[codex] Fix Kunlun Qwen2.5 greedy inference

[zhangyue207]

Summary

This PR fixes the InfiniLM-side issues found while bringing up Qwen2.5-7B-Instruct single-card inference on Kunlun XPU.

Changes

• Propagate CLI dtype into LLM / InferEngine / HF config loading so --dtype float32 and other overrides are honored.
• Convert floating safetensors to the requested dtype during model load.
• Map the Python-side kunlun device string to the underlying InfiniCore CUDA-compatible device mapping used by the Kunlun backend.
• Use int32 position IDs from the basic LLM processor for Kunlun RoPE compatibility.
• Replace the generic Qwen2 attention and MLP fused linears with explicit q_proj/k_proj/v_proj and gate_proj/up_proj modules, keeping checkpoint parameter names direct.
• Fix static attention RoPE by applying Q RoPE in-place instead of writing into a non-contiguous permuted output tensor.
• Make VideoNSA follow the updated common Attention members.
• Add a Kunlun greedy decoding fallback: when top_k == 1, copy logits to CPU and compute argmax there to avoid the incorrect Kunlun random sampling result.

Root Cause

The main correctness issue was in static attention: Q RoPE wrote into a non-contiguous permuted tensor. Embedding and input RMSNorm matched Hugging Face exactly, but layer 0 attention diverged immediately. Switching Q RoPE to the in-place single-tensor path made layer 0 match Hugging Face within about 1e-6 and restored the expected first generated token.

Kunlun sampling also selected the wrong token even when logits were correct, so greedy decoding now uses a CPU argmax fallback for top_k=1 on Kunlun.

Validation

Validated with Docker image fixing:v1, local InfiniCore/InfiniLM builds, and model /workspace/models/Qwen2.5-7B-Instruct.

• Built and installed _infinilm.
• CPU FP32 full prompt matched Hugging Face first-token reference:
  • top token 114602
  • score about 31.068
  • decoded token 昆仑
• Kunlun FP32 single-card max-new-tokens=1:
  • CPU argmax token 114602
  • score about 31.06
  • response 昆仑
• Kunlun FP32 single-card max-new-tokens=2:
  • response 昆仑芯

Runtime note: the verified path uses DISABLE_XPYTORCH=1 to avoid loading an incompatible XPU runtime from torch_xmlir. FP32 decode is currently slow, about 29 seconds per generated token in this setup.