[vLLM-ATOM] Fix GLM-4.7 MTP in vLLM plugin

[kliuae]

Motivation

In the current vLLM-ATOM upstream, GLM-4.7 crashes with arguments like lm_head and embed_tokens not found and conflicting compile cache key, when MTP is enabled. This PR adds fixes to MTP to address these issues.

Technical Details

• Expose speculative decode attributes, which is fine for qwen3next mtp too as its lm_head lives in model, not inner
• Mask inputs for glm4-mtp
• Separate main model and draft model configs so the draft model doesn't load from the main model's compiled artifact

Test Plan

lm_eval with gsm8k

Test Result

Server command

ATOM_ENABLE_QK_NORM_ROPE_CACHE_QUANT_FUSION=1 \
AITER_QUICK_REDUCE_QUANTIZATION=INT4 \
SAFETENSORS_FAST_GPU=1 \
ATOM_DISABLE_VLLM_PLUGIN=0 \
ATOM_DISABLE_VLLM_PLUGIN_ATTENTION=0 \
VLLM_LOGGING_LEVEL=INFO \
VLLM_USE_V1=1 \
VLLM_ROCM_USE_AITER=1 \
vllm serve zai-org/GLM-4.7-FP8 \
  -tp 4 \
  --gpu-memory-utilization 0.9 \
  --no-enable-prefix-caching \
  --disable-uvicorn-access-log \
  --trust-remote-code \
  --compilation-config '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' \
  --kv-cache-dtype fp8 \
  --speculative-config.method mtp \
  --speculative-config.num_speculative_tokens 1

Eval command

lm_eval --model local-completions   --model_args model=zai-org/GLM-4.7-FP8,base_url=http://localhost:8000/v1/completions,num_concurrent=64,tokenized_requests=False  --tasks gsm8k --num_fewshot 5

Eval score

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	_	0.9462	_	0.0062
		strict-match	5	exact_match	_	0.9454	_	0.0063

Submission Checklist

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