COMMITS
May 21, 2026
D
feat: Add support for LoRa delta in MOE mxint4 x bf16, MXFP8 & BF16 to trtllm backend (#3153)
Daniel Stokes committed
L
fix(cute_dsl): avoid MoE wrapper runner reference cycle (#3340)
Lee Nau committed
K
Reject EP configurations in b12x MoE with a clear error (#3302)
Ka-Hyun Nam committed
P
Add DeepSeek V4 sparse MLA TRTLLM-GEN kernels (#3269)
Perkz Zheng committed
Z
feat: Support 4over6 nvfp4 for quantizer and fused MoE (#3264)
Ziang Li committed
A
chore: Add list_api script (#3341)
Alex Yang committed
A
Add AGENTS.md shim (#3342)
Alex Yang committed
May 20, 2026
B
feat: Add 8x4 swizzle layout support to MXFP4 and MXFP8 CuTe-DSL kernels (#3357)
Brian K. Ryu committed
E
fix(trace): repair TGV and XQA MLA reference tests (#3365)
eigen committed
May 19, 2026
D
Update trtllm FMHA cubins (#3317)
Duncan Moss committed
A
Ameyn/gdn bf16 dispatcher and 4d pool (#3268)
ameynaik-hub committed
I
checkpointing_ssu kernel: fused replay + conditional state-write for Mamba2 (#3324)
Igor Shovkun committed
J
[feat] Add gemma RMS AR fusion (#3322)
Jiahan Chang (Cyrus) committed
May 18, 2026
A
Ep api design - Build Infra dependencies (#3315)
Anerudhan Gopal committed
L
test: enable bmm_mxfp8 cutlass backend coverage on SM12x (#3183)
leonardHONG committed
L
Fix/3170 dense blockscaled sm12x (#3180)
leonardHONG committed
May 17, 2026
L
E
May 16, 2026
K
M
feat: Reland support lse in trtllm paged attn kernels (#3116)
Matt Murphy committed
A
feat: Expose unpacked topk weights for routed moe (fp4) (#2425)
Alex Yang committed
May 15, 2026
Y
Use cudnn 9.23 new API to query workspace with override shape (#3291)
yanqinz2 committed
E
L
fix typo llama routing issue in trtllm-gen moe (#3303)
Lain committed
May 14, 2026
J
fix(fmha_v2): fix FP8 V-scratch pipeline and varlen scheduler on SM90 (#3276)
Jimmy Zhou committed
B
feat: enable glm5 router gemm (#3185)
Brayden Zhong committed
K
fix(gdn): allow importing gdn_decode without a CUDA device (#3293)
Ka-Hyun Nam committed
N
feat: Add CuTe DSL grouped-gemm + combine fusion support (#2944)
Nicolas Castet committed
D
Add to code owners (#3326)
Dhiraj Reddy (cuDNN/FlashInfer) committed