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

combine_start

combine_start()​

max.nn.ep.combine_start(config, buffers, outputs, route, *, total_tokens=None)

source

Launches each device’s combine; combine_finish() waits for it.

With config.combine_mx_min_tokens set, a step of at least that many tokens sends the outputs as MXFP8, and a smaller step sends them exact.

Parameters:

  • config (EPConfig) – The expert-parallel configuration.
  • buffers (EPBuffers) – The exchange’s buffers.
  • outputs (Tensor) – The experts’ outputs, as for combine().
  • route (Route) – The route that dispatch() returned.
  • total_tokens (Tensor | None) – The step’s token count over all devices, a uint32 host scalar. Needed with config.combine_mx_min_tokens.

Raises:

ValueError – If check_supported() rejects config, or if config sets combine_mx_min_tokens without total_tokens.

Return type:

None