Render untrusted repository text without controls or extra log lines.

Gather each token's expert results back and weight them by the gate values.

Place every routed token in its expert slot, returning the [e, c, m] buffer. Inverting the routing map first turns the copy into a single gather.

Replace image placeholder positions in *text_embeds* with *visual_embeds*. This is intentionally architecture-specific. The default raises

Split the tensor along its last dimension and keep the corresponding slice.

Split a rank's core list into (zf_affinity, pt_affinity): reserve the first ceil(pt_reserved_cores_perc * n) cores for the training thread and giv

Build a SplitPlan from a global per-expert token histogram (ep_size > 1). A single all-to-all exchanges the whole ``[ep_size, E_local]`` count ma

Rows of the matrix rank ``ep_rank`` receives from the counts all-to-all.

Read an NVMe-offloaded partition back into memory. The swapper leaves ``ds_tensor.data`` as a 0-dim placeholder while a partition is on NVMe.

Get a full non-partitioned state_dict with fp16 weights on cpu. Important: this function must be called on all ranks and not just ran

Adds a pointer offset in units of Element

proxy func to torch.distributed.all_reduce_coalesced, which is included in PyTorch 1.13 and above

Apply explicit AutoEP config overrides to a preset. Return the original preset object when there are no overrides. When overrides are present

Pre-multiply token representations by router scores before expert compute.

Apply `function` to every Tensor in `value`. Args: functional: The function class to apply. value (Any): Target object to ap

For ZeRO-3, assert every non-persistent param is released after backward. The recompute bug left frozen params gathered after backward, so the re

Ensure cached ``ep_size_N`` rank lists match the requested folding spec.

For ZeRO-3, assert every persistent param stays gathered (AVAILABLE) after backward. Persistent frozen params are added to a recompute owner duri

Queue parameter for optimization in the worker process.

Classify one folded TP/SP gradient as ``sum``, ``average``, or ``skip``. TP means Tensor Parallel and SP means Sequence Parallel. The parallel mo