[Bug] DeepSeek-V4 default prevents speculative decoding from resetting max-running-requests to 48

[junliu-mde]

Checklist

• I searched related issues but found no solution.
• The bug persists on current main (ae848116662ece923501131ce773a4602223237e).
• This report includes a minimal source-level reproduction.
• This report is in English.

Describe the bug

For DeepseekV4ForCausalLM, enabling speculative decoding without explicitly passing --max-running-requests leaves the effective value at 256.

This contradicts the DeepSeek-V4 cookbook:

Speculative decoding (MTP) is on — SGLang resets --max-running-requests to 48 when it is not set.

Cookbook: https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4

It also contradicts the warning and intended default in the speculative decoding hook:

sglang/python/sglang/srt/arg_groups/speculative_hook.py

Lines 396 to 400 in ae84811

	if server_args.max_running_requests is None:
	server_args.max_running_requests = 48
	logger.warning(
	"Max running requests is reset to 48 for speculative decoding. You can override this by explicitly setting --max-running-requests."
	)

The cause is default resolution order:

1. The DeepSeek-V4 model hook runs first and changes max_running_requests from None to 256:
   

   sglang/python/sglang/srt/arg_groups/deepseek_v4_hook.py

   Lines 141 to 145 in ae84811

   	if server_args.max_running_requests is None:
   	server_args.max_running_requests = 256
   	logger.warning(
   	f"Setting max_running_requests to {server_args.max_running_requests} for {model_arch}."
   	)

2. The speculative hook runs later, but only sets 48 when the field is still None.
3. The field is already 256, so the speculative default never applies.

The DeepSeek-V4 hook comment says that the speculative hook is a later writer of this field, but the is None guard means it cannot overwrite the model default:

sglang/python/sglang/srt/arg_groups/deepseek_v4_hook.py

Lines 106 to 115 in ae84811

	def apply_deepseek_v4_defaults(server_args: ServerArgs, model_arch: str) -> None:
	"""Residual imperative arm of the DeepSeek V4 defaults.
	The attention/page/window/MoE-runner declarations moved to the override
	registry (arg_groups/overrides.py: _deepseek_v4_overrides) and the
	kv-cache dtype default to the resolution pipeline
	(_deepseek_v4_kv_cache_dtype, invoked below at its legacy slot). This
	keeps, at the legacy slot: the ROCm env fill (env-write policy), the
	max_running_requests fill (the speculative hook is a later writer of
	that field) and the validations.

Explicit user input must keep its current precedence. Only the model-provided default should yield to the speculative-decoding default.

Reproduction

Start DeepSeek-V4 with speculative decoding and omit --max-running-requests, as in a cookbook-generated low-latency command:

sglang serve \
  --trust-remote-code \
  --model-path deepseek-ai/DeepSeek-V4-Flash \
  --tp 4 \
  --moe-runner-backend flashinfer_mxfp4 \
  --speculative-algorithm EAGLE \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 4

The same issue applies to --speculative-algorithm DSPARK, whose handler has the same is None condition.

Expected effective value:

max_running_requests = 48

Actual effective value:

max_running_requests = 256

Passing --max-running-requests 48 explicitly works around the issue.

Environment

• SGLang source: current main, commit ae848116662ece923501131ce773a4602223237e
• Model architecture: DeepseekV4ForCausalLM
• Reproduces through deterministic argument post-processing before GPU execution; it is not hardware-specific.