Use official CPU offload and verified pipeline signature for 48 GB evaluation

EVALUATION_PROVENANCE.json

@@ -1,7 +1,7 @@
 {
   "source": "Qwen/Qwen-Image-2.1",
   "source_revision": "d26bb61231c349cf6b7896fa83353113880e1ba3",
-  "handler_git": "1805a9dcb663b115221a5a55d72150d70f2968d7",
-  "handler_sha256": "d0dc0e84a39786a2de9f7808b8adae83ad0c66ba9661f7ca47ba2707e6eb892a",
+  "handler_git": "cf13cfff6d29586858d45a7e0d7972d0b88fd98f",
+  "handler_sha256": "126ca057f75fbc25185a75fb00cb9041942d9744f0489dd64f6685b5faa50877",
   "use": "Isolated research/evaluation only; no production promotion"
 }

handler/qwen21/README.md

@@ -2,6 +2,10 @@
 
 Research only. Root 2511 handler, Dockerfile and production endpoint stay unchanged.
 The official base model is tested at 40 steps, CFG 1 and BF16, with prefix KV cache.
+Reference conditioning uses the upstream `output_resolution=1024` default;
+the requested final image is 1024x1536. Full-resident BF16 failed with CUDA OOM
+on the actual 48 GB worker. The official `enable_model_cpu_offload()` method
+keeps BF16 and moves components in sequence without requiring a 96 GB GPU.
 This is a base-model quality comparison; it is not an 8-step Turbo benchmark.
 
 Build from the repository root using `qwen21/Dockerfile`. Set

handler/qwen21/handler.py

@@ -3,6 +3,7 @@ import io
 import json
 import os
 import time
+import traceback
 import uuid
 from pathlib import Path
 
@@ -32,7 +33,8 @@ def sampling(job):
         raise ValueError('An explicit editing prompt is required')
     return {'prompt': job['prompt'], 'negative_prompt': job.get('negative_prompt'),
             'width': width, 'height': height, 'num_inference_steps': steps,
-            'true_cfg_scale': cfg, 'use_kv_cache': True}
+            'true_cfg_scale': cfg, 'use_kv_cache': True,
+            'output_resolution': 1024}
 
 
 def cached_model_path():
@@ -51,7 +53,8 @@ def pipeline():
         from diffusers import QwenImage21Pipeline
         _pipe = QwenImage21Pipeline.from_pretrained(
             cached_model_path(), local_files_only=True,
-            torch_dtype=torch.bfloat16).to('cuda')
+            torch_dtype=torch.bfloat16)
+        _pipe.enable_model_cpu_offload()
     return _pipe
 
 
@@ -108,7 +111,7 @@ def handler(event):
                             'peak_reserved_gib': round(torch.cuda.max_memory_reserved() / 1024**3, 2),
                             'model_revision': os.environ['QWEN21_MODEL_REVISION']})
         except Exception as exc:
-            detail = str(exc)
+            detail = traceback.format_exc()
             for key in ('HF_TOKEN', 'JSON_PATH_KEY'):
                 if os.environ.get(key):
                     detail = detail.replace(os.environ[key], '[REDACTED]')

handler/qwen21/tests/pipeline_contract.json

@@ -0,0 +1,29 @@
+{
+  "diffusers_commit": "85c9fa4caee33d1c5b473395784471d2d7957c1b",
+  "source": "src/diffusers/pipelines/qwenimage21/pipeline_qwenimage21.py",
+  "source_sha256": "e75e4ce38d38e194edaac818bba99b2115b4bb2bfdbd15ce03a2813b0fe8856c",
+  "call_parameters": [
+    "prompt",
+    "image",
+    "negative_prompt",
+    "true_cfg_scale",
+    "height",
+    "width",
+    "num_inference_steps",
+    "sigmas",
+    "num_images_per_prompt",
+    "generator",
+    "latents",
+    "prompt_embeds",
+    "prompt_embeds_mask",
+    "negative_prompt_embeds",
+    "negative_prompt_embeds_mask",
+    "output_type",
+    "return_dict",
+    "attention_kwargs",
+    "callback_on_step_end",
+    "callback_on_step_end_tensor_inputs",
+    "output_resolution",
+    "use_kv_cache"
+  ]
+}

handler/qwen21/tests/test_sampling.py

@@ -1,6 +1,7 @@
 import unittest
 import os
 import tempfile
+import json
 from pathlib import Path
 from unittest.mock import patch
 
@@ -29,9 +30,17 @@ class SamplingTests(unittest.TestCase):
         self.assertEqual((args['width'], args['height'], args['num_inference_steps']), (1024, 1536, 40))
         self.assertEqual(args['true_cfg_scale'], 1.0)
         self.assertTrue(args['use_kv_cache'])
+        self.assertEqual(args['output_resolution'], 1024)
         self.assertNotIn('guidance_scale', args)
         self.assertNotIn('sigmas', args)
 
+    def test_sampling_matches_the_pinned_official_pipeline_signature(self):
+        root = Path(__file__).parents[1]
+        contract = json.loads((root / 'tests/pipeline_contract.json').read_text())
+        self.assertIn(contract['diffusers_commit'], (root / 'requirements.txt').read_text())
+        self.assertLessEqual(set(sampling({'prompt': 'Keep the actual bag.'})),
+                             set(contract['call_parameters']))
+
     def test_invalid_controls_fail_before_loading_a_model(self):
         with patch('qwen21.handler.pipeline') as load:
             result = handler({'input': {'task': 'edit', 'jobs': [