-
Notifications
You must be signed in to change notification settings - Fork 17
Independent Draft Ubatch
A separately loaded draft model can use a physical microbatch size different from the target context. This controls draft prompt/workspace geometry independently from the target's -ub value.
Use --spec-draft-ubatch-size N, --ubatch-size-draft N, or -ubd N. Zero inherits the target. A smaller draft ubatch can reduce draft workspace and make a separate drafter fit, but can lower its prompt throughput.
Integrated MTP deliberately does not accept a different physical ubatch: it must inherit or exactly equal the target ubatch. Long-output validation showed that changing recurrent prompt synchronization geometry can change later output even when short screens pass. Use phase-aware workspace to reduce MTP generation workspace without changing physical MTP geometry.
This feature does not choose a drafting algorithm and does not alter draft depth. Use the upstream speculative guide for ordinary drafting configuration.
| Branch | Implementation |
|---|---|
beellama/main |
common/speculative.cpp, common/arg.cpp
|
llama/main |
common/speculative.cpp, common/arg.cpp
|
llama/dev |
common/speculative.cpp, common/arg.cpp
|
moe-cache |
common/speculative.cpp, common/arg.cpp
|
Design history: PR 27, independent draft ubatch.
GenerelSchwerz llama.cpp
- Home
- Discord community
- Contributors
- Complete feature index
- Hardware setup guides
- Owner-verified evidence
- Notable runs
- Benchmark comparison showcase
- BeeLlama Main
- Llama Main
- Llama Dev
- MoE Cache
Feature groups
- Memory placement and workspace
- Validation and diagnostics
- CUDA MoE cache and helpers
- Grouped MoE drafting
Source branches