Allow torchao quantization in SiglipMLP #15575
Conversation
Summary: SiglipMLP is not quantized by default, so when we try to load a torchao quantized model it doesn't work because the mlps are not quantized, this PR enables quantization when we are loading a torchao quantized checkpoint. Depends on vllm-project#14231 Test Plan: python benchmarks/benchmark_latency.py --input-len 256 --output-len 256 --model jerryzh168/gemma3-4b-it-float8dq --batch-size 1 Reviewers: Subscribers: Tasks: Tags:
|
👋 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 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 🚀 |
|
Thanks, please address the CI failure. |
jeejeelee
left a comment
jeejeelee
left a comment
There was a problem hiding this comment.
Thank you, we can merge first
jeejeelee
added
ready
Signed-off-by: xinyuxiao <[email protected]>
Signed-off-by: Louis Ulmer <[email protected]>
Signed-off-by: Mu Huai <[email protected]>
3 participants
Summary:
SiglipMLP is not quantized by default, so when we try to load a torchao quantized model it doesn't work because the mlps are not quantized, this PR enables quantization when we are loading a torchao quantized checkpoint.
Depends on #14231
Test Plan:
python benchmarks/benchmark_latency.py --input-len 256 --output-len 256 --model jerryzh168/gemma3-4b-it-float8dq --batch-size 1
Reviewers:
Subscribers:
Tasks:
Tags: