[brandneway]

Purpose

Two independent startup/runtime fixes for MiniMax-H3 multi-degree parallel serving on NPU (e.g. 16 cards: DiT SP=16 + text encoder TP=8 + VAE patch parallel), both hit while bringing up Ulysses × AllGather-KV topologies:

1. Text encoder TP fails to start when text_encoder_tp_size < DiT world size.

_build_text_encoder_group builds a single GroupCoordinator covering only ranks [0, tp_size). GroupCoordinator.__init__ asserts that every constructing rank belongs to one of its groups, so with tp_size smaller than the DiT world — e.g. tp=8 on 16 devices, the only legal topology for Qwen3-VL's 64 Q / 8 KV heads — every rank outside [0, tp) fails the bare assert at startup.

Fix: tile the DiT world into dit_world / tp_size contiguous groups (_text_encoder_group_ranks). Every rank holds membership in exactly one group; only the first group encodes, and its output still reaches all DiT ranks through the existing world-group broadcast. tp_size == world yields the same single group as before, so existing deployments are unchanged. Non-divisible combinations now fail with a clear message.

2. Packed varlen attention rejects every request under AllGather-KV SP.

With allgather_degree > 1 the NPU attention backend sees local Q against global K/V. The square packed contract in _resolve_packed_seq_npu (used_q <= total_q) declines, and the mask-rebuild fallback then rejects the global valid_kv_length against the local query length, failing every packed H3 request with "packed metadata is unusable ... refusing to run unmasked attention over padding rows".

Fix: _resolve_packed_seq_npu accepts the query_ranges recorded by the AllGather-KV strategy. A single packed range means the local Q rows are a contiguous slice of the same global real document, so seq_q becomes the slice's [local_real, local_total] pair while seq_k keeps the global [real, pad] pair — the same "pad is a strict suffix, pad rows are never consumed" contract, with no mask materialized. Both packed NPU paths (varlen, laser prefix-slice) forward the ranges; square (non-SP) behavior is unchanged. The mask fallback now checks valid_kv_length against the K/V domain (key.shape[1]) so it stays a correct safety net for multi-range layouts.

Test Plan

• New CPU unit tests: tests/diffusion/models/minimax_h3/test_minimax_h3_encoder_group.py (group layout tiling, divisibility rejection) and tests/diffusion/attention/test_flash_attn.py group A2 (per-rank packed boundary resolution under AllGather-KV, unsupported-range rejection, square-contract preservation).
• pytest tests/diffusion/attention/test_flash_attn.py -k "packed_seq or cp_": 16 passed.
• NPU serving validation (Ascend 950PR): 4-card ulysses_degree=2 × allgather_degree=2 end-to-end video generation verified bitwise-size-identical outputs against pure Ulysses baselines.

Test Result

All new unit tests pass on CPU; the serving path was validated end-to-end on Ascend NPU with the composed topology.

---

This contribution is made by the author with DCO sign-off.

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/diffusion/index.md.

Module owners: @david6666666 @wtomin @Isotr0py

Routing: @david6666666 via module of the changed files, module named in the PR description, model owner; @wtomin via module of the changed files, module named in the PR description, CODEOWNERS; @Isotr0py via module of the changed files, module named in the PR description

@brandneway, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[brandneway]

Superseded by #7937, which includes both fixes from this PR on top of the composed Ulysses x AllGather-KV topology they were built for.