Range resume corrupts files: a 200 response that serves the requested range is treated as a restart

[ryanzhangtianran]

The ModelScope file endpoint answers a Range request with 200 OK instead of 206 Partial Content, even though it does serve only the requested slice:

curl -sL -o /dev/null -D -  
-H 'Range: bytes=1000000000-1000000099' 
'https://www.modelscope.cn/api/v1/models/Qwen/Qwen3-VL-8B-Instruct/repo？Revision=master&FilePath=model-00002-of-00004.safetensors'

Returned:

HTTP/1.1 200 OK
Server: nginx/1.24.0
Content-Length: 100
Accept-Ranges: bytes
Content-Range: bytes 1000000000-1000000099/4915962496

100 bytes requested, 100 bytes returned, correct Content-Range — the range is honoured. Only the status code is wrong. Per RFC 9110 §15.3.7 this should be 206.

And it may break the SDK

_download_with_resume decides whether a response is a resume from the status code alone.

is_resumed = resp.status_code == 206
...
mode = "ab" if is_resumed else "wb"
if not is_resumed:
    existing_size = 0

So on every resume is_resumed is False, and the .incomplete file is truncated (wb) — but the server is still sending from the requested offset, so mid-file bytes get written at offset 0. The result is a file that is both short and shifted. Nothing notices until the SHA-256 check rejects it, which then deletes the file and restarts from byte 0.

In one run over an unreliable link, /proc//io showed 30.9 GB written for a 17.5 GB model, with 6.1 GB retained.

[ryanzhangtianran]

I added and modified some code to resolve this issue temporarily .

Add:

_CONTENT_RANGE_START_RE = re.compile(r"\s*bytes\s+(\d+)\s*-")


# ModelScope answers Range requests with 200 instead of 206, so Content-Range
# is the only reliable signal that a resume actually took effect.
def _serves_range_from(resp: requests.Response, offset: int) -> bool:
    """Whether ``resp`` delivers the file starting at ``offset``."""
    if offset <= 0:
        return False
    match = _CONTENT_RANGE_START_RE.match(resp.headers.get("Content-Range", ""))
    if match is not None:
        return int(match.group(1)) == offset
    return resp.status_code == 206

And modify the code in _download_with_resume ():

content_length = resp.headers.get("Content-Length")
total_size = int(content_length) if content_length else None
is_resumed = _serves_range_from(resp, existing_size) // this line is modified

if is_resumed and total_size:
    total_size += existing_size