[Feature] Support block scheduler v1 for FD

[rainyfly]

Background

In the current version, blocks are devided into two tables. The table for prefilling is managed by server, and the one for decoding is managed by worker. Users should set kv_cache_ratio to determine how many blocks are used by each phase.
And when blocks for decoding are not enough, blocks scheduled for decoding are performed by ops step_paddle, which is easy to cause OOM when output length is large.

To make deployment more easier and stable, we start to develop scheduler v1. Only one block table exists and managed by server.
Note
The scheduler v1 is experimentally developed and only supported for mixed and sampling scenerio currently. In the feature, we will continue to build and improve it.

Perf

• when kv_cache_ratio is set appropriately，the performance with scheduler v1 is worse by 3-5% than before due to more communitcation overhead between engine and worker.
• when kv_cache_ratio is too high or too low, the performance with scheduler v1 is better.

Stability

• OOM did not appear in testing.

How to enable

Set environment variable ENABLE_V1_KVCACHE_SCHEDULER to 1 to enable scheduler v1.

export ENABLE_V1_KVCACHE_SCHEDULER=1

[paddle-bot]

Thanks for your contribution!

[Copilot]

Pull Request Overview

This PR introduces experimental support for block scheduler v1 in FastDeploy, which unifies block management under a single server-managed cache system. The new scheduler eliminates the need for manual kv_cache_ratio configuration and aims to improve stability by preventing OOM issues during long sequence generation.

• Adds a new unified block scheduling system managed entirely by the server
• Implements conditional logic throughout the codebase to support both legacy and v1 schedulers
• Introduces new CUDA operations for v1-specific input handling and task recovery

Reviewed Changes

Copilot reviewed 15 out of 15 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
fastdeploy/envs.py	Adds environment variable to enable v1 scheduler
fastdeploy/engine/engine.py	Integrates v1 resource manager and scheduler logic
fastdeploy/engine/sched/resource_manager_v1.py	Core v1 resource manager implementation
fastdeploy/worker/gpu_worker.py	Adds v1 task insertion logic
fastdeploy/worker/gpu_model_runner.py	Implements v1 task processing and input management
fastdeploy/model_executor/pre_and_post_process.py	Updates post-processing for v1 scheduler
fastdeploy/output/token_processor.py	Adds v1-specific resource recycling
fastdeploy/engine/request.py	Extends Request class with scheduler state
custom_ops/	New CUDA operations for v1 input handling

Comments suppressed due to low confidence (1)

fastdeploy/worker/gpu_model_runner.py:269

• The variable 'i' is used in a nested loop context where there's already an outer loop variable 'i' at line 199. This creates a naming conflict and potential confusion. Consider renaming the inner loop variable to 'j' or 'seq_idx'.

                for i in range(stop_seqs_num, self.model_config.max_stop_seqs_num):

[gongshaotian]

After merging #2924, we need to delete the branch.

[carryyu]

和上面的分支冗余了，可以提到下面去。