Repository Issue Activity (beta)

#15989 [Performance]: Compact pseudo-KV/topological sparse attention benchmark evidence and backend gap

#7085 [Bug]: BertForSequenceClassification outputs diverge from HuggingFace transformers when input length > 64

#6706 [BUG] Discrepancy in Gemma2 Prefill Hidden States During Inflight Batching (gemm_plugin bf16 numerical error)

#3277 Qwen2.5-72B-Instruct int4_wo can't use `--gather_generation_logits` on A100

#12877 [Bug]: Garbage output with modelopt FP8 checkpoint in 1.2.0: use_fp8_context_fmha forced True, regression from 1.1.0

#15537 [Bug]: XQA multi_block_mode crashes with CUDA_ERROR_INVALID_VALUE under concurrent inference (v1.0.0)

#14588 [Feature]: [AutoDeploy] Perf analysis & optimize DeekSeek R1 (lower concurrency)

#15941 [Bug]: DGX Spark playbook supported models fail in trtllm-serve: Nemotron-3 Nano Omni and GPT-OSS rejected as unsupported due to MODEL_MAP registry mismatch

#15793 [Bug]: Redundant KV host cache tier on DGX Spark unified-memory systems

#15896 Community benchmark tip

#15885 [Feature]: Support multiple input types

#13957 [Feature]: PyExecutor multi-stage pipelined scheduler (matching vLLM v1's step_with_batch_queue)

#10481 [Feature]: Adding support for Fuzzy Speculative Decoding (ACL Findings 2025)

#15853 [Bug]: B12x NVFP4 MoE JIT fails on SM120/SM121 with CUDA 12 CuTe DSL payload

#15792 SM120 skip-softmax FMHA bridge is missed by support probe

#15791 SM120 skip-softmax FMHA support probe misses bridge path

#14784 [Feature Request] Make sharding IR the AutoDeploy default with auto-detect fallback to legacy

#15747 SuperCompress: Free open-source LLM prompt compression - cut token costs by ~65%

#15735 [Feature]: General cross-instance KV-cache transfer/staging (kv_transfer)

#15706 [Bug]: Bad Words Missing Space-Prefixed Token Variants for BPE Tokenizers (e.g., GPT-2, Qwen)

#15715 [Bug]: Clamp very small non-zero temperature values to avoid numerical instability

#9628 [Bug]: trtllm-serve periodically hanging with "No free slots"

#12411 [SM120] NVFP4 static _input_prepare causes illegal memory access on RTX 5080

#15367 [Feature Request] Context-parallel (Helix) sharding for MLA in AutoDeploy

#15684 [DeepSeek-V4] Overlap scheduler + chunked prefill deadlocks in sparse-MLA ctx metadata (device→host sync at _compute_ctx_compressed_position_ids)