[HiCache] feat: add draft KV cache backing for L2/L3

[alphabetc1]

Motivation

see #16964

Root Cause

In speculative decoding, the draft and target models share a req_to_token_pool (KV cache index mapping) but maintain separate KV cache pools.

When both HiCache and speculative decoding are enabled, the target KV cache may be loaded back from L2/L3. The shared index table is updated accordingly, but the draft model’s KV cache is not updated. As a result, the draft model reads incorrect KV, leading to accept length regression.

A natural alternative would be to re-run draft prefill to regenerate its KV cache after the target load_back. However, in speculative decoding the draft prefill depends on target hidden states (fc(concat(embed(token_{i+1}), target_hidden_state_i))). When KV is loaded back from L2/L3, the target prefill stage is skipped and the full hidden states are not produced, so the draft model cannot perform prefill.

Solution

This PR supports paired offload/load_back of the target and draft KV caches to/from L2/L3 (based on #17338).

How wo reproduce:

1. launch with spec+hicache
SGLANG_HICACHE_FILE_BACKEND_STORAGE_DIR=/root/code/tmp/sglang_hicache_file_test \
python3 -m sglang.launch_server \
  --model /models/ZhipuAI/GLM-4.7-FP8/ \
  --enable-hierarchical-cache \
  --hicache-io-backend direct \
  --hicache-mem-layout page_first_direct \
  --hicache-write-policy write_through --host 0.0.0.0 \
  --mem-fraction-static 0.55 --page-size 64 \
  --port 7000 \
  --reasoning-parser glm45 \
  --served-model-name GLM-4.7 \
  --speculative-algorithm EAGLE \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 8 \
  --speculative-num-steps 7 \
  --tool-call-parser glm47 \
  --tp 8 \
  --trust-remote-code \
  --model-loader-extra-config '{"enable_multithread_load": true, "num_threads": 64}' \
  --hicache-size 60 \
  --hicache-storage-prefetch-policy wait_complete \
  --hicache-storage-backend file \
   --speculative-draft-model-quantization compressed-tensors

2. send request to write target kvcache
 python3 -m sglang.bench_serving \
    --backend sglang-oai \
    --dataset-name random \
    --num-prompts 1 \
    --model /models/ZhipuAI/GLM-4.7-FP8/ \
    --dataset-path /root/code/datasets/ShareGPT_V3_unfiltered_cleaned_split.json \
    --random-input-len 150000 \
    --random-output-len 1024 \
    --random-range-ratio 0.3 \
    --max-concurrency 1 \
    --warmup-requests 0 \
    --seed 77 \
    --host 0.0.0.0 --port 7000

3. cleanup L1 kvcache(relaunch sglang)

4. benchmark with long prefix hit
 python3 -m sglang.bench_serving \
    --backend sglang-oai \
    --dataset-name random \
    --num-prompts 1 \
    --model /models/ZhipuAI/GLM-4.7-FP8/ \
    --dataset-path /root/code/datasets/ShareGPT_V3_unfiltered_cleaned_split.json \
    --random-input-len 150000 \
    --random-output-len 1024 \
    --random-range-ratio 0.3 \
    --max-concurrency 1 \
    --warmup-requests 0 \
    --seed 77 \
    --host 0.0.0.0 --port 7000

benchmark result before this feature(draft L2/L3kvcache):

============ Serving Benchmark Result ============
Backend:                                 sglang-oai
Traffic request rate:                    inf
Max request concurrency:                 1
Successful requests:                     1
Benchmark duration (s):                  11.29
Total input tokens:                      92831
Total input text tokens:                 92831
Total generated tokens:                  914
Total generated tokens (retokenized):    914
Request throughput (req/s):              0.09
Input token throughput (tok/s):          8224.80
Output token throughput (tok/s):         80.98
Peak output token throughput (tok/s):    29.00
Peak concurrent requests:                1
Total token throughput (tok/s):          8305.78
Concurrency:                             1.00
Accept length:                           3.07
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   11278.09
Median E2E Latency (ms):                 11278.09
P90 E2E Latency (ms):                    11278.09
P99 E2E Latency (ms):                    11278.09
---------------Time to First Token----------------
Mean TTFT (ms):                          365.38
Median TTFT (ms):                        365.38
P99 TTFT (ms):                           365.38
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          11.95
Median TPOT (ms):                        11.95
P99 TPOT (ms):                           11.95
---------------Inter-Token Latency----------------
Mean ITL (ms):                           11.95
Median ITL (ms):                         11.65
P95 ITL (ms):                            18.40
P99 ITL (ms):                            18.80
Max ITL (ms):                            35.02
==================================================

benchmark result after this feature(draft L2/L3kvcache):
accelt length 3.07->6.94

============ Serving Benchmark Result ============
Backend:                                 sglang-oai
Traffic request rate:                    inf
Max request concurrency:                 1
Successful requests:                     1
Benchmark duration (s):                  5.29
Total input tokens:                      92831
Total input text tokens:                 92831
Total generated tokens:                  914
Total generated tokens (retokenized):    914
Request throughput (req/s):              0.19
Input token throughput (tok/s):          17533.77
Output token throughput (tok/s):         172.63
Peak output token throughput (tok/s):    29.00
Peak concurrent requests:                1
Total token throughput (tok/s):          17706.40
Concurrency:                             1.00
Accept length:                           6.94
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   5286.52
Median E2E Latency (ms):                 5286.52
P90 E2E Latency (ms):                    5286.52
P99 E2E Latency (ms):                    5286.52
---------------Time to First Token----------------
Mean TTFT (ms):                          409.57
Median TTFT (ms):                        409.57
P99 TTFT (ms):                           409.57
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          5.34
Median TPOT (ms):                        5.34
P99 TPOT (ms):                           5.34
---------------Inter-Token Latency----------------
Mean ITL (ms):                           5.34
Median ITL (ms):                         5.01
P95 ITL (ms):                            7.54
P99 ITL (ms):                            8.74
Max ITL (ms):                            54.88
==================================================

Modifications

Accuracy Tests

Benchmarking and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
4. After green CI and required approvals, ask Merge Oncalls to merge.

[gemini-code-assist]

Summary of Changes

Hello, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request introduces a draft KV cache backing for the HiCache system, specifically targeting L2/L3 caches. This enhancement addresses a regression issue in speculative decoding where the draft model could see stale or incorrect slot indices after a load_back operation. The changes ensure that the draft KV pool is properly synchronized with the main KV cache, improving the accuracy and stability of speculative decoding.

Highlights

• HiCache Draft KV Pool: Introduces a draft KV cache backing for L2/L3 HiCache to improve speculative decoding accuracy by ensuring draft KV is restored to correct indices after load_back operations.
• Integration with HiRadixCache and HiMambaRadixCache: The draft KV pool functionality is integrated into both HiRadixCache and HiMambaRadixCache via a mixin class, providing a unified approach for managing draft KV in hierarchical caches.
• Hooks for Write, Load, and Evict Operations: The changes include hooks into the HiCache write, load, and evict paths to manage the draft KV pool, ensuring consistency between the main and draft KV caches.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request introduces support for backing up the draft model's KV cache to L2 (host) and L3 (storage) when using HiCache with speculative decoding. This is achieved by a new HiCacheDraftMixin that hooks into HiCache's data movement operations (write, load, evict) to synchronize the draft KV cache with the main model's KV cache. The changes are well-integrated into HiRadixCache and HiMambaRadixCache. My main feedback is to refactor a piece of duplicated logic in the scheduler for better maintainability.

[alphabetc1]

/rerun-failed-ci 1

[ovowei]

can this work for eagle3？

[alphabetc1]

can this work for eagle3？

yes. but may not work well for multi-layer eagle

[alphabetc1]

/tag-and-rerun-ci

[ykwd]

Thanks for this work. Is this pr compatible with DSA model (#21259) ?

[alphabetc1]

/rerun-failed-ci

[alphabetc1]

/rerun-test test_hicache_storage.py test_hicache_variants.py test_hicache_storage_3fs_backend.py test_hicache_storage_file_backend.py test_hicache_storage_mooncake_backend.py test_hicache_storage_runtime_attach_detach.py test_hicache_spec_file_storage.py

[github-actions]

✅ 1-gpu-5090 (1 test): View workflow run

cd test/ && python3 registered/hicache/test_hicache_storage.py

✅ 1-gpu-h100 (2 tests): View workflow run

cd test/ && python3 registered/hicache/test_hicache_variants.py
cd test/ && python3 registered/hicache/test_hicache_spec_file_storage.py

✅ 2-gpu-h100 (4 tests): View workflow run

cd test/ && python3 registered/hicache/test_hicache_storage_3fs_backend.py
cd test/ && python3 registered/hicache/test_hicache_storage_file_backend.py
cd test/ && python3 registered/hicache/test_hicache_storage_mooncake_backend.py
cd test/ && python3 registered/hicache/test_hicache_storage_runtime_attach_detach.py

[alphabetc1]

✅ 1-gpu-5090 (1 test): View workflow run

cd test/ && python3 registered/hicache/test_hicache_storage.py

✅ 1-gpu-h100 (2 tests): View workflow run

cd test/ && python3 registered/hicache/test_hicache_variants.py
cd test/ && python3 registered/hicache/test_hicache_spec_file_storage.py

✅ 2-gpu-h100 (4 tests): View workflow run

cd test/ && python3 registered/hicache/test_hicache_storage_3fs_backend.py
cd test/ && python3 registered/hicache/test_hicache_storage_file_backend.py
cd test/ && python3 registered/hicache/test_hicache_storage_mooncake_backend.py
cd test/ && python3 registered/hicache/test_hicache_storage_runtime_attach_detach.py

ALL hicache ci passed & local ci passed