fix(http): serve a body the transport already read through the bounded reader

litellm/llms/custom_httpx/http_handler.py

@@ -630,6 +630,17 @@ async def _wire_bounded(response: httpx.Response, limit: int) -> AsyncGenerator[
         yield chunk
 
 
+def _already_read_within(response: httpx.Response, max_bytes: int) -> httpx.Response:
+    if len(response.content) > max_bytes:
+        raise HTTPResponseLimitError("Response exceeds the configured size limit")
+    return httpx.Response(
+        response.status_code,
+        headers=_headers_of_the_decoded_body(response.headers),
+        content=response.content,
+        request=response.request,
+    )
+
+
 async def _decoded_within(response: httpx.Response, wire: AsyncGenerator[bytes, None], max_bytes: int) -> bytes:
     decoding: Final = httpx.Response(
         response.status_code, headers=response.headers, content=wire, request=response.request
@@ -817,6 +828,8 @@ class AsyncHTTPHandler:
                 )
             if response.is_redirect or response.is_error:
                 return httpx.Response(response.status_code, headers=response.headers, request=response.request)
+            if response.is_stream_consumed:
+                return _already_read_within(response, max_bytes)
             wire_limit: Final = _wire_byte_limit(response.headers, max_bytes)
             if int(response.headers.get("content-length", "0")) > wire_limit:
                 raise HTTPResponseLimitError("Response exceeds the configured size limit")

tests/unit/llms/custom_httpx/test_http_handler.py

@@ -1839,6 +1839,45 @@ async def test_bounded_get_caps_the_wire_bytes_of_a_compressed_body_that_decodes
         await handler.close()
 
 
+@pytest.mark.asyncio
+async def test_bounded_get_serves_a_body_the_transport_already_read(monkeypatch):
+    monkeypatch.setenv("DISABLE_AIOHTTP_TRANSPORT", "True")
+    card: Final = {"data": [{"id": "served", "max_model_len": 8192}]}
+
+    def respond(request: httpx.Request) -> httpx.Response:
+        return httpx.Response(200, json=card)
+
+    handler = AsyncHTTPHandler()
+    handler.client = httpx.AsyncClient(transport=httpx.MockTransport(respond))
+    try:
+        response = await handler.get("https://one.test/v1/models", max_response_bytes=2 * 1024 * 1024)
+    finally:
+        await handler.close()
+    assert response.json() == card
+    assert response.headers["content-length"] == str(len(response.content))
+
+
+@pytest.mark.asyncio
+async def test_bounded_get_bounds_an_already_read_body_on_its_decoded_length(monkeypatch):
+    monkeypatch.setenv("DISABLE_AIOHTTP_TRANSPORT", "True")
+    document: Final = b"0" * 4096
+
+    def respond(request: httpx.Request) -> httpx.Response:
+        return httpx.Response(200, content=gzip.compress(document), headers={"content-encoding": "gzip"})
+
+    handler = AsyncHTTPHandler()
+    handler.client = httpx.AsyncClient(transport=httpx.MockTransport(respond))
+    try:
+        served = await handler.get("https://cdn.example/notes.txt", max_response_bytes=len(document))
+        assert served.content == document
+        assert "content-encoding" not in served.headers
+        assert served.headers["content-length"] == str(len(document))
+        with pytest.raises(HTTPResponseLimitError, match="size limit"):
+            await handler.get("https://cdn.example/notes.txt", max_response_bytes=len(document) - 1)
+    finally:
+        await handler.close()
+
+
 @pytest.mark.asyncio
 async def test_bounded_get_rejects_a_declared_compressed_length_over_the_wire_limit(respx_mock, monkeypatch):
     monkeypatch.setenv("DISABLE_AIOHTTP_TRANSPORT", "True")