Spaces:
Running
Running
| """Exercise the public viewer with real bundled captures, sequentially in one browser.""" | |
| import argparse,json,pathlib,tempfile | |
| from playwright.sync_api import sync_playwright,expect | |
| def main(): | |
| parser=argparse.ArgumentParser();parser.add_argument('--base',default='http://localhost:8000');parser.add_argument('--chromium');parser.add_argument('--out',default='reports');args=parser.parse_args() | |
| out=pathlib.Path(args.out);out.mkdir(parents=True,exist_ok=True);checks=[];errors=[];failed_requests=[] | |
| def passed(name):checks.append(name);print('PASS '+name,flush=True) | |
| with sync_playwright() as p: | |
| browser=p.chromium.launch(headless=True,**({'executable_path':args.chromium} if args.chromium else {}),args=['--disable-dev-shm-usage']) | |
| try: | |
| context=browser.new_context(viewport={'width':1440,'height':1100},permissions=['clipboard-read','clipboard-write'],accept_downloads=True) | |
| page=context.new_page();page.on('pageerror',lambda e:errors.append(str(e)));page.on('response',lambda r:failed_requests.append({'url':r.url,'status':r.status}) if r.status>=400 else None) | |
| page.goto(args.base);page.wait_for_function('typeof S!=="undefined"&&S.frames.length===59&&!S.loadingTrace',timeout=90000) | |
| assert page.evaluate('S.model.id==="flint"&&S.frames[0].layers.length===49&&S.frame===0');expect(page.locator('#expertGrid button')).to_have_count(256);passed('Flint example: 59 measured frames and 49 blocks') | |
| expected=page.evaluate('S.frames[S.frame].layers[S.layer].heads[S.head].toFixed(3)+" RMS"');expect(page.locator('#headValue')).to_have_text(expected);passed('Inspector displays actual selected-frame RMS') | |
| page.locator('#openBlock').click();expect(page.locator('#detailDialog')).to_be_visible();expect(page.locator('#dBlockExperts button')).to_have_count(256);expect(page.locator('#dBlockRoutes button')).to_have_count(6);page.keyboard.press('Escape');passed('Fullscreen block exposes every expert and six actual routes') | |
| page.locator('#expertList button').first.click();expect(page.locator('#dExpertStatus')).to_have_text('Selected');expect(page.locator('#dTimeline button')).to_have_count(59);page.locator('#detailClose').click();passed('Selected expert has a complete token timeline') | |
| page.locator('#playSpeed').select_option('250');page.locator('#runButton').click();page.wait_for_function('S.frame>=2',timeout=10000);page.locator('#playButton').click();assert page.evaluate('S.timer===null');page.locator('#liveButton').click();assert page.evaluate('S.frame===58');page.locator('#stepBack').click();assert page.evaluate('S.frame===57');passed('Replay, speed, pause, step and jump to end change selected frames') | |
| page.locator('#historyButton').click();page.wait_for_function('S.geometry.history===true');page.locator('#structureButton').click();page.wait_for_function('S.geometry.history===false');passed('Token-history heatmap and structure switch') | |
| with tempfile.TemporaryDirectory() as temp: | |
| export=pathlib.Path(temp)/'trace.json' | |
| with page.expect_download() as d:page.locator('#exportButton').click() | |
| d.value.save_as(export);trace=json.loads(export.read_text());assert len(trace['frames'])==59 | |
| page.locator('#importFile').set_input_files(str(export));page.wait_for_function('!S.loadingTrace&&S.trace.frames.length===59');assert page.evaluate('S.frame===58');passed('Export/import retains the complete measured trace') | |
| trace['frames'][0]['layers'][12]['heads'][0]=-1;bad=pathlib.Path(temp)/'invalid.json';bad.write_text(json.dumps(trace)) | |
| page.locator('#importFile').set_input_files(str(bad));expect(page.locator('#toast')).to_contain_text('invalid head measurements');assert page.evaluate('S.frames[0].layers[12].heads[0]>=0&&S.frame===58');expect(page.locator('#runButton')).to_be_enabled();page.locator('#runButton').click();page.wait_for_function('S.frame>=1',timeout=10000);page.locator('#playButton').click();passed('Invalid import preserves the loaded capture and working replay') | |
| page.locator('#scrubber').fill('12');page.locator('#modelSelect').select_option('qwen');page.wait_for_function('S.model.id==="qwen"&&S.frames.length===44&&!S.switching',timeout=90000) | |
| expect(page.locator('#headGrid button')).to_have_count(48);assert page.evaluate('S.frames[0].layers.length===64');page.locator('#flowAttention').click();expect(page.locator('#dMemory')).to_be_visible();assert 'Write gate' in page.locator('#dMemory').inner_text();page.keyboard.press('Escape');passed('Qwen has 64 blocks and measured recurrent memory') | |
| for _ in range(3):page.locator('#nextLayer').click() | |
| expect(page.locator('#headGrid button')).to_have_count(24);page.locator('#flowAttention').click();expect(page.locator('#dAttentionRows')).to_be_visible();page.keyboard.press('Escape');page.locator('#flowMlp').click();expect(page.locator('#dFfnWidth')).to_have_text('17408');page.keyboard.press('Escape');passed('Qwen full-attention and dense feed-forward inspectors') | |
| page.locator('#modelSelect').select_option('flint');page.wait_for_function('S.model.id==="flint"&&!S.switching');assert page.evaluate('S.frame===12');passed('Switching models preserves each selected trace and token') | |
| page.locator('#modelCanvas').focus();old=page.evaluate('S.layer');page.keyboard.press('ArrowDown');assert page.evaluate('S.layer')==old+1;page.keyboard.press('Enter');expect(page.locator('#detailDialog')).to_be_visible();page.keyboard.press('Escape');passed('Keyboard topology navigation and fullscreen access') | |
| page.locator('#shareButton').click();expect(page.locator('#toast')).to_contain_text('view link copied');url=page.evaluate('navigator.clipboard.readText()');assert 'model=flint' in url and 'token=13' in url;passed('Share link records model, layer and token') | |
| page.locator('#toast').wait_for(state='hidden');page.screenshot(path=str(out/'desktop.png'),full_page=True) | |
| for width in [390,320]: | |
| page.set_viewport_size({'width':width,'height':844});assert page.evaluate('document.documentElement.scrollWidth<=innerWidth'),width | |
| page.locator('#openBlock').click();assert page.locator('#detailDialog').bounding_box()['width']<=width+1;page.keyboard.press('Escape') | |
| page.screenshot(path=str(out/'phone.png'),full_page=True);passed('320px and 390px layouts and fullscreen fit their viewport') | |
| for width,height in [(1440,720),(1440,900),(1440,1100),(1024,768)]: | |
| page.set_viewport_size({'width':width,'height':height}) | |
| page.wait_for_function("document.querySelector('.model-panel').getBoundingClientRect().bottom<=document.querySelector('main').getBoundingClientRect().bottom+1") | |
| assert page.evaluate("document.querySelector('#modelCanvas').getBoundingClientRect().bottom<=document.querySelector('.model-panel').getBoundingClientRect().bottom+1") | |
| assert page.evaluate("document.querySelector('.saved-section').getBoundingClientRect().bottom<=document.querySelector('main').getBoundingClientRect().bottom+1") | |
| passed('All decoder blocks and trace controls fit short and tall desktop panels') | |
| page.goto(args.base.rstrip('/')+'/?model=qwen&layer=3&token=7');page.wait_for_function('S.model?.id==="qwen"&&S.frames.length===44&&!S.loadingTrace&&S.frame===6',timeout=90000);assert page.evaluate('S.layer===3');passed('Shared deep link opens the requested example, layer and token') | |
| page.locator('#aboutButton').click();expect(page.locator('#aboutContent')).to_contain_text('95th percentile');expect(page.locator('#aboutSource')).to_have_attribute('href','https://huggingface.co/Qwen/Qwen3.8-27B/blob/main/config.json');page.locator('#closeAbout').click();passed('Explanation and model configuration link follow model selection') | |
| assert not errors,errors;assert not failed_requests,failed_requests;passed('No uncaught browser errors or failed release requests') | |
| context.close() | |
| # Failure injection uses a fresh context so a successful cache cannot hide it. | |
| recovery=browser.new_context(viewport={'width':390,'height':844});page=recovery.new_page();blocked=[True] | |
| def capture_route(route): | |
| if blocked[0]:route.abort() | |
| else:route.continue_() | |
| page.route('**/traces/*.json.gz',capture_route);page.goto(args.base);expect(page.locator('#engineStatus')).to_have_text('Trace unavailable',timeout=30000);assert page.evaluate('S.config.num_hidden_layers===49');blocked[0]=False;page.locator('#runButton').click();page.wait_for_function('S.frames.length===59&&!S.loadingTrace',timeout=90000);expect(page.locator('#engineStatus')).to_contain_text('Measured Q8 replay');recovery.close();passed('Failed capture leaves topology visible and Retry recovers') | |
| touch=browser.new_context(viewport={'width':390,'height':844},has_touch=True,is_mobile=True);page=touch.new_page();page.goto(args.base);page.wait_for_function('S.frames.length===59&&!S.loadingTrace',timeout=90000);page.locator('#openBlock').tap();expect(page.locator('#detailDialog')).to_be_visible();page.locator('#detailClose').tap();touch.close();passed('Touch interaction opens and closes fullscreen inspection') | |
| finally:browser.close() | |
| (out/'results.json').write_text(json.dumps({'passed':len(checks),'checks':checks,'browser_errors':errors,'failed_requests':failed_requests},indent=2)+'\n') | |
| print(json.dumps({'passed':len(checks),'errors':errors,'failed_requests':failed_requests})) | |
| if __name__=='__main__':main() | |