[Kernel] Enable FP16 and BF16 CUTLASS MoE kernels

[ElizaWszola]

Implement BF16 and FP16 weight support in CUTLASS MoE kernels. Tested with

llm = LLM("mistralai/Mixtral-8x7B-Instruct-v0.1",
          tensor_parallel_size=2,
)

and

llm = LLM("mistralai/Mixtral-8x7B-Instruct-v0.1",
          tensor_parallel_size=2,
          dtype=torch.float16,
)

[github-actions[bot]]

👋 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.

🚀

[mergify[bot]]

This pull request has merge conflicts that must be resolved before it can be merged. Please rebase the PR, @ElizaWszola.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[tlrmchlsmth]

It appears we need to tune these before landing.

vllm serve mistralai/Mixtral-8x7B-Instruct-v0.1 --tensor_parallel_size=2 --max_model_len=4096 --port 8192 --disable-log-requests --no-enable-prefix-caching

python benchmarks/benchmark_serving.py --model mistralai/Mixtral-8x7B-Instruct-v0.1 --dataset-name random --random-input-len 1000 --random-output-len 100 --ignore-eos --port 8192 --request-rate 10

main

============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  103.86
Total input tokens:                      1000000
Total generated tokens:                  100000
Request throughput (req/s):              9.63
Output token throughput (tok/s):         962.84
Total Token throughput (tok/s):          10591.26
---------------Time to First Token----------------
Mean TTFT (ms):                          81.42
Median TTFT (ms):                        74.82
P99 TTFT (ms):                           151.31
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          25.38
Median TPOT (ms):                        25.33
P99 TPOT (ms):                           29.51
---------------Inter-token Latency----------------
Mean ITL (ms):                           25.38
Median ITL (ms):                         20.68
P99 ITL (ms):                            67.00
==================================================

this pr

============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  104.13
Total input tokens:                      1000000
Total generated tokens:                  100000
Request throughput (req/s):              9.60
Output token throughput (tok/s):         960.35
Total Token throughput (tok/s):          10563.86
---------------Time to First Token----------------
Mean TTFT (ms):                          152.00
Median TTFT (ms):                        138.80
P99 TTFT (ms):                           357.98
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          38.58
Median TPOT (ms):                        38.34
P99 TPOT (ms):                           50.13
---------------Inter-token Latency----------------
Mean ITL (ms):                           38.58
Median ITL (ms):                         21.81
P99 ITL (ms):                            161.14
==================================================

[mergify[bot]]

This pull request has merge conflicts that must be resolved before it can be merged. Please rebase the PR, @ElizaWszola.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[varun-sundar-rabindranath]

@tlrmchlsmth - I have the changes here https://github.com/neuralmagic/vllm/pull/57 waiting to be merged on the neuralmagic:cutlass-moe-bf16-weights branch. I am still getting the e2e and microbenchmarks.

[varun-sundar-rabindranath]

Factoring out expert_map support into a separate PR https://github.com/vllm-project/vllm/pull/16861

[mergify[bot]]

This pull request has merge conflicts that must be resolved before it can be merged. Please rebase the PR, @ElizaWszola.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify[bot]]

This pull request has merge conflicts that must be resolved before it can be merged. Please rebase the PR, @ElizaWszola.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[github-actions[bot]]

This pull request has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this pull request should remain open. Thank you!

[github-actions[bot]]

This pull request has been automatically closed due to inactivity. Please feel free to reopen if you intend to continue working on it. Thank you!