[Core] LoRA: V1 Scheduler optimization

[varun-sundar-rabindranath]

LoRA Scheduler Optimization

Running Example:
Let max_loras be set to 4
Let Waiting queue be : { R1, R2, R3, R4-L1, R5-L3, R6-L2, R7-L4, R8-L5, R9-L5, R10-L5, R11-L5, R12, R13, R14-L1, R15-L2, R16-L3 }
Rx - Request number x. Request doesn't need any LoRA
Rx-Ly - Request number x that needs LoRA number y

Why:
In V1 + LoRA, at the moment we stop scheduling waiting requests when we can no longer honor max_loras user inputs. This is not optimal as,

• Other requests that do not use LoRA, or,
• Other requests that use the already requested LoRAs.
  are blocked from scheduling.

In the example above,

• The scheduler schedules { R1, R2, R3, R4-L1, R5-L3, R6-L2, R7-L4 }
• In future iterations the scheduler can't schedule additional requests until one of { R4-L1, R5-L3, R6-L2, R7-L4 } completes.
• Lets say R4-L1 completes.
• Now, the scheduler will schedule {R1, R2, R3, R5-L3, R6-L2, R7-L4, R8-L5, R9-L5, R10-L5, R11-L5, R12, R13}
• and so on

What:
This PR updates the scheduling logic to continue scanning the waiting queue for requests that can still be scheduled.
With this PR,

• The scheduler will schedule { R1, R2, R3, R4-L1, R5-L3, R6-L2, R7-L4, R12, R13, R14-L1, R15-L2, R16-L3 } . The requests {R8-L5, R9-L5, R10-L5, R11-L5} will stay in the waiting queue.
• When one of the LoRA request set completes i.e. {R4-L1, R14-L1} or {R5-L3, R16-L3} or {R6-L2, R15-L2} the scheduler will continue scheduling the {R8-L5, R9-L5, R10-L5, R11-L5}

How:
We leverage the "skip waiting requests" logic introduced by structured decoding.

Benchmarks:

max_loras = 4 , number of lora modules = 8 , max_num_seqs = 256, max_num_batched_tokens = 4096

Server Command :

vllm serve  meta-llama/Llama-2-7b-hf  --enable-lora  --max-loras 4  --max-lora-rank 8  --lora-modules lora0=yard1/llama-2-7b-sql-lora-test lora1=yard1/llama-2-7b-sql-lora-test lora2=yard1/llama-2-7b-sql-lora-test lora3=yard1/llama-2-7b-sql-lora-test lora4=yard1/llama-2-7b-sql-lora-test lora5=yard1/llama-2-7b-sql-lora-test lora6=yard1/llama-2-7b-sql-lora-test lora7=yard1/llama-2-7b-sql-lora-test                       --max-num-seqs 256   --max-num-batched-tokens 4096  --no-enable-prefix-caching   --port 9002  --disable-log-stats

benchmark_serving.py command

python3 benchmarks/benchmark_serving.py   --model meta-llama/Llama-2-7b-hf                     --dataset-name sharegpt  --dataset-path ./ShareGPT_V3_unfiltered_cleaned_split.json   --sharegpt-output-len 120  --num-prompts 500 --request-rate inf  --lora-modules lora0 lora1 lora2 lora3 lora4 lora5 lora6 lora7  --ignore-eos --port 9002  --seed 2

main V1

============ Serving Benchmark Result ============
Successful requests:                     500       
Benchmark duration (s):                  200.71    
Total input tokens:                      124520    
Total generated tokens:                  60000     
Request throughput (req/s):              2.49      
Output token throughput (tok/s):         298.94    
Total Token throughput (tok/s):          919.35    
---------------Time to First Token----------------
Mean TTFT (ms):                          99070.16  
Median TTFT (ms):                        101450.02 
P99 TTFT (ms):                           195994.95 
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          15.13     
Median TPOT (ms):                        15.24     
P99 TPOT (ms):                           16.17     
---------------Inter-token Latency----------------
Mean ITL (ms):                           15.31     
Median ITL (ms):                         15.22     
P99 ITL (ms):                            16.88     
==================================================

main V0

============ Serving Benchmark Result ============
Successful requests:                     500       
Benchmark duration (s):                  29.19     
Total input tokens:                      124520    
Total generated tokens:                  60000     
Request throughput (req/s):              17.13     
Output token throughput (tok/s):         2055.28   
Total Token throughput (tok/s):          6320.66   
---------------Time to First Token----------------
Mean TTFT (ms):                          10400.52  
Median TTFT (ms):                        6083.83   
P99 TTFT (ms):                           26071.17  
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          79.85     
Median TPOT (ms):                        80.54     
P99 TPOT (ms):                           106.48    
---------------Inter-token Latency----------------
Mean ITL (ms):                           79.85     
Median ITL (ms):                         61.96     
P99 ITL (ms):                            389.40    
==================================================

This PR

============ Serving Benchmark Result ============
Successful requests:                     500       
Benchmark duration (s):                  29.84     
Total input tokens:                      124520    
Total generated tokens:                  60000     
Request throughput (req/s):              16.75     
Output token throughput (tok/s):         2010.46   
Total Token throughput (tok/s):          6182.82   
---------------Time to First Token----------------
Mean TTFT (ms):                          11862.28  
Median TTFT (ms):                        6406.40   
P99 TTFT (ms):                           24552.50  
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          73.55     
Median TPOT (ms):                        77.75     
P99 TPOT (ms):                           102.69    
---------------Inter-token Latency----------------
Mean ITL (ms):                           73.55     
Median ITL (ms):                         50.22     
P99 ITL (ms):                            346.09    
==================================================

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[comaniac]

Otherwise LGTM

[varun-sundar-rabindranath]

Requesting review from @russellb for the changes to the "Structured Outputs" land! 🙌

[varun-sundar-rabindranath]

Requesting reviews from @jeejeelee 🙌

[WoosukKwon]

Besides the variable name, the PR looks good to me!
@varun-sundar-rabindranath Do you have any performance numbers after this PR?

[varun-sundar-rabindranath]

Besides the variable name, the PR looks good to me! @varun-sundar-rabindranath Do you have any performance numbers after this PR?

I am running some benchmarks now. Ill add it to the PR 👍

[varun-sundar-rabindranath]

Besides the variable name, the PR looks good to me! @varun-sundar-rabindranath Do you have any performance numbers after this PR?

Hi @WoosukKwon - Added benchmark numbers to the PR description. It definitely helps V1 when max_loras < number of loras used. However V1 LoRA in this case does lag behind V0 - I think it has to do with the set_active_loras overhead in scheduling. I can handle it in a separate PR.