fix(scheduler): stop the min-free-slots delay once every waiting request fits

[jhinpan]

Motivation

For the DFlash family (DSpark included) the scheduler auto-enables MinFreeSlotsDelayer, which holds fresh prefills until min(4, max(2, (max_running_requests + 5) // 6)) request slots are free so they are admitted together. When fewer requests than that are waiting and all of them already fit, waiting cannot batch anything more; it only delays them until running requests finish.

On DeepSeek-V4.1-Flash with DSpark and 256 concurrent requests, the server settles at 254 running with 2 queued. With 2 free slots and a threshold of 4, those 2 requests wait for the whole decode of the batch (about 57 s), which is the 86-90 s maximum TTFT of every such burst.

Modifications

• MinFreeSlotsDelayer.should_delay takes the number of waiting requests and delays only while the free slots are fewer than both the threshold and the waiting requests. With at least threshold requests waiting the behavior is unchanged.
• The scheduler passes len(self.waiting_queue).
• Unit test: 254 running, 2 free, 2 waiting is admitted; 3 waiting still waits.

Accuracy Tests

GSM8K (64 fixed questions, natural EOS, concurrency 64) on every server: 61-63/64 in both arms.

Speed Tests and Profiling

MI355X x4, TP4/EP4, DeepSeek-V4.1-Flash dba1be0a, this branch at e2e824dc58, the cookbook MI350X Low-Latency cell (DSpark block 5) with --cuda-graph-max-bs-decode 256, AITER built as in this branch's Dockerfile, plus #8's int64 FlashMLA store fix in both arms. 256 distinct real-text 4096-token prompts submitted at once, 3072 forced output tokens, temperature 0; fresh servers in ABBA order, two scored bursts per server.

Metric	Base	This PR	Change (95% CI)
Burst output tok/s	6966	7315	+5.02% [+4.42, +5.65]
TTFT max	89.1 s	32.0 s	-64.1%
TTFT median	15.9 s	15.9 s	+0.08% [-0.07, +0.28]
Peak running requests	254	256

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): ❌ Missing run-ci label -- add it to run CI tests.
Latest PR Test (Extra): ❌ Blocked -- run-ci is required first.
Latest PR Test (AMD ROCm 7.2): ➖ No AMD PR run found for this commit.

[jhinpan]

Carried onto dsv41-amd-5-integration (the sgl-project#41308 head) in #18 and re-measured there. Closing in favor of #18.