All issues

Search results

[Bug][Spec Decode] probabilistic draft sampling corrupts output above max_num_seqs=2 at temperature > 0

Status: Open.

#656 In 1CatAI/1Cat-vLLM;

· Leonccaa opened on Sep 19, 2026

[Bug]: Engine startup deadlocks on V100 (SM70, TP=2): custom_all_reduce kernel spins forever (workaround: --disable-custom-all-reduce)

bugSomething isn't workingSomething isn't working

Status: Open.

#652 In 1CatAI/1Cat-vLLM;

· Zhang-Man opened on Sep 18, 2026

[Bug][SM70] QUASAR-QAT/Qwen3.8-27B-QUASAR-NVFP4 never closes <think> (restart loop, then EOS mid-think) - post-trained NVFP4 and AWQ close fine on the identical path

Status: Open.

#649 In 1CatAI/1Cat-vLLM;

· jsantos98 opened on Sep 17, 2026

[Bug][SM70] E4M3 KV: XQA page4 route rejects the fp16 tmp_out that qsa.py allocates (kernel expects fp32 since 8a10215d6)

Status: Open.

#648 In 1CatAI/1Cat-vLLM;

· areslp opened on Sep 17, 2026

[New Model]: Qwen3.8-27B-QUASAR-NVFP4 changed

Status: Open.

#644 In 1CatAI/1Cat-vLLM;

· uraraca-vituber opened on Sep 16, 2026

[Bug] NVFP4 MoE fails with intermediate_size=640: size_n=160 not divisible by tile_n_size=64

Status: Open.

#642 In 1CatAI/1Cat-vLLM;

· johnbluer opened on Sep 16, 2026

[Model] 512-experts NVFP4 MoE generates garbage at deep layers (L44/L46 decode outputs 400+, prefill hidden grows unbounded)

Status: Open.

#643 In 1CatAI/1Cat-vLLM;

· johnbluer opened on Sep 16, 2026

[Bug][SM70]: TurboMind GEMM tuner illegal memory access with checkpoint FP8 MTP experts and --max-num-seqs 1

Status: Open.

#641 In 1CatAI/1Cat-vLLM;

· Peuqui opened on Sep 15, 2026

[Backport]: Port vLLM main load_start_gpu_block_idx (stops 21:18 EngineDead)

Status: Open.

#633 In 1CatAI/1Cat-vLLM;

· DGUINAN opened on Sep 14, 2026

[Bug]: Native KV offload advances next_stored_block_idx for every KV group when no keys were stored

Status: Open.

#632 In 1CatAI/1Cat-vLLM;

· DGUINAN opened on Sep 14, 2026

[Bug]: GLM-5.3-Flash NVFP4 on V100 (SM70): --mm-processor-kwargs max_pixels ignored during dummy multimodal profiling, causing "max_pixels=47040000 is too small"

bugSomething isn't workingSomething isn't working

Status: Open.

#631 In 1CatAI/1Cat-vLLM;

· syvince opened on Sep 14, 2026

[Feature]: Arch Linux

Status: Open.

#627 In 1CatAI/1Cat-vLLM;

· crystalsighting opened on Sep 14, 2026