forked from sgl-project/SpecForge
-
Notifications
You must be signed in to change notification settings - Fork 0
Pull requests: Boreas618/SpecForge
Author
Label
Projects
Milestones
Reviews
Assignee
Sort
Pull requests list
fix(training): resume falls back to the newest checkpoint common to all ranks
#17
opened Aug 1, 2026 by
Boreas618
Owner
Loading…
fix(draft): validate GQA head divisibility in DFlash attention
#16
opened Aug 1, 2026 by
Boreas618
Owner
Loading…
fix(data): tokenized-cache key includes the data file's content hash
#15
opened Aug 1, 2026 by
Boreas618
Owner
Loading…
feat(inkling): colocated launcher + probe layer-ids fix
#14
opened Aug 1, 2026 by
Boreas618
Owner
Loading…
docs: two-stage RoPE recipe for long-context drafter training
#12
opened Aug 1, 2026 by
Boreas618
Owner
Loading…
feat(inkling): reference configs, data pipeline, launcher, and v2 results
#10
opened Aug 1, 2026 by
Boreas618
Owner
Loading…
feat(bench): standalone DSpark accept-length benchmark (server path)
#9
opened Aug 1, 2026 by
Boreas618
Owner
Loading…
docs: document PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True for CUDA runs
#8
opened Aug 1, 2026 by
Boreas618
Owner
Loading…
fix(dspark): checkpoint config carries BOTH rope schemas
#7
opened Aug 1, 2026 by
Boreas618
Owner
Loading…
fix(dspark): drain engine streams before FSDP forward AND backward
#6
opened Aug 1, 2026 by
Boreas618
Owner
Loading…
fix(dspark): evals disabled actually disables in-loop eval
#5
opened Aug 1, 2026 by
Boreas618
Owner
Loading…
feat(dspark): Inkling-era trainer v2 — new-target adaptation + config-level backbone
#4
opened Aug 1, 2026 by
Boreas618
Owner
Loading…
fix(data): native-list tools rows crashed json.loads
#3
opened Aug 1, 2026 by
Boreas618
Owner
Loading…
fix: don't hang every rank when a micro-batch has no valid anchors
#2
opened Aug 1, 2026 by
Boreas618
Owner
Loading…
ProTip!
Follow long discussions with comments:>50.