Motivation

PrefillDelayer is currently gated by assert disaggregation_mode == "null" in its constructor, so launching a disaggregated prefill engine with --enable-prefill-delayer aborts at scheduler init. However, the delayer's negotiation logic — per-iteration all_gather over DP ranks of an all/none/mixed prefillable status, with bounded max_delay_passes waiting — is exactly the mechanism a disaggregated-prefill + --enable-dp-attention deployment needs to fuse near-simultaneous prefills into a single forward pass instead of paying for repeated "1 real + N idle" passes through the EP all-to-all.

In a --tp 4 --dp 4 --ep 4 --enable-dp-attention --disaggregation-mode prefill setup, four concurrent client requests routed round-robin to four DP ranks currently end up as two scheduler iterations:

1. Iteration 1: DP0 prefills its req, DP1/2/3 run idle batches (still paying the EP collective)
2. Iteration 2: DP1/2/3 prefill their reqs

With this change, --enable-prefill-delayer can be used on the prefill engine to delay iteration 1 by a few passes so all four reqs land in the same forward, halving the number of EP all-to-alls and improving prefill throughput at small/medium prompt lengths where the per-pass collective overhead is non-trivial relative to the actual MoE compute.

Modifications

• python/sglang/srt/managers/prefill_delayer.py: drop the disaggregation_mode == "null" assertion. The delayer's logic only depends on the per-iteration prefill scheduling path, which exists in both null and prefill modes.
• python/sglang/srt/managers/scheduler.py: when constructing the delayer, skip it (and log) on a decode engine. A decode engine has no prefill scheduling path, so --enable-prefill-delayer would be silently unused; this makes that explicit and avoids allocating the gather buffer / cpu_group state on decode workers.

No behavior change for non-disaggregated deployments. Existing --prefill-delayer-max-delay-passes / --prefill-delayer-token-usage-low-watermark flags work unchanged on the prefill engine of a disaggregated PD setup.

Accuracy Tests

N/A — this is a scheduler-side gating change. No model forward / kernel code is touched. Generation outputs are unaffected.

Speed Tests and Profiling

Tested locally on 1×H200 node, Qwen/Qwen3-30B-A3B, prefill engine launched as:

python3 -m sglang.launch_server --model-path Qwen/Qwen3-30B-A3B \
  --tp 4 --dp 4 --ep 4 --enable-dp-attention \
  --disaggregation-mode prefill --disable-radix-cache \
  --load-balance-method round_robin --chunked-prefill-size 32768 \
  [--enable-prefill-delayer]

(Without this PR the --enable-prefill-delayer variant aborts at PrefillDelayer.__init__ with the assertion.)

Client: 4 concurrent requests, 8192-token prompt, 1 output token, round-robin over the 4 DP ranks, no client-side stagger.

Without --enable-prefill-delayer: requests split into two batches

python3 client.py --batch-size 4 --input-len 8192 --output-len 1 \
  --tokenizer Qwen/Qwen3-30B-A3B --dp-size 4 --flush-cache --stagger-ms 0
sent 4 reqs in 3.023s, ok=4/4

Per-DP Prefill batch log on the prefill engine — DP0 fires alone first, DPs 1/2/3 fuse in the next iteration:

DP0 ... Prefill batch, #new-seq: 1, #new-token: 8192   <- iter 1
DP1 ... Prefill batch, #new-seq: 1, #new-token: 8192   <- iter 2
DP2 ... Prefill batch, #new-seq: 1, #new-token: 8192   <- iter 2
DP3 ... Prefill batch, #new-seq: 1, #new-token: 8192   <- iter 2

With --enable-prefill-delayer: requests fused into one batch

python3 client.py --batch-size 4 --input-len 8192 --output-len 1 \
  --tokenizer Qwen/Qwen3-30B-A3B --dp-size 4 --flush-cache --stagger-ms 0
sent 4 reqs in 1.957s, ok=4/4

All four DPs prefill in the same iteration:

DP0 ... Prefill batch, #new-seq: 1, #new-token: 8192   <- iter 1
DP1 ... Prefill batch, #new-seq: 1, #new-token: 8192   <- iter 1
DP2 ... Prefill batch, #new-seq: 1, #new-token: 8192   <- iter 1
DP3 ... Prefill batch, #new-seq: 1, #new-token: 8192   <- iter 1

End-to-end wall clock for 4 concurrent reqs drops from 3.023 s → 1.957 s (~35% faster) because the EP all-to-all is amortized over one forward pass instead of two.

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.