feat: add --max-context-tokens for OOM prevention

[eloe]

Summary\nReject requests exceeding context window with 400 instead of OOM crash. --max-context-tokens flag. 5 tests.

[Copilot]

Pull request overview

This PR adds an optional prompt context-length guard to the FastAPI server to proactively reject overly-long requests (HTTP 400) instead of risking an OOM crash. It introduces a --max-context-tokens CLI flag (backed by MAX_CONTEXT_TOKENS) and basic unit tests for the new helpers.

Changes:

• Add get_max_context_tokens() and check_context_length() utilities to compute tokenized prompt length and reject requests beyond a configured limit.
• Wire the context-length check into both /responses and /chat/completions request flows after chat-template formatting.
• Add tests covering default/env configuration and helper behavior for within/over-limit cases.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
mlx_vlm/server.py	Adds max-context configuration + prompt token counting guard; integrates guard into request handlers; adds CLI flag/env wiring.
mlx_vlm/tests/test_server.py	Adds unit tests for max-context env parsing and context-length validation helper.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The new --max-context-tokens flag accepts negative values; those end up setting MAX_CONTEXT_TOKENS to a negative number, and check_context_length then treats it as “no limit” (max_context <= 0). Consider validating/clamping this argument to be non-negative (e.g., argparse custom type that raises on negatives, or enforce >= 0 after parsing) so misconfiguration doesn’t silently disable the protection.

Suggested change

	parser.add_argument(
	"--max-context-tokens",
	type=int,
	def non_negative_int(value):
	try:
	parsed_value = int(value)
	except ValueError as exc:
	raise argparse.ArgumentTypeError(
	f"invalid int value: {value!r}"
	) from exc
	if parsed_value < 0:
	raise argparse.ArgumentTypeError(
	"--max-context-tokens must be greater than or equal to 0"
	)
	return parsed_value
	parser.add_argument(
	"--max-context-tokens",
	type=non_negative_int,

[Copilot]

These tests mutate os.environ (pop/set MAX_CONTEXT_TOKENS) without restoring any pre-existing value. This can make the suite order-/environment-dependent when MAX_CONTEXT_TOKENS is set in the runner. Use pytest’s monkeypatch fixture (or save/restore the prior value) so environment changes are isolated to the test.

Suggested change

	def test_get_max_context_tokens_default():
	import os
	os.environ.pop("MAX_CONTEXT_TOKENS", None)
	assert server.get_max_context_tokens() == 0
	def test_get_max_context_tokens_from_env():
	import os
	os.environ["MAX_CONTEXT_TOKENS"] = "16384"
	assert server.get_max_context_tokens() == 16384
	os.environ.pop("MAX_CONTEXT_TOKENS")
	def test_get_max_context_tokens_default(monkeypatch):
	monkeypatch.delenv("MAX_CONTEXT_TOKENS", raising=False)
	assert server.get_max_context_tokens() == 0
	def test_get_max_context_tokens_from_env(monkeypatch):
	monkeypatch.setenv("MAX_CONTEXT_TOKENS", "16384")
	assert server.get_max_context_tokens() == 16384