtp.py	82.97%	6 Missing and 2 partials ⚠️
fastdeploy/worker/gpu_model_runner.py	71.42%	3 Missing and 3 partials ⚠️
fastdeploy/worker/input_batch.py	62.50%	6 Missing ⚠️
fastdeploy/model_executor/pre_and_post_process.py	55.55%	3 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #7001   +/-   ##
==========================================
  Coverage           ?   73.96%           
==========================================
  Files              ?      399           
  Lines              ?    56466           
  Branches           ?     8931           
==========================================
  Hits               ?    41765           
  Misses             ?    11720           
  Partials           ?     2981           

Flag	Coverage Δ
GPU	73.96% <75.75%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[fastdeploy-bot]

AI CI Agent | skill: pr_review_agent

This PR implements MTP (Multi-Token Prediction) overlap scheduling support with significant refactoring of CUDA kernels to use cooperative groups for better parallelization. The changes include removing CPU-GPU copies for better performance, adding defensive checks for negative batch indices, and modifying API signatures.

I found 1 P1 logic bug that needs to be addressed before merging.

[fastdeploy-bot]

P1 - API Mismatch Bug: The eagle_get_self_hidden_states kernel API was changed to expect seq_lens_encoder as the 4th parameter (see eagle_get_self_hidden_states.cu line 225), but this XPU branch is still passing step_idx. The kernel now uses seq_lens_encoder[t] > 0 to detect encoder phase instead of the old step_idx[i] == 1 check. This will produce incorrect results on XPU platform. Should be self.model_inputs.last_seq_lens_encoder (similar to the CUDA branch at line 1205) instead of self.model_inputs["step_idx"].

[fastdeploy-bot]

P1 - Test Inconsistency: The test passes step_idx_tensor but the kernel now expects seq_lens_encoder. While PaddlePaddle binds by position so this may not crash, the reference implementation computeOrderKernel (line 23-45) still uses step_idx == 1 logic while the actual CUDA kernel now uses seq_lens_encoder > 0 logic. The test should be updated to: 1) pass a proper seq_lens_encoder tensor, and 2) update the reference implementation to match the new kernel semantics.