Repository Issue Activity (beta)

#246 Tensor parallel shared-memory RPC lacks worker acknowledgement and can corrupt commands

#245 Low inference speed in A100 PCIE GPU

#244 Lightweight Go load-balancing proxy for nano-vllm

#240 [Bug] may_append allocates new block one token too late, causing KV cache write to unallocated block

#230 📋 Documentation Enhancement Suggestion

#228 [change] Int8 KV Cache + Async Pipeline + Head-major reordering for 22% Throughput Boost

#225 Introducing Quantized Inference to Nano-vLLM - A Lightweight FP8 Runtime for Qwen Models

#220 [Feature Request] Support weightless RMSNorm (for FlashNorm weight folding trick)

#221 [Discussion] As vLLM-Omni is also complex, I tried building a minimal version (~1k LOC) inspired by nano-vllm

#219 [Discussion] Duplicate blocks for same-step prefix sharing

#208 Prefix cache can prematurely hit in-flight blocks in same prefill batch

#212 prepare_prefill logic flaw: seqlen_k incorrectly uses total sequence length during Chunked Prefill

#64 Confusion about the RoPE implementation

#206 Implement support for Qwen3.5

#114 [BUG] Crashes when the prompt length exactly equals kvcache_block_size

#144 Two bugs

#125 RowParallelLayer with bias crash

#154 Suggestion to add a standard attention implementation

#156 Which vLLM version is used for the experimental comparison?

#91 the method of BlockManager.can_append is error

#189 Qwen models initialization fails with TypeError: unhashable type: 'dict' and AssertionError due to rope_scaling

#165 🚀 Introducing Nano-vLLM-v1 – A Lightweight, Re-engineered Nano vLLM Implementation with vLLM v1 Scheduler and Chunked Prefill

#190 CUDA graph replay can fail with block_tables shape mismatch

#167 rope argument inconsistent

#187 Potential TP correctness issue: BlockManager capacity may rely on per-rank local KV-cache estimation