[Qwen3.5] Enable MTP spec_v2 and add test for nvidia/Qwen3.5-397B-A17B-NVFP4

[hlu1]

Motivation

• Make MTP_v2 work for Qwen3.5 by passing mm_input_embeds to the MTP head.
• Add MTP_v1/v2 and non-MTP accuracy test for nvidia/Qwen3.5-397B-A17B-NVFP4 and check acceptance length in the MTP tests. Note that it uses the eval harness from sglang.test.run_eval import run_eval which applies the chat_template. Without the chat_template, the accuracy is very bad. The sampling parameters are based on the official recommendation from https://huggingface.co/Qwen/Qwen3.5-397B-A17B-FP8
• Remove the two extra_buffer and mtp_v2 checks. These two checks are incorrect because mtp_v2 only requires extra_buffer when radix_cache is on. To make it more user friendly, I changed the default behavior of turning off radix-cache silently when spec decoding, no_buffer, and radix cache are enabled at the same time to raising an exception, in case the user want to enable spec decoding (v1 or v2) and radix cache but forgot to enable extra_buffer.
• Remove duplicated code from server_args.py

co-author: @vincentzed #18906

Accuracy

Without radix cache

gpqa:
Repeat: 8, mean: 0.866
Scores: ['0.859', '0.869', '0.874', '0.869', '0.884', '0.848', '0.869', '0.859']

With radix cache

Repeat: 8, mean: 0.861
Scores: ['0.848', '0.843', '0.874', '0.854', '0.864', '0.859', '0.894', '0.854']

Benchmark

radix cache off

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.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
4. After green CI and required approvals, ask Merge Oncalls to merge.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[hlu1]

/tag-and-rerun-ci

[hlu1]

/tag-and-rerun-ci

[vincentzed]

I tested trtllm_mha under this as well:
| Latency (s) | Tokens | Acc Length | Speed (token/s) |
| 3.088 | 512 | 3.413 | 165.82 |

Triton
| 2.177 | 512 | 3.303 | 235.23 |

[ShangmingCai]

LGTM as long as the CI passes. CC: @yizhang2077 Please double check.

[hzh0425]

LGTM

[hlu1]

The gb200 CI is temporarily disabled. I ran the tests locally and they all pass.

[hlu1]

/rerun-failed-ci

[hlu1]

/rerun-failed-ci

[hlu1]

/rerun-failed-ci

[hlu1]

/rerun-failed-ci

[hlu1]

Both test/registered/4-gpu-models/test_qwen3_next_models_mtp.py and test/registered/4-gpu-models/test_qwen35_models.py have passed in the latest CI run.