Automatically use CPU for API-only evaluations

backend/app/api.py

@@ -84,9 +84,16 @@ def effective_config(incoming,user_id):
         except (ValueError,ValidationError):
             raise HTTPException(422,'Spec overrides conflict with site defaults or model panel') from None
         config['advanced_spec']=advanced.model_dump(exclude_unset=True,exclude_none=True)
-        enforce(config,db.effective_limits(user_id))
         return config
 
+    def select_compute(config,refs,user_id):
+        local = any(isinstance(ref,dict) for ref in refs.values())
+        if local and config['compute_type']=='cpu':
+            raise HTTPException(422,'CPU evaluations support API models only; Hugging Face models require GPU')
+        if not local:
+            config.update(compute_type='cpu',gpu_count=1)
+        enforce(config,db.effective_limits(user_id))
+
     @app.get('/admin')
     def admin_data(actor=Depends(administrator)):
         return db.admin_snapshot() | {'policy_schema':Policy.model_json_schema(),'limits_schema':Limits.model_json_schema()}
@@ -186,8 +193,7 @@ def spec_preview(incoming: EvaluationRequest, user_id=Depends(user)):
         import pprint
         config = effective_config(incoming,user_id)
         config['model_refs'] = resolve_models(incoming, catalog, incoming.hf_token.get_secret_value() or (cfg.hf_token if incoming.funding == 'service' else ''))
-        if incoming.compute_type == 'cpu' and any(isinstance(ref, dict) for ref in config['model_refs'].values()):
-            raise HTTPException(422, 'CPU evaluations support API models only; Hugging Face models require GPU')
+        select_compute(config,config['model_refs'],user_id)
         spec = build_spec(config, '.')
         return {'spec': spec, 'python': 'RUN_SPEC = ' + pprint.pformat(spec, sort_dicts=False) + '\n'}
 
@@ -218,12 +224,11 @@ async def submit(request: Request, user_id=Depends(user), idempotency_key: str =
             raise HTTPException(422, 'Supply both your OpenRouter and RunPod API keys')
         if incoming.funding == 'own_keys':
             await run_in_threadpool(verify_provider_keys, keys)
-        if incoming.funding == 'service' and db.member(user_id)['role']!='admin' and (incoming.gpu_type != cfg.runpod_gpu_type or incoming.disk_gb != cfg.runpod_disk_gb or incoming.gpu_count != 1 or incoming.compute_type != 'gpu' or incoming.cpu_count != 4 or incoming.cpu_flavor != 'cpu3g' or incoming.volume_gb != 0):
+        if incoming.funding == 'service' and db.member(user_id)['role']!='admin' and (incoming.gpu_type != cfg.runpod_gpu_type or incoming.disk_gb != cfg.runpod_disk_gb or incoming.gpu_count != 1 or incoming.cpu_count != 4 or incoming.cpu_flavor != 'cpu3g' or incoming.volume_gb != 0):
             raise HTTPException(422, 'Custom compute requires your own provider keys')
+        select_compute(config,refs,user_id)
         digest = hashlib.sha256(json.dumps(config, sort_keys=True).encode()).hexdigest()
         # Snapshot the catalog: queued jobs do not silently change if administrators update it.
-        if incoming.compute_type == 'cpu' and any(isinstance(ref, dict) for ref in refs.values()):
-            raise HTTPException(422, 'CPU evaluations support API models only. Select GPU for Hugging Face models.')
         config['model_refs'] = refs
         if incoming.funding != 'own_keys': keys = {}
         if incoming.hf_token.get_secret_value(): keys['hf_token'] = incoming.hf_token.get_secret_value()

backend/tests/test_backend.py

@@ -320,7 +320,8 @@ def create(self, job):
     assert db.get(job_id)['error_code'] == 'worker_start_timeout'
 
 
-def test_unavailable_progress_shows_attempt_and_next_retry(service):
+def test_unavailable_progress_shows_attempt_and_next_retry(service, monkeypatch):
+    monkeypatch.setattr('app.api.resolve_models',lambda *args:{'a':{'provider':'hf_local'},'b':'org/b'})
     from app.runpod import GPUUnavailable
     cfg,db,client=service
     job_id=submit(client).json()['id']
@@ -348,3 +349,40 @@ def test_compute_limits_and_cpu_local_models(service, monkeypatch):
     monkeypatch.setattr('app.api.resolve_models', lambda *args: {'a':{'provider':'hf_local'},'b':'org/b'})
     result=client.post('/spec-preview',json=payload(compute_type='cpu'),headers={'Authorization':'Bearer '+USER})
     assert result.status_code==422 and 'API models only' in result.json()['detail']
+
+
+@pytest.mark.parametrize('requested_compute',['gpu','cpu'])
+def test_api_only_panels_automatically_use_cpu_for_members(service,requested_compute,monkeypatch):
+    cfg,db,client=service
+    config=payload(compute_type=requested_compute)
+    response=submit(client,config)
+    assert response.status_code==202,response.text
+    job=db.get(response.json()['id'])
+    assert job['config']['compute_type']=='cpu'
+    assert job['config']['gpu_count']==1
+    assert response.json()['compute_type']=='cpu'
+    # Preview applies the same choice before compiling the spec.
+    from app.spec import build_spec as compile_spec
+    def capture(config,directory):
+        assert config['compute_type']=='cpu'
+        return compile_spec(config,directory)
+    monkeypatch.setattr('app.api.build_spec',capture)
+    assert client.post('/spec-preview',json=config,headers={'Authorization':'Bearer '+USER}).status_code==200
+
+
+def test_mixed_panel_keeps_gpu(service,monkeypatch):
+    _,db,client=service
+    monkeypatch.setattr('app.api.resolve_models',lambda *args:{'a':{'provider':'hf_local'},'b':'org/b'})
+    response=submit(client)
+    assert response.status_code==202
+    assert db.get(response.json()['id'])['config']['compute_type']=='gpu'
+
+
+def test_auto_cpu_respects_cpu_limit(service):
+    from app.governance import PolicyUpdate
+    _,db,client=service
+    policy=db.policy();policy['policy']['limits']['max_cpu_count']=2
+    db.set_policy(USER,PolicyUpdate(**policy))
+    response=submit(client,payload(compute_type='gpu'))
+    assert response.status_code==403
+    assert 'max_cpu_count' in response.json()['detail']