[Ronald1995]

Fix HY-V4 checkpoint-backed learnable sink loading in NVIDIA target/MTP and AMD target models. The AMD target path was omitted from the original series and is included here.

The runtime parameter already has local TP shape. load_weights takes the rank slice once and passes it to the parameter's current weight_loader; no padded helper or second sharding loader is needed. The AMD and NVIDIA target paths preserve their TP slicing, load the local shard through the current weight_loader, and use the same shared loaded_params path.

Validation

• This PR changes only three HY-V4 files. ruff check, ruff format --check, and git diff --check pass.

• 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--57803.org.readthedocs.build/en/57803/

[Ronald1995]

Verification on H20

Verified commit: 69031c5a1480db8eeb8eb1993befdc8a864c56a7 (this PR's head; it covers 3 HY-V4 call sites (amd/nvidia target, nvidia mtp))
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

• The AMD and NVIDIA target sink branches still fell through into the generic weight-remapping code after loading the sink. They now continue, matching the MTP branch.
• Both target branches also now add the sink to loaded_params, so callers that validate "all parameters loaded" see the truth. The MTP branch already did.
• All three sites use the shared load_padded_attn_sink helper instead of a second copy of the inline loader lookup. HY-V4's parameter already has the local head shape, so the helper's padding path is a no-op there.

[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.