fix pp for deepseek

[zhjc1124]

Motivation

run DeepSeek in pp.
#5724 #5925

$python3 -m sglang.bench_one_batch_server --model /data/modelscope/DeepSeek-Coder-V2-Lite-Instruct/ --batch-size 1 --trust-remote-code --base-gpu-id 4 --port 38884 --pp 2 --tp 2
batch size: 16
latency: 2.50 s
output throughput: 102.25 token/s
(input + output) throughput: 6645.98 token/s
[2025-05-09 17:37:07 TP0 PP0] Prefill batch. #new-seq: 1, #new-token: 1024, #cached-token: 0, token usage: 0.00, #running-req: 0, #queue-req: 0
[2025-05-09 17:37:07 TP0 PP1] Prefill batch. #new-seq: 1, #new-token: 1024, #cached-token: 0, token usage: 0.00, #running-req: 0, #queue-req: 0
[2025-05-09 17:37:07 TP0 PP0] Decode batch. #running-req: 1, #token: 1027, token usage: 0.00, gen throughput (token/s): 20.20, #queue-req: 0
[2025-05-09 17:37:07 TP0 PP1] Decode batch. #running-req: 1, #token: 1027, token usage: 0.00, gen throughput (token/s): 20.19, #queue-req: 0
[2025-05-09 17:37:07] INFO: 127.0.0.1:59888 - "POST /generate HTTP/1.1" 200 OK
batch size: 1
latency: 0.17 s
output throughput: 95.08 token/s
(input + output) throughput: 6180.45 token/s

Modifications

modify deepseep_v2.py to fit pp.

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://slack.sglang.ai to discuss your PR.

[Edenzzzz]

@zhjc1124 Can I ask why you closed it? Is it because deepseek should use EP instead?

[zhjc1124]

@zhjc1124 Can I ask why you closed it? Is it because deepseek should use EP instead?

I saw ur comment too.
Once I found I failed to run DeepSeek-R1 in three nodes with tp=8 pp=3. So I considered if there are some compatibility with mla backend need to handle.
In fact, I found there was something wrong in my machine nccl configuration. Now I run DeepSeek-R1 with tp=8 and pp=3 sucessfully.But I am still not sure if the compatibility with mla backend need to be handled