[TPU] [Perf] Improve Memory Usage Estimation

[robertgshaw2-redhat]

SUMMARY:

• TPU uses ~3GB loading Llama-8B + compiling XLA graphs, which is captured in peak memory
• This 3GB was not being used for KV cache as a result
• Tweak heuristic to be 2% of weights instead

This improves performance 4%, since we are able to fit more items in the batch

MODEL=meta-llama/Llama-3.1-8B-Instruct

VLLM_USE_V1=1 vllm serve $MODEL --max-model-len 2048 --disable-log-requests

• this pr:

python3 benchmark_serving.py --model $MODEL --dataset-name random --random-input-len 1000 --random-output-len 100

>> Throughput: 8.22

• main

python3 benchmark_serving.py --model $MODEL --dataset-name random --random-input-len 1000 --random-output-len 100

>> Throughput: 7.89

[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.

🚀

[bvrockwell]

SUMMARY:

• TPU uses ~3GB loading Llama-8B + compiling XLA graphs, which is captured in peak memory
• This 3GB was not being used for KV cache as a result
• Tweak heuristic to be 2% of weights instead

This improves performance 4%, since we are able to fit more items in the batch

MODEL=meta-llama/Llama-3.1-8B-Instruct

VLLM_USE_V1=1 vllm serve $MODEL --max-model-len 2048 --disable-log-requests

• main:

python3 benchmark_serving.py --model $MODEL --dataset-name random --random-input-len 1000 --random-output-len 100

>> Throughput: 8.22

• this pr

python3 benchmark_serving.py --model $MODEL --dataset-name random --random-input-len 1000 --random-output-len 100

>> Throughput: 7.89

@robertgshaw2-redhat this is great, thanks for the simple tweak, but is the throughput result reversed in this comment?

[robertgshaw2-redhat]

SUMMARY:

• TPU uses ~3GB loading Llama-8B + compiling XLA graphs, which is captured in peak memory
• This 3GB was not being used for KV cache as a result
• Tweak heuristic to be 2% of weights instead

This improves performance 4%, since we are able to fit more items in the batch
MODEL=meta-llama/Llama-3.1-8B-Instruct

VLLM_USE_V1=1 vllm serve $MODEL --max-model-len 2048 --disable-log-requests

• main:

python3 benchmark_serving.py --model $MODEL --dataset-name random --random-input-len 1000 --random-output-len 100

>> Throughput: 8.22

• this pr

python3 benchmark_serving.py --model $MODEL --dataset-name random --random-input-len 1000 --random-output-len 100

>> Throughput: 7.89

@robertgshaw2-redhat this is great, thanks for the simple tweak, but is the throughput result reversed in this comment?

yep, I updated the summary

[robertgshaw2-redhat]

SUMMARY:

• TPU uses ~3GB loading Llama-8B + compiling XLA graphs, which is captured in peak memory
• This 3GB was not being used for KV cache as a result
• Tweak heuristic to be 2% of weights instead

This improves performance 4%, since we are able to fit more items in the batch
MODEL=meta-llama/Llama-3.1-8B-Instruct

VLLM_USE_V1=1 vllm serve $MODEL --max-model-len 2048 --disable-log-requests

• main:

python3 benchmark_serving.py --model $MODEL --dataset-name random --random-input-len 1000 --random-output-len 100

>> Throughput: 8.22

• this pr

python3 benchmark_serving.py --model $MODEL --dataset-name random --random-input-len 1000 --random-output-len 100

>> Throughput: 7.89

@robertgshaw2-redhat this is great, thanks for the simple tweak, but is the throughput result reversed in this comment?

yep, I updated the summary

[yaochengji]

LGTM! It's a simple modification but just works!

I double checked and confirm there's no such reset peak memory size support.
I submitted a feature request in the pytorch/xla project for the this functionality. pytorch/xla#8900