bytedance / bytedance/InfiniStore
Unexpected hang when executing infinistore-based PD Disaggregation.
- Dominant language
- C++
- Stars
- 437
- Forks
- 44
- PR merge metrics
- No merged PRs in 30d
Description
I am benchmarking 1P1D performance of infinistore using vLLM benchmark_serving script that generates random requests. I successfully tested n=32 bs=8 and n=128 bs=32 cases with 1024 input and 256 output. But the program hangs with no errors when testing n=512 bs=128 case.
```shell
[2025-03-31 10:22:32.740] [infini] [warning] [infinistore.cpp:638] Key not found: vllm@LLM-Research/Meta-Llama-3.1-8B-Instruct@4@2@dcf8c97ee6ea445b069f7b4ce7c64ddb8d21f4f802b1cfa978b596e03b5000d6metadata
[2025-03-31 10:22:32.740] [infini] [warning] [infinistore.cpp:638] Key not found: vllm@LLM-Research/Meta-Llama-3.1-8B-Instruct@4@1@dcf8c97ee6ea445b069f7b4ce7c64ddb8d21f4f802b1cfa978b596e03b5000d6metadata
[2025-03-31 10:22:32.793] [infini] [warning] [infinistore.cpp:638] Key not found: vllm@LLM-Research/Meta-Llama-3.1-8B-Instruct@4@2@ea1fdde5b08b9b2a2b2c9ca5a1e55aadc7eba79b731125b420a994d58dee14fckv_bytes
[2025-03-31 10:22:32.804] [infini] [warning] [infinistore.cpp:638] Key not found: vllm@LLM-Research/Meta-Llama-3.1-8B-Instruct@4@1@7abd2ea5d154c86670a03e5c3e2d3f669cda2f606aea526a9cb3e5372fbf2b51kv_bytes
[2025-03-31 10:22:32.813] [infini] [warning] [infinistore.cpp:638] Key not found: vllm@LLM-Research/Meta-Llama-3.1-8B-Instruct@4@3@652524da94b364056564d62c899b2164a87b3dce1532027c2d3c66a2d360f654kv_bytes
[2025-03-31 10:22:32.822] [infini] [warning] [infinistore.cpp:638] Key not found: vllm@LLM-Research/Meta-Llama-3.1-8B-Instruct@4@0@8139aa78a94781ac1680c87db6a363077a619d86aa65aed83773ff99d75b1b29kv_bytes
[2025-03-31 10:22:32.935] [infini] [info] evict memory done, usage: from 0.95 => 0.80
[2025-03-31 10:22:34.065] [infini] [info] evict memory done, usage: from 0.95 => 0.80
```
My benchmarking scripts is
```
command = [
'python', 'benchmark_serving.py',
'--backend', 'vllm',
'--dataset-name', 'random',
'--model', r'LLM-Research/Meta-Llama-3.1-8B-Instruct',
'--tokenizer', r'/root/.cache/modelscope/hub/models/LLM-Research/Meta-Llama-3.1-8B-Instruct/',
'--dataset-path', 'random',
'--random-input-len', str(random_len),
'--random-output-len', str(random_output_len),
'--random-range-ratio', '1',
'--request-rate', str(request_rate),
'--max-concurrency', str(max_concurrency),
'--num-prompts', str(num_prompts),
'--host', '0.0.0.0',
'--port', '8080',
'--save-result',
'--result-dir', './log'
]
```
Does anyone have any ideas about this hanging?
Contributor guide
Research direction
Reproduce the issue with the shown benchmark_serving.py command, comparing the working n=32 and n=128 cases with n=512. Start by inspecting the infinistore.cpp:638 key-not-found warnings and the subsequent eviction messages. Done means the n=512, bs=128 benchmark completes without hanging and the relevant failure is reported clearly.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp, python
- Domain
- ai-infra-agents, distributed-systems
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 35/100