[PERF] Add conv1d metadata to GDN attn

[vadiklyutiy]

Purpose

The on of the main gap in performance for Qwen3-next model is low utilisation of GPU in GDN attn.
Low utilization caused by several GPU<->host memory transfers.
These transfers caused by causal_conv1d_fn.
Qwen3-next haven't supported this specific convolution metadata. The purpose of the metadata is avoid these memory transfer.

Add conv metadata (similar to mamba2).

+ corrected tensor.tensor -> tensor.Tensor types annotation.

Test Result

H200, tp=4

vllm bench serve --backend vllm --model Qwen/Qwen3-Next-80B-A3B-Instruct   --endpoint /v1/completions --dataset-name random --random-input 32768   --random-output 1 --max-concurrency 8 --num-prompt 8

Before

============ Serving Benchmark Result ============
Successful requests:                     8
Maximum request concurrency:             8
Benchmark duration (s):                  4.56
Total input tokens:                      262144
Total generated tokens:                  8
Request throughput (req/s):              1.76
Output token throughput (tok/s):         1.76
Total Token throughput (tok/s):          57510.95
---------------Time to First Token----------------
Mean TTFT (ms):                          2600.46
Median TTFT (ms):                        2601.40
P99 TTFT (ms):                           4511.95
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          0.00
Median TPOT (ms):                        0.00
P99 TPOT (ms):                           0.00
---------------Inter-token Latency----------------
Mean ITL (ms):                           0.00
Median ITL (ms):                         0.00
P99 ITL (ms):                            0.00
==================================================

After

============ Serving Benchmark Result ============
Successful requests:                     8
Maximum request concurrency:             8
Benchmark duration (s):                  3.61
Total input tokens:                      262144
Total generated tokens:                  8
Request throughput (req/s):              2.22
Output token throughput (tok/s):         2.22
Total Token throughput (tok/s):          72623.02
---------------Time to First Token----------------
Mean TTFT (ms):                          2067.73
Median TTFT (ms):                        2068.62
P99 TTFT (ms):                           3570.84
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          0.00
Median TPOT (ms):                        0.00
P99 TPOT (ms):                           0.00
---------------Inter-token Latency----------------
Mean ITL (ms):                           0.00
Median ITL (ms):                         0.00
P99 ITL (ms):                            0.00
==================================================

Speedup is 26%

[gemini-code-assist]

Code Review

This pull request introduces an optimization for the Qwen3-next model by adding mamba2-style convolution metadata to the GDN attention mechanism. This change aims to reduce GPU-host memory transfers in causal_conv1d_fn, leading to a significant performance improvement as demonstrated by the benchmark results. The implementation correctly extends GDNAttentionMetadata and integrates the metadata preparation and usage within the model's forward pass. The changes are well-targeted and effective. I have one minor suggestion to improve type hint correctness.

[vadiklyutiy]

Could someone please run the CI?