[ROCm][Bugfix] Ignore redundant qzeros in symmetric RDNAHybridW4A16

[tangzzycc]

Purpose

Symmetric GPTQ uint4b8 uses a fixed zero point of 8, but some checkpoints still contain redundant packed qzeros. RDNAHybridW4A16LinearKernel previously passed this tensor to the kernel, causing startup failures such as:

AssertionError: zp shape mismatch: torch.Size([40, 4352]) vs (34816, 40)

This PR drops redundant qzeros after loading symmetric GPTQ weights and adds regression coverage. The existing asymmetric zero-point path is unchanged.

Duplicate-work check

Open PR searches for RDNAHybridW4A16LinearKernel qzeros and symmetric qzeros ROCm W4A16 found no duplicate. PRs #47770 and #48998 only address TritonW4A16LinearKernel; neither changes the RDNA hybrid kernel.

Test Plan

Component	Value
GPU	AMD Radeon 8060S (gfx1151)
Memory	128 GB unified memory (96 GiB VRAM)
Operating system	Ubuntu 24.04.4 LTS
ROCm	7.2.3

I ran the focused and complete hybrid W4A16 tests, applicable pre-commit hooks, and an end-to-end before/after serving test with the official Qwen/Qwen3.5-27B-GPTQ-Int4 checkpoint.

Test commands

VLLM_ROCM_USE_AITER=0 VLLM_ENABLE_V1_MULTIPROCESSING=0 \
.venv/bin/vllm serve ~/models/Qwen3.5-27B-GPTQ-Int4 \
  --language-model-only --attention-backend ROCM_ATTN \
  --max-model-len 2048 --max-num-seqs 1 \
  --gpu-memory-utilization 0.3 --enforce-eager --port 8001

Test Result

Revision	Result
Before this PR	Startup failed during the profiling forward pass with the qzeros shape mismatch above
With this PR	Engine startup completed and the Chat Completions API generated 32 tokens successfully

• Focused weight-processing tests: 6 passed, 77 deselected, 14 warnings
• Complete hybrid W4A16 test file: 83 passed, 14 warnings
• Applicable pre-commit hooks: passed
• git diff --check: passed

AI assistance

OpenAI Codex assisted with this change. I reviewed and understand every submitted line and ran the validation reported above.

No documentation update is required because this change does not alter any user-facing API or configuration.

---

Essential Elements of an Effective PR Description Checklist

• Purpose and failure mode documented.
• Duplicate PRs checked.
• Test commands and results included.
• End-to-end model serving validated.
• Documentation impact considered.
• AI assistance disclosed.

[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.

[mgehre-amd]

Can this be removed in the GPTQ loader instead of the RDNAHybridW4A16 kernel? I guess the same mismatch between c.zero_points and self.w_zp_name might also confuse other kernels.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @tangzzycc.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork