fix: unpack FA4 auxiliary outputs

[TomerBN-Nvidia]

Summary

• accept the four-value return contract of the packaged FA4/CuTe _flash_attn_fwd
• backport the relevant hunk from upstream vLLM commit 7cd1d57b749f / PR [CI/Build][Docker] Bump nvidia-cutlass-dsl to 4.6.0 and drop packaging workarounds vllm-project/vllm#47442
• unblock GB200 multimodal encoder warmup on the Super v0.25.1 custom branch

Failure

HSG job 5915987 loaded all 63 checkpoint shards, then failed during RADIO vision-encoder warmup:

File "vllm/vllm_flash_attn/flash_attn_interface.py", line 391
  out, softmax_lse = _flash_attn_fwd(...)
ValueError: too many values to unpack (expected 2)

The SQSH's packaged CuTe implementation returns (out, lse, p, row_max) on this path. The failure occurs entirely in vLLM FA4 before entering FlashInfer.

Duplicate check

No open PR matching this fix was found in this fork or upstream. Upstream already carries the fix as part of merged PR vllm-project#47442; this PR is the minimal backport to tbn/super-vl-bf16-btk-prep.

Tests

• .venv/bin/python -m compileall -q vllm/vllm_flash_attn/flash_attn_interface.py
• git diff --check
• uv run --no-sync pre-commit run --files vllm/vllm_flash_attn/flash_attn_interface.py — all applicable hooks passed
• HSG diagnostic job 5916309: one 4xGB200 node, TP4+EP, same model and r7 SQSH with only --mm-encoder-attn-backend TORCH_SDPA; server started and 3/3 image requests returned HTTP 200. This confirms the failure is isolated to the FA4 multimodal-encoder path.

A rebuilt SQSH should be validated on one GB200 node with TP4+EP and an image request using the default FLASH_ATTN multimodal encoder backend.

AI assistance

OpenAI Codex assisted with root-cause analysis and preparing this one-line backport. A human reviewer should review and understand the changed line before merge.

[TomerBN-Nvidia]

Closing because this was mistakenly targeted at the BTK image-preparation branch. Reopening the same minimal backport against the canonical custom release branch super_vl_rl_v0.25.1.