GitHub · Where software is built

E5CA GitHub · Where software is built

Roadmap (2025 Q4)
#1770 · sricketts opened on Sep 25, 2025
11

Labels Milestones New issue

NVFP4 CPU RAM leak caused by dynamic creation of TuningConfig objects

#2139

· juju812 opened

on Nov 24, 2025

[Documentation fix] flashinfer.mla API documentation on flashinfer blog

#2124

· kmrao-nv opened

on Nov 20, 2025

FlashInfer MXFP4 moe crashes with GPT-OSS-120B

#2122

· xhejtman opened

on Nov 20, 2025

Question about default filter_apply_order="top_k_first" in top-k + top-p sampling

#2120

· Xseventh opened

on Nov 20, 2025

[Feature Request] <code>trtllm_batch_decode_with_kv_cache_mla</code> Batch Invariant support

#2107

· yewentao256 opened

on Nov 18, 2025

[Feature Request] Support all head dims in BatchQKApplyRotaryPosIdsCosSinCache

#2104

· merrymercy opened

on Nov 18, 2025

[feature request] add optimal gb200 moe comm kernels

#2094

· nvrohanv opened

on Nov 14, 2025

[BUG] Flashinfer with jit-cache and cubin still attempts to jit

#2093

· bbartels opened

on Nov 14, 2025

【Bug】Test prefill attention kv_dtype=fp8 error

#2085

· heliqi opened

on Nov 13, 2025

MoE autotune print a lot failed kernel on SM120

#2077

· weireweire opened

on Nov 11, 2025

[Feature Request] FP8 block quant support for AR+RMSNOrm , AR+RMSNorm+Quant.

feature request

#2073

· kmrao-nv opened

on Nov 10, 2025

[Installation failed] Is <code>nvidia-cudnn-frontend>=1.13.0</code> required?

#2068

· ziminz19 opened

on Nov 10, 2025

0