Views
No views yet
BV=32, num_warps=4) instead of launching NV=4 CTAs
per head. Tuned on NVIDIA GB10 for
AxionML/Qwen3.5-0.8B-NVFP4.kernel repo, not a model checkpoint.1from kernels import get_kernel
2
3k = get_kernel("<this-repo>", trust_remote_code=True)
4# k.fused_recurrent_gated_delta_rule_packed_decode(...)trust_remote_code=True is required unless the publisher is on the
kernels trusted list. Replace <this-repo> with this repository id.1hf download <this-repo> --local-dir ./gdn-packed-decode
2python3 ./gdn-packed-decode/apply_kernel.py apply
3# restart vLLM so EngineCore reimports, then:
4# vllm serve AxionML/Qwen3.5-0.8B-NVFP4| Kernel | Acc | Median decode tok/s |
|---|---|---|
| Upstream vLLM | 5/10 | 205.45 |
| This kernel (run 1) | 5/10 | 219.90 |
| This kernel (run 2) | 5/10 | 219.89 |
fused_recurrent_gated_delta_rule_packed_decode(out, initial_state).