cdpearlman commited on
Commit
3f991b4
·
1 Parent(s): 718dfcb

Added full sequence to analysis, still need to do output scrubber

Browse files
Files changed (3) hide show
  1. app.py +41 -14
  2. tests/test_model_patterns.py +87 -0
  3. todo.md +8 -0
app.py CHANGED
@@ -370,10 +370,14 @@ def run_generation(n_clicks, model_name, prompt, max_new_tokens, beam_width, pat
370
  return (html.Div("Please select modules in the sidebar.", style={'color': 'red'}),
371
  results, {'display': 'none'}, {'display': 'none'}, {}, {}, original_prompt_data, {})
372
 
373
- # AGENT F KEY CHANGE: Run analysis on ORIGINAL PROMPT only, not generated text
374
- # This ensures pipeline stages show how the model processes the user's input,
375
- # regardless of what tokens are generated.
376
- activation_data = execute_forward_pass(model, tokenizer, prompt, config)
 
 
 
 
377
 
378
  results_ui = []
379
  if max_new_tokens > 1:
@@ -410,25 +414,27 @@ def run_generation(n_clicks, model_name, prompt, max_new_tokens, beam_width, pat
410
 
411
  @app.callback(
412
  [Output('session-selected-beam-store', 'data', allow_duplicate=True),
413
- Output('generation-results-container', 'children', allow_duplicate=True)],
 
414
  Input({'type': 'result-item', 'index': ALL}, 'n_clicks'),
415
- [State('generation-results-store', 'data')],
 
416
  prevent_initial_call=True
417
  )
418
- def store_selected_beam(n_clicks_list, results_data):
419
  """
420
- Agent F: Store selected beam for post-experiment comparison.
421
 
422
- The pipeline analysis already runs on the original prompt. This callback
423
- stores the selected beam text and updates the UI to show only the selected
424
- sequence, clearing all other options.
425
  """
426
  if not any(n_clicks_list) or not results_data:
427
- return no_update, no_update
428
 
429
  ctx = dash.callback_context
430
  if not ctx.triggered:
431
- return no_update, no_update
432
 
433
  triggered_id = json.loads(ctx.triggered[0]['prop_id'].split('.')[0])
434
  index = triggered_id['index']
@@ -470,8 +476,29 @@ def store_selected_beam(n_clicks_list, results_data):
470
  })
471
  ])
472
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
473
  # Store the selected beam for comparison after experiments
474
- return {'text': result['text'], 'score': result.get('score', 0), 'index': index}, selected_ui
 
475
 
476
 
477
  # ============================================================================
 
370
  return (html.Div("Please select modules in the sidebar.", style={'color': 'red'}),
371
  results, {'display': 'none'}, {'display': 'none'}, {}, {}, original_prompt_data, {})
372
 
373
+ # For single-token generation, analyze the full output (prompt + generated token)
374
+ # so attention covers the entire sequence. For multi-token, start with prompt only;
375
+ # the full-sequence analysis runs when the user selects a beam.
376
+ if max_new_tokens == 1:
377
+ full_text = results[0]['text']
378
+ else:
379
+ full_text = prompt
380
+ activation_data = execute_forward_pass(model, tokenizer, full_text, config)
381
 
382
  results_ui = []
383
  if max_new_tokens > 1:
 
414
 
415
  @app.callback(
416
  [Output('session-selected-beam-store', 'data', allow_duplicate=True),
417
+ Output('generation-results-container', 'children', allow_duplicate=True),
418
+ Output('session-activation-store', 'data', allow_duplicate=True)],
419
  Input({'type': 'result-item', 'index': ALL}, 'n_clicks'),
420
+ [State('generation-results-store', 'data'),
421
+ State('session-activation-store', 'data')],
422
  prevent_initial_call=True
423
  )
424
+ def store_selected_beam(n_clicks_list, results_data, existing_activation_data):
425
  """
426
+ Store selected beam and re-run forward pass on the full sequence.
427
 
428
+ When a beam is selected, this re-runs execute_forward_pass on the complete
429
+ generated text (prompt + output) so the attention visualization covers
430
+ the entire chosen output, not just the input.
431
  """
432
  if not any(n_clicks_list) or not results_data:
433
+ return no_update, no_update, no_update
434
 
435
  ctx = dash.callback_context
436
  if not ctx.triggered:
437
+ return no_update, no_update, no_update
438
 
439
  triggered_id = json.loads(ctx.triggered[0]['prop_id'].split('.')[0])
440
  index = triggered_id['index']
 
476
  })
477
  ])
478
 
479
+ # Re-run forward pass on the full beam text so attention covers entire output
480
+ new_activation_data = no_update
481
+ if existing_activation_data:
482
+ try:
483
+ from transformers import AutoModelForCausalLM, AutoTokenizer
484
+ model_name = existing_activation_data['model']
485
+ config = {
486
+ 'attention_modules': existing_activation_data['attention_modules'],
487
+ 'block_modules': existing_activation_data['block_modules'],
488
+ 'norm_parameters': existing_activation_data.get('norm_parameters', [])
489
+ }
490
+ model = AutoModelForCausalLM.from_pretrained(model_name, attn_implementation='eager')
491
+ tokenizer = AutoTokenizer.from_pretrained(model_name)
492
+ model.eval()
493
+ new_activation_data = execute_forward_pass(model, tokenizer, result['text'], config)
494
+ except Exception as e:
495
+ import traceback
496
+ traceback.print_exc()
497
+ print(f"Warning: Could not re-run forward pass for full sequence: {e}")
498
+
499
  # Store the selected beam for comparison after experiments
500
+ return ({'text': result['text'], 'score': result.get('score', 0), 'index': index},
501
+ selected_ui, new_activation_data)
502
 
503
 
504
  # ============================================================================
tests/test_model_patterns.py CHANGED
@@ -261,3 +261,90 @@ class TestMultiLayerHeadAblation:
261
 
262
  assert 'error' in result
263
  assert '99' in result['error'] # Should mention the invalid layer
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
261
 
262
  assert 'error' in result
263
  assert '99' in result['error'] # Should mention the invalid layer
264
+
265
+
266
+ class TestFullSequenceAttentionData:
267
+ """
268
+ Tests verifying that activation data for full sequences (prompt + generated output)
269
+ has correctly-sized attention matrices matching the full token count.
270
+ """
271
+
272
+ def _make_activation_data(self, num_tokens, num_layers=2, num_heads=1):
273
+ """Helper to build mock activation data for a given sequence length."""
274
+ input_ids = list(range(1, num_tokens + 1))
275
+ # Build uniform attention weights: [batch=1, heads, seq, seq]
276
+ uniform_val = 1.0 / num_tokens
277
+ attn_row = [uniform_val] * num_tokens
278
+ head_matrix = [attn_row] * num_tokens
279
+ head_block = [head_matrix] * num_heads
280
+ attn_weights = [head_block] # batch dim
281
+
282
+ attention_outputs = {}
283
+ for layer in range(num_layers):
284
+ module_name = f'model.layers.{layer}.self_attn'
285
+ attention_outputs[module_name] = {
286
+ 'output': [
287
+ [[0.1] * num_tokens], # hidden states placeholder
288
+ attn_weights
289
+ ]
290
+ }
291
+
292
+ return {
293
+ 'model': 'mock-model',
294
+ 'prompt': 'x ' * num_tokens,
295
+ 'input_ids': [input_ids],
296
+ 'attention_modules': list(attention_outputs.keys()),
297
+ 'attention_outputs': attention_outputs,
298
+ 'block_modules': [f'model.layers.{i}' for i in range(num_layers)],
299
+ 'block_outputs': {f'model.layers.{i}': {'output': [[[0.1] * num_tokens]]} for i in range(num_layers)},
300
+ 'norm_parameters': [],
301
+ 'norm_data': [],
302
+ 'actual_output': {'token': 'tok', 'probability': 0.5},
303
+ 'global_top5_tokens': []
304
+ }
305
+
306
+ def test_short_prompt_attention_dimensions(self):
307
+ """Attention matrix for a 4-token prompt should be 4x4."""
308
+ data = self._make_activation_data(num_tokens=4)
309
+ attn = data['attention_outputs']['model.layers.0.self_attn']['output'][1]
310
+ seq_len = len(attn[0][0]) # batch -> head -> rows
311
+ assert seq_len == 4
312
+ assert len(attn[0][0][0]) == 4 # columns
313
+
314
+ def test_full_sequence_attention_dimensions(self):
315
+ """Attention matrix for a 10-token full sequence (prompt+output) should be 10x10."""
316
+ data = self._make_activation_data(num_tokens=10)
317
+ attn = data['attention_outputs']['model.layers.0.self_attn']['output'][1]
318
+ seq_len = len(attn[0][0])
319
+ assert seq_len == 10
320
+ assert len(attn[0][0][0]) == 10
321
+
322
+ def test_full_sequence_has_more_tokens_than_prompt(self):
323
+ """Full-sequence activation data should have larger dimensions than prompt-only."""
324
+ prompt_data = self._make_activation_data(num_tokens=4)
325
+ full_data = self._make_activation_data(num_tokens=10)
326
+
327
+ prompt_ids = prompt_data['input_ids'][0]
328
+ full_ids = full_data['input_ids'][0]
329
+ assert len(full_ids) > len(prompt_ids)
330
+
331
+ prompt_attn = prompt_data['attention_outputs']['model.layers.0.self_attn']['output'][1]
332
+ full_attn = full_data['attention_outputs']['model.layers.0.self_attn']['output'][1]
333
+ assert len(full_attn[0][0]) > len(prompt_attn[0][0])
334
+
335
+ def test_input_ids_match_attention_seq_len(self):
336
+ """input_ids length should match attention matrix dimensions."""
337
+ for n in [3, 7, 15]:
338
+ data = self._make_activation_data(num_tokens=n)
339
+ num_ids = len(data['input_ids'][0])
340
+ attn = data['attention_outputs']['model.layers.0.self_attn']['output'][1]
341
+ assert num_ids == len(attn[0][0]) == n
342
+
343
+ def test_all_layers_have_same_dimensions(self):
344
+ """All layers should have the same attention matrix size for a given sequence."""
345
+ data = self._make_activation_data(num_tokens=8, num_layers=3)
346
+ for layer in range(3):
347
+ module = f'model.layers.{layer}.self_attn'
348
+ attn = data['attention_outputs'][module]['output'][1]
349
+ assert len(attn[0][0]) == 8
350
+ assert len(attn[0][0][0]) == 8
todo.md CHANGED
@@ -200,3 +200,11 @@
200
  - [x] Add CSS for `.chat-resize-handle` (cursor, hover highlight) in style.css
201
  - [x] Add `assets/chat_resize.js` for mousedown/mousemove/mouseup drag logic
202
  - Default size unchanged (25vw, 320px–450px); drag overrides max-width up to 80vw
 
 
 
 
 
 
 
 
 
200
  - [x] Add CSS for `.chat-resize-handle` (cursor, hover highlight) in style.css
201
  - [x] Add `assets/chat_resize.js` for mousedown/mousemove/mouseup drag logic
202
  - Default size unchanged (25vw, 320px–450px); drag overrides max-width up to 80vw
203
+
204
+ ## Completed: Full-Sequence Attention Visualization
205
+
206
+ - [x] Modified `run_generation()` in app.py: for single-token generation, run forward pass on full beam text (prompt + generated token) instead of prompt only
207
+ - [x] Enhanced `store_selected_beam()` in app.py: added `session-activation-store` output and `session-activation-store` State; re-runs `execute_forward_pass()` on selected beam's full text when user picks a beam
208
+ - [x] Added 5 tests in `test_model_patterns.py` (`TestFullSequenceAttentionData`) verifying attention matrix dimensions match full sequence length
209
+ - Attention visualization now covers the entire chosen output (input + generated tokens), not just the input prompt
210
+ - No changes needed in `model_patterns.py`, `beam_search.py`, `pipeline.py`, or `head_detection.py`