[vllm-atom] Fix GLM-5 accuracy in vLLM plugin

[kliuae-amd]

Motivation

This PR fixes the accuracy drop of GLM-5 in vLLM-ATOM mode.

Technical Details

Build ragged layout in every layer.
Use top_k_per_row_prefill from vLLM that handles indexing more consistently across short/long context lengths.
In combination with bugfix for cp_gather_indexer_k_quant_cache in aiter ROCm/aiter#2954, accuracy is restored to baseline.

Test Plan

Accuracy test with lm_eval

Model: zai-org/GLM-5-FP8

Server command

ATOM_DISABLE_VLLM_PLUGIN=0 \
ATOM_DISABLE_VLLM_PLUGIN_ATTENTION=0 \
vllm serve zai-org/GLM-5-FP8 \
  -tp 8 \
  --gpu-memory-utilization 0.7 \
  --no-enable-prefix-caching \
  --disable-uvicorn-access-log \
  --trust-remote-code \
  --compilation-config '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' \
  --kv-cache-dtype fp8 \
  --block-size 1

lm_eval command:

lm_eval --model local-completions   --model_args model=zai-org/GLM-5-FP8,base_url=http://localhost:8000/v1/completions,num_concurrent=64,tokenized_requests=False  --tasks gsm8k --num_fewshot 20

Test Result

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	20	exact_match	_	0.9431	_	0.0064
		strict-match	20	exact_match	_	0.9431	_	0.0064

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.