[ROCM][KERNEL] Paged attention for V1

[maleksan85]

Adopting ROCM Paged Attention to be use in V1 FA as alternative to Triton kernel. Perf I see:

Baseline (VLLM_ROCM_CUSTOM_PAGED_ATTN=0 and --no-enable-prefix-caching):

============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  27.51
Total input tokens:                      215196
Total generated tokens:                  197090
Request throughput (req/s):              36.35
Output token throughput (tok/s):         7164.57
Total Token throughput (tok/s):          14987.32
---------------Time to First Token----------------
Mean TTFT (ms):                          5068.86
Median TTFT (ms):                        4666.36
P99 TTFT (ms):                           10794.20
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          59.98
Median TPOT (ms):                        58.69
P99 TPOT (ms):                           87.44
---------------Inter-token Latency----------------
Mean ITL (ms):                           43.91
Median ITL (ms):                         34.99
P99 ITL (ms):                            95.14
==================================================

With change (--no-enable-prefix-caching):

============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  23.97
Total input tokens:                      215196
Total generated tokens:                  197119
Request throughput (req/s):              41.73
Output token throughput (tok/s):         8225.24
Total Token throughput (tok/s):          17204.79
---------------Time to First Token----------------
Mean TTFT (ms):                          4853.87
Median TTFT (ms):                        4488.42
P99 TTFT (ms):                           10301.76
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          56.14
Median TPOT (ms):                        54.89
P99 TPOT (ms):                           82.79
---------------Inter-token Latency----------------
Mean ITL (ms):                           39.92
Median ITL (ms):                         32.00
P99 ITL (ms):                            94.70
==================================================

latency:

python benchmarks/benchmark_latency.py --model /data/models/Llama-3.1-8B-Instruct --input-len 2048 --output-len 2048 --batch-size 64 --num-iters-warmup 1 --num-iters 3

V0
Avg latency: 40.9744499316439 seconds

V1
upstream (VLLM_ROCM_CUSTOM_PAGED_ATTN=0):
Avg latency: 46.44388557784259 seconds

with change:
Avg latency: 34.401718624557056 seconds

correctness:
2025-03-31:21:23:17,108 INFO [lm_eval.loggers.evaluation_tracker:272] Output path not provided, skipping saving results aggregated
vllm (pretrained=/data/models/Llama-3.1-8B-Instruct), gen_kwargs: (None), limit: 500.0, num_fewshot: 5, batch_size: auto

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	↑	0.800	±	0.0179
		strict-match	5	exact_match	↑	0.782	±	0.0185

cc @SageMoore

[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.

🚀

[maleksan85]

llama 70B TP2 same test:
upstream (--no-enable-prefix-caching)

============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  96.75
Total input tokens:                      215196
Total generated tokens:                  125080
Request throughput (req/s):              10.34
Output token throughput (tok/s):         1292.76
Total Token throughput (tok/s):          3516.92
---------------Time to First Token----------------
Mean TTFT (ms):                          24240.22
Median TTFT (ms):                        22927.67
P99 TTFT (ms):                           51727.34
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          267.65
Median TPOT (ms):                        233.12
P99 TPOT (ms):                           446.70
---------------Inter-token Latency----------------
Mean ITL (ms):                           172.88
Median ITL (ms):                         104.39
P99 ITL (ms):                            476.44
==================================================

this PR (--no-enable-prefix-caching):

============ Serving Benchmark Result ============
Successful requests:                     1000
Benchmark duration (s):                  91.33
Total input tokens:                      215196
Total generated tokens:                  127292
Request throughput (req/s):              10.95
Output token throughput (tok/s):         1393.71
Total Token throughput (tok/s):          3749.88
---------------Time to First Token----------------
Mean TTFT (ms):                          23905.00
Median TTFT (ms):                        22544.27
P99 TTFT (ms):                           50916.09
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          258.73
Median TPOT (ms):                        220.76
P99 TPOT (ms):                           440.69
---------------Inter-token Latency----------------
Mean ITL (ms):                           164.74
Median ITL (ms):                         97.24
P99 ITL (ms):                            470.45
==================================================

[maleksan85]

Output heavy load with llama 3.1 70B

python vllm/benchmarks/benchmark_serving.py --backend vllm --model /data/models/Llama-3.1-70B-Instruct --dataset-name random --random-input-len 150 --random-output-len 16384 --max-concurrency 32 --num-prompts 512 --seed 42

upstream (--no-enable-prefix-caching):

============ Serving Benchmark Result ============
Successful requests:                     512
Benchmark duration (s):                  2765.66
Total input tokens:                      76800
Total generated tokens:                  1782512
Request throughput (req/s):              0.19
Output token throughput (tok/s):         644.52
Total Token throughput (tok/s):          672.29
---------------Time to First Token----------------
Mean TTFT (ms):                          340.30
Median TTFT (ms):                        284.30
P99 TTFT (ms):                           1063.48
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          57.06
Median TPOT (ms):                        48.44
P99 TPOT (ms):                           196.84
---------------Inter-token Latency----------------
Mean ITL (ms):                           48.18
Median ITL (ms):                         48.23
P99 ITL (ms):                            60.74
==================================================

this PR (--no-enable-prefix-caching):

============ Serving Benchmark Result ============
Successful requests:                     512
Benchmark duration (s):                  2045.44
Total input tokens:                      76800
Total generated tokens:                  1782991
Request throughput (req/s):              0.25
Output token throughput (tok/s):         871.69
Total Token throughput (tok/s):          909.24
---------------Time to First Token----------------
Mean TTFT (ms):                          298.00
Median TTFT (ms):                        258.42
P99 TTFT (ms):                           1071.87
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          44.30
Median TPOT (ms):                        35.77
P99 TPOT (ms):                           189.64
---------------Inter-token Latency----------------
Mean ITL (ms):                           35.65
Median ITL (ms):                         35.31
P99 ITL (ms):                            39.86
==================================================

[ProExpertProg]

So if I understand correctly, this adds support for chunked prefill to the existing ROCm custom Paged Attention, and then for prefill we still use the _fwd_kernel?

[maleksan85]

So if I understand correctly, this adds support for chunked prefill to the existing ROCm custom Paged Attention, and then for prefill we still use the _fwd_kernel?

correct, we had two kernels, one for chunked prefills in a query and another kernel, which used to be only triton, for decodes in the query. This PR adds alternative to triton decode kernel for set of cases, please see limitations in use_rocm_custom_paged_attention function. As for the rocm paged attention kernel itself, the kernel (which in fact 3 C++ kernels), just skips prefills in the query now.

PS of course there might be more tailored way to utilize rocm paged attention for V1, however it will take more time to update the kernel.

[robertgshaw2-redhat]

Nice. This is consistent with my profiling of where the issues are on V1. Glad to see the update. This generally looks fine to me. I will follow up with a refactor for attn backend.

@SageMoore @ProExpertProg - can you guys look through the cpp and let me know if its okay to merge

[ProExpertProg]

CPP looks good to me, very contained and straightforward change 😃

[robertgshaw2-redhat]

^ Mistral Small, TP=1, 1000 In | 100 Out

Nice step forward but still juice to squeeze
cc @SageMoore @ProExpertProg

[mgoin]

keep the static cast?

[maleksan85]

should this be conversion similar to what static cast provides? I mean when assignment happens to int64_t from int?