What causes "This browser cannot encode H.264 MP4" in sgl-project/sglang?

Fix the config string so every segment is exactly 'L-R|value', e.g. '2-4|8'

sgl-project/sglang · error · Error

This browser cannot encode H.264 MP4

Error message

This browser cannot encode H.264 MP4

What it means

Thrown by ngram_corpus Param::parse while parsing a spec-configuration segment. Each segment must split on '|' into exactly two parts: a positional range and a config value. Seeing part's size != 2 means the segment is malformed — missing the '|' separator or containing extra ones.

Source

Thrown at python/sglang/multimodal_gen/apps/realtime_webui/app.js:669

    { codec: "avc1.4d4028", width, height, bitrate, framerate: fps },
    { codec: "avc1.42e028", width, height, bitrate, framerate: fps },
  ];
  let supported = null;
  for (const config of configs) {
    const candidate = {
      ...config,
      avc: { format: "avc" },
      bitrateMode: "variable",
      hardwareAcceleration: "prefer-hardware",
      latencyMode: "realtime",
    };
    const result = await VideoEncoder.isConfigSupported(candidate);
    if (result.supported) {
      supported = result.config;
      break;
    }
  }
  if (!supported) throw new Error("This browser cannot encode H.264 MP4");
  recordingEncoderConfig = supported;
  recordingEncoder = new VideoEncoder({
    output: (chunk, metadata) => recordEncodedChunk(chunk, metadata),
    error: (error) => {
      recordingActive = false;
      addHistory(error.message || "recording encoder failed");
      updateRecordButton();
    },
  });
  recordingEncoder.configure(supported);
}

function recordEncodedChunk(chunk, metadata) {
  if (metadata?.decoderConfig?.description) {
    recordingEncoderConfig.description = metadata.decoderConfig.description;
  }
  const data = new Uint8Array(chunk.byteLength);
  chunk.copyTo(data);

View on GitHub (pinned to 0132848349)

Solutions

Fix the config string so every segment is exactly 'L-R|value', e.g. '2-4|8'
Print/inspect the segments the parser sees (the code already dumps each 'part' to stderr) to find the malformed one
Validate the config format in Python before passing it into the C++ API

Example fix

# before
param.resetBatchReturnTokenNum("2-4 5-8|3")

# after
param.resetBatchReturnTokenNum("2-4|3 5-8|3")

Defensive patterns

Strategy: validation

Validate before calling

import re
SEG = re.compile(r"^\d+-\d+\|\d+$")
assert all(SEG.match(s) for s in cfg.split()), f"bad ngram config segment in {cfg!r}"

Type guard

def is_valid_ngram_config(cfg: str) -> bool:
    segs = cfg.split()
    return bool(segs) and all(re.fullmatch(r"\d+-\d+\|\d+", s) for s in segs)

Try / catch

try:
    param.resetBatchReturnTokenNum(cfg)
except RuntimeError as e:
    if "invalid config" in str(e):
        raise ValueError(f"ngram config malformed: {cfg!r} (expected 'L-R|value' segments)") from e
    raise

Prevention

Build config strings from structured (L, R, value) tuples via a single formatter f"{L}-{R}|{v}"
Strip whitespace and trailing separators before passing to C++
Unit-test the config parser against valid and invalid samples

When it happens

Trigger: resetBatchReturnTokenNum is called with a config string whose pipe-separated segments don't each contain exactly one '|', e.g. "2-4" (no config part), "2-4|8|9" (extra part), or an empty/trailing segment like "2-4|8|".

Common situations: Typos in the speculative n-gram config string passed via CLI/server args, config strings assembled programmatically with a missing or duplicated '|' delimiter, or trailing separators after string joins.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/3188e757d98d56b6. Report an issue: GitHub.