[DFlash] Support grammar constrained decoding support for DFlash

[dcw02]

• Enables grammar-constrained decoding during DFlash speculative verification.
• Adds reusable DFlash grammar mask handling for JSON/grammar token filtering.
• Covers DFlash spec-v2 constrained decoding through the existing spec feature kit.

Server command:

SGLANG_ENABLE_OVERLAP_PLAN_STREAM=1 venv/bin/python -m sglang.launch_server \
    --model-path Qwen/Qwen3-8B \
    --host 127.0.0.1 --port 20554 \
    --trust-remote-code \
    --attention-backend trtllm_mha \
    --tp-size 1 \
    --dtype bfloat16 \
    --max-running-requests 32 \
    --cuda-graph-max-bs 32 \
    --mamba-scheduler-strategy no_buffer \
    --enforce-piecewise-cuda-graph \
    --speculative-algorithm DFLASH \
    --speculative-draft-model-path z-lab/Qwen3-8B-DFlash-b16 \
    --speculative-draft-attention-backend fa4 \
    --grammar-backend xgrammar

Question:

Janet's ducks lay 16 eggs per day. She eats 3 for breakfast and uses 4 to bake muffins. She sells the rest for $2 per egg. How much money does she make every day?

Unconstrained request:

sampling_params = {
    "temperature": 0.0,
    "top_p": 1.0,
    "top_k": 1,
    "max_new_tokens": 1024
}

Unconstrained output:

... Janet makes \boxed{18} dollars every day.

Grammar-constrained request:

sampling_params = {
    "temperature": 0.0,
    "top_p": 1.0,
    "top_k": 1,
    "max_new_tokens": 1024,
    "json_schema": "{\"type\":\"object\",\"properties\":{\"reasoning\":{\"type\":\"string\"},\"answer\":{\"type\":\"integer\"}},\"required\":
    [\"reasoning\",\"answer\"],\"additionalProperties\":false}"
}

Grammar-constrained output:

{
    "reasoning": "Janet's ducks lay 16 eggs per day. She uses 3 eggs for breakfast and 4 eggs for muffins, so she uses 3 + 4 = 7 eggs. The remaining eggs
    are 16 - 7 = 9 eggs. She sells these 9 eggs at $2 per egg, so the total money made is 9 * 2 = $18.",
    "answer": 18
}

---

CI States

Latest PR Test (Base): ❌ Run #27791182071
Latest PR Test (Extra): ❌ Run #27791181896
Latest PR Test (AMD ROCm 10): ❌ Run #27791182028

[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!

[dcw02]

/rerun-test test/registered/spec/dflash/test_dflash.py

[github-actions]

Results for /rerun-test test/registered/spec/dflash/test_dflash.py:

🚀 1-gpu-5090 (1 test): ❌ View workflow run

cd test/ && python3 registered/spec/dflash/test_dflash.py

[dcw02]

/rerun-test test/registered/spec/dflash/test_dflash.py

[github-actions]

Results for /rerun-test test/registered/spec/dflash/test_dflash.py:

🚀 1-gpu-5090 (1 test): ✅ View workflow run

cd test/ && python3 registered/spec/dflash/test_dflash.py