Repository Issue Activity (beta)

#5107 Split-KV decode rounds partial outputs to the output dtype, in a workspace already sized for fp32

#2861 mm_fp4 trtllm backend leaks padding scales into real rows (use_8x4_sf_layout=True)

#3864 [Bug] trtllm-gen spec-as-decode (q_len_per_req=8) with BF16 query + FP8 KV cache silently mis-computes sliding-window attention on SM100

#4254 [CAKE] Long-term CAKE-generated Kernel Progress Tracker

#5053 trtllm-gen NVFP4 MoE: the SiTU (siTuGlu) GEMM1 epilogue writes no output block scale factors for some configurations

#2511 [Bug] Cuda Graph Issues with TRTLLM-GEN Backend with BatchDecodeWithPagedKVCacheWrapper

#5058 top_k_page_table_transform(dsa_graph_safe=True) fails with cudaErrorNotSupported on SM120 (RTX 6000D), graph_safe=False works even under CUDA graph capture

#3617 [feat] support fp8 per-token-head kv cache with inline scale

#5162 [Bug] Nightly Release: BF16 rank-major session CPU tests assume a source checkout

#5157 H100 CI broken on main since #5131: cudnn-frontend 1.28 vs backend 9.24 rejects non-ragged Stats layout (134 nodes, blocks all PRs)

#5156 [Feature]: CuTe DSL W4A8 MXFP4 SiTU MoE for Kimi K3

#5095 [Bug] SM120 sparse-MLA prefill (DSv4 dual-cache) returns corrupted output for >64 query tokens; masking lanes beyond topk_length to -1 fixes it

#4973 [Bug] DSV4 Vision: illegal memory access on SM120 (2x RTX PRO 6000, TP=2) with long text prompts

#3620 [Bug] BatchPrefillWithPagedKVCache fails on SM75 (Turing / Tesla T4) with CUDA "invalid argument"

#4936 [KDA] Unify the public `recurrent_kda` API, and settle the unreleased KDA surface before FI v0.7

#3800 [Bug] cudnn_batch_prefill_with_kv_cache: one-token Q batch stride silently corrupts every batch after the first when batch_offsets_q is not passed

#3511 Trtllm-gen kernels not found: `headDim=512`, `tileSizeQ=128`

#3971 SM103 serving hang caused by TRTLLM_GEN_BMM artifact regeneration in #3708 (batched_gemm-dd6d23e-721ae60, 0.6.14): FP4 batched-GEMM clusters stuck in mbarrier phase wait

#4930 [Feature]: On SM107, the MXFP8 Quantization Only does not support FP32

#4999 [Bug] test_sparse_mla_sm120_cpb_model assert 0.0042064361572265625 <= (1.25 * 0.0028502678871154784)

#4957 [Bug] test_multistream_overlap.py CUDA error: unspecified launch failure Search for `cudaErrorLaunchFailure'

#5151 [Feature]: MNNVL AllReduce does not support TriggerCompletionAtEnd

#5141 [Feature]: Expose NVFP4 4over6 configuration through the public API

#4367 [Bug][v0.6.17] tests.trace.test_template_registry::test_registration_module_inventory_is_complete AssertionError: assert {'flashinfer....cat_ops', ...} == {'flashinfer....cat_ops', ...}

#5145 [Bug] JIT sampler built with CUDA 13 is reused by CUDA 12.9 Ray workers and fails at runtime