[Spec Decode] DSpark confidence-scheduled verification #47808
vllm-bot
merged
45 commits into
vllm-project:main
from
neuralmagic:codex/dspark-capacity-realloc
LucasWilkinson
changed the title
claude
commented
on 2026-07-16
depthfirst-app
commented
on 2026-07-29
LucasWilkinson
changed the title
depthfirst-app
commented
on 2026-07-30
benchislett
commented
on 2026-08-03
[Spec Decode] DSpark adaptive verification
0d35edb2
[Spec Decode] Keep non-adaptive request ordering identical to main
a772e95f
[Spec Decode] Source decode max_q_len from attention metadata
c4c8eec0
[Spec Decode] Prune implementation-focused unit tests
ae28ecf5
[Spec Decode] Rename adaptive verification config knobs
f8c57028
[Spec Decode] Fall back to fixed-length verification without cost pro…
9c708552
[Spec Decode] Cap draft budget to one rejection-sampler chunk
cb068aac
[Spec Decode] Do not treat async draft placeholders as grammar-rejected
4cbe277a
[Spec Decode] Gate adaptive verification on LoRA, confidence head, mo…
75b15226
cleanup
48daecce
[Spec Decode] Drop cudagraph prefill-padding trim from SWA builder
5105a571
[Spec Decode] Resolve profiling step timings without a per-step sync
5b6ce974
cleanup
1ea4f64f
cleanup
4db5223b
fix precommit
aad37e15
[Spec Decode] Only require a topk buffer on compressed DSV4 layers
ea368ee3
Revert "[Spec Decode] Only require a topk buffer on compressed DSV4 l…
45a82012
[Spec Decode] Require a topk buffer only where it is read
e5655838
minor cleanup
e6e8bea2
cleanup
da62b61d
wip: review comments
af4751b5
[Spec Decode] Inherit target quantization for the DSV4 DSpark draft
70596c82
[Spec Decode] Compute draft confidences in the topk sampling path too
eed2eefe
TheEpicDolphin
commented
on 2026-08-04
[Spec Decode] Rename max_chunk_logits helper to avoid shadowing
3c59d1e2
[Spec Decode] Align DSpark GSM8K eval config with the vLLM recipe
7ea349d6
benchislett
commented
on 2026-08-05
benchislett
commented
on 2026-08-05
benchislett
commented
on 2026-08-05
benchislett
commented
on 2026-08-06
benchislett
commented
on 2026-08-06
benchislett
commented
on 2026-08-06
review comments
387b0840
cleanup
e82c2b47
review comments
49f9571b
benchislett
commented
on 2026-08-06
benchislett
commented
on 2026-08-06
benchislett
commented
on 2026-08-06
benchislett
commented
on 2026-08-06
benchislett
commented
on 2026-08-06
Update vllm/config/speculative.py
e5933514
Merge remote-tracking branch 'upstream/main' into dspark-final
f7c227c8
review comments
01db35ac
review comments
3becde2b
remove slop tests
d80ff5b6
CI fixes
6636dfed
Merge branch 'main' into codex/dspark-capacity-realloc
10c132ff
benchislett
commented
on 2026-08-10
Apply suggestion from @benchislett
1160bceb
benchislett
approved these changes
on 2026-08-10
[Doc] Fix adaptive verification default in docs
1854fdfb
njhill
commented
on 2026-08-10
review comments
a6ef5acc
Merge remote-tracking branch 'upstream/main' into pr47808
24fbcfb1
review comments
b67f2319
njhill
approved these changes
on 2026-08-11
Merge branch 'main' into codex/dspark-capacity-realloc
a48eccc0
Merge branch 'main' into codex/dspark-capacity-realloc
aa72154a
fux test
7b032b5d
review refactors
5cd5b8a0
[Bugfix] Skip adaptive sizing for no-draft batches
e2e33533
mgoin
approved these changes
on 2026-08-12
Reviewers
Assignees
No one assigned
Labels
documentation
performance
speculative-decoding
ready
ci/build
v1
qwen
nvidia
mrv2
Milestone
No milestone
Login to write a write a comment.
Login via GitHub