Issues

• [Bug][KV Offload][Qwen3.8] OffloadingConnector crashes EngineCore on pending transfer_jobs — upstream fix #46231

  bug
  1CatAI/1Cat-vLLM#1249 · Filter by author gaetanbahl opened on Oct 22, 2026

• [Feature]: 能支持SystemOne模型吗？

  1CatAI/1Cat-vLLM#1239 · Filter by author hpx502766238 opened on Oct 20, 2026

• [Bug][Hybrid KV][LMCache] EngineCore crashes on failed KV loads / async completion (ValueError, AssertionError)

  bug
  1CatAI/1Cat-vLLM#1238 · Filter by author gaetanbahl opened on Oct 19, 2026

• [Bug]: MTP + align Mamba: a prefix restored from the CPU offload tier is not reusable on the GPU, so every later request with that prefix reloads it

  1CatAI/1Cat-vLLM#1233 · Filter by author Leonccaa opened on Oct 18, 2026

• [Bug][SM70][Qwen3.8][TP8] 双 NV2 岛 8 卡启动在 NCCL 初始化后卡住

  1CatAI/1Cat-vLLM#1224 · Filter by author buyi1net opened on Oct 16, 2026

• [Perf][Qwen3.8-27B] Mixed prefill+decode: what decode rate survives a 96K–192K prefill with 6–8 concurrent sessions (4×V100-SXM2-32G, TP4)?

  1CatAI/1Cat-vLLM#1223 · Filter by author superniker opened on Oct 16, 2026

• [Bug] TP2 on 2×V100-32GB: deterministic CUDA OOM during CUDA-graph capture (main post-1.5.1 build)

  1CatAI/1Cat-vLLM#1220 · Filter by author johnbluer opened on Oct 15, 2026

• [Community benchmark] 2× V100 PCIe 32GB TP2: QUASAR NVFP4 DFlash2 q7, 4K–256K, replay/shared cache and C1/C2/C4

  1CatAI/1Cat-vLLM#1202 · Filter by author AiRC-ai opened on Oct 13, 2026

• SM70 (V100) TP4: --async-scheduling deadlocks with DFlash2 ngram_assist (GPU spin, no CUDA error)

  1CatAI/1Cat-vLLM#1173 · Filter by author ChainZeaxion opened on Oct 11, 2026

• [Feature]: Is it possible to support Qwen3.8-Flash-Next on 8× Tesla V100 16GB?

  1CatAI/1Cat-vLLM#1158 · Filter by author buggg-hfc opened on Oct 10, 2026

• [Performance][SM70] Flash-V100 D=128 prefill is 8-12% faster on long prompts with two small kernel changes (measured, bit-identical, reference patch)

  1CatAI/1Cat-vLLM#1132 · Filter by author christopherrobertbrooks-tech opened on Oct 9, 2026

• [Experiment] VLLM_SM70_AWQ_MLP_DOWN_TILE_OVERLAP + _AR: no perf change on 2x V100-PCIE (TP2, AWQ)

  1CatAI/1Cat-vLLM#1106 · Filter by author johnbluer opened on Oct 8, 2026

• [Feature][SM70][Qwen3.8]: Run the FP16 GEMV/HC/GDN-input lanes when Flash-Next is served with vision

  1CatAI/1Cat-vLLM#1100 · Filter by author squibloads opened on Oct 8, 2026
  1

• [Bug][Qwen4Exp]: pinned-host PLE table is page-locked at the next power of two (64 GiB pinned for a 47.7 GiB table at TP4)

  1CatAI/1Cat-vLLM#1099 · Filter by author squibloads opened on Oct 8, 2026
  2

• [Bug][FlashQLA][SM70]: NVRTC >= 12.8 writes its precompiled header into the working directory; unprivileged engines die in the GDN warmup

  1CatAI/1Cat-vLLM#1098 · Filter by author squibloads opened on Oct 8, 2026

• [Benchmark] Qwen3.8-27B NVFP4 on 1x V100-PCIE-32GB (TP1) vs llama.cpp: 1Cat ahead on concurrency and prefill, behind on single-stream decode

  1CatAI/1Cat-vLLM#1087 · Filter by author christopherrobertbrooks-tech opened on Oct 8, 2026

• [Benchmark] Qwen3.8-27B-AWQ-MTP on 2x V100-PCIE (TP2): MAX_SEQS scaling & single/multi-user throughput

  1CatAI/1Cat-vLLM#1043 · Filter by author johnbluer opened on Oct 8, 2026

• [Perf Regression] main: concurrent throughput drops up to -59% vs v1.5.1 on V100 (cudagraph shapes truncated?)

  1CatAI/1Cat-vLLM#1042 · Filter by author johnbluer opened on Oct 7, 2026

• AutoRound W4A16 (per-expert GPTQ qweight) fails at fused_moe _load_w2 on SM70: packed 80 vs unpacked 640

  1CatAI/1Cat-vLLM#1029 · Filter by author greconevi978-rgb opened on Oct 7, 2026

• [Usage]: Is it possible to use only Two V100, and yes which model did you test

  1CatAI/1Cat-vLLM#1027 · Filter by author dummys opened on Oct 7, 2026

• [Bug][SM70][Qwen3.8] Retrieval correctness depends on max-num-batched-tokens / chunked-prefill boundaries

  1CatAI/1Cat-vLLM#1003 · Filter by author kardapolov opened on Oct 6, 2026
  2

• [V100/SM70] Kimi K2.5 Tower native-INT4 on 2×V100-SXM2: EDSQ subquant + vllm_volta plugin port

  1CatAI/1Cat-vLLM#987 · Filter by author febuz opened on Oct 5, 2026
  2

• Mixed-prefill latency control (#860): the 250 ms default stalls new long prompts ~10x on a MoE model over PCIe V100

  1CatAI/1Cat-vLLM#880 · Filter by author Leonccaa opened on Oct 4, 2026

• [Bug] SM70 MTP decode deadlocks with >=2 concurrent requests (Qwen3.8-Flash-Next NVFP4, TP4)

  1CatAI/1Cat-vLLM#865 · Filter by author dannyadams06-dotcom opened on Oct 3, 2026

• [Perf][SM70][PP=2] Disabling async scheduling gives +55% single-stream decode; SM70 async tuning is inert when pp_size > 1

  1CatAI/1Cat-vLLM#852 · Filter by author zambalee opened on Oct 3, 2026
  2