Search results

[Bug] trtllm_bf16_moe: missing SM103 cubin in fused_moe_trtllm_sm100.so causes cudaErrorNoKernelImageForDevice on B300 / Blackwell Ultra (compute capability 10.3)

needs-triage

Status: Open.

#3466 In flashinfer-ai/flashinfer;

· janbernloehr opened on May 29, 2026

install instructions future improvements suggestions

needs-triage

Status: Open.

#3462 In flashinfer-ai/flashinfer;

· aleozlx opened on May 29, 2026

[Feature request] TRT-LLM-Gen fused MoE cubins for GeGLU activation (mixed-precision formats)

needs-triage

Status: Open.

#3454 In flashinfer-ai/flashinfer;

· murphymatt opened on May 29, 2026

Support fp4 x bf16 -> bf16 mixed precision GEMM for W4A16 inference

feature requestNew feature or requestNew feature or request

needs-triage

op: gemm

Status: Open.

#3438 In flashinfer-ai/flashinfer;

· bkryu opened on May 28, 2026

Autotuning is often silently a no-op for cuBLAS bmm_fp8 and cuDNN bmm_fp8/mm_fp4 in common configurations

op: gemm

Status: Open.

#3436 In flashinfer-ai/flashinfer;

· bkryu opened on May 28, 2026

[trtllm-gen] Expose a separate counter workspace for multiCtasKvCounterPtr

needs-triage

Status: Open.

#3433 In flashinfer-ai/flashinfer;

· yihwang-nv opened on May 28, 2026

[GLM5] support sparse absorbed-MLA for prefill stage

needs-triage

Status: Open.

#3429 In flashinfer-ai/flashinfer;

· forrestl111 opened on May 28, 2026

FlashInfer TRTLLM BF16 MoE IMA on B200 with SGLang RL rollout, correlated with t128x8x128 + schPd2x1x2x3 BMM kernel at effbs=1024

op: moe

Status: Open.

#3427 In flashinfer-ai/flashinfer;

· kaixih opened on May 28, 2026

[Bug]Regression: backend="cute-dsl" MLA decode via API rejects page_size=128

needs-triage

Status: Open.

#3425 In flashinfer-ai/flashinfer;

· cindyzxq opened on May 28, 2026

0.6.12rc1 accuracy regression TRTLLM FP4 MoE involving bias

v0.6.13

Status: Open.

#3423 In flashinfer-ai/flashinfer;

· aleozlx opened on May 27, 2026

[Perf] qo_len=2 (MTP / spec-decode draft=1) dispatches to prefill kernel instead of decode kernel, 10× slower than FA on H20

needs-triage

Status: Open.

#3420 In flashinfer-ai/flashinfer;

· JackWang2120 opened on May 27, 2026

CuteDslMoEWrapper autotuner picks tactics ~10-25% slower than TRT-LLM equivalent at EP=16/32, small-mid N (B200)

op: moe

Status: Open.

#3409 In flashinfer-ai/flashinfer;

· leejnau opened on May 26, 2026