Studio: simplify the inference backend (#6490)

studio/backend/tests/test_mtp_vram_budget.py

@@ -878,8 +878,10 @@ class TestExtraArgsMtpDetection:
         # f16/f16 on the layer fallback) (#6312).
         load = "".join(inspect.getsource(LlamaCppBackend.load_model).split())
         assert "_tensor_dropped_extra_args=list(extra_args)" in load
-        # All three tensor->layer downgrade points restore the saved originals.
-        assert load.count("strip_split_mode_only(_tensor_dropped_extra_argsif") == 3
+        # The original extras are restored via one shared closure, called at all
+        # three tensor->layer downgrade points.
+        assert "strip_split_mode_only(_tensor_dropped_extra_argsif" in load
+        assert load.count("_restore_after_tensor_downgrade()") >= 3
 
     def test_load_model_tensor_skips_reserve_for_cpu_drafter(self):
         # A separate CPU-offloaded drafter (no embedded head) uses no GPU, so the

studio/backend/tests/test_sse_streaming_headers.py

@@ -0,0 +1,22 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Regression tests for the shared SSE streaming-response helper.
+
+Streaming endpoints must disable proxy buffering (``X-Accel-Buffering: no``);
+without it a reverse proxy (nginx / cloudflare tunnel) buffers the response and
+tokens stop appearing in real time. The native ``/generate/stream`` and legacy
+``/v1/completions`` streams historically omitted it and now route through the
+shared helper, so locking the helper's headers guards every standard path.
+"""
+
+import routes.inference as inference_route
+
+
+def test_sse_helper_sets_no_proxy_buffering_headers():
+    resp = inference_route._sse_streaming_response(iter(()))
+    assert resp.media_type == "text/event-stream"
+    # Starlette lowercases header keys in init_headers.
+    assert resp.headers["cache-control"] == "no-cache"
+    assert resp.headers["connection"] == "close"
+    assert resp.headers["x-accel-buffering"] == "no"

studio/backend/tests/test_tensor_parallel.py

@@ -1153,4 +1153,7 @@ def test_load_model_restores_quantized_kv_on_tensor_downgrade():
     # GPU-count and capacity-gate downgrades.
     compact = "".join(inspect.getsource(LlamaCppBackend.load_model).split())
     assert "_tensor_dropped_cache_type_kv=cache_type_kv" in compact  # captured pre-null
-    assert compact.count("cache_type_kv=_tensor_dropped_cache_type_kv") >= 2  # restored
+    # Restore is shared in one closure, called at every tensor->layer downgrade.
+    assert "cache_type_kv=_tensor_dropped_cache_type_kv" in compact  # restored in the closure
+    assert "def_restore_after_tensor_downgrade():" in compact  # one shared restore helper
+    assert compact.count("_restore_after_tensor_downgrade()") >= 3  # called at each downgrade