Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.
Python cliconsumer-gpudpofine-tuningggufhuggingfacellmllmopslocal-ailocal-llmloralow-vrammachine-learningollamapeftpythonpytorchqlorasfttransformers
"""Bound decoded HTTP response bodies before JSON parsing or buffering."""
from __future__ import annotations
from typing import TYPE_CHECKING, Any
if TYPE_CHECKING:
import httpx
JUDGE_RESPONSE_LIMIT_BYTES = 1024 * 1024
GENERATION_RESPONSE_LIMIT_BYTES = 8 * 1024 * 1024
_READ_CHUNK_BYTES = 64 * 1024
class ResponseTooLargeError(ValueError):
"""A selected endpoint exceeded its decoded response budget."""
def post_capped(
url: str, *, max_bytes: int, endpoint_label: str, **kwargs: Any,
) -> httpx.Response:
"""POST and return a bounded, decoded response, always closing the stream.
Callers retain their endpoint validation, request headers and status/retry
policies. Error text contains only the caller's setting label and limit,
never the URL, credentials, request headers or response body.
"""
import httpx
if isinstance(max_bytes, bool) or not isinstance(max_bytes, int) or max_bytes < 1:
raise ValueError("max_bytes must be a positive integer")
def too_large() -> ResponseTooLargeError:
return ResponseTooLargeError(
f"Response for {endpoint_label} exceeds the {max_bytes} bytes limit"
)
with httpx.stream("POST", url, **kwargs) as response:
try:
declared_length = int(response.headers.get("Content-Length", ""))
except ValueError:
declared_length = 0
if declared_length > max_bytes:
raise too_large()
chunks: list[bytes] = []
size = 0
# iter_bytes counts decompressed content, including gzip replies.
for chunk in response.iter_bytes(chunk_size=_READ_CHUNK_BYTES):
size += len(chunk)
if size > max_bytes:
raise too_large()
chunks.append(chunk)
headers = response.headers.copy()
for name in ("Content-Encoding", "Content-Length", "Transfer-Encoding"):
headers.pop(name, None)
headers["Content-Length"] = str(size)
return httpx.Response(
response.status_code, headers=headers, content=b"".join(chunks),
request=response.request, extensions=response.extensions,
)