Repository navigation

[Kernel] Add FlashInfer GDN and KDA prefill backend selection - #61250

Open

Anerudhan wants to merge 15 commits into

vllm-project:mainfrom

Anerudhan:flashinfer/linear-attention-cleanup

Open

[Kernel] Add FlashInfer GDN and KDA prefill backend selection#61250
Anerudhan wants to merge 15 commits into
vllm-project:mainfrom
Anerudhan:flashinfer/linear-attention-cleanup

Commits on Oct 12, 2026
- [Kernel] Add FlashInfer auto/cuDNN GDN and KDA prefill
  Anerudhan
  committed
  8abfe15
- docs: Record current QwenNext GB200 serving results
  
  Anerudhan
  and
  codex
  committed
  bad642f
- docs: Record current GLM GB300 prefill failure
  Anerudhan
  committed
  5762add
- docs: Record GLM current-serving retry gate failure
  Anerudhan
  committed
  fe9ff3e
- docs: Publish current Kimi and Qwen3.5 serving results
  Anerudhan
  committed
  a57b382
- Revert AMD Kimi cache dtype override
  
  Anerudhan
  and
  codex
  committed
  3e74150
- Record GLM FlashInfer runtime failure and cuDNN continuation
  
  Anerudhan
  and
  codex
  committed
  eef505d
- docs: report GB300 Qwen and standalone GLM cuDNN measurements
  
  Anerudhan
  and
  codex
  committed
  341a1f2
- Use cuDNN default additive KDA normalization
  Anerudhan
  committed
  16948f8
- docs: report FE1454 GLM GB200 TTFT and KDA CPU/kernel measurements
  Anerudhan
  committed
  4e3701f
- docs: complete FE1454 Kimi GB200 TTFT and KDA CPU/kernel results
  Anerudhan
  committed
  682065e
- docs: add latest GB200 Qwen3-Next full-serving results
  Anerudhan
  committed
  2296604
- docs: publish latest GB200 Kimi full serving results
  Anerudhan
  committed
  932292a
- Remove linear attention benchmark result artifacts
  
  Anerudhan
  and
  codex
  committed
  ef77587
- Expose all FlashInfer KDA prefill backends
  
  Anerudhan
  and
  codex
  committed
  6886d79