[Ronald1995]

Fix DeepSeek V4 target, MTP, and DSpark sink loading in one model-family PR. The 10 changed call-site files cover CPU/XPU/AMD/NVIDIA target models and XPU/AMD/NVIDIA draft models.

Each site passes its checkpoint tensor and rank head range to load_padded_attn_sink from #57797. TP slicing remains in that helper and happens once. The helper builds the full padded runtime tensor and calls the current parameter loader.

Dependency: #57797. Because the shared helper branch is in a fork, GitHub currently compares this PR with upstream main and also displays the two helper files. The V4-only delta is here. After #57797 merges, this PR can be rebased onto main and its helper files will disappear from Files changed.

Validation

• The V4-only diff contains 10 model call-site files; ruff check, ruff format --check, and git diff --check pass on the reconstructed combined tree.

• Static validation on the integrated final tree: ruff check, ruff format --check, git diff --check, and test tethering pass. The source audit passes 28 tests. The new runtime tests have not been executed, as requested.

AI assistance

AI assistance was used.

[mergify]

Documentation preview: https://vllm--57798.org.readthedocs.build/en/57798/

[Ronald1995]

Verification on H20

Verified commit: 842d6412c1ac5508c141a5d573c09e1c22996771 (this PR's head; it covers 10 DeepSeek V4 call sites (cpu/xpu/amd/nvidia target, xpu/amd/nvidia mtp, xpu/amd/nvidia dspark))
Integration commit: c623df9e4bfe1133a3205b94e2d771ba43876ec7 (attn-sink-integration = #57797 + #57798 + #57801 + #57803 + #57804 + #57805 + #57806)

Environment

• Cluster: H20 Slurm, node vllm-h20-01, 1x NVIDIA H20-3e (--partition=h20 --gres=gpu:h20:1)
• Python: $HOME/vllm-shared/prebuilt-venv/bin/python (torch 2.13.0+cu130 + compiled vllm from the shared prebuilt baseline)
• Pure-Python change, so the verified revision was overlaid onto the prebuilt package with make-overlay.sh; nothing was recompiled.

Results

Suite	Where	Result
test_attention_sink_load_audit.py (from #57806)	head node + GPU node	39 passed
test_attn_sink_reload.py (from #57797)	head node (CPU)	18 passed, 2 skipped (CUDA-only)
test_attn_sink_reload.py	H20 GPU	20 passed
GPU end-to-end sink reload script	H20 GPU	8/8 checks passed

The GPU script drives a real layerwise reload (record_metadata_for_reloading -> initialize_layerwise_reload -> load_weights -> finalize_layerwise_reload) on CUDA tensors and asserts the live update lands in the kernel tensor storage with the -inf padding tail re-established and the storage pointer unchanged:

PASS  live update reaches kernel tensor
PASS  live update re-establishes -inf padding
PASS  kernel tensor storage is stable
PASS  rank 1 live update loads its own heads
ALL GPU SINK CHECKS PASSED

Pre-commit

pre-commit run --files <changed files>
pass1 rc=0 dirty=0
pass2 rc=0 dirty=0

No hook rewrote a file. run-precommit.sh skips the pip-compile hooks because the sandbox cannot reach PyPI; no requirements file is touched here.

Not verified

• No full-model checkpoint was loaded (no matching weights available), so this is loader/contract-level plus a real GPU reload, not a full-model accuracy run.
• Single GPU; TP>1 was exercised in-process via distinct rank head ranges, not as a multi-process TP run.

What changed in this revision

• Imports point at the shared loader module; the stale vllm/models/deepseek_v4/sink.py copy is removed, so this PR no longer re-adds the helper and its Files changed will shrink to the ten call sites once [Bugfix][RL] Load attention sinks through weight_loader for layerwise reload #57797 merges.
• The test file is kept in sync with [Bugfix][RL] Load attention sinks through weight_loader for layerwise reload #57797's, including the new CUDA case.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.