[Cherry-Pick][CI]Support multi-step mtp with cudagraph(#5886)

[freeliuzc]

Motivation

💡 If this PR is a Cherry Pick, the PR title needs to follow the format by adding the [Cherry-Pick] label at the very beginning and appending the original PR ID at the end. For example, [Cherry-Pick][CI] Add check trigger and logic(#5191)

💡 如若此PR是Cherry Pick，PR标题需遵循格式，在最开始加上[Cherry-Pick]标签，以及最后面加上原PR ID，例如[Cherry-Pick][CI] Add check trigger and logic(#5191)

Modifications

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!

[Copilot]

Pull request overview

This cherry-pick PR aims to support multi-step MTP (Multi-Token Prediction) with CUDA Graph by modifying the capture process and fixing CUDA Graph compatibility issues.

• Simplified CUDA Graph capture by removing separate draft model capture logic in gpu_model_runner.py
• Modified the expected decode length calculation for MTP warmup
• Enhanced _initialize_forward_meta to conditionally enable CUDA Graph based on substep during dummy runs
• Fixed CUDA error 700 by replacing paddle.clone with copy_ in CUDA Graph mode

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
fastdeploy/worker/gpu_model_runner.py	Removed complex draft model CUDA Graph capture logic, updated expected_decode_len calculation for MTP, simplified warmup logging
fastdeploy/spec_decode/mtp.py	Added parameters to _initialize_forward_meta for multi-step CUDA Graph support, replaced paddle.clone with copy_ to avoid CUDA error 700, added documentation about CUDA Graph capture requirements

[Copilot]

The variable 'is_dummy_run' is used but not defined in the _propose method. This will cause a NameError at runtime when _initialize_forward_meta is called. The _propose method signature only includes 'step_use_cudagraph' as a parameter, but 'is_dummy_run' is being passed to _initialize_forward_meta. You need to either add 'is_dummy_run' as a parameter to the _propose method or determine it from existing state/attributes.

Suggested change

	step_use_cudagraph=step_use_cudagraph, is_dummy_run=is_dummy_run, substep=substep
	step_use_cudagraph=step_use_cudagraph, substep=substep

[Copilot]

The log message is inconsistent with the actual expected_decode_len value being passed to _dummy_run. The log says 'expected_decode_len:{self.speculative_config.num_speculative_tokens}' but the actual parameter passed on line 1953 is 'self.speculative_config.num_speculative_tokens * 2 + 1'. The log message should reflect the actual value being used to avoid confusion during debugging.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 85.71429% with 1 line in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (release/online/20251131@7aea651). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/spec_decode/mtp.py	85.71%	1 Missing ⚠️

Additional details and impacted files

@@                    Coverage Diff                     @@
##             release/online/20251131    #5897   +/-   ##
==========================================================
  Coverage                           ?   58.50%           
==========================================================
  Files                              ?      320           
  Lines                              ?    39181           
  Branches                           ?     5909           
==========================================================
  Hits                               ?    22923           
  Misses                             ?    14425           
  Partials                           ?     1833           

Flag	Coverage Δ
GPU	58.50% <85.71%> (?)

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.

[gongshaotian]

LGTM