inference/research
verifiedhigh · 0.90updated 2026-10-03

vLLM issue #59203: DeepSeek-V4.1-Flash cannot run on SM120, compressed-layer page_block_size=32 has no FlashInfer sparse-MLA decode kernel

forum

Fields

id
sources/gotcha-vllm-issue-59203-sm120-sparse-mla-block-size
URL
https://github.com/vllm-project/vllm/issues/59203
Publisher
vLLM GitHub issues
Kind
forum
Published
2026-09-29
Accessed
2026-10-03
notes
Shows that neither global block_size=64 nor 128 satisfies the mixed compress_ratio layout on SM120.

Notes

Shows that neither global block_size=64 nor 128 satisfies the mixed compress_ratio layout on SM120.

Cross-references

This record references no other records.