[Perf] Vectorize static_scaled_int8_quant_kernel

[ztang2370]

FIX partially #18866.

I noticed the issue was assigned but hasn’t seen activity yet, so I went ahead and implemented vectorization for weights quantization, i.e. static_scaled_int8_quant_kernel. Totally happy to close this if the assignee is actively working on it — or to collaborate if helpful. Just wanted to contribute in case it helps move things forward!

Evaluations are conducted on an RTX3090.

// Accuracy
lm_eval --model vllm --model_args pretrained=RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8,max_model_len=32768 --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size auto

// before
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.7779|±  |0.0114|
|     |       |strict-match    |     5|exact_match|↑  |0.7597|±  |0.0118|

// after
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.7779|±  |0.0114|
|     |       |strict-match    |     5|exact_match|↑  |0.7597|±  |0.0118|


// e2e throughput benchmark
vllm bench throughput --model RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8 --load-format dummy --input-len 1000 --output-len 100 --max-model-len 32768

// before
Throughput: 7.34 requests/s, 8067.01 total tokens/s, 734.31 output tokens/s

// after
Throughput: 7.54 requests/s, 8284.22 total tokens/s, 753.97 output tokens/s

[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.

🚀