Context
HybridEngineRollout supports both shared prompt prefill and continuous
batching, but their interaction is currently undefined.
When continuous batching is enabled through
SamplingConfig.continuous_batch_size, the continuous path is selected before
the shared-prefill logic. As a result, HybridEngineRolloutConfig.use_shared_prefill
is silently ignored.
Continuous batching currently requires n_samples_per_prompt=1, so shared
prefill does not provide a clear benefit in this mode. However, silently
ignoring the configuration can be confusing and may hide an invalid combination
of options.
Proposed work
Define and document the behavior of use_shared_prefill when continuous
batching is enabled.
Possible choices include:
- explicitly reject the combination with a clear
ValueError; or
- define and implement compatible shared-prefill semantics for continuous
batching.
Acceptance criteria
- The interaction between
use_shared_prefill and continuous batching is
explicitly defined.
- The configuration is not silently ignored.
- Tests cover both compatible and incompatible configurations.
- The default non-continuous generation path remains unchanged.
Follow-up from DeepSpeed PR #8368. cc @delock
Context
HybridEngineRolloutsupports both shared prompt prefill and continuousbatching, but their interaction is currently undefined.
When continuous batching is enabled through
SamplingConfig.continuous_batch_size, the continuous path is selected beforethe shared-prefill logic. As a result,
HybridEngineRolloutConfig.use_shared_prefillis silently ignored.
Continuous batching currently requires
n_samples_per_prompt=1, so sharedprefill does not provide a clear benefit in this mode. However, silently
ignoring the configuration can be confusing and may hide an invalid combination
of options.
Proposed work
Define and document the behavior of
use_shared_prefillwhen continuousbatching is enabled.
Possible choices include:
ValueError; orbatching.
Acceptance criteria
use_shared_prefilland continuous batching isexplicitly defined.
Follow-up from DeepSpeed PR #8368. cc @delock