[V1] Enable V1 Fp8 cache for FA3 in the oracle

[LucasWilkinson]

Now that FA3 Fp8 support has landed (#14570) we can enable Fp8 KV-caches for Hopper devices in V1

Update the oracle to allow this plus basic refactors

Test script

# SPDX-License-Identifier: Apache-2.0

from vllm import LLM, SamplingParams

# Sample prompts.
prompts = [
    "Hello, my name is",
    "The president of the United States is",
    "The capital of France is",
    "The future of AI is",
]
# Create a sampling params object.
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)

if __name__ == '__main__':
    # Create an LLM.
    llm = LLM(model="meta-llama/Llama-3.2-1B-Instruct", kv_cache_dtype="fp8")
    # Generate texts from the prompts. The output is a list of RequestOutput objects
    # that contain the prompt, generated text, and other information.
    outputs = llm.generate(prompts, sampling_params)
    # Print the outputs.
    for output in outputs:
        prompt = output.prompt
        generated_text = output.outputs[0].text
        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")

Results:

(vllm) lwilkinson@beaker:~/code/vllm$ python examples/offline_inference/basic/basic.py 
....
Prompt: 'Hello, my name is', Generated text: " Rachel and I'm a software developer. I'm new to the tech industry and"
Prompt: 'The president of the United States is', Generated text: ' the head of state and government of the United States. The president is elected by'
Prompt: 'The capital of France is', Generated text: ' Paris. You can visit the Eiffel Tower, the Louvre, and'
Prompt: 'The future of AI is', Generated text: ' a complex and multifaceted topic that spans various fields, including philosophy, ethics'
[rank0]:[W320 06:15:16.475614640 ProcessGroupNCCL.cpp:1496] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())


(vllm) lwilkinson@beaker:~/code/vllm$ VLLM_USE_V1=0 python examples/offline_inference/basic/basic.py 
....
Prompt: 'Hello, my name is', Generated text: " Alex. I've been spending a lot of time online lately, and I'm"
Prompt: 'The president of the United States is', Generated text: ' elected by the people through the Electoral College system. This system was established by the'
Prompt: 'The capital of France is', Generated text: ' Paris.\nThe capital of Germany is Berlin.\nThe capital of the United States is'
Prompt: 'The future of AI is', Generated text: " here, and it's making a huge impact on our lives. From intelligent machines"
[rank0]:[W320 06:18:31.628198327 ProcessGroupNCCL.cpp:1496] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

[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.

🚀

[houseroad]

Looks good to me.

[mergify]

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

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

[WoosukKwon]

@LucasWilkinson This is amazing! Could you please provide accuracy benchmark?

[LucasWilkinson]

@LucasWilkinson This is amazing! Could you please provide accuracy benchmark?

lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.3-70B-Instruct,tensor_parallel_size=2,dtype=auto,gpu_memory_utilization=0.9,trust_remote_code=True,max_model_len=16384 --task gsm8k --num_fewshot=5 --limit 200
....
INFO 03-22 02:59:18 [kv_cache_utils.py:537] GPU KV cache size: 347,136 tokens
INFO 03-22 02:59:18 [kv_cache_utils.py:540] Maximum concurrency for 16,384 tokens per request: 21.19x
INFO 03-22 02:59:18 [kv_cache_utils.py:537] GPU KV cache size: 347,136 tokens
INFO 03-22 02:59:18 [kv_cache_utils.py:540] Maximum concurrency for 16,384 tokens per request: 21.19x
....
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.935|±  |0.0175|
|     |       |strict-match    |     5|exact_match|↑  |0.900|±  |0.0213|

fp8

lm_eval --model vllm --model_args pretrained=meta-llama/Llama-3.3-70B-Instruct,tensor_parallel_size=2,dtype=auto,gpu_memory_utilization=0.9,trust_remote_code=True,max_model_len=16384,kv_cache_dtype=fp8 --task gsm8k --num_fewshot=5 --limit 200
....
INFO 03-22 03:13:26 [kv_cache_utils.py:537] GPU KV cache size: 694,272 tokens
INFO 03-22 03:13:26 [kv_cache_utils.py:540] Maximum concurrency for 16,384 tokens per request: 42.38x
INFO 03-22 03:13:26 [kv_cache_utils.py:537] GPU KV cache size: 694,272 tokens
INFO 03-22 03:13:26 [kv_cache_utils.py:540] Maximum concurrency for 16,384 tokens per request: 42.38x
....
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.930|±  |0.0181|
|     |       |strict-match    |     5|exact_match|↑  |0.895|±  |0.0217|

[renjie0]

What is the expected perf difference