For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Python class
SupportsSSMStateWarmup
SupportsSSMStateWarmup
class max.pipelines.lib.SupportsSSMStateWarmup(*args, **kwargs)
Bases: Protocol
Protocol for pipeline models holding state slots of their own.
A model whose per-request SSM/conv state is a slot pool outside the KV
cache must implement release_warmup_state(); nothing else frees
those slots. A model whose state lives in KV pages needs none of this.
The overlap pipeline’s _warmup_model_inputs context manager calls
release_warmup_state after each probe’s capture completes, with the
same request IDs that were passed to prepare_initial_token_inputs
during that probe.
release_warmup_state()
release_warmup_state(request_ids)
Release state pool slots claimed for the given warmup request IDs.
Called once per _warmup_model_inputs probe, after graph capture
for that probe completes. request_ids matches the list of request
IDs from the warmup batch constructed inside _warmup_model_inputs.
Implementations should release slots without zeroing the underlying
pool memory (zeroing happens at claim time, not release time).