[Refactor] Rename NSA → DSA: user-facing aliases, file/class/import rename by ch-wan · Pull Request #25821 · sgl-project/sglang

ch-wan · 2026-05-19T23:47:56Z

Motivation

NSA (Native Sparse Attention) is a misnomer — this attention variant is specific to DeepSeek and should be called DSA (DeepSeek Sparse Attention). This PR performs a full rename from nsa to dsa across user-facing CLI/env/registry surfaces and internal files, while preserving backward-compatible aliases for the user-facing layer.

Modifications

User-facing (with deprecated backward-compat aliases):

ServerArgs fields: nsa_prefill_backend → dsa_prefill_backend, nsa_decode_backend → dsa_decode_backend, enable_nsa_prefill_context_parallel → enable_dsa_prefill_context_parallel, nsa_prefill_cp_mode → dsa_prefill_cp_mode
CLI flags: --dsa-* as canonical; --nsa-* kept as deprecated aliases (emit logger warning via DeprecatedAliasStoreAction)
Attention backend key: "dsa" as canonical; "nsa" kept with DeprecationWarning
Env vars: SGLANG_DSA_* canonical; SGLANG_NSA_* as deprecated fallbacks via _DeprecatedEnvFallback mixin

Internal renames (no compat layer needed):

layers/attention/nsa/ → dsa/; nsa_backend.py → dsa_backend.py; nsa_indexer.py → dsa_indexer.py
NativeSparseAttnBackend → DeepseekSparseAttnBackend (one-release alias kept); NSAMetadata → DSAMetadata; NSAIndexerMetadata → DSAIndexerMetadata; etc.
Env var reads updated to SGLANG_DSA_* inside DSA backend
All 33+ import sites updated; test files renamed: test_nsa_indexer.py → test_dsa_indexer.py, test_nsa_pool_host_unit.py → test_dsa_pool_host_unit.py
handle_attention_nsa → handle_attention_dsa; _create_nsa_{decode,prefill}_backend → _create_dsa_*
attach_hybrid_nsa_pool_to_hiradix_cache → attach_hybrid_dsa_pool_to_hiradix_cache
Backend lists (MLA_ATTENTION_BACKENDS, FORWARD_ABSORB_CORE_ATTENTION_BACKENDS, SEPARATE_ROPE_BACKENDS) updated to include "dsa" as canonical entry

Backward-compat shims (to be removed in N+1 release):

layers/attention/nsa_backend.py — thin re-export shim
layers/attention/nsa/__init__.py — thin re-export shim

Note: the directory mv (nsa/ → dsa/) may rebase-conflict with upstream PR #23906 (Cuda Graph Runner refactor). Coordination recommended before merge.

Accuracy Tests

Pure rename refactor — no model forward or kernel code changed. Verified with:

test_dsa_indexer.py: 9 passed
test_dsa_alias_cli_registry_env.py: 24 passed
test_dsa_pool_host_unit.py: 2 passed

Speed Tests and Profiling

N/A — no computation path changes.

Checklist

Format your code according to the Format code with pre-commit.
Add unit tests according to the Run and add unit tests.
Update documentation according to Write documentations.
Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
Follow the SGLang code style guidance.

Review and Merge Process

Ping Merge Oncalls to start the process. See the PR Merge Process.
Get approvals from CODEOWNERS and other reviewers.
Trigger CI tests with comments or contact authorized users to do so.
- Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

CI States

Latest PR Test (Base): 🚫 Run #26137426290
Latest PR Test (Extra): ❌ Run #26137426228

gemini-code-assist · 2026-05-19T23:47:59Z

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

- Rename C++ struct fields nsa_cache_seqlens → dsa_cache_seqlens in fused_metadata_copy.cuh - Update Python wrapper fused_metadata_copy.py param names (nsa_* → dsa_*) - Rename get_nsa_kv_pool_cls → get_dsa_kv_pool_cls in platforms/interface.py and callers - Update comments/docstrings: NSA → DSA across 20 files (jit_kernel, mem_cache, srt layers, disaggregation, server_args) - Rename DeepSeekNSAAlgorithm → DeepSeekDSAAlgorithm and update imports - Update .claude/skills references (overlap-catalog, fuse-overlap-catalog, triage_kernel_helpers) Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

Update ~25 stray NSA references in comments and docstrings across dsa_indexer.py, cp_utils.py, pool_configurator.py, model_runner*.py, deepseek_v4.py, forward_mha.py, forward_mla.py, model_config.py, hisparse_hook.py, and deepseek_v2_attention_mla_npu.py. Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

- plugin.md: get_nsa_kv_pool_cls → get_dsa_kv_pool_cls (matches code rename) - attention_backend.md: remove (NSA) from section header, fix anchor link Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

- NSATokenToKVPool → DSATokenToKVPool (memory_pool.py + all callers) - HiSparseNSATokenToKVPool → HiSparseDSATokenToKVPool - NSAIndexerPoolHost → DSAIndexerPoolHost - nsa_kv_cache_store_fp8 → dsa_kv_cache_store_fp8 - nsa_impl / nsa_impl_for_batch → dsa_impl / dsa_impl_for_batch (dsa_backend.py) - _nsa_payload → _dsa_payload (disaggregation/prefill.py, decode.py) - _send_swa_nsa_state → _send_swa_dsa_state (mori/conn.py) - MockNSATokenToKVPool → MockDSATokenToKVPool (test file) - _jit_nsa_fused_store_module → _jit_dsa_fused_store_module Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

… names - docs_new/docs/: attention_backend.mdx, server_arguments.mdx, deepseek_v32.mdx, hisparse_guide.mdx, ascend docs, plugin.mdx, environment_variables.mdx — all --nsa-* → --dsa-* - docs_new/cookbook/: GLM-5, GLM-5.1, DeepSeek-V3.2 mdx files - docs_new/src/snippets/: 4 JSX deployment snippet files - .github/workflows/: nightly AMD test job names NSA → DSA - environ.py: comment update Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

- StateType.NSA = "nsa" → StateType.DSA = "dsa" (symbol + wire value) - Update all callers: disaggregation/utils.py, prefill.py, decode.py, nixl/conn.py, mooncake/conn.py, mori/conn.py ("nsa" string literal) - attention_registry.py: create_nsa_backend → _create_nsa_compat (keeps @register_attention_backend("nsa") user-facing key + DeprecationWarning) Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

…OPK, get_nsa_kv_pool_cls - test_fused_metadata_copy.py: nsa_cache_seqlens_*/nsa_cu_seqlens_k_* → dsa_* (dict keys + vars) - environ.py: remove stale SGLANG_NSA_* EnvBool definitions (now covered by DSA canonical + alias) - test_platform_interface.py: get_nsa_kv_pool_cls → get_dsa_kv_pool_cls - skills/: fused_nsa_cache_seqlens → fused_dsa_cache_seqlens in profiler catalog - docs/references/environment_variables.md: SGLANG_NSA_* → SGLANG_DSA_* as canonical + deprecated alias note - docs_new/docs/references/environment_variables.mdx: same Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

…nment_variables.mdx

…eepseek_v4.py - AttnTPContext.is_nsa → is_dsa (attr + init_context param) - _nsa_needs_bf16 local var → _dsa_needs_bf16 (×2 in communicator.py) - deepseek_v4.py: init_context(is_nsa=True) → init_context(is_dsa=True) Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

- test_fused_metadata_copy.py: nsa_seqlens_expanded → dsa_seqlens_expanded - dsa_indexer.py: get_nsa_extend_len_cpu → get_dsa_extend_len_cpu - test_dsa_indexer.py: mock get_nsa_extend_len_cpu → get_dsa_extend_len_cpu - dequant_k_cache.py: comment nsa_backend → dsa_backend - dsa_backend.py: comment nsa_prefill_impl → dsa_prefill_impl - factory.py: "deepseek_nsa" dict key → "deepseek_dsa" - model_runner_kv_cache_mixin.py: is_nsa_model → is_dsa_model (×9) - forward_mha.py: _get_mla_kv_buffer_from_fp8_for_nsa → _for_dsa - forward_mla.py: _skip_rope_for_nsa_tilelang_fused → _for_dsa (def + 3 calls) - communicator.py: is_nsa/init_context param already fixed in prior commit Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

…ning SGLANG_NSA_* as canonical - environ.py: add SGLANG_DSA_HIP_DISABLE_PRESHUFFLE with deprecated_name fallback - dsa/utils.py: use envs.SGLANG_DSA_HIP_DISABLE_PRESHUFFLE; update docstring - server_args.py: update comment to SGLANG_DSA_HIP_DISABLE_PRESHUFFLE - docs/basic_usage/deepseek_v32.md: SGLANG_NSA_PREFILL_DENSE_ATTN_KV_LEN_THRESHOLD → DSA canonical - docs_new/docs/basic_usage/deepseek_v32.mdx: same - skills/fuse-overlap-catalog.md: SGLANG_NSA_FUSE_TOPK → SGLANG_DSA_FUSE_TOPK Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

…iCacheTransfer → TestDSAHiCacheTransfer Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

Conflicts: .github/workflows/nightly-test-amd.yml and .github/workflows/nightly-test-amd-rocm720.yml. The only main commit that touched these two files was PR #25821 "[Refactor] Rename NSA → DSA: user-facing aliases, file/class/import rename" (8131641), which changed 4 step-name strings: "Accuracy Test (8-GPU GLM-5.1 NSA)" -> "...DSA)" "Accuracy Test MI35x (8-GPU GLM-5.1 NSA)" -> "...DSA)" "Accuracy Test ROCm 7.2 (8-GPU GLM-5.1 NSA)" -> "...DSA)" "Accuracy Test MI35x ROCm 7.2 (8-GPU GLM-5.1 NSA)" -> "...DSA)" (2 occurrences in nightly-test-amd.yml, 2 in nightly-test-amd-rocm720.yml.) This PR's "Reorder nightly-test-amd*.yml jobs by model" commits moved the surrounding GLM-5.1 job blocks to different physical line ranges in the YAML. Git's 3-way merge could not reconcile "we moved a 500-line range" with "they changed a single word in that range", so it dumped both entire spans into one large conflict region per file. Resolved by: 1. taking this PR's reordered content for both nightly files (git checkout --ours), since the reorder is a core change of this PR and is content-equivalent to main's; 2. then mechanically applying the 4 NSA -> DSA string renames on top so main's rename intent is preserved. After resolution `rg "GLM-5.1 NSA" .github/workflows/nightly-test-amd*.yml` returns 0 hits; the 4 expected "GLM-5.1 DSA" step names are present. Verified post-merge: - yaml.safe_load passes on all 4 AMD workflows - per-workflow: jobs <-> dropdown <-> contains() gating <-> finish-needs sets are mutually consistent - no conflict markers anywhere All other NSA -> DSA changes from #25821 are in files this PR does not touch (docs, attention backend code, server args, etc.) and auto-merged cleanly. Co-authored-by: Bingxu Chen <Bingxu.Chen@amd.com>

Resolve conflicts from NSA→DSA rename (sgl-project#25821): - Migrate SM120 guards from nsa_indexer.py → dsa_indexer.py - Migrate SM120 guards from nsa_backend.py → dsa_backend.py - Move sm120_mqa_fallback.py, sm120_mqa_triton.py to dsa/ directory - Update test import path to dsa.sm120_mqa_fallback - Use is_sm120_supported() directly (no module-level _is_sm120 var)

…ctor The NSA→DSA rename (PR sgl-project#25821) changed all backend attributes from nsa_prefill_backend/nsa_decode_backend to dsa_prefill_backend/dsa_decode_backend, but the SM120 branches in _set_default_dsa_backends() still used the old names, silently creating orphan attributes while leaving dsa_* as None. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…ename (sgl-project#25821) Co-authored-by: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

ch-wan requested review from Fridge003, Qiaolin-Yu, Ying1123, fzyzcjy, hnyls2002, ispobock and merrymercy as code owners May 19, 2026 23:47

ch-wan requested review from 1am9trash, BBuf, DarkSharpness, Edwardf0t1, HaiShaw, HydraQYH, ShangmingCai, celve, hanming-lu, hebiao064, hlu1, hubertlu-tw, hzh0425, iforgetmyname, kkHuang-amd, ping1jing2, xiezhq-hermann, yizhang2077 and yuan-luo as code owners May 19, 2026 23:47

github-actions Bot added quant LLM Quantization deepseek npu labels May 19, 2026

ch-wan requested review from ByronHsu and YAMY1234 as code owners May 20, 2026 01:20

ch-wan and others added 2 commits May 19, 2026 18:23

ch-wan removed the run-ci label May 20, 2026

ch-wan and others added 2 commits May 19, 2026 18:30

ch-wan requested review from JustinTong0323, Kangyan-Zhou and bingxche as code owners May 20, 2026 01:39

ch-wan and others added 8 commits May 19, 2026 18:44

style: remove extra parentheses from deprecated alias notes in enviro…

4d1716b

…nment_variables.mdx

refactor(dsa): fix nsa example string in nightly_utils.py docstring

7e592cc

refactor(dsa): rename NSABackendAdaptor → DSABackendAdaptor; TestNSAH…

5fe2394

…iCacheTransfer → TestDSAHiCacheTransfer Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

ch-wan added the run-ci label May 20, 2026

style: fix isort and black lint failures

6f59532

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

ch-wan mentioned this pull request May 20, 2026

[DSA] Make MQA logits free memory ratio configurable #25859

Merged

5 tasks

ch-wan merged commit 8131641 into main May 20, 2026
226 of 254 checks passed

ch-wan deleted the cheng/refactor/step01-nsa-dsa-rename branch May 20, 2026 07:18

Fridge003 mentioned this pull request May 20, 2026

Support batch size > 1 when enable CP #23269

Merged

5 tasks

zijiexia mentioned this pull request May 23, 2026

[docs] DeepSeek-V4 cookbook: balanced MegaMoE cap, H200 Pro FP4 mem-frac, nsa-* compat, PD-disagg fixes #26164

Merged

7 tasks

Shunkangz pushed a commit to Shunkangz/sglang that referenced this pull request May 27, 2026

[Refactor] Rename NSA → DSA: user-facing aliases, file/class/import r…

fab7d4d

…ename (sgl-project#25821) Co-authored-by: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Refactor] Rename NSA → DSA: user-facing aliases, file/class/import rename#25821

[Refactor] Rename NSA → DSA: user-facing aliases, file/class/import rename#25821
ch-wan merged 37 commits into
mainfrom
cheng/refactor/step01-nsa-dsa-rename

ch-wan commented May 19, 2026 •

edited by github-actions Bot

Loading

Uh oh!

gemini-code-assist Bot commented May 19, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

ch-wan commented May 19, 2026 • edited by github-actions Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Modifications

Accuracy Tests

Speed Tests and Profiling

Checklist

Review and Merge Process

CI States

Uh oh!

gemini-code-assist Bot commented May 19, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

ch-wan commented May 19, 2026 •

edited by github-actions Bot

Loading