feat: add xlite as third npu backend (support qwen3-dense/qwen3-moe/glm-4.7/glm-5.1/glm-5.2)

[banxiaduhuo]

Description

This PR adds xlite as a third NPU backend for xLLM, alongside the existing ATB backend. xlite is a lightweight NPU inference runtime that directly calls CANN ops without the aclgraph abstraction layer, providing lower launch overhead and higher TTFT.

The adapter integrates xlite into xLLM's model layer through a thin C++ adapter (xlite/core/layers/xlite/), supporting five model architectures out of the box:

• Qwen3-Dense
• Qwen3-MoE
• GLM-4.7
• GLM-5.1
• GLM-5.2

About xlite

xlite is a lightweight XPU virtualization frontend and backend inference runtime. It provides a minimalist and efficient heterogeneous computing environment, supporting diverse computing power collaboration. See https://gitcode.com/openeuler/GVirt/blob/master/README.en.md.

Key changes

• Adapter layer (xllm/core/layers/xlite/, 9 files ~1700 lines): xlite_causal_lm_base / xlite_config_builder / xlite_weight_utils / xlite_attn_meta_builder / xlite_freqs_cis / xlite_init_utils / xlite_model_adapter / xlite_register_macros. Bridges xlite's XModel runtime into xLLM's CausalLMBase lifecycle (load → init → forward), mapping xLLM weight tensors to xlite MatmulWeight/MoEWeight with per-model quantization field binding.
• Per-model adapters (xllm/models/llm/xlite/adapters/, 5 files): qwen3_adapter / qwen3_moe_adapter / glm4_moe_adapter / glm5_adapter / deepseek_v3_adapter + register_all.h (299 lines, model registry wiring).
• Runtime (xlite_executor_impl.cpp/h): XliteExecutorImpl implements init/forward/sync over xlite's XModel::Forward, with node_rank % device_count auto device selection (no --devices needed).
• Build (cmake/xlite.cmake, CMakeLists.txt): USE_XLITE=ON links libxlite.so from site-packages, propagates -L/-rpath-link to all executables.
• W8A8 quantization support: IsW8A8(context) guard (derived from quant_model_description.json) replaces hardcoded quantAttnWeightTrans=true, fixing a BF16 shared-expert layout regression (_isSharedExpertWeightFull is read unconditionally at construction). MLA four STATIC projections (mlaQKVA/mlaQB/attnOut/indexQB) + down_proj deqScale + shared indexer binding for GLM-5.2.

Verification

功能验证

All milestones verified by smoke (script/smoke/verify.sh, EP8 减层 functional) + EvalScope perf + msprof per-step 算子级对比 vs ATB:

• Functional: Qwen3-0.6B TP1 / Qwen3-32B TP2 / Qwen3-30B-A3B MoE EP4 / GLM-4.7-W8A8 EP8 / GLM-5.1 BF16 / GLM-5.2-W8A8 EP8+DP2×EP8 — all pass.
• Correctness: 三方对比 (ref_xlite / xllm / ATB), L0 input 完全匹配 ref; 三方 (ref_xlite / xllm / ATB) 逐 token 完全一致.
• Performance: 28 层减层 GLM-5.2-W8A8 EP8 vs ATB graph (各自最优可用配置), 见下方性能对比数据.

EvalScope perf — input_32 (decode 为主, prompt=32tok, output=128tok)

并发	指标	xlite	ATB graph	差异
1	TPOT(ms) P50/Avg	12.3 / 12.3	11.9 / 11.9	+3% (持平)
1	TTFT(ms) P50/Avg	30.2 / 30.5	71.9 / 75.4	-58% / -59%
1	Gen/s	80.40	61.50	+31%
4	TPOT(ms) P50/Avg	14.6 / 14.6	15.1 / 15.1	-3% (xlite 略优)
4	TTFT(ms) P50/Avg	55.2 / 51.2	98.0 / 115.6	-44% / -56%
4	Gen/s	269.35	250.86	+7%
16	TPOT(ms) P50/Avg	20.0 / 20.0	20.2 / 20.4	-1% / -2% (持平)
16	TTFT(ms) P50/Avg	78.2 / 74.0	214.8 / 196.8	-64% / -62%
16	Gen/s	782.44	730.40	+7%

EvalScope perf — input_2000 (prefill 重 + DSA 稀疏, prompt=2000tok, output=128tok)

并发	指标	xlite	ATB graph	差异
1	TPOT(ms) P50/Avg	13.8 / 13.8	13.9 / 13.9	-1% (持平)
1	TTFT(ms) P50/Avg	128.7 / 129.0	319.4 / 319.5	-60%
1	Gen/s	68.01	49.59	+37%
4	TPOT(ms) P50/Avg	17.3 / 17.6	20.9 / 19.4	-17% / -9%
4	TTFT(ms) P50/Avg	314.9 / 278.2	1144.9 / 809.3	-73% / -66%
4	Gen/s	204.20	155.50	+31%
16	TPOT(ms) P50/Avg	31.8 / 31.8	41.3 / 42.8	-23% / -26%
16	TTFT(ms) P50/Avg	421.8 / 421.8	1678.7 / 1444.1	-75% / -71%
16	Gen/s	458.94	284.03	+62%

Related Issues

Change Type

• Bug fix
• New feature
• Performance improvement
• Refactor
• Documentation
• Test
• Build or CI

Pull Request Checklist

Thank you for contributing to xLLM. Before requesting review, please make sure the following items are complete.

PR Title and Commit Messages

• The PR title and each commit message follow the xLLM commit format: <type>: <subject>.

Pre-commit Checks

• I have installed pre-commit by running pip install pre-commit or an equivalent command.
• I have installed the hooks with pre-commit install.
• I have run pre-commit run --all-files and fixed any reported issues.

If you are unsure how to set up pre-commit, see the pre-commit documentation.

Self Review

• I have self-reviewed the code according to .agents/skills/code-review/references/custom-code-style.md, especially code written or assisted by AI.
• I have rebased this PR onto the latest main branch.

Build and Test Coverage

• Tests have been added or updated as needed.
• CUDA: python setup.py build test has passed on a CUDA machine.
• NPU: python setup.py build test has passed on an NPU machine.
• MLU: python setup.py build test has passed on an MLU machine.

Reviewer Notes