[PD Disaggregation] pd + cache_storage support vl model

[juncaipeng]

Motivation

pd + cache_storage support vl model

💡 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

Use get_block_hash_extra_keys to cal cache keys

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

该 PR 旨在让 PD Disaggregation + cache_storage 在 VL（多模态）模型场景下能够生成一致的 cache key：通过在 PD 分离传输时透传 multimodal 的位置/哈希信息，并在 Decode 侧写回存储时把这些 extra keys 纳入 block hash 计算。

Changes:

• Request.to_dict() 在 V1 调度模式下，除 position_ids 外额外保留 mm_positions / mm_hashes 用于解码侧。
• PrefixCacheManager.write_cache_to_storage_decode() 生成 chained hash key 时引入 get_block_hash_extra_keys()（将 mm_hashes 参与每个 block 的 hash 输入）。
• 扩充 write_cache_to_storage() 的 docstring 说明写回流程。

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
fastdeploy/engine/request.py	PD 分离传输时放宽 multimodal_inputs 白名单，确保解码侧拿到 mm_positions/mm_hashes
fastdeploy/cache_manager/prefix_cache_manager.py	Decode 侧写回存储的 key 生成逻辑加入 multimodal extra keys，使其与 RadixTree/P 侧一致

[codecov-commenter]

Codecov Report

❌ Patch coverage is 50.00000% with 3 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@c9f7f52). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/cache_manager/prefix_cache_manager.py	50.00%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #6906   +/-   ##
==========================================
  Coverage           ?   72.83%           
==========================================
  Files              ?      399           
  Lines              ?    55716           
  Branches           ?     8776           
==========================================
  Hits               ?    40578           
  Misses             ?    12232           
  Partials           ?     2906           

Flag	Coverage Δ
GPU	72.83% <50.00%> (?)

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.

[Copilot]

Pull request overview

该 PR 旨在让 PD 分离（PD Disaggregation）+ cache_storage 在多模态/VL 请求下生成一致的缓存 key，从而支持多模态内容参与 prefix cache / storage cache 的命中与回写。

Changes:

• 在 PrefixCacheManager 的 storage 预取与 D 节点回写路径中，引入 get_block_hash_extra_keys() 将多模态 hash 纳入 block key 计算。
• 在 Request.to_dict() 的 V1 过滤逻辑中，额外透传 mm_positions/mm_hashes 以支持 decode 节点侧计算一致的 cache key。
• 更新 prefix cache manager 的相关测试用例构造，补齐 multimodal_inputs 字段。

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
tests/cache_manager/test_prefix_cache_manager.py	更新测试 request 构造以适配多模态字段接入
fastdeploy/engine/request.py	V1 模式下 multimodal_inputs 过滤白名单加入 mm_positions/mm_hashes
fastdeploy/cache_manager/prefix_cache_manager.py	storage 预取与 decode 回写的 key 生成链路加入多模态 extra keys；扩充写回函数说明

Comments suppressed due to low confidence (1)

fastdeploy/engine/request.py:470

• PR 描述里 Usage/Accuracy Tests 仍是占位符（“-”），Checklist 也未勾选；但该改动会影响 PD 分离场景下多模态请求在 cache/storage 的 key 计算与数据透传。建议补充：1) 触发该功能的配置/环境变量与使用方式；2) 至少一条可复现的验证步骤或准确性/回归测试结果；3) 若暂不加单测请说明原因。

        if isinstance(self.multimodal_inputs, dict):
            # Optimize multimodal data transfer during PD separation:
            # - V1 mode (ENABLE_V1_KVCACHE_SCHEDULER=1): position_ids, mm_positions and mm_hashes needed for decode nodes
            # - V0 mode (ENABLE_V1_KVCACHE_SCHEDULER=0): Full field set required for compatibility
            # This filtering significantly reduces serialized data size for large numpy arrays
            allowed_keys = {"position_ids", "mm_positions", "mm_hashes"}
            if not envs.ENABLE_V1_KVCACHE_SCHEDULER:
                allowed_keys.update(["input_ids", "token_type_ids", "images", "image_type_ids", "grid_thw"])

            data["multimodal_inputs"] = {
                key: value for key, value in self.multimodal_inputs.items() if key in allowed_keys