[Feature] Faster Custom Paged Attention kernels#385

Draft

tjtanaa wants to merge 18 commits intoROCm:llama_fp8_12062024from EmbeddedLLM:pg_attn_to_llama_fp8

+825-338

Commits on Nov 13, 2024

corrected types for strides in triton FA (ROCm#274 ) (ROCm#276 )
maleksan85
authored

Commits on Dec 4, 2024

Update test-template.j2 (ROCm#283 )
dhonnappa-amd
authored

Commits on Dec 9, 2024

Fix max_seqlens_q/k initialization for Navi GPUs (ROCm#310 )
hyoon1
authored
Merge remote-tracking branch 'origin/develop'
gshtras
committed
Setting the value for the scpecilative decoding worker class on rocm platform (ROCm#313 )
gshtras
authored

Commits on Dec 24, 2024

fp8 cpa
sanyalington
committed

Commits on Dec 26, 2024

load nt optional for kv
sanyalington
committed

Commits on Jan 14, 2025

enable alibi; fix gfx90a compile
sanyalington
committed

Commits on Jan 15, 2025

Commits on Jan 16, 2025

Commits on Jan 17, 2025

Commits on Jan 20, 2025

fix unit test for rocm custom attention kernel
vllmellm
committed

Commits on Jan 22, 2025

fix benchmark paged attention
vllmellm
committed

Commits on Jan 23, 2025

[Bugfix]: fix v1/v2 paged attention kernel unit test.
vllmellm
committed