[NVIDIA] Support Flashinfer TRTLLM FP8-q/kv/out Attention Kernel

[elvischenv]

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

Purpose

• Refactor the AttentionStaticQuantPattern in fusion_attn pass, which will fuse the attn+fp8_quant pattern for using the TRTLLM FP8-in FP8-out kernel.
• Align with other backends, Flashinfer will insert FP8 quant op for query if kv cache dtype is set to FP8.

Test Plan && Test Result

Functional:

• FP8 TRTLLM Prefill/Decode kernel unit test: tests/kernels/attention/test_flashinfer_trtllm_attention.py

====== 192 passed, 10 warnings in 143.25s (0:02:23) ======

• AttentionStaticQuantPattern fusion_attn pass: tests/compile/test_fusion_attn.py::test_attention_quant_pattern

====== 6 passed, 4 warnings in 33.01s ========

E2E Performance: nvidia/Llama-4-Scout-17B-16E-Instruct-FP8

main

--kv-cache-dtype = auto                                                      --kv-cache-dtype = fp8
============ Serving Benchmark Result ============                           ============ Serving Benchmark Result ============
Successful requests:                     640                                 Successful requests:                     640
Maximum request concurrency:             128                                 Maximum request concurrency:             128
Benchmark duration (s):                  209.72                              Benchmark duration (s):                  229.07
Total input tokens:                      653975                              Total input tokens:                      653975
Total generated tokens:                  655360                              Total generated tokens:                  655360
Request throughput (req/s):              3.05                                Request throughput (req/s):              2.79
Output token throughput (tok/s):         3124.98                             Output token throughput (tok/s):         2860.98
Total Token throughput (tok/s):          6243.37                             Total Token throughput (tok/s):          5715.91
---------------Time to First Token----------------                           ---------------Time to First Token----------------
Mean TTFT (ms):                          1367.30                             Mean TTFT (ms):                          1413.01
Median TTFT (ms):                        1007.55                             Median TTFT (ms):                        1050.07
P99 TTFT (ms):                           5422.35                             P99 TTFT (ms):                           5486.11
-----Time per Output Token (excl. 1st token)------                           -----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          39.61                               Mean TPOT (ms):                          43.35
Median TPOT (ms):                        39.88                               Median TPOT (ms):                        43.60
P99 TPOT (ms):                           40.60                               P99 TPOT (ms):                           44.49
---------------Inter-token Latency----------------                           ---------------Inter-token Latency----------------
Mean ITL (ms):                           39.61                               Mean ITL (ms):                           43.35
Median ITL (ms):                         36.09                               Median ITL (ms):                         39.67
P99 ITL (ms):                            326.60                              P99 ITL (ms):                            317.85
==================================================                           ==================================================
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|    |Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|    |-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9181|±  |0.0076|    |gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9234|±  |0.0073|
|     |       |strict-match    |     5|exact_match|↑  |0.8999|±  |0.0083|    |     |       |strict-match    |     5|exact_match|↑  |0.8999|±  |0.0083|

PR

--kv-cache-dtype = auto                                                      --kv-cache-dtype = fp8
============ Serving Benchmark Result ============                           ============ Serving Benchmark Result ============
Successful requests:                     640                                 Successful requests:                     640
Maximum request concurrency:             128                                 Maximum request concurrency:             128
Benchmark duration (s):                  213.18                              Benchmark duration (s):                  231.03
Total input tokens:                      653975                              Total input tokens:                      653975
Total generated tokens:                  655360                              Total generated tokens:                  655360
Request throughput (req/s):              3.00                                Request throughput (req/s):              2.77
Output token throughput (tok/s):         3074.24                             Output token throughput (tok/s):         2836.64
Total Token throughput (tok/s):          6141.98                             Total Token throughput (tok/s):          5667.28
---------------Time to First Token----------------                           ---------------Time to First Token----------------
Mean TTFT (ms):                          1375.34                             Mean TTFT (ms):                          1439.79
Median TTFT (ms):                        1019.52                             Median TTFT (ms):                        1061.45
P99 TTFT (ms):                           5287.76                             P99 TTFT (ms):                           5711.34
-----Time per Output Token (excl. 1st token)------                           -----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          40.28                               Mean TPOT (ms):                          43.71
Median TPOT (ms):                        40.52                               Median TPOT (ms):                        43.82
P99 TPOT (ms):                           41.37                               P99 TPOT (ms):                           44.86
---------------Inter-token Latency----------------                           ---------------Inter-token Latency----------------
Mean ITL (ms):                           40.28                               Mean ITL (ms):                           43.71
Median ITL (ms):                         36.71                               Median ITL (ms):                         40.04
P99 ITL (ms):                            309.98                              P99 ITL (ms):                            342.09
==================================================                           ==================================================
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|    |Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|    |-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9181|±  |0.0076|    |gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9234|±  |0.0073|
|     |       |strict-match    |     5|exact_match|↑  |0.9007|±  |0.0082|    |     |       |strict-match    |     5|exact_match|↑  |0.9007|±  |0.0082|

--kv-cache-dtype = fp8, enable_attn_fusion
============ Serving Benchmark Result ============
Successful requests:                     640
Maximum request concurrency:             128
Benchmark duration (s):                  199.56
Total input tokens:                      653975
Total generated tokens:                  655360
Request throughput (req/s):              3.21
Output token throughput (tok/s):         3284.02
Total Token throughput (tok/s):          6561.10
---------------Time to First Token----------------
Mean TTFT (ms):                          1347.81
Median TTFT (ms):                        999.97
P99 TTFT (ms):                           5192.63
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          37.65
Median TPOT (ms):                        37.89
P99 TPOT (ms):                           38.71
---------------Inter-token Latency----------------
Mean ITL (ms):                           37.65
Median ITL (ms):                         34.13
P99 ITL (ms):                            302.65
==================================================
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9128|±  |0.0078|
|     |       |strict-match    |     5|exact_match|↑  |0.8923|±  |0.0085|

(Optional) Documentation Update

[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.

🚀

[gemini-code-assist]

Code Review

This pull request adds support for the Flashinfer TRT-LLM FP8-query/output attention kernel. The changes span across benchmarks, tests, and core attention backend logic. The main changes involve updating the Flashinfer API calls to support an out parameter for in-place operations, and adding logic to handle FP8 quantization for queries and outputs. The PR also includes a significant refactoring of CUDA graph support for attention backends.

My review identifies two main issues. First, a critical bug in vllm/attention/layer.py where the query_scale for FP8 quantization is not being correctly propagated to the attention implementation. Second, a high-severity issue in vllm/v1/attention/backends/flashinfer.py where the usage of the TRT-LLM attention kernel is hardcoded, which limits flexibility.

[mergify]

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

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

[nvpohanh]

not related to this PR, but I think has_nvidia_artifactory() can be removed because FlashInfer now supports downloading all cubins at once.

[nvpohanh]

can do this in another PR

[elvischenv]

Got it.

[ProExpertProg]

This PR is looking really good! Thanks for all your hard work

[mergify]

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

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

[Sekri0]

May I ask why TTFT increases when qkv fp8 is enabled？I assume fp8 Tensor Core should be used for accelerating qk and pv matmul when qkv is quantized to fp8.

[nvpohanh]

May I ask why TTFT increases when qkv fp8 is enabled？I assume fp8 Tensor Core should be used for accelerating qk and pv matmul when qkv is quantized to fp8.

Using FP8 kv-cache introduces an additional FP8-Quant kernel for the Query tensor, so the performance may have a small drop if the attention speed up is too small. Ideally, that Quant should be fused with RoPE and that work is tracked in #24678