/te-ci pytorch L1

Greptile Summary

This PR enables packed THD tensors with hierarchical context parallelism by removing stale runtime, backend-selection, and test-suite exclusions.

• Allows thd inputs through the hierarchical A2A+P2P context-parallel dispatcher.
• Permits FusedAttention backend selection for THD with hierarchical context parallelism while retaining existing mask, bias, FP8, sliding-window, topology, and head-count guards.
• Extends FlashAttention and FusedAttention distributed numerical test matrices to cover the newly enabled combination.

Confidence Score: 5/5

The PR appears safe to merge because the newly enabled hierarchical THD path composes existing THD-aware A2A and P2P implementations and remains protected by the relevant backend compatibility checks.

The removed gates expose an implementation with dedicated THD sequence reordering, padded-sequence metadata handling, output correction, and backward gradient correction, and the distributed tests compare its forward and backward results against the non-context-parallel reference.

Important Files Changed

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A["THD Q/K/V and sequence metadata"] --> B["Backend selection"]
  B --> C["Hierarchical CP dispatcher"]
  C --> D["A2A head exchange"]
  D --> E["P2P context ring"]
  E --> F["FlashAttention or FusedAttention"]
  F --> G["A2A output exchange"]
  G --> H["THD output"]

_{Reviews (1): Last reviewed commit: "Enable THD with hierarchical context par..." | Re-trigger Greptile}