---
cited_by:
  - '[[thoughts/inference/deepseek-v4.1-serving#4-pr-reading-order|dsv41 §4]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/cmake/external_projects/flashmla.cmake
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/attention/backends/mla/sparse_swa.py
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/attention/backends/mla/flashmla_sparse.py
merged: '2026-09-15'
number: 56893
source:
  - https://github.com/vllm-project/vllm/blob/e6eb0d120c480aa5116aafc2d39f19c8ebcf3181/vllm/models/deepseek_v41/common/ops/cache_utils.py
  - https://github.com/vllm-project/vllm/blob/e6eb0d120c480aa5116aafc2d39f19c8ebcf3181/csrc/libtorch_stable/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu
  - https://github.com/vllm-project/vllm/blob/e6eb0d120c480aa5116aafc2d39f19c8ebcf3181/cmake/external_projects/flashmla.cmake
  - https://github.com/vllm-project/vllm/blob/e6eb0d120c480aa5116aafc2d39f19c8ebcf3181/vllm/models/deepseek_v41/common/ops/fused_compress_quant_cache.py
  - https://github.com/vllm-project/vllm/blob/e6eb0d120c480aa5116aafc2d39f19c8ebcf3181/vllm/models/deepseek_v4/nvidia/ops/dequant_gather_k_cutedsl.py
  - https://github.com/vllm-project/vllm/blob/e6eb0d120c480aa5116aafc2d39f19c8ebcf3181/vllm/models/deepseek_v41/attention.py
  - https://github.com/vllm-project/vllm/blob/e6eb0d120c480aa5116aafc2d39f19c8ebcf3181/vllm/v1/attention/backends/mla/sparse_swa.py
  - https://github.com/vllm-project/vllm/blob/e6eb0d120c480aa5116aafc2d39f19c8ebcf3181/vllm/v1/attention/backends/mla/flashmla_sparse.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#56893 [Model][DSv4.1] Store the whole KV in MXFP8 (FlashMLA V4.1 record)'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/56893
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/56893
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/56893.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
DeepSeek-V4.1 has been reusing DeepSeek-V4’s paged `fp8_ds_mla` record. FlashMLA has a V4.1-specific one that quantizes the RoPE dims too; this switches DeepSeek-V4.1 onto it on SM100. · full dump: [[thoughts/inference/vllm-prs/56893|#56893]]

