[NPU] Enable non-greedy MTP sampling

[cl-vv-h]

Motivation

NPU MTP verification currently uses greedy argmax for non-greedy requests.
Consequently, temperature, top_k, and top_p do not take effect during
speculative verification.

With GLM-4.7-Flash on GPQA, enabling non-greedy MTP reduced the observed
accuracy from approximately ~70% to ~50%. This PR enables
probability-based non-greedy MTP sampling on NPU.

Modifications

• Add an explicit NPU non-greedy branch in eagle_sample.
• Keep NPU-specific sampling logic in
  srt/hardware_backend/npu/speculative/mtp_sampling.py.
• Apply temperature scaling and sequential top-k/top-p probability
  renormalization.
• Support target-only sampling for general EAGLE trees and classic rejection
  sampling for tree_topk=1 chains.
• Validate input topology and probability tensors before dispatching kernels.
• Preserve the existing greedy path and fail explicitly for unsupported
  non-greedy configurations.

Sampling algorithm and kernel unit tests are covered in sgl-kernel-npu; no
duplicate SGLang-local test is added.

Accuracy Tests

Model: GLM-4.7-Flash
Dataset: GPQA

Configuration	Sampling	Before	After
NPU TP/TP+EP without MTP	temperature=1, top_p=0.95	~70%	~70%
NPU TP+MTP/TP+EP+MTP	temperature=1, top_p=0.95	~50%	~70%
NPU MTP greedy	temperature=0, top_p=1	~50%	~50%

The non-greedy MTP accuracy recovered to the expected range, while existing
non-MTP and greedy behavior remained unchanged.
before:

after:

Speed Tests and Profiling

Check	Result
End-to-end throughput	No measurable performance impact
End-to-end latency	No measurable performance impact

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #33240931809
Latest PR Test (Extra): ❌ Run #33240931714
Latest PR Test (AMD ROCm 7.2): ❌ Run #33240931803

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[liupeng374]

Is there too much isolation here? Are there too many duplicate codes in the newly added files?

[cl-vv-h]

Updated. NPU now shares the existing non-greedy sampling flow, with only backend-specific operator selection and tensor preparation kept in the NPU branch. The duplicated implementation have been removed.

[Hexq0210]

same as 797

[Hexq0210]

Run the import as command to normalize the names of the methods invoked by the NPU and GPU. That is, you can write the names only once.
sampling_fn = (
chain_speculative_sampling_triton
if use_rejection_sampling
else tree_speculative_sampling_target_only

[cl-vv-h]

Updated. Changes here have been minimized.

[sglang-npu-bot]

/tag-and-rerun-ci

[cl-vv-h]

/rerun-failed-ci

[cl-vv-h]

/rerun-failed-ci

[cl-vv-h]

/rerun-failed-ci

[cl-vv-h]

/rerun-failed-ci

[cl-vv-h]

/rerun-failed-ci

[cl-vv-h]

/rerun-failed-ci