Overview

Adds AttentionConfig.sparse_mla_prefill_workspace_factor. It replaces the fixed 5 * max_model_len FlashMLA sparse prefill workspace and the 40 * max_model_len DSA indexer prefill buffer sized from it. Unset, VllmConfig resolves it to 5 up to a 256K max_model_len and to 2 above. At a 1M max_model_len the fixed reservations hold several GiB per GPU that a long-context deployment can return to the KV cache.

Related: #60874 (right-sizes the DeepSeek-V4.1 copies of these buffers), #42856 (fixed lower factors on SM120 only), #59319 (skips the FlashMLA workspace on the fp8 mixed-batch path).

Claims

• Up to a 256K max_model_len the default keeps every size identical to main.
• Above 256K the default is 2: workspace 2 * max_model_len, indexer 16 * max_model_len. --attention-config.sparse_mla_prefill_workspace_factor=5 restores the main sizes.
• The default resolves once, in VllmConfig.__post_init__. KV cache auto-fit can lower max_model_len after the workspace is reserved, so a factor derived later could switch from 2 to 5 and size a step past the reservation.
• One knob sizes both buffers. The indexer buffer stays (576 * 2 // 132) * factor rows, so it still fits inside the FlashMLA workspace.
• Both chunkers keep each request's context in one gather, so a lower factor changes only how many requests share a gather.
• The floor is 2 (ge=2). Under PCP+DCP the per-rank budget is factor * max_model_len // dcp, and a full-length request's shard can round up past max_model_len // dcp, which split_prefill_chunks asserts against.
• GLM-5.3 FP8 on H100 at a 1M max_model_len (PCP8, DCP8, TP1): factor 2 raises available KV cache from 9.07 to 18.36 GiB per GPU.

Validation

Default, CPU. On 06a24c3, a real VllmConfig (Qwen/Qwen2.5-0.5B, CPU device) resolves the factor and both buffer sizes as follows:

An explicit factor of 3 at 1M is kept, and AttentionConfig(sparse_mla_prefill_workspace_factor=1) raises a validation error. test_indexer_prefill_budget_matches_compressed_workspace reads the resolved factor through create_vllm_config; it needs a GPU, so CI runs it in the v1/attention job. pre-commit (ruff, mypy, validate-config, test tethering) passes on the changed files.

GLM-5.3 on H100, same sizes. We run GLM-5.3 FP8 in production with these exact sizes (workspace 2x, indexer 16x) applied to the same two functions on our vLLM build. The prefill engines are PCP8, DCP8, TP1, --max-model-len 1048576, fp8_ds_mla, block size 64, --gpu-memory-utilization 0.93.

All 28 production prefill engines have run factor 2 since 2026-10-10 18:08 UTC, and every one booted without OOM.

This branch on GPU: not run. The production numbers come from the same sizes on our build.

Details

• Why a config field. The neighbouring sparse-MLA and indexer knobs (sparse_mla_force_mqa, indexer_kv_dtype, hisparse_config) live in AttentionConfig. get_prefill_workspace_size now takes the VllmConfig, like get_max_prefill_buffer_size already does.
• Scope. The FlashMLA workspace part applies to FLASHMLA_SPARSE with a quantized DS-MLA cache. get_max_prefill_buffer_size also sizes the indexer in deepseek_v2 (DeepSeek-V3.2, GLM-5.x), deepseek_v32, deepseek_v4, deepseek_v41, hy_v4 and glm5next, so the factor applies there too. The DeepSeek-V4.1 FlashMLA workspace is sized in models/deepseek_v41 and does not read it.
• Cost. A step gathers in more passes only when its prefill requests hold more than factor * max_model_len context tokens in total. At the default above 256K (factor 2), that takes three or more concurrent near-full-length prefills.
• Not a duplicate. Open PRs found with gh pr list --state open --search for "sparse prefill workspace", "prefill workspace size configurable", "flashmla sparse workspace", "indexer prefill buffer" and "workspace factor":
  • [Bugfix] Reduce sparse-MLA + indexer workspace bounds on SM_120 (consumer Blackwell) #42856 lowers the factors to 2 and 8 on SM120 only, stacked on [New Model][Nvidia] Add SM12x support for DeepSeek V4 Flash with essential fixes #41834.
  • [Perf][DSv4.1] Size the long-context prefill workspaces by what a chunk can use #60874 changes only the DeepSeek-V4.1 builders.
  • [Bugfix][Attention] Skip the FlashMLA sparse bf16 prefill workspace on the fp8 mixed-batch path #59319 drops the reservation on the mixed-batch path. On the separate prefill path this knob still sizes it, and the two PRs touch one shared call line in FlashMLASparseImpl.__init__.
  • None makes the size configurable.