feat(MODEL-DSV4-EXL3): the real DeepSeek-V4 DSA geometry loads, and the forward refuses it by name (#1970) by localai-bot · Pull Request #1971 · mudler/vllm.cpp

record(MODEL-DSV4-EXL3): the doubled DSA tensors are the coff=2 overlapping-window pair, and dense MLA was never exact (#1961)

mudler

committed

e68f57b

merge: origin/main into row/DSV4-DSA-GEOMETRY

mudler

committed

d8d9f6c

merge: origin/main into row/DSV4-DSA-GEOMETRY

mudler

committed

c006251

spec(MODEL-DSV4-EXL3): option C — the loader takes the real DSA geometry and the forward refuses by name (#1970)

mudler

committed

d92d7d4

feat(MODEL-DSV4-EXL3): the real DeepSeek-V4 DSA geometry loads, and the forward refuses it by name (#1970)

mudler

committed

60f49d4

merge: origin/main into row/DSV4-DSA-GEOMETRY

mudler

committed

6094c3d

fix(MODEL-DSV4-EXL3): derive the DSA width strictly, and withdraw the MatVec claim the refusal was justified with (#1970)

mudler

committed

6bf313c

record(MODEL-DSV4-EXL3): the cr == 128 EXL3 compressor gets its own issue, not #1964's (#1976)

mudler

committed

1d42674

fix(MODEL-DSV4-EXL3): the sixth copy of the wrong norm anchor, in the forward suite (#1970)

mudler

committed

7231a5e

merge: origin/main into row/DSV4-DSA-GEOMETRY

mudler

committed

17dab65

docs(MODEL-DSV4-EXL3): say which half-widening cases are gated and which are only structural (#1970)

mudler

committed

0acf014

merge: origin/main into row/DSV4-DSA-GEOMETRY

mudler

committed

8846e76

fix(MODEL-DSV4-EXL3): the seventh copy of the wrong anchor, and the three checks a mutation walked through (#1970)

mudler

committed

14b6f63

record(MODEL-DSV4-EXL3): finish the sweep at the second spec and the #1961 row, which round 1 never opened (#1961, #1970)

mudler

committed

3021119

record(MODEL-DSV4-EXL3): re-derive the whole mutation table at the branch head (#1970)

mudler

committed

a66e7c2

merge: origin/main into row/DSV4-DSA-GEOMETRY

mudler

committed

bcd2b1e

record(MODEL-DSV4-EXL3): the ninth and tenth copies of the withdrawn MatVec claim (#1970)

mudler

committed

f6aaf6f

merge: origin/main into row/DSV4-DSA-GEOMETRY

mudler

committed

6ee0195

record(MODEL-DSV4-EXL3): the index rows this branch adds carried stale anchors too (#1961, #1970, #1976)

mudler

committed

23f6ef8

merge: origin/main into row/DSV4-DSA-GEOMETRY

mudler

committed

60d6767

fix(#1970): a twelfth copy of the withdrawn claim, and the two counts that said it could not exist

mudler

committed

0bdd275

fix(#1970): the anchor-uniqueness count was measured, and it is six rather than two

mudler

committed

c365275

Navigation Menu

feat(MODEL-DSV4-EXL3): the real DeepSeek-V4 DSA geometry loads, and the forward refuses it by name (#1970) - #1971

feat(MODEL-DSV4-EXL3): the real DeepSeek-V4 DSA geometry loads, and the forward refuses it by name (#1970)#1971
localai-bot merged 22 commits into
mainfrom
row/DSV4-DSA-GEOMETRY

Commits on Aug 26, 2026

record(MODEL-DSV4-EXL3): the doubled DSA tensors are the coff=2 overlapping-window pair, and dense MLA was never exact (#1961)

merge: origin/main into row/DSV4-DSA-GEOMETRY

merge: origin/main into row/DSV4-DSA-GEOMETRY

spec(MODEL-DSV4-EXL3): option C — the loader takes the real DSA geometry and the forward refuses by name (#1970)

feat(MODEL-DSV4-EXL3): the real DeepSeek-V4 DSA geometry loads, and the forward refuses it by name (#1970)

merge: origin/main into row/DSV4-DSA-GEOMETRY

fix(MODEL-DSV4-EXL3): derive the DSA width strictly, and withdraw the MatVec claim the refusal was justified with (#1970)

record(MODEL-DSV4-EXL3): the cr == 128 EXL3 compressor gets its own issue, not #1964's (#1976)

fix(MODEL-DSV4-EXL3): the sixth copy of the wrong norm anchor, in the forward suite (#1970)

merge: origin/main into row/DSV4-DSA-GEOMETRY

docs(MODEL-DSV4-EXL3): say which half-widening cases are gated and which are only structural (#1970)

merge: origin/main into row/DSV4-DSA-GEOMETRY

fix(MODEL-DSV4-EXL3): the seventh copy of the wrong anchor, and the three checks a mutation walked through (#1970)

record(MODEL-DSV4-EXL3): finish the sweep at the second spec and the #1961 row, which round 1 never opened (#1961, #1970)

record(MODEL-DSV4-EXL3): re-derive the whole mutation table at the branch head (#1970)

merge: origin/main into row/DSV4-DSA-GEOMETRY

record(MODEL-DSV4-EXL3): the ninth and tenth copies of the withdrawn MatVec claim (#1970)

merge: origin/main into row/DSV4-DSA-GEOMETRY

record(MODEL-DSV4-EXL3): the index rows this branch adds carried stale anchors too (#1961, #1970, #1976)

merge: origin/main into row/DSV4-DSA-GEOMETRY

fix(#1970): a twelfth copy of the withdrawn claim, and the two counts that said it could not exist

fix(#1970): the anchor-uniqueness count was measured, and it is six rather than two