[NVIDIA] Support Flashinfer TRTLLM FP8-q/kv NVFP4-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

This PR based on the previous attn + FP8-quant fusion #21716, adding another attn + NVFP4-quant fusion for supporting TRTLLM-gen attn kernel.

Test Plan && Test Result

Functional:

• NVFP4 TRTLLM Prefill/Decode kernel unit test: tests/kernels/attention/test_flashinfer_trtllm_attention.py

======= 44 passed, 4 skipped, 4 warnings in 64.13s (0:01:04) ======

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

======== 12 passed, 4 warnings in 68.53s (0:01:08) ========

E2E Performance: nvidia/Llama-4-Scout-17B-16E-Instruct-FP4

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):                  143.36                              Benchmark duration (s):                  164.40
Total input tokens:                      653975                              Total input tokens:                      653975
Total generated tokens:                  655360                              Total generated tokens:                  655360
Request throughput (req/s):              4.46                                Request throughput (req/s):              3.89
Output token throughput (tok/s):         4571.39                             Output token throughput (tok/s):         3986.50
Total Token throughput (tok/s):          9133.12                             Total Token throughput (tok/s):          7964.57
---------------Time to First Token----------------                           ---------------Time to First Token----------------
Mean TTFT (ms):                          677.91                              Mean TTFT (ms):                          751.94
Median TTFT (ms):                        475.19                              Median TTFT (ms):                        521.30
P99 TTFT (ms):                           2509.39                             P99 TTFT (ms):                           2859.41
-----Time per Output Token (excl. 1st token)------                           -----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          27.32                               Mean TPOT (ms):                          31.36
Median TPOT (ms):                        27.32                               Median TPOT (ms):                        31.40
P99 TPOT (ms):                           28.20                               P99 TPOT (ms):                           32.38
---------------Inter-token Latency----------------                           ---------------Inter-token Latency----------------
Mean ITL (ms):                           27.32                               Mean ITL (ms):                           31.36
Median ITL (ms):                         25.66                               Median ITL (ms):                         29.73
P99 ITL (ms):                            146.45                              P99 ITL (ms):                            168.08
==================================================                           ==================================================
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|    |Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|    |-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9037|±  |0.0081|    |gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.8969|±  |0.0084|
|     |       |strict-match    |     5|exact_match|↑  |0.8908|±  |0.0086|    |     |       |strict-match    |     5|exact_match|↑  |0.8817|±  |0.0089|

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):                  143.82                              Benchmark duration (s):                  164.89
Total input tokens:                      653975                              Total input tokens:                      653975
Total generated tokens:                  655360                              Total generated tokens:                  655360
Request throughput (req/s):              4.45                                Request throughput (req/s):              3.88
Output token throughput (tok/s):         4556.72                             Output token throughput (tok/s):         3974.57
Total Token throughput (tok/s):          9103.80                             Total Token throughput (tok/s):          7940.75
---------------Time to First Token----------------                           ---------------Time to First Token----------------
Mean TTFT (ms):                          694.81                              Mean TTFT (ms):                          750.19
Median TTFT (ms):                        476.66                              Median TTFT (ms):                        522.09
P99 TTFT (ms):                           2595.24                             P99 TTFT (ms):                           2847.96
-----Time per Output Token (excl. 1st token)------                           -----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          27.40                               Mean TPOT (ms):                          31.46
Median TPOT (ms):                        27.38                               Median TPOT (ms):                        31.51
P99 TPOT (ms):                           28.52                               P99 TPOT (ms):                           32.46
---------------Inter-token Latency----------------                           ---------------Inter-token Latency----------------
Mean ITL (ms):                           27.40                               Mean ITL (ms):                           31.46
Median ITL (ms):                         25.68                               Median ITL (ms):                         29.84
P99 ITL (ms):                            153.03                              P99 ITL (ms):                            171.23
==================================================                           ==================================================
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|    |Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|    |-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9075|±  |0.0080|    |gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.8984|±  |0.0083|
|     |       |strict-match    |     5|exact_match|↑  |0.8954|±  |0.0084|    |     |       |strict-match    |     5|exact_match|↑  |0.8832|±  |0.0088|

--kv-cache-dtype = fp8, enable_attn_fusion
============ Serving Benchmark Result ============
Successful requests:                     640
Maximum request concurrency:             128
Benchmark duration (s):                  130.05
Total input tokens:                      653975
Total generated tokens:                  655360
Request throughput (req/s):              4.92
Output token throughput (tok/s):         5039.38
Total Token throughput (tok/s):          10068.10
---------------Time to First Token----------------
Mean TTFT (ms):                          658.54
Median TTFT (ms):                        456.26
P99 TTFT (ms):                           2400.84
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          24.74
Median TPOT (ms):                        24.74
P99 TPOT (ms):                           25.72
---------------Inter-token Latency----------------
Mean ITL (ms):                           24.74
Median ITL (ms):                         23.13
P99 ITL (ms):                            140.43
==================================================
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.9014|±  |0.0082|
|     |       |strict-match    |     5|exact_match|↑  |0.8863|±  |0.0087|

(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 introduces support for the Flashinfer TRTLLM FP8-q/kv NVFP4-out Attention Kernel, which is a significant enhancement for NVIDIA Blackwell GPUs. The changes are extensive, touching benchmarks, tests, and core attention and compilation logic. The implementation of the new QuantAttentionQuantPattern fusion pass is particularly noteworthy, enabling more advanced quantization-aware optimizations. The code appears well-structured, and the necessary updates to tests and benchmarks have been included to validate the new functionality. The approach to handle framework limitations, such as using a cache for fusion status and offline scale extraction, is pragmatic. Overall, this is a high-quality contribution that should bring performance improvements for supported hardware.

[nvpohanh]

This will be blocked by FlashInfer fix: flashinfer-ai/flashinfer#1460

[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

[mgoin]

Model access should be resolved now, thanks!