IMPORTANT: To view this page as Markdown, append `.md` to the URL (e.g. /get-started.md). For the complete documentation index, see llms.txt.
Skip to main content
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 function

replay_conv_pools

replay_conv_pools()​

max.pipelines.speculative.gated_delta_rollback.replay_conv_pools(captures, live_conv_pools, conv_row_ids, row_indices, replay_offsets, signal_buffers, verify_width)

source

Writes each conv window at the accepted length.

Re-runs the conv kernel over the accepted rows of the verify pass’s copy of its input, carrying forward from the pre-verify window still in the pool.

Parameters:

  • captures (Sequence[Sequence[ConvReplayInputs]]) – Per-device, per-layer inputs captured by the verify pass.
  • live_conv_pools (Sequence[BufferValue]) – Per-device conv pool, still pre-verify.
  • conv_row_ids (Sequence[TensorValue]) – Per-device [num_layers, batch_size] conv rows.
  • row_indices (TensorValue) – Rows of the verify tensors the replay consumes.
  • replay_offsets (TensorValue) – [batch + 1] ragged offsets over those rows.
  • signal_buffers (Sequence[BufferValue]) – Used only to place the plan on each device.
  • verify_width (TensorValue) – The verify’s width operand, which skips every launch at width zero.

Return type:

None