ref(upstream): FULL TREE — Deep-Spark xllm (1470) + ds_vllm csrc/models (703)

Replaces cherry-picked upstream_ref with complete source trees.

xllm/ — Iluvatar official C++ inference engine (15MB, 1470 files)
  Complete: kernels → layers → models → runtime → scheduler → api
  Excluded: .git, binary images, third_party submodule checkouts

ds_vllm/ — Iluvatar official vllm fork (8MB, 703 files)
  Included: csrc/ (ALL CUDA kernels), fused_moe/, qwen3_5 model, _custom_ops
  Excluded: tests, benchmarks, docs, examples (not needed for reference)

Critical call chains now fully traceable:
  MoE: moe_topk_softmax_kernels.cuh → ixformer.h → fused_moe.cpp → layer
  GDN: qwen3_gated_delta_net_base.cpp → qwen3_5_gated_delta_net.cpp
  Attention: ixformer.h → xllm_paged_attention → attention.cpp

upstream_ref/xllm/examples/sample.py

@@ -0,0 +1,51 @@
+# python examples/sample.py --model='/path/models/Qwen3-8B' --devices='npu:0'
+# python examples/sample.py --model='/path/models/Qwen3-8B' --devices='npu:0,npu:1'
+
+from xllm import ArgumentParser, LLM, RequestParams
+
+# Create an LLM.
+parser = ArgumentParser()
+llm = LLM(**vars(parser.parse_args()))
+
+# selector must be a stable single special token in the target tokenizer.
+selector = "masked"
+
+prompts = [
+    f"candidate_a={selector}, candidate_b={selector}",
+    f"user_feature={selector}",
+]
+
+# RequestParams can still carry generic sampling knobs.
+# sample() will enforce:
+# - max_tokens=1
+# - n=1
+# - best_of=1
+# - logprobs=True
+request_params_list = []
+for _ in prompts:
+    request_params = RequestParams()
+    request_params.temperature = 0.0
+    request_params.top_p = 1.0
+    request_params_list.append(request_params)
+
+outputs = llm.sample(
+    prompts,
+    selector=selector,
+    request_params=request_params_list,
+    logprobs=5,
+    wait_schedule_done=True,
+)
+
+# One RequestOutput per input prompt.
+# output.outputs is expanded by selector hits in that prompt.
+for i, output in enumerate(outputs):
+    print(f"[prompt-{i}] {output.prompt!r}")
+    for sample_output in output.outputs:
+        print(
+            f"  sample_id={sample_output.index}, "
+            f"token={sample_output.text!r}, "
+            f"token_ids={sample_output.token_ids}, "
+            f"finish_reason={sample_output.finish_reason}"
+        )
+
+llm.finish()