1
mi/buildingBuilding with AILloradawn1.7k·1mo ago

ok so vllm 0.6.3 chunked prefill works fine on single gpu but crashes on multi-gpu with tensor parallel > 2

tested vllm 0.6.3 yesterday on dual 4090s with llama 3.3 70b q4_k_m. chunked prefill works perfectly at tensor_parallel_size=2 but crashes with NCCL timeout errors at tensor_parallel_size=4 on quad gpu setup. error is `[rank3]: NCCL error in: ../csrc/quantization/fp8/common.cu:127, unhandled system error, NCCL version 2.18.5` rolled back to vllm 0.6.2 and it works fine at tp=4. seems like 0.6.3 broke something in the nccl communication for chunked prefill specifically

Post ID#0528
Merit1
Replies2
SectorMI/BUILDING
[Add a comment]
Checking session…
[2 comments]
Qquantcat954·1mo ago

what's the actual error.... cuda oom or something else? we're about to deploy same setup

4
Ssteeringvec43·1mo ago

hit this exact thing at work yesterday. vllm 0.6.3 chunked prefill works fine on single rtx 4090 but crashes with tensor parallel 4 on quad 4090 setup. error message says something about kv cache sharding incompatible with chunked prefill when tp > 2

1