fix: increase gpu memory for MOSS with gplang-omni to .87

1 files changed, 23 insertions, 17 deletions

diff --git a/app/tests/test_backends_sglomni.py b/app/tests/test_backends_sglomni.py
index e61d51e..54659d6 100644
--- a/app/tests/test_backends_sglomni.py
+++ b/app/tests/test_backends_sglomni.py
@@ -447,30 +447,36 @@ class EngineMemoryBudgetTests(unittest.TestCase):
     VRAM when mem_fraction_static is unset (qwen3_tts, moss_tts, voxtral):
     on a 24 GB card the first /v1/audio/speech request aborts with
     "CUDA out of memory. Tried to allocate ~100 MiB" and every retry fails
-    identically. These vendored configs pin the pool at 0.70, leaving ~7 GB
-    for the colocated vocoder, CUDA graphs, and other GPU processes.
+    identically. These vendored configs pin the pool below auto-size,
+    leaving several GB for the colocated vocoder, CUDA graphs, and other
+    GPU processes — 0.70 for the small models. moss_tts pins 0.87 instead:
+    its ~17.1 GB of v1.5 bf16 weights alone exceed the 0.70 budget on the
+    ~20.7 GB the engine profiler sees on a 24 GB card (its colocated
+    tokenizer stage is already resident), so a 0.70 pin would abort the
+    boot with "Loaded weights leave no GPU memory for the KV cache"
+    instead of fixing the request-time OOM.
     """
 
-    # (catalog key, engine stage name). Voxtral's engine stage is
-    # tts_generation; every other engine-bearing pipeline names it
-    # tts_engine. moss_tts_local is intentionally absent — its upstream
-    # config class already budgets its colocated stages explicitly
-    # (0.15 preprocessing / 0.67 AR / 0.18 vocoder), and pinning would
-    # fight that logic (the codec reserve derives from the fractions).
+    # (catalog key, engine stage name, pinned mem_fraction_static).
+    # Voxtral's engine stage is tts_generation; every other engine-bearing
+    # pipeline names it tts_engine. moss_tts_local is intentionally absent —
+    # its upstream config class already budgets its colocated stages
+    # explicitly (0.15 preprocessing / 0.67 AR / 0.18 vocoder), and pinning
+    # would fight that logic (the codec reserve derives from the fractions).
     # s2-pro's OOM is model size, not budgeting (upstream issue #359).
     PINNED = (
-        ("qwen3_tts_0_6b_base", "tts_engine"),
-        ("qwen3_tts_0_6b_customvoice", "tts_engine"),
-        ("qwen3_tts_1_7b_base", "tts_engine"),
-        ("qwen3_tts_1_7b_voicedesign", "tts_engine"),
-        ("moss_tts", "tts_engine"),
-        ("voxtral_tts", "tts_generation"),
+        ("qwen3_tts_0_6b_base", "tts_engine", 0.70),
+        ("qwen3_tts_0_6b_customvoice", "tts_engine", 0.70),
+        ("qwen3_tts_1_7b_base", "tts_engine", 0.70),
+        ("qwen3_tts_1_7b_voicedesign", "tts_engine", 0.70),
+        ("moss_tts", "tts_engine", 0.87),
+        ("voxtral_tts", "tts_generation", 0.70),
     )
     UNPINNED = ("moss_tts_local", "fish_s2_pro")
 
     def test_engine_budget_is_pinned_under_the_engine_stage(self):
-        for key, stage in self.PINNED:
-            with self.subTest(key=key, stage=stage):
+        for key, stage, fraction in self.PINNED:
+            with self.subTest(key=key, stage=stage, fraction=fraction):
                 path = config_path(entry_by_key(key))
                 self.assertIsNotNone(path)
                 self.assertTrue(path.is_file(), f"missing {path}")
@@ -479,7 +485,7 @@ class EngineMemoryBudgetTests(unittest.TestCase):
                     "stages:\n"
                     f"  {stage}:\n"
                     "    engine:\n"
-                    "      mem_fraction_static: 0.70\n",
+                    f"      mem_fraction_static: {fraction:.2f}\n",
                     text)
 
     def test_already_budgeted_and_oversized_models_stay_unpinned(self):